dump.co

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-27727916
// Unknown Toolkit Version
// Based on LLVM 3.4svn
//

.version 6.5
.target sm_61, texmode_independent
.address_size 64

	// .globl	verus_gpu_hash
// verus_gpu_hash$sharedMemory1 has been demoted
// verus_gpu_hash$sharedrand has been demoted
// verus_gpu_hash$sharedrandex has been demoted
.const .align 4 .b8 saes_table[4096] = {198, 99, 99, 165, 248, 124, 124, 132, 238, 119, 119, 153, 246, 123, 123, 141, 255, 242, 242, 13, 214, 107, 107, 189, 222, 111, 111, 177, 145, 197, 197, 84, 96, 48, 48, 80, 2, 1, 1, 3, 206, 103, 103, 169, 86, 43, 43, 125, 231, 254, 254, 25, 181, 215, 215, 98, 77, 171, 171, 230, 236, 118, 118, 154, 143, 202, 202, 69, 31, 130, 130, 157, 137, 201, 201, 64, 250, 125, 125, 135, 239, 250, 250, 21, 178, 89, 89, 235, 142, 71, 71, 201, 251, 240, 240, 11, 65, 173, 173, 236, 179, 212, 212, 103, 95, 162, 162, 253, 69, 175, 175, 234, 35, 156, 156, 191, 83, 164, 164, 247, 228, 114, 114, 150, 155, 192, 192, 91, 117, 183, 183, 194, 225, 253, 253, 28, 61, 147, 147, 174, 76, 38, 38, 106, 108, 54, 54, 90, 126, 63, 63, 65, 245, 247, 247, 2, 131, 204, 204, 79, 104, 52, 52, 92, 81, 165, 165, 244, 209, 229, 229, 52, 249, 241, 241, 8, 226, 113, 113, 147, 171, 216, 216, 115, 98, 49, 49, 83, 42, 21, 21, 63, 8, 4, 4, 12, 149, 199, 199, 82, 70, 35, 35, 101, 157, 195, 195, 94, 48, 24, 24, 40, 55, 150, 150, 161, 10, 5, 5, 15, 47, 154, 154, 181, 14, 7, 7, 9, 36, 18, 18, 54, 27, 128, 128, 155, 223, 226, 226, 61, 205, 235, 235, 38, 78, 39, 39, 105, 127, 178, 178, 205, 234, 117, 117, 159, 18, 9, 9, 27, 29, 131, 131, 158, 88, 44, 44, 116, 52, 26, 26, 46, 54, 27, 27, 45, 220, 110, 110, 178, 180, 90, 90, 238, 91, 160, 160, 251, 164, 82, 82, 246, 118, 59, 59, 77, 183, 214, 214, 97, 125, 179, 179, 206, 82, 41, 41, 123, 221, 227, 227, 62, 94, 47, 47, 113, 19, 132, 132, 151, 166, 83, 83, 245, 185, 209, 209, 104, 0, 0, 0, 0, 193, 237, 237, 44, 64, 32, 32, 96, 227, 252, 252, 31, 121, 177, 177, 200, 182, 91, 91, 237, 212, 106, 106, 190, 141, 203, 203, 70, 103, 190, 190, 217, 114, 57, 57, 75, 148, 74, 74, 222, 152, 76, 76, 212, 176, 88, 88, 232, 133, 207, 207, 74, 187, 208, 208, 107, 197, 239, 239, 42, 79, 170, 170, 229, 237, 251, 251, 22, 134, 67, 67, 197, 154, 77, 77, 215, 102, 51, 51, 85, 17, 133, 133, 148, 138, 69, 69, 207, 233, 249, 249, 16, 4, 2, 2, 6, 254, 127, 127, 129, 160, 80, 80, 240, 120, 60, 60, 68, 37, 159, 159, 186, 75, 168, 168, 227, 162, 81, 81, 243, 93, 163, 163, 254, 128, 64, 64, 192, 5, 143, 143, 138, 63, 146, 146, 173, 33, 157, 157, 188, 112, 56, 56, 72, 241, 245, 245, 4, 99, 188, 188, 223, 119, 182, 182, 193, 175, 218, 218, 117, 66, 33, 33, 99, 32, 16, 16, 48, 229, 255, 255, 26, 253, 243, 243, 14, 191, 210, 210, 109, 129, 205, 205, 76, 24, 12, 12, 20, 38, 19, 19, 53, 195, 236, 236, 47, 190, 95, 95, 225, 53, 151, 151, 162, 136, 68, 68, 204, 46, 23, 23, 57, 147, 196, 196, 87, 85, 167, 167, 242, 252, 126, 126, 130, 122, 61, 61, 71, 200, 100, 100, 172, 186, 93, 93, 231, 50, 25, 25, 43, 230, 115, 115, 149, 192, 96, 96, 160, 25, 129, 129, 152, 158, 79, 79, 209, 163, 220, 220, 127, 68, 34, 34, 102, 84, 42, 42, 126, 59, 144, 144, 171, 11, 136, 136, 131, 140, 70, 70, 202, 199, 238, 238, 41, 107, 184, 184, 211, 40, 20, 20, 60, 167, 222, 222, 121, 188, 94, 94, 226, 22, 11, 11, 29, 173, 219, 219, 118, 219, 224, 224, 59, 100, 50, 50, 86, 116, 58, 58, 78, 20, 10, 10, 30, 146, 73, 73, 219, 12, 6, 6, 10, 72, 36, 36, 108, 184, 92, 92, 228, 159, 194, 194, 93, 189, 211, 211, 110, 67, 172, 172, 239, 196, 98, 98, 166, 57, 145, 145, 168, 49, 149, 149, 164, 211, 228, 228, 55, 242, 121, 121, 139, 213, 231, 231, 50, 139, 200, 200, 67, 110, 55, 55, 89, 218, 109, 109, 183, 1, 141, 141, 140, 177, 213, 213, 100, 156, 78, 78, 210, 73, 169, 169, 224, 216, 108, 108, 180, 172, 86, 86, 250, 243, 244, 244, 7, 207, 234, 234, 37, 202, 101, 101, 175, 244, 122, 122, 142, 71, 174, 174, 233, 16, 8, 8, 24, 111, 186, 186, 213, 240, 120, 120, 136, 74, 37, 37, 111, 92, 46, 46, 114, 56, 28, 28, 36, 87, 166, 166, 241, 115, 180, 180, 199, 151, 198, 198, 81, 203, 232, 232, 35, 161, 221, 221, 124, 232, 116, 116, 156, 62, 31, 31, 33, 150, 75, 75, 221, 97, 189, 189, 220, 13, 139, 139, 134, 15, 138, 138, 133, 224, 112, 112, 144, 124, 62, 62, 66, 113, 181, 181, 196, 204, 102, 102, 170, 144, 72, 72, 216, 6, 3, 3, 5, 247, 246, 246, 1, 28, 14, 14, 18, 194, 97, 97, 163, 106, 53, 53, 95, 174, 87, 87, 249, 105, 185, 185, 208, 23, 134, 134, 145, 153, 193, 193, 88, 58, 29, 29, 39, 39, 158, 158, 185, 217, 225, 225, 56, 235, 248, 248, 19, 43, 152, 152, 179, 34, 17, 17, 51, 210, 105, 105, 187, 169, 217, 217, 112, 7, 142, 142, 137, 51, 148, 148, 167, 45, 155, 155, 182, 60, 30, 30, 34, 21, 135, 135, 146, 201, 233, 233, 32, 135, 206, 206, 73, 170, 85, 85, 255, 80, 40, 40, 120, 165, 223, 223, 122, 3, 140, 140, 143, 89, 161, 161, 248, 9, 137, 137, 128, 26, 13, 13, 23, 101, 191, 191, 218, 215, 230, 230, 49, 132, 66, 66, 198, 208, 104, 104, 184, 130, 65, 65, 195, 41, 153, 153, 176, 90, 45, 45, 119, 30, 15, 15, 17, 123, 176, 176, 203, 168, 84, 84, 252, 109, 187, 187, 214, 44, 22, 22, 58, 165, 198, 99, 99, 132, 248, 124, 124, 153, 238, 119, 119, 141, 246, 123, 123, 13, 255, 242, 242, 189, 214, 107, 107, 177, 222, 111, 111, 84, 145, 197, 197, 80, 96, 48, 48, 3, 2, 1, 1, 169, 206, 103, 103, 125, 86, 43, 43, 25, 231, 254, 254, 98, 181, 215, 215, 230, 77, 171, 171, 154, 236, 118, 118, 69, 143, 202, 202, 157, 31, 130, 130, 64, 137, 201, 201, 135, 250, 125, 125, 21, 239, 250, 250, 235, 178, 89, 89, 201, 142, 71, 71, 11, 251, 240, 240, 236, 65, 173, 173, 103, 179, 212, 212, 253, 95, 162, 162, 234, 69, 175, 175, 191, 35, 156, 156, 247, 83, 164, 164, 150, 228, 114, 114, 91, 155, 192, 192, 194, 117, 183, 183, 28, 225, 253, 253, 174, 61, 147, 147, 106, 76, 38, 38, 90, 108, 54, 54, 65, 126, 63, 63, 2, 245, 247, 247, 79, 131, 204, 204, 92, 104, 52, 52, 244, 81, 165, 165, 52, 209, 229, 229, 8, 249, 241, 241, 147, 226, 113, 113, 115, 171, 216, 216, 83, 98, 49, 49, 63, 42, 21, 21, 12, 8, 4, 4, 82, 149, 199, 199, 101, 70, 35, 35, 94, 157, 195, 195, 40, 48, 24, 24, 161, 55, 150, 150, 15, 10, 5, 5, 181, 47, 154, 154, 9, 14, 7, 7, 54, 36, 18, 18, 155, 27, 128, 128, 61, 223, 226, 226, 38, 205, 235, 235, 105, 78, 39, 39, 205, 127, 178, 178, 159, 234, 117, 117, 27, 18, 9, 9, 158, 29, 131, 131, 116, 88, 44, 44, 46, 52, 26, 26, 45, 54, 27, 27, 178, 220, 110, 110, 238, 180, 90, 90, 251, 91, 160, 160, 246, 164, 82, 82, 77, 118, 59, 59, 97, 183, 214, 214, 206, 125, 179, 179, 123, 82, 41, 41, 62, 221, 227, 227, 113, 94, 47, 47, 151, 19, 132, 132, 245, 166, 83, 83, 104, 185, 209, 209, 0, 0, 0, 0, 44, 193, 237, 237, 96, 64, 32, 32, 31, 227, 252, 252, 200, 121, 177, 177, 237, 182, 91, 91, 190, 212, 106, 106, 70, 141, 203, 203, 217, 103, 190, 190, 75, 114, 57, 57, 222, 148, 74, 74, 212, 152, 76, 76, 232, 176, 88, 88, 74, 133, 207, 207, 107, 187, 208, 208, 42, 197, 239, 239, 229, 79, 170, 170, 22, 237, 251, 251, 197, 134, 67, 67, 215, 154, 77, 77, 85, 102, 51, 51, 148, 17, 133, 133, 207, 138, 69, 69, 16, 233, 249, 249, 6, 4, 2, 2, 129, 254, 127, 127, 240, 160, 80, 80, 68, 120, 60, 60, 186, 37, 159, 159, 227, 75, 168, 168, 243, 162, 81, 81, 254, 93, 163, 163, 192, 128, 64, 64, 138, 5, 143, 143, 173, 63, 146, 146, 188, 33, 157, 157, 72, 112, 56, 56, 4, 241, 245, 245, 223, 99, 188, 188, 193, 119, 182, 182, 117, 175, 218, 218, 99, 66, 33, 33, 48, 32, 16, 16, 26, 229, 255, 255, 14, 253, 243, 243, 109, 191, 210, 210, 76, 129, 205, 205, 20, 24, 12, 12, 53, 38, 19, 19, 47, 195, 236, 236, 225, 190, 95, 95, 162, 53, 151, 151, 204, 136, 68, 68, 57, 46, 23, 23, 87, 147, 196, 196, 242, 85, 167, 167, 130, 252, 126, 126, 71, 122, 61, 61, 172, 200, 100, 100, 231, 186, 93, 93, 43, 50, 25, 25, 149, 230, 115, 115, 160, 192, 96, 96, 152, 25, 129, 129, 209, 158, 79, 79, 127, 163, 220, 220, 102, 68, 34, 34, 126, 84, 42, 42, 171, 59, 144, 144, 131, 11, 136, 136, 202, 140, 70, 70, 41, 199, 238, 238, 211, 107, 184, 184, 60, 40, 20, 20, 121, 167, 222, 222, 226, 188, 94, 94, 29, 22, 11, 11, 118, 173, 219, 219, 59, 219, 224, 224, 86, 100, 50, 50, 78, 116, 58, 58, 30, 20, 10, 10, 219, 146, 73, 73, 10, 12, 6, 6, 108, 72, 36, 36, 228, 184, 92, 92, 93, 159, 194, 194, 110, 189, 211, 211, 239, 67, 172, 172, 166, 196, 98, 98, 168, 57, 145, 145, 164, 49, 149, 149, 55, 211, 228, 228, 139, 242, 121, 121, 50, 213, 231, 231, 67, 139, 200, 200, 89, 110, 55, 55, 183, 218, 109, 109, 140, 1, 141, 141, 100, 177, 213, 213, 210, 156, 78, 78, 224, 73, 169, 169, 180, 216, 108, 108, 250, 172, 86, 86, 7, 243, 244, 244, 37, 207, 234, 234, 175, 202, 101, 101, 142, 244, 122, 122, 233, 71, 174, 174, 24, 16, 8, 8, 213, 111, 186, 186, 136, 240, 120, 120, 111, 74, 37, 37, 114, 92, 46, 46, 36, 56, 28, 28, 241, 87, 166, 166, 199, 115, 180, 180, 81, 151, 198, 198, 35, 203, 232, 232, 124, 161, 221, 221, 156, 232, 116, 116, 33, 62, 31, 31, 221, 150, 75, 75, 220, 97, 189, 189, 134, 13, 139, 139, 133, 15, 138, 138, 144, 224, 112, 112, 66, 124, 62, 62, 196, 113, 181, 181, 170, 204, 102, 102, 216, 144, 72, 72, 5, 6, 3, 3, 1, 247, 246, 246, 18, 28, 14, 14, 163, 194, 97, 97, 95, 106, 53, 53, 249, 174, 87, 87, 208, 105, 185, 185, 145, 23, 134, 134, 88, 153, 193, 193, 39, 58, 29, 29, 185, 39, 158, 158, 56, 217, 225, 225, 19, 235, 248, 248, 179, 43, 152, 152, 51, 34, 17, 17, 187, 210, 105, 105, 112, 169, 217, 217, 137, 7, 142, 142, 167, 51, 148, 148, 182, 45, 155, 155, 34, 60, 30, 30, 146, 21, 135, 135, 32, 201, 233, 233, 73, 135, 206, 206, 255, 170, 85, 85, 120, 80, 40, 40, 122, 165, 223, 223, 143, 3, 140, 140, 248, 89, 161, 161, 128, 9, 137, 137, 23, 26, 13, 13, 218, 101, 191, 191, 49, 215, 230, 230, 198, 132, 66, 66, 184, 208, 104, 104, 195, 130, 65, 65, 176, 41, 153, 153, 119, 90, 45, 45, 17, 30, 15, 15, 203, 123, 176, 176, 252, 168, 84, 84, 214, 109, 187, 187, 58, 44, 22, 22, 99, 165, 198, 99, 124, 132, 248, 124, 119, 153, 238, 119, 123, 141, 246, 123, 242, 13, 255, 242, 107, 189, 214, 107, 111, 177, 222, 111, 197, 84, 145, 197, 48, 80, 96, 48, 1, 3, 2, 1, 103, 169, 206, 103, 43, 125, 86, 43, 254, 25, 231, 254, 215, 98, 181, 215, 171, 230, 77, 171, 118, 154, 236, 118, 202, 69, 143, 202, 130, 157, 31, 130, 201, 64, 137, 201, 125, 135, 250, 125, 250, 21, 239, 250, 89, 235, 178, 89, 71, 201, 142, 71, 240, 11, 251, 240, 173, 236, 65, 173, 212, 103, 179, 212, 162, 253, 95, 162, 175, 234, 69, 175, 156, 191, 35, 156, 164, 247, 83, 164, 114, 150, 228, 114, 192, 91, 155, 192, 183, 194, 117, 183, 253, 28, 225, 253, 147, 174, 61, 147, 38, 106, 76, 38, 54, 90, 108, 54, 63, 65, 126, 63, 247, 2, 245, 247, 204, 79, 131, 204, 52, 92, 104, 52, 165, 244, 81, 165, 229, 52, 209, 229, 241, 8, 249, 241, 113, 147, 226, 113, 216, 115, 171, 216, 49, 83, 98, 49, 21, 63, 42, 21, 4, 12, 8, 4, 199, 82, 149, 199, 35, 101, 70, 35, 195, 94, 157, 195, 24, 40, 48, 24, 150, 161, 55, 150, 5, 15, 10, 5, 154, 181, 47, 154, 7, 9, 14, 7, 18, 54, 36, 18, 128, 155, 27, 128, 226, 61, 223, 226, 235, 38, 205, 235, 39, 105, 78, 39, 178, 205, 127, 178, 117, 159, 234, 117, 9, 27, 18, 9, 131, 158, 29, 131, 44, 116, 88, 44, 26, 46, 52, 26, 27, 45, 54, 27, 110, 178, 220, 110, 90, 238, 180, 90, 160, 251, 91, 160, 82, 246, 164, 82, 59, 77, 118, 59, 214, 97, 183, 214, 179, 206, 125, 179, 41, 123, 82, 41, 227, 62, 221, 227, 47, 113, 94, 47, 132, 151, 19, 132, 83, 245, 166, 83, 209, 104, 185, 209, 0, 0, 0, 0, 237, 44, 193, 237, 32, 96, 64, 32, 252, 31, 227, 252, 177, 200, 121, 177, 91, 237, 182, 91, 106, 190, 212, 106, 203, 70, 141, 203, 190, 217, 103, 190, 57, 75, 114, 57, 74, 222, 148, 74, 76, 212, 152, 76, 88, 232, 176, 88, 207, 74, 133, 207, 208, 107, 187, 208, 239, 42, 197, 239, 170, 229, 79, 170, 251, 22, 237, 251, 67, 197, 134, 67, 77, 215, 154, 77, 51, 85, 102, 51, 133, 148, 17, 133, 69, 207, 138, 69, 249, 16, 233, 249, 2, 6, 4, 2, 127, 129, 254, 127, 80, 240, 160, 80, 60, 68, 120, 60, 159, 186, 37, 159, 168, 227, 75, 168, 81, 243, 162, 81, 163, 254, 93, 163, 64, 192, 128, 64, 143, 138, 5, 143, 146, 173, 63, 146, 157, 188, 33, 157, 56, 72, 112, 56, 245, 4, 241, 245, 188, 223, 99, 188, 182, 193, 119, 182, 218, 117, 175, 218, 33, 99, 66, 33, 16, 48, 32, 16, 255, 26, 229, 255, 243, 14, 253, 243, 210, 109, 191, 210, 205, 76, 129, 205, 12, 20, 24, 12, 19, 53, 38, 19, 236, 47, 195, 236, 95, 225, 190, 95, 151, 162, 53, 151, 68, 204, 136, 68, 23, 57, 46, 23, 196, 87, 147, 196, 167, 242, 85, 167, 126, 130, 252, 126, 61, 71, 122, 61, 100, 172, 200, 100, 93, 231, 186, 93, 25, 43, 50, 25, 115, 149, 230, 115, 96, 160, 192, 96, 129, 152, 25, 129, 79, 209, 158, 79, 220, 127, 163, 220, 34, 102, 68, 34, 42, 126, 84, 42, 144, 171, 59, 144, 136, 131, 11, 136, 70, 202, 140, 70, 238, 41, 199, 238, 184, 211, 107, 184, 20, 60, 40, 20, 222, 121, 167, 222, 94, 226, 188, 94, 11, 29, 22, 11, 219, 118, 173, 219, 224, 59, 219, 224, 50, 86, 100, 50, 58, 78, 116, 58, 10, 30, 20, 10, 73, 219, 146, 73, 6, 10, 12, 6, 36, 108, 72, 36, 92, 228, 184, 92, 194, 93, 159, 194, 211, 110, 189, 211, 172, 239, 67, 172, 98, 166, 196, 98, 145, 168, 57, 145, 149, 164, 49, 149, 228, 55, 211, 228, 121, 139, 242, 121, 231, 50, 213, 231, 200, 67, 139, 200, 55, 89, 110, 55, 109, 183, 218, 109, 141, 140, 1, 141, 213, 100, 177, 213, 78, 210, 156, 78, 169, 224, 73, 169, 108, 180, 216, 108, 86, 250, 172, 86, 244, 7, 243, 244, 234, 37, 207, 234, 101, 175, 202, 101, 122, 142, 244, 122, 174, 233, 71, 174, 8, 24, 16, 8, 186, 213, 111, 186, 120, 136, 240, 120, 37, 111, 74, 37, 46, 114, 92, 46, 28, 36, 56, 28, 166, 241, 87, 166, 180, 199, 115, 180, 198, 81, 151, 198, 232, 35, 203, 232, 221, 124, 161, 221, 116, 156, 232, 116, 31, 33, 62, 31, 75, 221, 150, 75, 189, 220, 97, 189, 139, 134, 13, 139, 138, 133, 15, 138, 112, 144, 224, 112, 62, 66, 124, 62, 181, 196, 113, 181, 102, 170, 204, 102, 72, 216, 144, 72, 3, 5, 6, 3, 246, 1, 247, 246, 14, 18, 28, 14, 97, 163, 194, 97, 53, 95, 106, 53, 87, 249, 174, 87, 185, 208, 105, 185, 134, 145, 23, 134, 193, 88, 153, 193, 29, 39, 58, 29, 158, 185, 39, 158, 225, 56, 217, 225, 248, 19, 235, 248, 152, 179, 43, 152, 17, 51, 34, 17, 105, 187, 210, 105, 217, 112, 169, 217, 142, 137, 7, 142, 148, 167, 51, 148, 155, 182, 45, 155, 30, 34, 60, 30, 135, 146, 21, 135, 233, 32, 201, 233, 206, 73, 135, 206, 85, 255, 170, 85, 40, 120, 80, 40, 223, 122, 165, 223, 140, 143, 3, 140, 161, 248, 89, 161, 137, 128, 9, 137, 13, 23, 26, 13, 191, 218, 101, 191, 230, 49, 215, 230, 66, 198, 132, 66, 104, 184, 208, 104, 65, 195, 130, 65, 153, 176, 41, 153, 45, 119, 90, 45, 15, 17, 30, 15, 176, 203, 123, 176, 84, 252, 168, 84, 187, 214, 109, 187, 22, 58, 44, 22, 99, 99, 165, 198, 124, 124, 132, 248, 119, 119, 153, 238, 123, 123, 141, 246, 242, 242, 13, 255, 107, 107, 189, 214, 111, 111, 177, 222, 197, 197, 84, 145, 48, 48, 80, 96, 1, 1, 3, 2, 103, 103, 169, 206, 43, 43, 125, 86, 254, 254, 25, 231, 215, 215, 98, 181, 171, 171, 230, 77, 118, 118, 154, 236, 202, 202, 69, 143, 130, 130, 157, 31, 201, 201, 64, 137, 125, 125, 135, 250, 250, 250, 21, 239, 89, 89, 235, 178, 71, 71, 201, 142, 240, 240, 11, 251, 173, 173, 236, 65, 212, 212, 103, 179, 162, 162, 253, 95, 175, 175, 234, 69, 156, 156, 191, 35, 164, 164, 247, 83, 114, 114, 150, 228, 192, 192, 91, 155, 183, 183, 194, 117, 253, 253, 28, 225, 147, 147, 174, 61, 38, 38, 106, 76, 54, 54, 90, 108, 63, 63, 65, 126, 247, 247, 2, 245, 204, 204, 79, 131, 52, 52, 92, 104, 165, 165, 244, 81, 229, 229, 52, 209, 241, 241, 8, 249, 113, 113, 147, 226, 216, 216, 115, 171, 49, 49, 83, 98, 21, 21, 63, 42, 4, 4, 12, 8, 199, 199, 82, 149, 35, 35, 101, 70, 195, 195, 94, 157, 24, 24, 40, 48, 150, 150, 161, 55, 5, 5, 15, 10, 154, 154, 181, 47, 7, 7, 9, 14, 18, 18, 54, 36, 128, 128, 155, 27, 226, 226, 61, 223, 235, 235, 38, 205, 39, 39, 105, 78, 178, 178, 205, 127, 117, 117, 159, 234, 9, 9, 27, 18, 131, 131, 158, 29, 44, 44, 116, 88, 26, 26, 46, 52, 27, 27, 45, 54, 110, 110, 178, 220, 90, 90, 238, 180, 160, 160, 251, 91, 82, 82, 246, 164, 59, 59, 77, 118, 214, 214, 97, 183, 179, 179, 206, 125, 41, 41, 123, 82, 227, 227, 62, 221, 47, 47, 113, 94, 132, 132, 151, 19, 83, 83, 245, 166, 209, 209, 104, 185, 0, 0, 0, 0, 237, 237, 44, 193, 32, 32, 96, 64, 252, 252, 31, 227, 177, 177, 200, 121, 91, 91, 237, 182, 106, 106, 190, 212, 203, 203, 70, 141, 190, 190, 217, 103, 57, 57, 75, 114, 74, 74, 222, 148, 76, 76, 212, 152, 88, 88, 232, 176, 207, 207, 74, 133, 208, 208, 107, 187, 239, 239, 42, 197, 170, 170, 229, 79, 251, 251, 22, 237, 67, 67, 197, 134, 77, 77, 215, 154, 51, 51, 85, 102, 133, 133, 148, 17, 69, 69, 207, 138, 249, 249, 16, 233, 2, 2, 6, 4, 127, 127, 129, 254, 80, 80, 240, 160, 60, 60, 68, 120, 159, 159, 186, 37, 168, 168, 227, 75, 81, 81, 243, 162, 163, 163, 254, 93, 64, 64, 192, 128, 143, 143, 138, 5, 146, 146, 173, 63, 157, 157, 188, 33, 56, 56, 72, 112, 245, 245, 4, 241, 188, 188, 223, 99, 182, 182, 193, 119, 218, 218, 117, 175, 33, 33, 99, 66, 16, 16, 48, 32, 255, 255, 26, 229, 243, 243, 14, 253, 210, 210, 109, 191, 205, 205, 76, 129, 12, 12, 20, 24, 19, 19, 53, 38, 236, 236, 47, 195, 95, 95, 225, 190, 151, 151, 162, 53, 68, 68, 204, 136, 23, 23, 57, 46, 196, 196, 87, 147, 167, 167, 242, 85, 126, 126, 130, 252, 61, 61, 71, 122, 100, 100, 172, 200, 93, 93, 231, 186, 25, 25, 43, 50, 115, 115, 149, 230, 96, 96, 160, 192, 129, 129, 152, 25, 79, 79, 209, 158, 220, 220, 127, 163, 34, 34, 102, 68, 42, 42, 126, 84, 144, 144, 171, 59, 136, 136, 131, 11, 70, 70, 202, 140, 238, 238, 41, 199, 184, 184, 211, 107, 20, 20, 60, 40, 222, 222, 121, 167, 94, 94, 226, 188, 11, 11, 29, 22, 219, 219, 118, 173, 224, 224, 59, 219, 50, 50, 86, 100, 58, 58, 78, 116, 10, 10, 30, 20, 73, 73, 219, 146, 6, 6, 10, 12, 36, 36, 108, 72, 92, 92, 228, 184, 194, 194, 93, 159, 211, 211, 110, 189, 172, 172, 239, 67, 98, 98, 166, 196, 145, 145, 168, 57, 149, 149, 164, 49, 228, 228, 55, 211, 121, 121, 139, 242, 231, 231, 50, 213, 200, 200, 67, 139, 55, 55, 89, 110, 109, 109, 183, 218, 141, 141, 140, 1, 213, 213, 100, 177, 78, 78, 210, 156, 169, 169, 224, 73, 108, 108, 180, 216, 86, 86, 250, 172, 244, 244, 7, 243, 234, 234, 37, 207, 101, 101, 175, 202, 122, 122, 142, 244, 174, 174, 233, 71, 8, 8, 24, 16, 186, 186, 213, 111, 120, 120, 136, 240, 37, 37, 111, 74, 46, 46, 114, 92, 28, 28, 36, 56, 166, 166, 241, 87, 180, 180, 199, 115, 198, 198, 81, 151, 232, 232, 35, 203, 221, 221, 124, 161, 116, 116, 156, 232, 31, 31, 33, 62, 75, 75, 221, 150, 189, 189, 220, 97, 139, 139, 134, 13, 138, 138, 133, 15, 112, 112, 144, 224, 62, 62, 66, 124, 181, 181, 196, 113, 102, 102, 170, 204, 72, 72, 216, 144, 3, 3, 5, 6, 246, 246, 1, 247, 14, 14, 18, 28, 97, 97, 163, 194, 53, 53, 95, 106, 87, 87, 249, 174, 185, 185, 208, 105, 134, 134, 145, 23, 193, 193, 88, 153, 29, 29, 39, 58, 158, 158, 185, 39, 225, 225, 56, 217, 248, 248, 19, 235, 152, 152, 179, 43, 17, 17, 51, 34, 105, 105, 187, 210, 217, 217, 112, 169, 142, 142, 137, 7, 148, 148, 167, 51, 155, 155, 182, 45, 30, 30, 34, 60, 135, 135, 146, 21, 233, 233, 32, 201, 206, 206, 73, 135, 85, 85, 255, 170, 40, 40, 120, 80, 223, 223, 122, 165, 140, 140, 143, 3, 161, 161, 248, 89, 137, 137, 128, 9, 13, 13, 23, 26, 191, 191, 218, 101, 230, 230, 49, 215, 66, 66, 198, 132, 104, 104, 184, 208, 65, 65, 195, 130, 153, 153, 176, 41, 45, 45, 119, 90, 15, 15, 17, 30, 176, 176, 203, 123, 84, 84, 252, 168, 187, 187, 214, 109, 22, 22, 58, 44};

.entry verus_gpu_hash(
	.param .u64 .ptr .const .align 4 verus_gpu_hash_param_0,
	.param .u64 .ptr .const .align 16 verus_gpu_hash_param_1,
	.param .u64 .ptr .global .align 16 verus_gpu_hash_param_2,
	.param .u64 .ptr .const .align 16 verus_gpu_hash_param_3,
	.param .u64 .ptr .global .align 4 verus_gpu_hash_param_4,
	.param .u64 .ptr .global .align 4 verus_gpu_hash_param_5
)
.reqntid 64, 1, 1
{
	.local .align 16 .b8 	__local_depot0[128];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<93>;
	.reg .b16 	%rs<239>;
	.reg .b32 	%r<6734>;
	.reg .b64 	%rd<6140>;
	// demoted variable
	.shared .align 4 .b8 verus_gpu_hash$sharedMemory1[4096];
	// demoted variable
	.shared .align 2 .b8 verus_gpu_hash$sharedrand[4096];
	// demoted variable
	.shared .align 2 .b8 verus_gpu_hash$sharedrandex[4096];

	mov.u64 	%SPL, __local_depot0;
	ld.param.u64 	%rd207, [verus_gpu_hash_param_0];
	ld.param.u64 	%rd208, [verus_gpu_hash_param_1];
	ld.param.u64 	%rd203, [verus_gpu_hash_param_2];
	ld.param.u64 	%rd6090, [verus_gpu_hash_param_3];
	add.u64 	%rd1, %SPL, 64;
	mov.u32 	%r740, %ctaid.x;
	mov.u32 	%r741, %ntid.x;
	mov.b32	%r742, %envreg3;
	mad.lo.s32 	%r1, %r740, %r741, %r742;
	mov.u32 	%r2, %tid.x;
	add.s32 	%r743, %r1, %r2;
	cvt.u64.u32	%rd210, %r743;
	cvt.s64.s32	%rd2, %r2;
	ld.const.u32 	%r3, [%rd207];
	add.s32 	%r4, %r743, %r3;
	and.b64  	%rd211, %rd210, 65535;
	mul.lo.s64 	%rd3, %rd211, 552;
	ld.const.v4.u32 	{%r744, %r745, %r746, %r747}, [%rd208];
	st.local.v4.u32 	[%rd1], {%r744, %r745, %r746, %r747};
	ld.const.v4.u32 	{%r748, %r749, %r750, %r751}, [%rd208+16];
	st.local.v4.u32 	[%rd1+16], {%r748, %r749, %r750, %r751};
	ld.const.v4.u32 	{%r752, %r753, %r754, %r755}, [%rd208+32];
	st.local.v4.u32 	[%rd1+32], {%r752, %r753, %r754, %r755};
	ld.const.v4.u32 	{%r756, %r757, %r758, %r759}, [%rd208+48];
	st.local.v4.u32 	[%rd1+48], {%r756, %r757, %r758, %r759};
	setp.gt.s32	%p1, %r2, 255;
	@%p1 bra 	BB0_9;

	mov.u32 	%r760, 255;
	sub.s32 	%r761, %r760, %r2;
	shr.u32 	%r762, %r761, 6;
	add.s32 	%r21, %r762, 1;
	and.b32  	%r22, %r21, 3;
	setp.eq.s32	%p2, %r22, 0;
	@%p2 bra 	BB0_7;

	setp.eq.s32	%p3, %r22, 1;
	@%p3 bra 	BB0_6;

	setp.eq.s32	%p4, %r22, 2;
	@%p4 bra 	BB0_5;

	shl.b64 	%rd212, %rd2, 2;
	mov.u64 	%rd213, saes_table;
	add.s64 	%rd214, %rd213, %rd212;
	ld.const.u32 	%r763, [%rd214];
	mov.u64 	%rd215, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd216, %rd215, %rd212;
	st.shared.u32 	[%rd216], %r763;
	ld.const.u32 	%r764, [%rd214+1024];
	st.shared.u32 	[%rd216+1024], %r764;
	ld.const.u32 	%r765, [%rd214+2048];
	st.shared.u32 	[%rd216+2048], %r765;
	ld.const.u32 	%r766, [%rd214+3072];
	st.shared.u32 	[%rd216+3072], %r766;
	add.s32 	%r2, %r2, 64;

BB0_5:
	mul.wide.s32 	%rd217, %r2, 4;
	mov.u64 	%rd218, saes_table;
	add.s64 	%rd219, %rd218, %rd217;
	ld.const.u32 	%r767, [%rd219];
	mov.u64 	%rd220, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd221, %rd220, %rd217;
	st.shared.u32 	[%rd221], %r767;
	ld.const.u32 	%r768, [%rd219+1024];
	st.shared.u32 	[%rd221+1024], %r768;
	ld.const.u32 	%r769, [%rd219+2048];
	st.shared.u32 	[%rd221+2048], %r769;
	ld.const.u32 	%r770, [%rd219+3072];
	st.shared.u32 	[%rd221+3072], %r770;
	add.s32 	%r2, %r2, 64;

BB0_6:
	mul.wide.s32 	%rd222, %r2, 4;
	mov.u64 	%rd223, saes_table;
	add.s64 	%rd224, %rd223, %rd222;
	ld.const.u32 	%r771, [%rd224];
	mov.u64 	%rd225, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd226, %rd225, %rd222;
	st.shared.u32 	[%rd226], %r771;
	ld.const.u32 	%r772, [%rd224+1024];
	st.shared.u32 	[%rd226+1024], %r772;
	ld.const.u32 	%r773, [%rd224+2048];
	st.shared.u32 	[%rd226+2048], %r773;
	ld.const.u32 	%r774, [%rd224+3072];
	st.shared.u32 	[%rd226+3072], %r774;
	add.s32 	%r2, %r2, 64;

BB0_7:
	setp.lt.u32	%p5, %r21, 4;
	@%p5 bra 	BB0_9;

BB0_8:
	mul.wide.s32 	%rd227, %r2, 4;
	mov.u64 	%rd228, saes_table;
	add.s64 	%rd229, %rd228, %rd227;
	ld.const.u32 	%r775, [%rd229];
	mov.u64 	%rd230, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd231, %rd230, %rd227;
	st.shared.u32 	[%rd231], %r775;
	ld.const.u32 	%r776, [%rd229+1024];
	st.shared.u32 	[%rd231+1024], %r776;
	ld.const.u32 	%r777, [%rd229+2048];
	st.shared.u32 	[%rd231+2048], %r777;
	ld.const.u32 	%r778, [%rd229+3072];
	st.shared.u32 	[%rd231+3072], %r778;
	ld.const.u32 	%r779, [%rd229+256];
	st.shared.u32 	[%rd231+256], %r779;
	ld.const.u32 	%r780, [%rd229+1280];
	st.shared.u32 	[%rd231+1280], %r780;
	ld.const.u32 	%r781, [%rd229+2304];
	st.shared.u32 	[%rd231+2304], %r781;
	ld.const.u32 	%r782, [%rd229+3328];
	st.shared.u32 	[%rd231+3328], %r782;
	ld.const.u32 	%r783, [%rd229+512];
	st.shared.u32 	[%rd231+512], %r783;
	ld.const.u32 	%r784, [%rd229+1536];
	st.shared.u32 	[%rd231+1536], %r784;
	ld.const.u32 	%r785, [%rd229+2560];
	st.shared.u32 	[%rd231+2560], %r785;
	ld.const.u32 	%r786, [%rd229+3584];
	st.shared.u32 	[%rd231+3584], %r786;
	ld.const.u32 	%r787, [%rd229+768];
	st.shared.u32 	[%rd231+768], %r787;
	ld.const.u32 	%r788, [%rd229+1792];
	st.shared.u32 	[%rd231+1792], %r788;
	ld.const.u32 	%r789, [%rd229+2816];
	st.shared.u32 	[%rd231+2816], %r789;
	ld.const.u32 	%r790, [%rd229+3840];
	st.shared.u32 	[%rd231+3840], %r790;
	add.s32 	%r30, %r2, 256;
	setp.lt.s32	%p6, %r2, 0;
	mov.u32 	%r2, %r30;
	@%p6 bra 	BB0_8;

BB0_9:
	setp.ne.s32	%p7, %r3, 0;
	@%p7 bra 	BB0_12;

	and.b32  	%r793, %r743, 65535;
	mul.wide.u32 	%rd232, %r793, 552;
	shl.b64 	%rd233, %rd232, 4;
	add.s64 	%rd6091, %rd203, %rd233;
	mov.u32 	%r6471, -552;

BB0_11:
	ld.const.v4.u32 	{%r794, %r795, %r796, %r797}, [%rd6090];
	st.global.v4.u32 	[%rd6091], {%r794, %r795, %r796, %r797};
	ld.const.v4.u32 	{%r802, %r803, %r804, %r805}, [%rd6090+16];
	st.global.v4.u32 	[%rd6091+16], {%r802, %r803, %r804, %r805};
	ld.const.v4.u32 	{%r810, %r811, %r812, %r813}, [%rd6090+32];
	st.global.v4.u32 	[%rd6091+32], {%r810, %r811, %r812, %r813};
	ld.const.v4.u32 	{%r818, %r819, %r820, %r821}, [%rd6090+48];
	st.global.v4.u32 	[%rd6091+48], {%r818, %r819, %r820, %r821};
	ld.const.v4.u32 	{%r826, %r827, %r828, %r829}, [%rd6090+64];
	st.global.v4.u32 	[%rd6091+64], {%r826, %r827, %r828, %r829};
	ld.const.v4.u32 	{%r834, %r835, %r836, %r837}, [%rd6090+80];
	st.global.v4.u32 	[%rd6091+80], {%r834, %r835, %r836, %r837};
	ld.const.v4.u32 	{%r842, %r843, %r844, %r845}, [%rd6090+96];
	st.global.v4.u32 	[%rd6091+96], {%r842, %r843, %r844, %r845};
	ld.const.v4.u32 	{%r850, %r851, %r852, %r853}, [%rd6090+112];
	st.global.v4.u32 	[%rd6091+112], {%r850, %r851, %r852, %r853};
	ld.const.v4.u32 	{%r858, %r859, %r860, %r861}, [%rd6090+128];
	st.global.v4.u32 	[%rd6091+128], {%r858, %r859, %r860, %r861};
	ld.const.v4.u32 	{%r866, %r867, %r868, %r869}, [%rd6090+144];
	st.global.v4.u32 	[%rd6091+144], {%r866, %r867, %r868, %r869};
	ld.const.v4.u32 	{%r874, %r875, %r876, %r877}, [%rd6090+160];
	st.global.v4.u32 	[%rd6091+160], {%r874, %r875, %r876, %r877};
	ld.const.v4.u32 	{%r882, %r883, %r884, %r885}, [%rd6090+176];
	st.global.v4.u32 	[%rd6091+176], {%r882, %r883, %r884, %r885};
	add.s64 	%rd6091, %rd6091, 192;
	add.s64 	%rd6090, %rd6090, 192;
	add.s32 	%r6471, %r6471, 12;
	setp.ne.s32	%p8, %r6471, 0;
	@%p8 bra 	BB0_11;

BB0_12:
	mov.u32 	%r6729, 0;
	st.local.v4.u32 	[%rd1+32], {%r4, %r753, %r754, %r755};
	shl.b64 	%rd234, %rd2, 5;
	and.b64  	%rd9, %rd234, 4294967264;
	shl.b64 	%rd235, %rd3, 4;
	add.s64 	%rd236, %rd235, %rd203;
	ld.global.v4.u32 	{%r6730, %r6731, %r6732, %r6733}, [%rd236+8208];
	cvt.u64.u32	%rd237, %r6731;
	xor.b32  	%r895, %r744, %r4;
	xor.b32  	%r896, %r745, %r753;
	xor.b32  	%r897, %r746, %r754;
	xor.b32  	%r898, %r747, %r755;
	st.local.v4.u32 	[%rd1], {%r895, %r896, %r897, %r898};
	xor.b32  	%r899, %r748, %r756;
	xor.b32  	%r900, %r749, %r757;
	xor.b32  	%r901, %r750, %r758;
	xor.b32  	%r902, %r751, %r759;
	st.local.v4.u32 	[%rd1+16], {%r899, %r900, %r901, %r902};
	shr.u32 	%r903, %r6730, 5;
	cvt.u64.u32	%rd238, %r903;
	cvt.u16.u32	%rs61, %r903;
	and.b16  	%rs236, %rs61, 511;
	cvt.u16.u32	%rs62, %r6731;
	and.b16  	%rs237, %rs62, 511;
	and.b64  	%rd239, %rd238, 511;
	add.s64 	%rd240, %rd239, %rd3;
	shl.b64 	%rd241, %rd240, 4;
	add.s64 	%rd242, %rd203, %rd241;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd242];
	and.b64  	%rd243, %rd237, 511;
	add.s64 	%rd244, %rd243, %rd3;
	shl.b64 	%rd245, %rd244, 4;
	add.s64 	%rd246, %rd203, %rd245;
	ld.global.v4.u32 	{%r6667, %r6668, %r6669, %r6670}, [%rd246];
	mov.b64	%rd6123, {%r6730, %r6731};
	cvt.u16.u64	%rs63, %rd6123;
	and.b16  	%rs238, %rs63, 28;
	mul.wide.u32 	%rd247, %r6730, 16;
	and.b64  	%rd248, %rd247, 48;
	add.s64 	%rd6122, %rd1, %rd248;
	add.u64 	%rd12, %SPL, 0;

BB0_13:
	add.s64 	%rd15, %rd12, 16;
	and.b16  	%rs64, %rs238, 255;
	setp.ne.s16	%p9, %rs64, 20;
	@%p9 bra 	BB0_14;

	and.b64  	%rd251, %rd6123, 1;
	mov.u64 	%rd252, 1;
	shl.b64 	%rd253, %rd251, 1;
	sub.s64 	%rd254, %rd252, %rd253;
	shl.b64 	%rd255, %rd254, 4;
	add.s64 	%rd16, %rd6122, %rd255;
	shr.u64 	%rd6094, %rd6123, 61;
	cvt.u64.u16	%rd256, %rs236;
	add.s64 	%rd257, %rd256, %rd3;
	shl.b64 	%rd258, %rd257, 4;
	add.s64 	%rd6095, %rd203, %rd258;
	mov.u64 	%rd6096, 0;

BB0_16:
	cvt.u32.u64	%r912, %rd6094;
	and.b32  	%r913, %r912, 63;
	mov.u64 	%rd259, 268435456;
	shl.b64 	%rd260, %rd259, %r913;
	and.b64  	%rd261, %rd260, %rd6123;
	setp.gt.u64	%p10, %rd261, 268435455;
	ld.global.v4.u32 	{%r914, %r915, %r916, %r917}, [%rd6095];
	and.b64  	%rd22, %rd6094, 1;
	@%p10 bra 	BB0_18;
	bra.uni 	BB0_17;

BB0_18:
	setp.eq.s64	%p12, %rd22, 0;
	selp.b64	%rd441, %rd16, %rd6122, %p12;
	ld.local.v4.u32 	{%r1139, %r1140, %r1141, %r1142}, [%rd441];
	xor.b32  	%r73, %r917, %r1142;
	xor.b32  	%r72, %r916, %r1141;
	xor.b32  	%r71, %r915, %r1140;
	xor.b32  	%r70, %r914, %r1139;
	mov.b64	%rd442, {%r72, %r73};
	mov.u64 	%rd443, 0;
	st.local.v2.u64 	[%rd12], {%rd443, %rd442};
	shl.b64 	%rd446, %rd442, 1;
	xor.b64  	%rd447, %rd446, %rd442;
	st.local.v2.u64 	[%rd15], {%rd446, %rd447};
	shl.b64 	%rd448, %rd442, 2;
	xor.b64  	%rd449, %rd448, %rd442;
	st.local.v2.u64 	[%rd15+16], {%rd448, %rd449};
	shl.b64 	%rd450, %rd447, 1;
	xor.b64  	%rd451, %rd450, %rd442;
	st.local.v2.u64 	[%rd15+32], {%rd450, %rd451};
	mov.b64	%rd452, {%r70, %r71};
	shl.b64 	%rd453, %rd452, 3;
	and.b64  	%rd454, %rd453, 56;
	add.s64 	%rd455, %rd12, %rd454;
	and.b64  	%rd456, %rd452, 56;
	add.s64 	%rd457, %rd12, %rd456;
	ld.local.u64 	%rd458, [%rd457];
	shl.b64 	%rd459, %rd458, 3;
	ld.local.u64 	%rd460, [%rd455];
	xor.b64  	%rd461, %rd460, %rd459;
	shr.u64 	%rd462, %rd458, 61;
	mov.u32 	%r1147, 0;
	mov.b64	%rd463, {%r1147, %r1147};
	xor.b64  	%rd464, %rd463, %rd462;
	shr.u64 	%rd465, %rd452, 3;
	and.b64  	%rd466, %rd465, 56;
	add.s64 	%rd467, %rd12, %rd466;
	ld.local.u64 	%rd468, [%rd467];
	shl.b64 	%rd469, %rd468, 6;
	mov.b64	{%r1148, %r1149}, %rd461;
	mov.b64	%rd470, {%r1148, %r1149};
	xor.b64  	%rd471, %rd470, %rd469;
	shr.u64 	%rd472, %rd468, 58;
	mov.b64	{%r1150, %r1151}, %rd464;
	mov.b64	%rd473, {%r1150, %r1151};
	xor.b64  	%rd474, %rd473, %rd472;
	shr.u64 	%rd475, %rd452, 6;
	and.b64  	%rd476, %rd475, 56;
	add.s64 	%rd477, %rd12, %rd476;
	ld.local.u64 	%rd478, [%rd477];
	shl.b64 	%rd479, %rd478, 9;
	mov.b64	{%r1152, %r1153}, %rd471;
	mov.b64	%rd480, {%r1152, %r1153};
	xor.b64  	%rd481, %rd480, %rd479;
	shr.u64 	%rd482, %rd478, 55;
	mov.b64	{%r1154, %r1155}, %rd474;
	mov.b64	%rd483, {%r1154, %r1155};
	xor.b64  	%rd484, %rd483, %rd482;
	shr.u64 	%rd485, %rd452, 9;
	and.b64  	%rd486, %rd485, 56;
	add.s64 	%rd487, %rd12, %rd486;
	ld.local.u64 	%rd488, [%rd487];
	shl.b64 	%rd489, %rd488, 12;
	mov.b64	{%r1156, %r1157}, %rd481;
	mov.b64	%rd490, {%r1156, %r1157};
	xor.b64  	%rd491, %rd490, %rd489;
	shr.u64 	%rd492, %rd488, 52;
	mov.b64	{%r1158, %r1159}, %rd484;
	mov.b64	%rd493, {%r1158, %r1159};
	xor.b64  	%rd494, %rd493, %rd492;
	shr.u64 	%rd495, %rd452, 12;
	and.b64  	%rd496, %rd495, 56;
	add.s64 	%rd497, %rd12, %rd496;
	ld.local.u64 	%rd498, [%rd497];
	shl.b64 	%rd499, %rd498, 15;
	mov.b64	{%r1160, %r1161}, %rd491;
	mov.b64	%rd500, {%r1160, %r1161};
	xor.b64  	%rd501, %rd500, %rd499;
	shr.u64 	%rd502, %rd498, 49;
	mov.b64	{%r1162, %r1163}, %rd494;
	mov.b64	%rd503, {%r1162, %r1163};
	xor.b64  	%rd504, %rd503, %rd502;
	shr.u64 	%rd505, %rd452, 15;
	and.b64  	%rd506, %rd505, 56;
	add.s64 	%rd507, %rd12, %rd506;
	ld.local.u64 	%rd508, [%rd507];
	shl.b64 	%rd509, %rd508, 18;
	mov.b64	{%r1164, %r1165}, %rd501;
	mov.b64	%rd510, {%r1164, %r1165};
	xor.b64  	%rd511, %rd510, %rd509;
	shr.u64 	%rd512, %rd508, 46;
	mov.b64	{%r1166, %r1167}, %rd504;
	mov.b64	%rd513, {%r1166, %r1167};
	xor.b64  	%rd514, %rd513, %rd512;
	shr.u64 	%rd515, %rd452, 18;
	and.b64  	%rd516, %rd515, 56;
	add.s64 	%rd517, %rd12, %rd516;
	ld.local.u64 	%rd518, [%rd517];
	shl.b64 	%rd519, %rd518, 21;
	mov.b64	{%r1168, %r1169}, %rd511;
	mov.b64	%rd520, {%r1168, %r1169};
	xor.b64  	%rd521, %rd520, %rd519;
	shr.u64 	%rd522, %rd518, 43;
	mov.b64	{%r1170, %r1171}, %rd514;
	mov.b64	%rd523, {%r1170, %r1171};
	xor.b64  	%rd524, %rd523, %rd522;
	shr.u64 	%rd525, %rd452, 21;
	and.b64  	%rd526, %rd525, 56;
	add.s64 	%rd527, %rd12, %rd526;
	ld.local.u64 	%rd528, [%rd527];
	shl.b64 	%rd529, %rd528, 24;
	mov.b64	{%r1172, %r1173}, %rd521;
	mov.b64	%rd530, {%r1172, %r1173};
	xor.b64  	%rd531, %rd530, %rd529;
	shr.u64 	%rd532, %rd528, 40;
	mov.b64	{%r1174, %r1175}, %rd524;
	mov.b64	%rd533, {%r1174, %r1175};
	xor.b64  	%rd534, %rd533, %rd532;
	shr.u64 	%rd535, %rd452, 24;
	and.b64  	%rd536, %rd535, 56;
	add.s64 	%rd537, %rd12, %rd536;
	ld.local.u64 	%rd538, [%rd537];
	shl.b64 	%rd539, %rd538, 27;
	mov.b64	{%r1176, %r1177}, %rd531;
	mov.b64	%rd540, {%r1176, %r1177};
	xor.b64  	%rd541, %rd540, %rd539;
	shr.u64 	%rd542, %rd538, 37;
	mov.b64	{%r1178, %r1179}, %rd534;
	mov.b64	%rd543, {%r1178, %r1179};
	xor.b64  	%rd544, %rd543, %rd542;
	shr.u64 	%rd545, %rd452, 27;
	and.b64  	%rd546, %rd545, 56;
	add.s64 	%rd547, %rd12, %rd546;
	ld.local.u64 	%rd548, [%rd547];
	shl.b64 	%rd549, %rd548, 30;
	mov.b64	{%r1180, %r1181}, %rd541;
	mov.b64	%rd550, {%r1180, %r1181};
	xor.b64  	%rd551, %rd550, %rd549;
	shr.u64 	%rd552, %rd548, 34;
	mov.b64	{%r1182, %r1183}, %rd544;
	mov.b64	%rd553, {%r1182, %r1183};
	xor.b64  	%rd554, %rd553, %rd552;
	shr.u64 	%rd555, %rd452, 30;
	and.b64  	%rd556, %rd555, 56;
	add.s64 	%rd557, %rd12, %rd556;
	ld.local.u64 	%rd558, [%rd557];
	shl.b64 	%rd559, %rd558, 33;
	mov.b64	{%r1184, %r1185}, %rd551;
	mov.b64	%rd560, {%r1184, %r1185};
	xor.b64  	%rd561, %rd560, %rd559;
	shr.u64 	%rd562, %rd558, 31;
	mov.b64	{%r1186, %r1187}, %rd554;
	mov.b64	%rd563, {%r1186, %r1187};
	xor.b64  	%rd564, %rd563, %rd562;
	shr.u64 	%rd565, %rd452, 33;
	and.b64  	%rd566, %rd565, 56;
	add.s64 	%rd567, %rd12, %rd566;
	ld.local.u64 	%rd568, [%rd567];
	shl.b64 	%rd569, %rd568, 36;
	mov.b64	{%r1188, %r1189}, %rd561;
	mov.b64	%rd570, {%r1188, %r1189};
	xor.b64  	%rd571, %rd570, %rd569;
	shr.u64 	%rd572, %rd568, 28;
	mov.b64	{%r1190, %r1191}, %rd564;
	mov.b64	%rd573, {%r1190, %r1191};
	xor.b64  	%rd574, %rd573, %rd572;
	shr.u64 	%rd575, %rd452, 36;
	and.b64  	%rd576, %rd575, 56;
	add.s64 	%rd577, %rd12, %rd576;
	ld.local.u64 	%rd578, [%rd577];
	shl.b64 	%rd579, %rd578, 39;
	mov.b64	{%r1192, %r1193}, %rd571;
	mov.b64	%rd580, {%r1192, %r1193};
	xor.b64  	%rd581, %rd580, %rd579;
	shr.u64 	%rd582, %rd578, 25;
	mov.b64	{%r1194, %r1195}, %rd574;
	mov.b64	%rd583, {%r1194, %r1195};
	xor.b64  	%rd584, %rd583, %rd582;
	shr.u64 	%rd585, %rd452, 39;
	and.b64  	%rd586, %rd585, 56;
	add.s64 	%rd587, %rd12, %rd586;
	ld.local.u64 	%rd588, [%rd587];
	shl.b64 	%rd589, %rd588, 42;
	mov.b64	{%r1196, %r1197}, %rd581;
	mov.b64	%rd590, {%r1196, %r1197};
	xor.b64  	%rd591, %rd590, %rd589;
	shr.u64 	%rd592, %rd588, 22;
	mov.b64	{%r1198, %r1199}, %rd584;
	mov.b64	%rd593, {%r1198, %r1199};
	xor.b64  	%rd594, %rd593, %rd592;
	shr.u64 	%rd595, %rd452, 42;
	and.b64  	%rd596, %rd595, 56;
	add.s64 	%rd597, %rd12, %rd596;
	ld.local.u64 	%rd598, [%rd597];
	shl.b64 	%rd599, %rd598, 45;
	mov.b64	{%r1200, %r1201}, %rd591;
	mov.b64	%rd600, {%r1200, %r1201};
	xor.b64  	%rd601, %rd600, %rd599;
	shr.u64 	%rd602, %rd598, 19;
	mov.b64	{%r1202, %r1203}, %rd594;
	mov.b64	%rd603, {%r1202, %r1203};
	xor.b64  	%rd604, %rd603, %rd602;
	shr.u64 	%rd605, %rd452, 45;
	and.b64  	%rd606, %rd605, 56;
	add.s64 	%rd607, %rd12, %rd606;
	ld.local.u64 	%rd608, [%rd607];
	shl.b64 	%rd609, %rd608, 48;
	mov.b64	{%r1204, %r1205}, %rd601;
	mov.b64	%rd610, {%r1204, %r1205};
	xor.b64  	%rd611, %rd610, %rd609;
	shr.u64 	%rd612, %rd608, 16;
	mov.b64	{%r1206, %r1207}, %rd604;
	mov.b64	%rd613, {%r1206, %r1207};
	xor.b64  	%rd614, %rd613, %rd612;
	shr.u64 	%rd615, %rd452, 48;
	and.b64  	%rd616, %rd615, 56;
	add.s64 	%rd617, %rd12, %rd616;
	ld.local.u64 	%rd618, [%rd617];
	shl.b64 	%rd619, %rd618, 51;
	mov.b64	{%r1208, %r1209}, %rd611;
	mov.b64	%rd620, {%r1208, %r1209};
	xor.b64  	%rd621, %rd620, %rd619;
	shr.u64 	%rd622, %rd618, 13;
	mov.b64	{%r1210, %r1211}, %rd614;
	mov.b64	%rd623, {%r1210, %r1211};
	xor.b64  	%rd624, %rd623, %rd622;
	shr.u64 	%rd625, %rd452, 51;
	and.b64  	%rd626, %rd625, 56;
	add.s64 	%rd627, %rd12, %rd626;
	ld.local.u64 	%rd628, [%rd627];
	shl.b64 	%rd629, %rd628, 54;
	mov.b64	{%r1212, %r1213}, %rd621;
	mov.b64	%rd630, {%r1212, %r1213};
	xor.b64  	%rd631, %rd630, %rd629;
	shr.u64 	%rd632, %rd628, 10;
	mov.b64	{%r1214, %r1215}, %rd624;
	mov.b64	%rd633, {%r1214, %r1215};
	xor.b64  	%rd634, %rd633, %rd632;
	shr.u64 	%rd635, %rd452, 54;
	and.b64  	%rd636, %rd635, 56;
	add.s64 	%rd637, %rd12, %rd636;
	ld.local.u64 	%rd638, [%rd637];
	shl.b64 	%rd639, %rd638, 57;
	mov.b64	{%r1216, %r1217}, %rd631;
	mov.b64	%rd640, {%r1216, %r1217};
	xor.b64  	%rd641, %rd640, %rd639;
	shr.u64 	%rd642, %rd638, 7;
	mov.b64	{%r1218, %r1219}, %rd634;
	mov.b64	%rd643, {%r1218, %r1219};
	xor.b64  	%rd644, %rd643, %rd642;
	shr.u64 	%rd645, %rd452, 57;
	and.b64  	%rd646, %rd645, 56;
	add.s64 	%rd647, %rd12, %rd646;
	ld.local.u64 	%rd648, [%rd647];
	shl.b64 	%rd649, %rd648, 60;
	mov.b64	{%r1220, %r1221}, %rd641;
	mov.b64	%rd650, {%r1220, %r1221};
	xor.b64  	%rd651, %rd650, %rd649;
	shr.u64 	%rd652, %rd648, 4;
	mov.b64	{%r1222, %r1223}, %rd644;
	mov.b64	%rd653, {%r1222, %r1223};
	xor.b64  	%rd654, %rd653, %rd652;
	shr.u64 	%rd655, %rd452, 60;
	and.b64  	%rd656, %rd655, 8;
	add.s64 	%rd657, %rd12, %rd656;
	ld.local.u64 	%rd658, [%rd657];
	shl.b64 	%rd659, %rd658, 63;
	mov.b64	{%r1224, %r1225}, %rd651;
	mov.b64	%rd660, {%r1224, %r1225};
	xor.b64  	%rd661, %rd660, %rd659;
	shr.u64 	%rd662, %rd658, 1;
	mov.b64	{%r1226, %r1227}, %rd654;
	mov.b64	%rd663, {%r1226, %r1227};
	xor.b64  	%rd664, %rd663, %rd662;
	mov.b64	{%r6491, %r6492}, %rd664;
	mov.b64	{%r6493, %r6494}, %rd661;
	setp.gt.s32	%p13, %r73, -1;
	@%p13 bra 	BB0_20;

	shr.u32 	%r1228, %r70, 1;
	shl.b32 	%r1229, %r71, 31;
	shr.u32 	%r1230, %r71, 1;
	and.b32  	%r1231, %r1228, 1533916891;
	or.b32  	%r1232, %r1231, %r1229;
	xor.b32  	%r6491, %r6491, %r1232;
	and.b32  	%r1233, %r1230, 920350134;
	xor.b32  	%r6492, %r6492, %r1233;

BB0_20:
	and.b32  	%r1234, %r73, 1073741824;
	setp.eq.s32	%p14, %r1234, 0;
	@%p14 bra 	BB0_22;

	shr.u32 	%r1235, %r70, 2;
	shl.b32 	%r1236, %r71, 30;
	or.b32  	%r1237, %r1235, %r1236;
	shr.u32 	%r1238, %r71, 2;
	and.b32  	%r1239, %r1237, 1227133513;
	xor.b32  	%r6491, %r6491, %r1239;
	and.b32  	%r1240, %r1238, 306783378;
	xor.b32  	%r6492, %r6492, %r1240;

BB0_22:
	xor.b32  	%r6730, %r6493, %r6730;
	xor.b32  	%r6731, %r6494, %r6731;
	xor.b32  	%r6732, %r6491, %r6732;
	xor.b32  	%r6733, %r6492, %r6733;
	bra.uni 	BB0_23;

BB0_17:
	setp.eq.s64	%p11, %rd22, 0;
	selp.b64	%rd262, %rd6122, %rd16, %p11;
	add.s64 	%rd23, %rd6096, 1;
	shl.b64 	%rd263, %rd6096, 6;
	add.s64 	%rd264, %rd263, %rd6095;
	and.b32  	%r918, %r914, 255;
	mul.wide.u32 	%rd265, %r918, 4;
	mov.u64 	%rd266, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd267, %rd266, %rd265;
	shr.u32 	%r919, %r914, 8;
	mul.wide.u32 	%rd268, %r915, 4;
	and.b64  	%rd269, %rd268, 1020;
	add.s64 	%rd270, %rd266, %rd269;
	shr.u32 	%r920, %r915, 8;
	mul.wide.u32 	%rd271, %r916, 4;
	and.b64  	%rd272, %rd271, 1020;
	add.s64 	%rd273, %rd266, %rd272;
	shr.u32 	%r921, %r916, 8;
	mul.wide.u32 	%rd274, %r917, 4;
	and.b64  	%rd275, %rd274, 1020;
	add.s64 	%rd276, %rd266, %rd275;
	shr.u32 	%r922, %r917, 8;
	mul.wide.u32 	%rd277, %r920, 4;
	and.b64  	%rd278, %rd277, 1020;
	add.s64 	%rd279, %rd278, %rd266;
	ld.shared.u32 	%r923, [%rd279+1024];
	ld.shared.u32 	%r924, [%rd267];
	xor.b32  	%r925, %r923, %r924;
	shr.u32 	%r926, %r915, 16;
	mul.wide.u32 	%rd280, %r921, 4;
	and.b64  	%rd281, %rd280, 1020;
	add.s64 	%rd282, %rd281, %rd266;
	ld.shared.u32 	%r927, [%rd282+1024];
	ld.shared.u32 	%r928, [%rd270];
	xor.b32  	%r929, %r927, %r928;
	shr.u32 	%r930, %r916, 16;
	mul.wide.u32 	%rd283, %r922, 4;
	and.b64  	%rd284, %rd283, 1020;
	add.s64 	%rd285, %rd284, %rd266;
	ld.shared.u32 	%r931, [%rd285+1024];
	ld.shared.u32 	%r932, [%rd273];
	xor.b32  	%r933, %r931, %r932;
	shr.u32 	%r934, %r917, 16;
	mul.wide.u32 	%rd286, %r919, 4;
	and.b64  	%rd287, %rd286, 1020;
	add.s64 	%rd288, %rd287, %rd266;
	ld.shared.u32 	%r935, [%rd288+1024];
	ld.shared.u32 	%r936, [%rd276];
	xor.b32  	%r937, %r935, %r936;
	shr.u32 	%r938, %r914, 16;
	mul.wide.u32 	%rd289, %r930, 4;
	and.b64  	%rd290, %rd289, 1020;
	add.s64 	%rd291, %rd290, %rd266;
	ld.shared.u32 	%r939, [%rd291+2048];
	xor.b32  	%r940, %r925, %r939;
	shr.u32 	%r941, %r916, 24;
	mul.wide.u32 	%rd292, %r934, 4;
	and.b64  	%rd293, %rd292, 1020;
	add.s64 	%rd294, %rd293, %rd266;
	ld.shared.u32 	%r942, [%rd294+2048];
	xor.b32  	%r943, %r929, %r942;
	shr.u32 	%r944, %r917, 24;
	mul.wide.u32 	%rd295, %r938, 4;
	and.b64  	%rd296, %rd295, 1020;
	add.s64 	%rd297, %rd296, %rd266;
	ld.shared.u32 	%r945, [%rd297+2048];
	xor.b32  	%r946, %r933, %r945;
	shr.u32 	%r947, %r914, 24;
	mul.wide.u32 	%rd298, %r926, 4;
	and.b64  	%rd299, %rd298, 1020;
	add.s64 	%rd300, %rd299, %rd266;
	ld.shared.u32 	%r948, [%rd300+2048];
	xor.b32  	%r949, %r937, %r948;
	shr.u32 	%r950, %r915, 24;
	mul.wide.u32 	%rd301, %r944, 4;
	add.s64 	%rd302, %rd301, %rd266;
	ld.shared.u32 	%r951, [%rd302+3072];
	mul.wide.u32 	%rd303, %r947, 4;
	add.s64 	%rd304, %rd303, %rd266;
	ld.shared.u32 	%r952, [%rd304+3072];
	mul.wide.u32 	%rd305, %r950, 4;
	add.s64 	%rd306, %rd305, %rd266;
	ld.shared.u32 	%r953, [%rd306+3072];
	mul.wide.u32 	%rd307, %r941, 4;
	add.s64 	%rd308, %rd307, %rd266;
	ld.shared.u32 	%r954, [%rd308+3072];
	xor.b32  	%r955, %r949, %r954;
	xor.b32  	%r956, %r946, %r953;
	xor.b32  	%r957, %r943, %r952;
	xor.b32  	%r958, %r940, %r951;
	ld.global.v4.u32 	{%r959, %r960, %r961, %r962}, [%rd264+16];
	ld.local.v4.u32 	{%r967, %r968, %r969, %r970}, [%rd262];
	mul.wide.u32 	%rd309, %r967, 4;
	and.b64  	%rd310, %rd309, 1020;
	add.s64 	%rd311, %rd266, %rd310;
	shr.u32 	%r972, %r967, 8;
	mul.wide.u32 	%rd312, %r968, 4;
	and.b64  	%rd313, %rd312, 1020;
	add.s64 	%rd314, %rd266, %rd313;
	shr.u32 	%r974, %r968, 8;
	mul.wide.u32 	%rd315, %r969, 4;
	and.b64  	%rd316, %rd315, 1020;
	add.s64 	%rd317, %rd266, %rd316;
	shr.u32 	%r976, %r969, 8;
	mul.wide.u32 	%rd318, %r970, 4;
	and.b64  	%rd319, %rd318, 1020;
	add.s64 	%rd320, %rd266, %rd319;
	shr.u32 	%r978, %r970, 8;
	mul.wide.u32 	%rd321, %r974, 4;
	and.b64  	%rd322, %rd321, 1020;
	add.s64 	%rd323, %rd322, %rd266;
	ld.shared.u32 	%r979, [%rd323+1024];
	ld.shared.u32 	%r980, [%rd311];
	xor.b32  	%r981, %r979, %r980;
	shr.u32 	%r982, %r968, 16;
	mul.wide.u32 	%rd324, %r976, 4;
	and.b64  	%rd325, %rd324, 1020;
	add.s64 	%rd326, %rd325, %rd266;
	ld.shared.u32 	%r983, [%rd326+1024];
	ld.shared.u32 	%r984, [%rd314];
	xor.b32  	%r985, %r983, %r984;
	shr.u32 	%r986, %r969, 16;
	mul.wide.u32 	%rd327, %r978, 4;
	and.b64  	%rd328, %rd327, 1020;
	add.s64 	%rd329, %rd328, %rd266;
	ld.shared.u32 	%r987, [%rd329+1024];
	ld.shared.u32 	%r988, [%rd317];
	xor.b32  	%r989, %r987, %r988;
	shr.u32 	%r990, %r970, 16;
	mul.wide.u32 	%rd330, %r972, 4;
	and.b64  	%rd331, %rd330, 1020;
	add.s64 	%rd332, %rd331, %rd266;
	ld.shared.u32 	%r991, [%rd332+1024];
	ld.shared.u32 	%r992, [%rd320];
	xor.b32  	%r993, %r991, %r992;
	shr.u32 	%r994, %r967, 16;
	mul.wide.u32 	%rd333, %r986, 4;
	and.b64  	%rd334, %rd333, 1020;
	add.s64 	%rd335, %rd334, %rd266;
	ld.shared.u32 	%r995, [%rd335+2048];
	xor.b32  	%r996, %r981, %r995;
	shr.u32 	%r997, %r969, 24;
	mul.wide.u32 	%rd336, %r990, 4;
	and.b64  	%rd337, %rd336, 1020;
	add.s64 	%rd338, %rd337, %rd266;
	ld.shared.u32 	%r998, [%rd338+2048];
	xor.b32  	%r999, %r985, %r998;
	shr.u32 	%r1000, %r970, 24;
	mul.wide.u32 	%rd339, %r994, 4;
	and.b64  	%rd340, %rd339, 1020;
	add.s64 	%rd341, %rd340, %rd266;
	ld.shared.u32 	%r1001, [%rd341+2048];
	xor.b32  	%r1002, %r989, %r1001;
	shr.u32 	%r1003, %r967, 24;
	mul.wide.u32 	%rd342, %r982, 4;
	and.b64  	%rd343, %rd342, 1020;
	add.s64 	%rd344, %rd343, %rd266;
	ld.shared.u32 	%r1004, [%rd344+2048];
	xor.b32  	%r1005, %r993, %r1004;
	shr.u32 	%r1006, %r968, 24;
	mul.wide.u32 	%rd345, %r1000, 4;
	add.s64 	%rd346, %rd345, %rd266;
	ld.shared.u32 	%r1007, [%rd346+3072];
	mul.wide.u32 	%rd347, %r1003, 4;
	add.s64 	%rd348, %rd347, %rd266;
	ld.shared.u32 	%r1008, [%rd348+3072];
	mul.wide.u32 	%rd349, %r1006, 4;
	add.s64 	%rd350, %rd349, %rd266;
	ld.shared.u32 	%r1009, [%rd350+3072];
	mul.wide.u32 	%rd351, %r997, 4;
	add.s64 	%rd352, %rd351, %rd266;
	ld.shared.u32 	%r1010, [%rd352+3072];
	xor.b32  	%r1011, %r1005, %r1010;
	xor.b32  	%r1012, %r1002, %r1009;
	xor.b32  	%r1013, %r999, %r1008;
	xor.b32  	%r1014, %r996, %r1007;
	ld.global.v4.u32 	{%r1015, %r1016, %r1017, %r1018}, [%rd264+32];
	xor.b32  	%r1023, %r958, %r959;
	mul.wide.u32 	%rd353, %r1023, 4;
	and.b64  	%rd354, %rd353, 1020;
	add.s64 	%rd355, %rd266, %rd354;
	shr.u32 	%r1024, %r1023, 8;
	xor.b32  	%r1025, %r957, %r960;
	mul.wide.u32 	%rd356, %r1025, 4;
	and.b64  	%rd357, %rd356, 1020;
	add.s64 	%rd358, %rd266, %rd357;
	shr.u32 	%r1026, %r1025, 8;
	xor.b32  	%r1027, %r956, %r961;
	mul.wide.u32 	%rd359, %r1027, 4;
	and.b64  	%rd360, %rd359, 1020;
	add.s64 	%rd361, %rd266, %rd360;
	shr.u32 	%r1028, %r1027, 8;
	xor.b32  	%r1029, %r955, %r962;
	mul.wide.u32 	%rd362, %r1029, 4;
	and.b64  	%rd363, %rd362, 1020;
	add.s64 	%rd364, %rd266, %rd363;
	shr.u32 	%r1030, %r1029, 8;
	mul.wide.u32 	%rd365, %r1026, 4;
	and.b64  	%rd366, %rd365, 1020;
	add.s64 	%rd367, %rd366, %rd266;
	ld.shared.u32 	%r1031, [%rd367+1024];
	ld.shared.u32 	%r1032, [%rd355];
	xor.b32  	%r1033, %r1031, %r1032;
	shr.u32 	%r1034, %r1025, 16;
	mul.wide.u32 	%rd368, %r1028, 4;
	and.b64  	%rd369, %rd368, 1020;
	add.s64 	%rd370, %rd369, %rd266;
	ld.shared.u32 	%r1035, [%rd370+1024];
	ld.shared.u32 	%r1036, [%rd358];
	xor.b32  	%r1037, %r1035, %r1036;
	shr.u32 	%r1038, %r1027, 16;
	mul.wide.u32 	%rd371, %r1030, 4;
	and.b64  	%rd372, %rd371, 1020;
	add.s64 	%rd373, %rd372, %rd266;
	ld.shared.u32 	%r1039, [%rd373+1024];
	ld.shared.u32 	%r1040, [%rd361];
	xor.b32  	%r1041, %r1039, %r1040;
	shr.u32 	%r1042, %r1029, 16;
	mul.wide.u32 	%rd374, %r1024, 4;
	and.b64  	%rd375, %rd374, 1020;
	add.s64 	%rd376, %rd375, %rd266;
	ld.shared.u32 	%r1043, [%rd376+1024];
	ld.shared.u32 	%r1044, [%rd364];
	xor.b32  	%r1045, %r1043, %r1044;
	shr.u32 	%r1046, %r1023, 16;
	mul.wide.u32 	%rd377, %r1038, 4;
	and.b64  	%rd378, %rd377, 1020;
	add.s64 	%rd379, %rd378, %rd266;
	ld.shared.u32 	%r1047, [%rd379+2048];
	xor.b32  	%r1048, %r1033, %r1047;
	shr.u32 	%r1049, %r1027, 24;
	mul.wide.u32 	%rd380, %r1042, 4;
	and.b64  	%rd381, %rd380, 1020;
	add.s64 	%rd382, %rd381, %rd266;
	ld.shared.u32 	%r1050, [%rd382+2048];
	xor.b32  	%r1051, %r1037, %r1050;
	shr.u32 	%r1052, %r1029, 24;
	mul.wide.u32 	%rd383, %r1046, 4;
	and.b64  	%rd384, %rd383, 1020;
	add.s64 	%rd385, %rd384, %rd266;
	ld.shared.u32 	%r1053, [%rd385+2048];
	xor.b32  	%r1054, %r1041, %r1053;
	shr.u32 	%r1055, %r1023, 24;
	mul.wide.u32 	%rd386, %r1034, 4;
	and.b64  	%rd387, %rd386, 1020;
	add.s64 	%rd388, %rd387, %rd266;
	ld.shared.u32 	%r1056, [%rd388+2048];
	xor.b32  	%r1057, %r1045, %r1056;
	shr.u32 	%r1058, %r1025, 24;
	mul.wide.u32 	%rd389, %r1052, 4;
	add.s64 	%rd390, %rd389, %rd266;
	ld.shared.u32 	%r1059, [%rd390+3072];
	mul.wide.u32 	%rd391, %r1055, 4;
	add.s64 	%rd392, %rd391, %rd266;
	ld.shared.u32 	%r1060, [%rd392+3072];
	mul.wide.u32 	%rd393, %r1058, 4;
	add.s64 	%rd394, %rd393, %rd266;
	ld.shared.u32 	%r1061, [%rd394+3072];
	mul.wide.u32 	%rd395, %r1049, 4;
	add.s64 	%rd396, %rd395, %rd266;
	ld.shared.u32 	%r1062, [%rd396+3072];
	xor.b32  	%r1063, %r1057, %r1062;
	xor.b32  	%r1064, %r1051, %r1060;
	xor.b32  	%r1065, %r1048, %r1059;
	xor.b32  	%r1066, %r1054, %r1061;
	ld.global.v4.u32 	{%r1067, %r1068, %r1069, %r1070}, [%rd264+48];
	xor.b32  	%r1075, %r1066, %r1069;
	xor.b32  	%r1076, %r1065, %r1067;
	xor.b32  	%r1077, %r1014, %r1015;
	mul.wide.u32 	%rd397, %r1077, 4;
	and.b64  	%rd398, %rd397, 1020;
	add.s64 	%rd399, %rd266, %rd398;
	shr.u32 	%r1078, %r1077, 8;
	xor.b32  	%r1079, %r1013, %r1016;
	mul.wide.u32 	%rd400, %r1079, 4;
	and.b64  	%rd401, %rd400, 1020;
	add.s64 	%rd402, %rd266, %rd401;
	shr.u32 	%r1080, %r1079, 8;
	xor.b32  	%r1081, %r1012, %r1017;
	mul.wide.u32 	%rd403, %r1081, 4;
	and.b64  	%rd404, %rd403, 1020;
	add.s64 	%rd405, %rd266, %rd404;
	shr.u32 	%r1082, %r1081, 8;
	xor.b32  	%r1083, %r1011, %r1018;
	mul.wide.u32 	%rd406, %r1083, 4;
	and.b64  	%rd407, %rd406, 1020;
	add.s64 	%rd408, %rd266, %rd407;
	shr.u32 	%r1084, %r1083, 8;
	mul.wide.u32 	%rd409, %r1080, 4;
	and.b64  	%rd410, %rd409, 1020;
	add.s64 	%rd411, %rd410, %rd266;
	ld.shared.u32 	%r1085, [%rd411+1024];
	ld.shared.u32 	%r1086, [%rd399];
	xor.b32  	%r1087, %r1085, %r1086;
	shr.u32 	%r1088, %r1079, 16;
	mul.wide.u32 	%rd412, %r1082, 4;
	and.b64  	%rd413, %rd412, 1020;
	add.s64 	%rd414, %rd413, %rd266;
	ld.shared.u32 	%r1089, [%rd414+1024];
	ld.shared.u32 	%r1090, [%rd402];
	xor.b32  	%r1091, %r1089, %r1090;
	shr.u32 	%r1092, %r1081, 16;
	mul.wide.u32 	%rd415, %r1084, 4;
	and.b64  	%rd416, %rd415, 1020;
	add.s64 	%rd417, %rd416, %rd266;
	ld.shared.u32 	%r1093, [%rd417+1024];
	ld.shared.u32 	%r1094, [%rd405];
	xor.b32  	%r1095, %r1093, %r1094;
	shr.u32 	%r1096, %r1083, 16;
	mul.wide.u32 	%rd418, %r1078, 4;
	and.b64  	%rd419, %rd418, 1020;
	add.s64 	%rd420, %rd419, %rd266;
	ld.shared.u32 	%r1097, [%rd420+1024];
	ld.shared.u32 	%r1098, [%rd408];
	xor.b32  	%r1099, %r1097, %r1098;
	shr.u32 	%r1100, %r1077, 16;
	mul.wide.u32 	%rd421, %r1092, 4;
	and.b64  	%rd422, %rd421, 1020;
	add.s64 	%rd423, %rd422, %rd266;
	ld.shared.u32 	%r1101, [%rd423+2048];
	xor.b32  	%r1102, %r1087, %r1101;
	shr.u32 	%r1103, %r1081, 24;
	mul.wide.u32 	%rd424, %r1096, 4;
	and.b64  	%rd425, %rd424, 1020;
	add.s64 	%rd426, %rd425, %rd266;
	ld.shared.u32 	%r1104, [%rd426+2048];
	xor.b32  	%r1105, %r1091, %r1104;
	shr.u32 	%r1106, %r1083, 24;
	mul.wide.u32 	%rd427, %r1100, 4;
	and.b64  	%rd428, %rd427, 1020;
	add.s64 	%rd429, %rd428, %rd266;
	ld.shared.u32 	%r1107, [%rd429+2048];
	xor.b32  	%r1108, %r1095, %r1107;
	shr.u32 	%r1109, %r1077, 24;
	mul.wide.u32 	%rd430, %r1088, 4;
	and.b64  	%rd431, %rd430, 1020;
	add.s64 	%rd432, %rd431, %rd266;
	ld.shared.u32 	%r1110, [%rd432+2048];
	xor.b32  	%r1111, %r1099, %r1110;
	shr.u32 	%r1112, %r1079, 24;
	mul.wide.u32 	%rd433, %r1106, 4;
	add.s64 	%rd434, %rd433, %rd266;
	ld.shared.u32 	%r1113, [%rd434+3072];
	mul.wide.u32 	%rd435, %r1109, 4;
	add.s64 	%rd436, %rd435, %rd266;
	ld.shared.u32 	%r1114, [%rd436+3072];
	mul.wide.u32 	%rd437, %r1112, 4;
	add.s64 	%rd438, %rd437, %rd266;
	ld.shared.u32 	%r1115, [%rd438+3072];
	mul.wide.u32 	%rd439, %r1103, 4;
	add.s64 	%rd440, %rd439, %rd266;
	ld.shared.u32 	%r1116, [%rd440+3072];
	xor.b32  	%r1117, %r1102, %r1113;
	xor.b32  	%r1118, %r1105, %r1114;
	xor.b32  	%r1119, %r1111, %r1116;
	xor.b32  	%r1120, %r1108, %r1115;
	ld.global.v4.u32 	{%r1121, %r1122, %r1123, %r1124}, [%rd264+64];
	xor.b32  	%r1129, %r1120, %r1123;
	xor.b32  	%r1130, %r1119, %r1124;
	xor.b32  	%r1131, %r1118, %r1122;
	xor.b32  	%r1132, %r1117, %r1121;
	xor.b32  	%r1133, %r1064, %r1068;
	xor.b32  	%r1134, %r1063, %r1070;
	xor.b32  	%r1135, %r1132, %r6731;
	xor.b32  	%r1136, %r1131, %r6733;
	xor.b32  	%r1137, %r1133, %r6732;
	xor.b32  	%r1138, %r1076, %r6730;
	xor.b32  	%r6733, %r1130, %r1136;
	xor.b32  	%r6731, %r1129, %r1135;
	xor.b32  	%r6730, %r1075, %r1138;
	xor.b32  	%r6732, %r1134, %r1137;
	mov.u64 	%rd6096, %rd23;

BB0_23:
	add.s64 	%rd6095, %rd6095, 16;
	add.s64 	%rd26, %rd6094, -1;
	setp.ne.s64	%p15, %rd6094, 0;
	mov.u64 	%rd6094, %rd26;
	@%p15 bra 	BB0_16;

	cvt.u64.u16	%rd6074, %rs236;
	add.s64 	%rd6073, %rd6074, %rd3;
	shl.b64 	%rd6072, %rd6073, 4;
	add.s64 	%rd6071, %rd203, %rd6072;
	mov.b64	%rd665, {%r6732, %r6733};
	mov.b64	%rd666, {%r6730, %r6731};
	cvt.u32.u64	%r1241, %rd666;
	cvt.s32.s16 	%r1242, %r1241;
	mov.b64	%rd667, {%r6723, %r6724};
	mov.b64	%rd668, {%r6721, %r6722};
	cvt.u32.u64	%r1243, %rd668;
	cvt.s32.s16 	%r1244, %r1243;
	mad.lo.s32 	%r1245, %r1242, %r1244, 16384;
	bfe.u32 	%r1246, %r1245, 15, 16;
	shr.s32 	%r1247, %r1241, 16;
	shr.s32 	%r1248, %r1243, 16;
	mad.lo.s32 	%r1249, %r1247, %r1248, 16384;
	mov.u32 	%r1250, 0;
	mov.b64	%rd669, {%r1246, %r1250};
	shl.b32 	%r1251, %r1249, 1;
	cvt.u64.u32	%rd670, %r1251;
	and.b64  	%rd671, %rd669, 65535;
	shr.u64 	%rd672, %rd665, 16;
	shr.u64 	%rd673, %rd666, 16;
	cvt.u32.u64	%r1252, %rd673;
	shr.s32 	%r1253, %r1252, 16;
	shr.u64 	%rd674, %rd667, 16;
	shr.u64 	%rd675, %rd668, 16;
	cvt.u32.u64	%r1254, %rd675;
	shr.s32 	%r1255, %r1254, 16;
	mad.lo.s32 	%r1256, %r1253, %r1255, 16384;
	mul.wide.u32 	%rd676, %r1256, 131072;
	shr.u64 	%rd677, %rd666, 32;
	cvt.u32.u64	%r1257, %rd677;
	shr.s32 	%r1258, %r1257, 16;
	shr.u64 	%rd678, %rd668, 32;
	cvt.u32.u64	%r1259, %rd678;
	shr.s32 	%r1260, %r1259, 16;
	mad.lo.s32 	%r1261, %r1258, %r1260, 16384;
	shr.u32 	%r1262, %r1261, 15;
	cvt.u64.u32	%rd679, %r1262;
	shl.b64 	%rd680, %rd679, 48;
	and.b64  	%rd681, %rd676, 281470681743360;
	shl.b64 	%rd682, %rd665, 16;
	cvt.u32.u64	%r1263, %rd682;
	shr.s32 	%r1264, %r1263, 16;
	shl.b64 	%rd683, %rd667, 16;
	cvt.u32.u64	%r1265, %rd683;
	shr.s32 	%r1266, %r1265, 16;
	mad.lo.s32 	%r1267, %r1264, %r1266, 16384;
	shr.u32 	%r1268, %r1267, 15;
	cvt.u64.u32	%rd684, %r1268;
	cvt.u32.u64	%r1269, %rd665;
	shr.s32 	%r1270, %r1269, 16;
	cvt.u32.u64	%r1271, %rd667;
	shr.s32 	%r1272, %r1271, 16;
	mad.lo.s32 	%r1273, %r1270, %r1272, 16384;
	and.b64  	%rd685, %rd684, 65535;
	and.b64  	%rd686, %rd670, 4294901760;
	or.b64  	%rd687, %rd686, %rd671;
	or.b64  	%rd688, %rd687, %rd681;
	shl.b32 	%r1274, %r1273, 1;
	cvt.u64.u32	%rd689, %r1274;
	cvt.u32.u64	%r1275, %rd672;
	shr.s32 	%r1276, %r1275, 16;
	cvt.u32.u64	%r1277, %rd674;
	shr.s32 	%r1278, %r1277, 16;
	mad.lo.s32 	%r1279, %r1276, %r1278, 16384;
	mul.wide.u32 	%rd690, %r1279, 131072;
	and.b64  	%rd691, %rd689, 4294901760;
	or.b64  	%rd692, %rd691, %rd685;
	shr.u64 	%rd693, %rd665, 32;
	cvt.u32.u64	%r1280, %rd693;
	shr.s32 	%r1281, %r1280, 16;
	shr.u64 	%rd694, %rd667, 32;
	cvt.u32.u64	%r1282, %rd694;
	shr.s32 	%r1283, %r1282, 16;
	mad.lo.s32 	%r1284, %r1281, %r1283, 16384;
	shr.u32 	%r1285, %r1284, 15;
	cvt.u64.u32	%rd695, %r1285;
	shl.b64 	%rd696, %rd695, 48;
	and.b64  	%rd697, %rd690, 281470681743360;
	or.b64  	%rd698, %rd697, %rd696;
	or.b64  	%rd699, %rd698, %rd692;
	or.b64  	%rd700, %rd688, %rd680;
	mov.b64	{%r1286, %r1287}, %rd699;
	mov.b64	{%r1288, %r1289}, %rd700;
	xor.b32  	%r6502, %r6722, %r1289;
	xor.b32  	%r6501, %r6721, %r1288;
	xor.b32  	%r6504, %r6724, %r1287;
	xor.b32  	%r6503, %r6723, %r1286;
	cvt.u64.u32	%rd701, %r6729;
	add.s64 	%rd702, %rd701, %rd9;
	shl.b64 	%rd703, %rd702, 1;
	mov.u64 	%rd704, verus_gpu_hash$sharedrand;
	add.s64 	%rd705, %rd704, %rd703;
	st.shared.u16 	[%rd705], %rs236;
	mov.u64 	%rd706, verus_gpu_hash$sharedrandex;
	add.s64 	%rd707, %rd706, %rd703;
	st.shared.u16 	[%rd707], %rs237;
	st.global.v4.u32 	[%rd6071], {%r6667, %r6668, %r6669, %r6670};
	cvt.u64.u16	%rd708, %rs237;
	add.s64 	%rd709, %rd708, %rd3;
	shl.b64 	%rd710, %rd709, 4;
	add.s64 	%rd711, %rd203, %rd710;
	st.global.v4.u32 	[%rd711], {%r6501, %r6502, %r6503, %r6504};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p16, %r6729, 32;
	@%p16 bra 	BB0_134;

	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p17, %r6729, 0;
	mov.u16 	%rs238, 20;
	@%p17 bra 	BB0_26;
	bra.uni 	BB0_27;

BB0_26:
	mov.u32 	%r6721, %r6667;
	mov.u32 	%r6722, %r6668;
	mov.u32 	%r6723, %r6669;
	mov.u32 	%r6724, %r6670;
	mov.u32 	%r6729, %r1250;
	bra.uni 	BB0_28;

BB0_14:
	mov.u32 	%r6501, %r6667;
	mov.u32 	%r6502, %r6668;
	mov.u32 	%r6503, %r6669;
	mov.u32 	%r6504, %r6670;
	bra.uni 	BB0_28;

BB0_27:
	cvt.u64.u32	%rd712, %r6731;
	shr.u32 	%r1291, %r6730, 5;
	cvt.u64.u32	%rd713, %r1291;
	cvt.u16.u32	%rs66, %r1291;
	and.b16  	%rs236, %rs66, 511;
	cvt.u16.u32	%rs67, %r6731;
	and.b16  	%rs237, %rs67, 511;
	cvt.u16.u64	%rs68, %rd6123;
	and.b16  	%rs238, %rs68, 28;
	and.b64  	%rd714, %rd713, 511;
	add.s64 	%rd715, %rd714, %rd3;
	shl.b64 	%rd716, %rd715, 4;
	add.s64 	%rd717, %rd203, %rd716;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd717];
	and.b64  	%rd718, %rd712, 511;
	add.s64 	%rd719, %rd718, %rd3;
	shl.b64 	%rd720, %rd719, 4;
	add.s64 	%rd721, %rd203, %rd720;
	ld.global.v4.u32 	{%r6501, %r6502, %r6503, %r6504}, [%rd721];
	mul.wide.u32 	%rd722, %r6730, 16;
	and.b64  	%rd723, %rd722, 48;
	add.s64 	%rd6122, %rd1, %rd723;

BB0_28:
	and.b16  	%rs69, %rs238, 255;
	setp.ne.s16	%p18, %rs69, 0;
	@%p18 bra 	BB0_41;

	and.b64  	%rd724, %rd6123, 1;
	mov.u64 	%rd725, 1;
	shl.b64 	%rd726, %rd724, 1;
	sub.s64 	%rd727, %rd725, %rd726;
	shl.b64 	%rd728, %rd727, 4;
	add.s64 	%rd729, %rd6122, %rd728;
	ld.local.v4.u32 	{%r1300, %r1301, %r1302, %r1303}, [%rd729];
	xor.b32  	%r132, %r6504, %r1303;
	xor.b32  	%r131, %r6503, %r1302;
	xor.b32  	%r130, %r6502, %r1301;
	xor.b32  	%r129, %r6501, %r1300;
	mov.b64	%rd730, {%r131, %r132};
	mov.u64 	%rd731, 0;
	st.local.v2.u64 	[%rd12], {%rd731, %rd730};
	shl.b64 	%rd734, %rd730, 1;
	xor.b64  	%rd735, %rd734, %rd730;
	st.local.v2.u64 	[%rd15], {%rd734, %rd735};
	shl.b64 	%rd736, %rd730, 2;
	xor.b64  	%rd737, %rd736, %rd730;
	st.local.v2.u64 	[%rd15+16], {%rd736, %rd737};
	shl.b64 	%rd738, %rd735, 1;
	xor.b64  	%rd739, %rd738, %rd730;
	st.local.v2.u64 	[%rd15+32], {%rd738, %rd739};
	mov.b64	%rd740, {%r129, %r130};
	shl.b64 	%rd741, %rd740, 3;
	and.b64  	%rd742, %rd741, 56;
	add.s64 	%rd743, %rd12, %rd742;
	and.b64  	%rd744, %rd740, 56;
	add.s64 	%rd745, %rd12, %rd744;
	ld.local.u64 	%rd746, [%rd745];
	shl.b64 	%rd747, %rd746, 3;
	ld.local.u64 	%rd748, [%rd743];
	xor.b64  	%rd749, %rd748, %rd747;
	shr.u64 	%rd750, %rd746, 61;
	mov.u32 	%r1308, 0;
	mov.b64	%rd751, {%r1308, %r1308};
	xor.b64  	%rd752, %rd751, %rd750;
	shr.u64 	%rd753, %rd740, 3;
	and.b64  	%rd754, %rd753, 56;
	add.s64 	%rd755, %rd12, %rd754;
	ld.local.u64 	%rd756, [%rd755];
	shl.b64 	%rd757, %rd756, 6;
	mov.b64	{%r1309, %r1310}, %rd749;
	mov.b64	%rd758, {%r1309, %r1310};
	xor.b64  	%rd759, %rd758, %rd757;
	shr.u64 	%rd760, %rd756, 58;
	mov.b64	{%r1311, %r1312}, %rd752;
	mov.b64	%rd761, {%r1311, %r1312};
	xor.b64  	%rd762, %rd761, %rd760;
	shr.u64 	%rd763, %rd740, 6;
	and.b64  	%rd764, %rd763, 56;
	add.s64 	%rd765, %rd12, %rd764;
	ld.local.u64 	%rd766, [%rd765];
	shl.b64 	%rd767, %rd766, 9;
	mov.b64	{%r1313, %r1314}, %rd759;
	mov.b64	%rd768, {%r1313, %r1314};
	xor.b64  	%rd769, %rd768, %rd767;
	shr.u64 	%rd770, %rd766, 55;
	mov.b64	{%r1315, %r1316}, %rd762;
	mov.b64	%rd771, {%r1315, %r1316};
	xor.b64  	%rd772, %rd771, %rd770;
	shr.u64 	%rd773, %rd740, 9;
	and.b64  	%rd774, %rd773, 56;
	add.s64 	%rd775, %rd12, %rd774;
	ld.local.u64 	%rd776, [%rd775];
	shl.b64 	%rd777, %rd776, 12;
	mov.b64	{%r1317, %r1318}, %rd769;
	mov.b64	%rd778, {%r1317, %r1318};
	xor.b64  	%rd779, %rd778, %rd777;
	shr.u64 	%rd780, %rd776, 52;
	mov.b64	{%r1319, %r1320}, %rd772;
	mov.b64	%rd781, {%r1319, %r1320};
	xor.b64  	%rd782, %rd781, %rd780;
	shr.u64 	%rd783, %rd740, 12;
	and.b64  	%rd784, %rd783, 56;
	add.s64 	%rd785, %rd12, %rd784;
	ld.local.u64 	%rd786, [%rd785];
	shl.b64 	%rd787, %rd786, 15;
	mov.b64	{%r1321, %r1322}, %rd779;
	mov.b64	%rd788, {%r1321, %r1322};
	xor.b64  	%rd789, %rd788, %rd787;
	shr.u64 	%rd790, %rd786, 49;
	mov.b64	{%r1323, %r1324}, %rd782;
	mov.b64	%rd791, {%r1323, %r1324};
	xor.b64  	%rd792, %rd791, %rd790;
	shr.u64 	%rd793, %rd740, 15;
	and.b64  	%rd794, %rd793, 56;
	add.s64 	%rd795, %rd12, %rd794;
	ld.local.u64 	%rd796, [%rd795];
	shl.b64 	%rd797, %rd796, 18;
	mov.b64	{%r1325, %r1326}, %rd789;
	mov.b64	%rd798, {%r1325, %r1326};
	xor.b64  	%rd799, %rd798, %rd797;
	shr.u64 	%rd800, %rd796, 46;
	mov.b64	{%r1327, %r1328}, %rd792;
	mov.b64	%rd801, {%r1327, %r1328};
	xor.b64  	%rd802, %rd801, %rd800;
	shr.u64 	%rd803, %rd740, 18;
	and.b64  	%rd804, %rd803, 56;
	add.s64 	%rd805, %rd12, %rd804;
	ld.local.u64 	%rd806, [%rd805];
	shl.b64 	%rd807, %rd806, 21;
	mov.b64	{%r1329, %r1330}, %rd799;
	mov.b64	%rd808, {%r1329, %r1330};
	xor.b64  	%rd809, %rd808, %rd807;
	shr.u64 	%rd810, %rd806, 43;
	mov.b64	{%r1331, %r1332}, %rd802;
	mov.b64	%rd811, {%r1331, %r1332};
	xor.b64  	%rd812, %rd811, %rd810;
	shr.u64 	%rd813, %rd740, 21;
	and.b64  	%rd814, %rd813, 56;
	add.s64 	%rd815, %rd12, %rd814;
	ld.local.u64 	%rd816, [%rd815];
	shl.b64 	%rd817, %rd816, 24;
	mov.b64	{%r1333, %r1334}, %rd809;
	mov.b64	%rd818, {%r1333, %r1334};
	xor.b64  	%rd819, %rd818, %rd817;
	shr.u64 	%rd820, %rd816, 40;
	mov.b64	{%r1335, %r1336}, %rd812;
	mov.b64	%rd821, {%r1335, %r1336};
	xor.b64  	%rd822, %rd821, %rd820;
	shr.u64 	%rd823, %rd740, 24;
	and.b64  	%rd824, %rd823, 56;
	add.s64 	%rd825, %rd12, %rd824;
	ld.local.u64 	%rd826, [%rd825];
	shl.b64 	%rd827, %rd826, 27;
	mov.b64	{%r1337, %r1338}, %rd819;
	mov.b64	%rd828, {%r1337, %r1338};
	xor.b64  	%rd829, %rd828, %rd827;
	shr.u64 	%rd830, %rd826, 37;
	mov.b64	{%r1339, %r1340}, %rd822;
	mov.b64	%rd831, {%r1339, %r1340};
	xor.b64  	%rd832, %rd831, %rd830;
	shr.u64 	%rd833, %rd740, 27;
	and.b64  	%rd834, %rd833, 56;
	add.s64 	%rd835, %rd12, %rd834;
	ld.local.u64 	%rd836, [%rd835];
	shl.b64 	%rd837, %rd836, 30;
	mov.b64	{%r1341, %r1342}, %rd829;
	mov.b64	%rd838, {%r1341, %r1342};
	xor.b64  	%rd839, %rd838, %rd837;
	shr.u64 	%rd840, %rd836, 34;
	mov.b64	{%r1343, %r1344}, %rd832;
	mov.b64	%rd841, {%r1343, %r1344};
	xor.b64  	%rd842, %rd841, %rd840;
	shr.u64 	%rd843, %rd740, 30;
	and.b64  	%rd844, %rd843, 56;
	add.s64 	%rd845, %rd12, %rd844;
	ld.local.u64 	%rd846, [%rd845];
	shl.b64 	%rd847, %rd846, 33;
	mov.b64	{%r1345, %r1346}, %rd839;
	mov.b64	%rd848, {%r1345, %r1346};
	xor.b64  	%rd849, %rd848, %rd847;
	shr.u64 	%rd850, %rd846, 31;
	mov.b64	{%r1347, %r1348}, %rd842;
	mov.b64	%rd851, {%r1347, %r1348};
	xor.b64  	%rd852, %rd851, %rd850;
	shr.u64 	%rd853, %rd740, 33;
	and.b64  	%rd854, %rd853, 56;
	add.s64 	%rd855, %rd12, %rd854;
	ld.local.u64 	%rd856, [%rd855];
	shl.b64 	%rd857, %rd856, 36;
	mov.b64	{%r1349, %r1350}, %rd849;
	mov.b64	%rd858, {%r1349, %r1350};
	xor.b64  	%rd859, %rd858, %rd857;
	shr.u64 	%rd860, %rd856, 28;
	mov.b64	{%r1351, %r1352}, %rd852;
	mov.b64	%rd861, {%r1351, %r1352};
	xor.b64  	%rd862, %rd861, %rd860;
	shr.u64 	%rd863, %rd740, 36;
	and.b64  	%rd864, %rd863, 56;
	add.s64 	%rd865, %rd12, %rd864;
	ld.local.u64 	%rd866, [%rd865];
	shl.b64 	%rd867, %rd866, 39;
	mov.b64	{%r1353, %r1354}, %rd859;
	mov.b64	%rd868, {%r1353, %r1354};
	xor.b64  	%rd869, %rd868, %rd867;
	shr.u64 	%rd870, %rd866, 25;
	mov.b64	{%r1355, %r1356}, %rd862;
	mov.b64	%rd871, {%r1355, %r1356};
	xor.b64  	%rd872, %rd871, %rd870;
	shr.u64 	%rd873, %rd740, 39;
	and.b64  	%rd874, %rd873, 56;
	add.s64 	%rd875, %rd12, %rd874;
	ld.local.u64 	%rd876, [%rd875];
	shl.b64 	%rd877, %rd876, 42;
	mov.b64	{%r1357, %r1358}, %rd869;
	mov.b64	%rd878, {%r1357, %r1358};
	xor.b64  	%rd879, %rd878, %rd877;
	shr.u64 	%rd880, %rd876, 22;
	mov.b64	{%r1359, %r1360}, %rd872;
	mov.b64	%rd881, {%r1359, %r1360};
	xor.b64  	%rd882, %rd881, %rd880;
	shr.u64 	%rd883, %rd740, 42;
	and.b64  	%rd884, %rd883, 56;
	add.s64 	%rd885, %rd12, %rd884;
	ld.local.u64 	%rd886, [%rd885];
	shl.b64 	%rd887, %rd886, 45;
	mov.b64	{%r1361, %r1362}, %rd879;
	mov.b64	%rd888, {%r1361, %r1362};
	xor.b64  	%rd889, %rd888, %rd887;
	shr.u64 	%rd890, %rd886, 19;
	mov.b64	{%r1363, %r1364}, %rd882;
	mov.b64	%rd891, {%r1363, %r1364};
	xor.b64  	%rd892, %rd891, %rd890;
	shr.u64 	%rd893, %rd740, 45;
	and.b64  	%rd894, %rd893, 56;
	add.s64 	%rd895, %rd12, %rd894;
	ld.local.u64 	%rd896, [%rd895];
	shl.b64 	%rd897, %rd896, 48;
	mov.b64	{%r1365, %r1366}, %rd889;
	mov.b64	%rd898, {%r1365, %r1366};
	xor.b64  	%rd899, %rd898, %rd897;
	shr.u64 	%rd900, %rd896, 16;
	mov.b64	{%r1367, %r1368}, %rd892;
	mov.b64	%rd901, {%r1367, %r1368};
	xor.b64  	%rd902, %rd901, %rd900;
	shr.u64 	%rd903, %rd740, 48;
	and.b64  	%rd904, %rd903, 56;
	add.s64 	%rd905, %rd12, %rd904;
	ld.local.u64 	%rd906, [%rd905];
	shl.b64 	%rd907, %rd906, 51;
	mov.b64	{%r1369, %r1370}, %rd899;
	mov.b64	%rd908, {%r1369, %r1370};
	xor.b64  	%rd909, %rd908, %rd907;
	shr.u64 	%rd910, %rd906, 13;
	mov.b64	{%r1371, %r1372}, %rd902;
	mov.b64	%rd911, {%r1371, %r1372};
	xor.b64  	%rd912, %rd911, %rd910;
	shr.u64 	%rd913, %rd740, 51;
	and.b64  	%rd914, %rd913, 56;
	add.s64 	%rd915, %rd12, %rd914;
	ld.local.u64 	%rd916, [%rd915];
	shl.b64 	%rd917, %rd916, 54;
	mov.b64	{%r1373, %r1374}, %rd909;
	mov.b64	%rd918, {%r1373, %r1374};
	xor.b64  	%rd919, %rd918, %rd917;
	shr.u64 	%rd920, %rd916, 10;
	mov.b64	{%r1375, %r1376}, %rd912;
	mov.b64	%rd921, {%r1375, %r1376};
	xor.b64  	%rd922, %rd921, %rd920;
	shr.u64 	%rd923, %rd740, 54;
	and.b64  	%rd924, %rd923, 56;
	add.s64 	%rd925, %rd12, %rd924;
	ld.local.u64 	%rd926, [%rd925];
	shl.b64 	%rd927, %rd926, 57;
	mov.b64	{%r1377, %r1378}, %rd919;
	mov.b64	%rd928, {%r1377, %r1378};
	xor.b64  	%rd929, %rd928, %rd927;
	shr.u64 	%rd930, %rd926, 7;
	mov.b64	{%r1379, %r1380}, %rd922;
	mov.b64	%rd931, {%r1379, %r1380};
	xor.b64  	%rd932, %rd931, %rd930;
	shr.u64 	%rd933, %rd740, 57;
	and.b64  	%rd934, %rd933, 56;
	add.s64 	%rd935, %rd12, %rd934;
	ld.local.u64 	%rd936, [%rd935];
	shl.b64 	%rd937, %rd936, 60;
	mov.b64	{%r1381, %r1382}, %rd929;
	mov.b64	%rd938, {%r1381, %r1382};
	xor.b64  	%rd939, %rd938, %rd937;
	shr.u64 	%rd940, %rd936, 4;
	mov.b64	{%r1383, %r1384}, %rd932;
	mov.b64	%rd941, {%r1383, %r1384};
	xor.b64  	%rd942, %rd941, %rd940;
	shr.u64 	%rd943, %rd740, 60;
	and.b64  	%rd944, %rd943, 8;
	add.s64 	%rd945, %rd12, %rd944;
	ld.local.u64 	%rd946, [%rd945];
	shl.b64 	%rd947, %rd946, 63;
	mov.b64	{%r1385, %r1386}, %rd939;
	mov.b64	%rd948, {%r1385, %r1386};
	xor.b64  	%rd949, %rd948, %rd947;
	shr.u64 	%rd950, %rd946, 1;
	mov.b64	{%r1387, %r1388}, %rd942;
	mov.b64	%rd951, {%r1387, %r1388};
	xor.b64  	%rd952, %rd951, %rd950;
	mov.b64	{%r6516, %r6517}, %rd952;
	mov.b64	{%r6518, %r6519}, %rd949;
	setp.gt.s32	%p19, %r132, -1;
	@%p19 bra 	BB0_31;

	shr.u32 	%r1389, %r129, 1;
	shl.b32 	%r1390, %r130, 31;
	shr.u32 	%r1391, %r130, 1;
	and.b32  	%r1392, %r1389, 1533916891;
	or.b32  	%r1393, %r1392, %r1390;
	xor.b32  	%r6516, %r6516, %r1393;
	and.b32  	%r1394, %r1391, 920350134;
	xor.b32  	%r6517, %r6517, %r1394;

BB0_31:
	and.b32  	%r1395, %r132, 1073741824;
	setp.eq.s32	%p20, %r1395, 0;
	@%p20 bra 	BB0_33;

	shr.u32 	%r1396, %r129, 2;
	shl.b32 	%r1397, %r130, 30;
	or.b32  	%r1398, %r1396, %r1397;
	shr.u32 	%r1399, %r130, 2;
	and.b32  	%r1400, %r1398, 1227133513;
	xor.b32  	%r6516, %r6516, %r1400;
	and.b32  	%r1401, %r1399, 306783378;
	xor.b32  	%r6517, %r6517, %r1401;

BB0_33:
	mov.u64 	%rd6088, 0;
	mov.u32 	%r6462, 0;
	xor.b32  	%r154, %r6518, %r6730;
	xor.b32  	%r155, %r6519, %r6731;
	xor.b32  	%r156, %r6516, %r6732;
	xor.b32  	%r157, %r6517, %r6733;
	mov.b64	%rd953, {%r156, %r157};
	mov.b64	%rd954, {%r154, %r155};
	cvt.u32.u64	%r1402, %rd954;
	cvt.s32.s16 	%r1403, %r1402;
	mov.b64	%rd955, {%r6503, %r6504};
	mov.b64	%rd956, {%r6501, %r6502};
	cvt.u32.u64	%r1404, %rd956;
	cvt.s32.s16 	%r1405, %r1404;
	mad.lo.s32 	%r1406, %r1403, %r1405, 16384;
	bfe.u32 	%r1407, %r1406, 15, 16;
	shr.s32 	%r1408, %r1402, 16;
	shr.s32 	%r1409, %r1404, 16;
	mad.lo.s32 	%r1410, %r1408, %r1409, 16384;
	mov.b64	%rd957, {%r1407, %r6462};
	shl.b32 	%r1412, %r1410, 1;
	cvt.u64.u32	%rd958, %r1412;
	and.b64  	%rd959, %rd957, 65535;
	shr.u64 	%rd960, %rd953, 16;
	shr.u64 	%rd961, %rd954, 16;
	cvt.u32.u64	%r1413, %rd961;
	shr.s32 	%r1414, %r1413, 16;
	shr.u64 	%rd962, %rd955, 16;
	shr.u64 	%rd963, %rd956, 16;
	cvt.u32.u64	%r1415, %rd963;
	shr.s32 	%r1416, %r1415, 16;
	mad.lo.s32 	%r1417, %r1414, %r1416, 16384;
	mul.wide.u32 	%rd964, %r1417, 131072;
	shr.u64 	%rd965, %rd954, 32;
	cvt.u32.u64	%r1418, %rd965;
	shr.s32 	%r1419, %r1418, 16;
	shr.u64 	%rd966, %rd956, 32;
	cvt.u32.u64	%r1420, %rd966;
	shr.s32 	%r1421, %r1420, 16;
	mad.lo.s32 	%r1422, %r1419, %r1421, 16384;
	shr.u32 	%r1423, %r1422, 15;
	cvt.u64.u32	%rd967, %r1423;
	shl.b64 	%rd968, %rd967, 48;
	and.b64  	%rd969, %rd964, 281470681743360;
	shl.b64 	%rd970, %rd953, 16;
	cvt.u32.u64	%r1424, %rd970;
	shr.s32 	%r1425, %r1424, 16;
	shl.b64 	%rd972, %rd955, 16;
	cvt.u32.u64	%r1426, %rd972;
	shr.s32 	%r1427, %r1426, 16;
	mad.lo.s32 	%r1428, %r1425, %r1427, 16384;
	shr.u32 	%r1429, %r1428, 15;
	cvt.u64.u32	%rd973, %r1429;
	cvt.u32.u64	%r1430, %rd953;
	shr.s32 	%r1431, %r1430, 16;
	cvt.u32.u64	%r1432, %rd955;
	shr.s32 	%r1433, %r1432, 16;
	mad.lo.s32 	%r1434, %r1431, %r1433, 16384;
	and.b64  	%rd974, %rd973, 65535;
	and.b64  	%rd975, %rd958, 4294901760;
	or.b64  	%rd976, %rd975, %rd959;
	or.b64  	%rd977, %rd976, %rd969;
	shl.b32 	%r1435, %r1434, 1;
	cvt.u64.u32	%rd978, %r1435;
	cvt.u32.u64	%r1436, %rd960;
	shr.s32 	%r1437, %r1436, 16;
	cvt.u32.u64	%r1438, %rd962;
	shr.s32 	%r1439, %r1438, 16;
	mad.lo.s32 	%r1440, %r1437, %r1439, 16384;
	mul.wide.u32 	%rd979, %r1440, 131072;
	and.b64  	%rd980, %rd978, 4294901760;
	or.b64  	%rd981, %rd980, %rd974;
	shr.u64 	%rd982, %rd953, 32;
	cvt.u32.u64	%r1441, %rd982;
	shr.s32 	%r1442, %r1441, 16;
	shr.u64 	%rd983, %rd955, 32;
	cvt.u32.u64	%r1443, %rd983;
	shr.s32 	%r1444, %r1443, 16;
	mad.lo.s32 	%r1445, %r1442, %r1444, 16384;
	shr.u32 	%r1446, %r1445, 15;
	cvt.u64.u32	%rd984, %r1446;
	shl.b64 	%rd985, %rd984, 48;
	and.b64  	%rd986, %rd979, 281470681743360;
	or.b64  	%rd987, %rd986, %rd985;
	or.b64  	%rd988, %rd987, %rd981;
	or.b64  	%rd989, %rd977, %rd968;
	mov.b64	{%r1447, %r1448}, %rd988;
	mov.b64	{%r1449, %r1450}, %rd989;
	xor.b32  	%r161, %r1448, %r6504;
	xor.b32  	%r160, %r1447, %r6503;
	xor.b32  	%r159, %r1450, %r6502;
	xor.b32  	%r158, %r1449, %r6501;
	ld.local.v4.u32 	{%r1451, %r1452, %r1453, %r1454}, [%rd6122];
	xor.b32  	%r164, %r6723, %r1453;
	xor.b32  	%r163, %r6722, %r1452;
	xor.b32  	%r162, %r6721, %r1451;
	xor.b32  	%r165, %r6724, %r1454;
	mov.b64	%rd990, {%r164, %r165};
	st.local.v2.u64 	[%rd12], {%rd6088, %rd990};
	shl.b64 	%rd993, %rd990, 1;
	xor.b64  	%rd994, %rd993, %rd990;
	st.local.v2.u64 	[%rd15], {%rd993, %rd994};
	shl.b64 	%rd995, %rd990, 2;
	xor.b64  	%rd996, %rd995, %rd990;
	st.local.v2.u64 	[%rd15+16], {%rd995, %rd996};
	shl.b64 	%rd997, %rd994, 1;
	xor.b64  	%rd998, %rd997, %rd990;
	st.local.v2.u64 	[%rd15+32], {%rd997, %rd998};
	mov.b64	%rd999, {%r162, %r163};
	shl.b64 	%rd1000, %rd999, 3;
	and.b64  	%rd1001, %rd1000, 56;
	add.s64 	%rd1002, %rd12, %rd1001;
	and.b64  	%rd1003, %rd999, 56;
	add.s64 	%rd1004, %rd12, %rd1003;
	ld.local.u64 	%rd1005, [%rd1004];
	shl.b64 	%rd1006, %rd1005, 3;
	ld.local.u64 	%rd1007, [%rd1002];
	xor.b64  	%rd1008, %rd1007, %rd1006;
	shr.u64 	%rd1009, %rd1005, 61;
	mov.b64	%rd1010, {%r6462, %r6462};
	xor.b64  	%rd1011, %rd1010, %rd1009;
	shr.u64 	%rd1012, %rd999, 3;
	and.b64  	%rd1013, %rd1012, 56;
	add.s64 	%rd1014, %rd12, %rd1013;
	ld.local.u64 	%rd1015, [%rd1014];
	shl.b64 	%rd1016, %rd1015, 6;
	mov.b64	{%r1459, %r1460}, %rd1008;
	mov.b64	%rd1017, {%r1459, %r1460};
	xor.b64  	%rd1018, %rd1017, %rd1016;
	shr.u64 	%rd1019, %rd1015, 58;
	mov.b64	{%r1461, %r1462}, %rd1011;
	mov.b64	%rd1020, {%r1461, %r1462};
	xor.b64  	%rd1021, %rd1020, %rd1019;
	shr.u64 	%rd1022, %rd999, 6;
	and.b64  	%rd1023, %rd1022, 56;
	add.s64 	%rd1024, %rd12, %rd1023;
	ld.local.u64 	%rd1025, [%rd1024];
	shl.b64 	%rd1026, %rd1025, 9;
	mov.b64	{%r1463, %r1464}, %rd1018;
	mov.b64	%rd1027, {%r1463, %r1464};
	xor.b64  	%rd1028, %rd1027, %rd1026;
	shr.u64 	%rd1029, %rd1025, 55;
	mov.b64	{%r1465, %r1466}, %rd1021;
	mov.b64	%rd1030, {%r1465, %r1466};
	xor.b64  	%rd1031, %rd1030, %rd1029;
	shr.u64 	%rd1032, %rd999, 9;
	and.b64  	%rd1033, %rd1032, 56;
	add.s64 	%rd1034, %rd12, %rd1033;
	ld.local.u64 	%rd1035, [%rd1034];
	shl.b64 	%rd1036, %rd1035, 12;
	mov.b64	{%r1467, %r1468}, %rd1028;
	mov.b64	%rd1037, {%r1467, %r1468};
	xor.b64  	%rd1038, %rd1037, %rd1036;
	shr.u64 	%rd1039, %rd1035, 52;
	mov.b64	{%r1469, %r1470}, %rd1031;
	mov.b64	%rd1040, {%r1469, %r1470};
	xor.b64  	%rd1041, %rd1040, %rd1039;
	shr.u64 	%rd1042, %rd999, 12;
	and.b64  	%rd1043, %rd1042, 56;
	add.s64 	%rd1044, %rd12, %rd1043;
	ld.local.u64 	%rd1045, [%rd1044];
	shl.b64 	%rd1046, %rd1045, 15;
	mov.b64	{%r1471, %r1472}, %rd1038;
	mov.b64	%rd1047, {%r1471, %r1472};
	xor.b64  	%rd1048, %rd1047, %rd1046;
	shr.u64 	%rd1049, %rd1045, 49;
	mov.b64	{%r1473, %r1474}, %rd1041;
	mov.b64	%rd1050, {%r1473, %r1474};
	xor.b64  	%rd1051, %rd1050, %rd1049;
	shr.u64 	%rd1052, %rd999, 15;
	and.b64  	%rd1053, %rd1052, 56;
	add.s64 	%rd1054, %rd12, %rd1053;
	ld.local.u64 	%rd1055, [%rd1054];
	shl.b64 	%rd1056, %rd1055, 18;
	mov.b64	{%r1475, %r1476}, %rd1048;
	mov.b64	%rd1057, {%r1475, %r1476};
	xor.b64  	%rd1058, %rd1057, %rd1056;
	shr.u64 	%rd1059, %rd1055, 46;
	mov.b64	{%r1477, %r1478}, %rd1051;
	mov.b64	%rd1060, {%r1477, %r1478};
	xor.b64  	%rd1061, %rd1060, %rd1059;
	shr.u64 	%rd1062, %rd999, 18;
	and.b64  	%rd1063, %rd1062, 56;
	add.s64 	%rd1064, %rd12, %rd1063;
	ld.local.u64 	%rd1065, [%rd1064];
	shl.b64 	%rd1066, %rd1065, 21;
	mov.b64	{%r1479, %r1480}, %rd1058;
	mov.b64	%rd1067, {%r1479, %r1480};
	xor.b64  	%rd1068, %rd1067, %rd1066;
	shr.u64 	%rd1069, %rd1065, 43;
	mov.b64	{%r1481, %r1482}, %rd1061;
	mov.b64	%rd1070, {%r1481, %r1482};
	xor.b64  	%rd1071, %rd1070, %rd1069;
	shr.u64 	%rd1072, %rd999, 21;
	and.b64  	%rd1073, %rd1072, 56;
	add.s64 	%rd1074, %rd12, %rd1073;
	ld.local.u64 	%rd1075, [%rd1074];
	shl.b64 	%rd1076, %rd1075, 24;
	mov.b64	{%r1483, %r1484}, %rd1068;
	mov.b64	%rd1077, {%r1483, %r1484};
	xor.b64  	%rd1078, %rd1077, %rd1076;
	shr.u64 	%rd1079, %rd1075, 40;
	mov.b64	{%r1485, %r1486}, %rd1071;
	mov.b64	%rd1080, {%r1485, %r1486};
	xor.b64  	%rd1081, %rd1080, %rd1079;
	shr.u64 	%rd1082, %rd999, 24;
	and.b64  	%rd1083, %rd1082, 56;
	add.s64 	%rd1084, %rd12, %rd1083;
	ld.local.u64 	%rd1085, [%rd1084];
	shl.b64 	%rd1086, %rd1085, 27;
	mov.b64	{%r1487, %r1488}, %rd1078;
	mov.b64	%rd1087, {%r1487, %r1488};
	xor.b64  	%rd1088, %rd1087, %rd1086;
	shr.u64 	%rd1089, %rd1085, 37;
	mov.b64	{%r1489, %r1490}, %rd1081;
	mov.b64	%rd1090, {%r1489, %r1490};
	xor.b64  	%rd1091, %rd1090, %rd1089;
	shr.u64 	%rd1092, %rd999, 27;
	and.b64  	%rd1093, %rd1092, 56;
	add.s64 	%rd1094, %rd12, %rd1093;
	ld.local.u64 	%rd1095, [%rd1094];
	shl.b64 	%rd1096, %rd1095, 30;
	mov.b64	{%r1491, %r1492}, %rd1088;
	mov.b64	%rd1097, {%r1491, %r1492};
	xor.b64  	%rd1098, %rd1097, %rd1096;
	shr.u64 	%rd1099, %rd1095, 34;
	mov.b64	{%r1493, %r1494}, %rd1091;
	mov.b64	%rd1100, {%r1493, %r1494};
	xor.b64  	%rd1101, %rd1100, %rd1099;
	shr.u64 	%rd1102, %rd999, 30;
	and.b64  	%rd1103, %rd1102, 56;
	add.s64 	%rd1104, %rd12, %rd1103;
	ld.local.u64 	%rd1105, [%rd1104];
	shl.b64 	%rd1106, %rd1105, 33;
	mov.b64	{%r1495, %r1496}, %rd1098;
	mov.b64	%rd1107, {%r1495, %r1496};
	xor.b64  	%rd1108, %rd1107, %rd1106;
	shr.u64 	%rd1109, %rd1105, 31;
	mov.b64	{%r1497, %r1498}, %rd1101;
	mov.b64	%rd1110, {%r1497, %r1498};
	xor.b64  	%rd1111, %rd1110, %rd1109;
	shr.u64 	%rd1112, %rd999, 33;
	and.b64  	%rd1113, %rd1112, 56;
	add.s64 	%rd1114, %rd12, %rd1113;
	ld.local.u64 	%rd1115, [%rd1114];
	shl.b64 	%rd1116, %rd1115, 36;
	mov.b64	{%r1499, %r1500}, %rd1108;
	mov.b64	%rd1117, {%r1499, %r1500};
	xor.b64  	%rd1118, %rd1117, %rd1116;
	shr.u64 	%rd1119, %rd1115, 28;
	mov.b64	{%r1501, %r1502}, %rd1111;
	mov.b64	%rd1120, {%r1501, %r1502};
	xor.b64  	%rd1121, %rd1120, %rd1119;
	shr.u64 	%rd1122, %rd999, 36;
	and.b64  	%rd1123, %rd1122, 56;
	add.s64 	%rd1124, %rd12, %rd1123;
	ld.local.u64 	%rd1125, [%rd1124];
	shl.b64 	%rd1126, %rd1125, 39;
	mov.b64	{%r1503, %r1504}, %rd1118;
	mov.b64	%rd1127, {%r1503, %r1504};
	xor.b64  	%rd1128, %rd1127, %rd1126;
	shr.u64 	%rd1129, %rd1125, 25;
	mov.b64	{%r1505, %r1506}, %rd1121;
	mov.b64	%rd1130, {%r1505, %r1506};
	xor.b64  	%rd1131, %rd1130, %rd1129;
	shr.u64 	%rd1132, %rd999, 39;
	and.b64  	%rd1133, %rd1132, 56;
	add.s64 	%rd1134, %rd12, %rd1133;
	ld.local.u64 	%rd1135, [%rd1134];
	shl.b64 	%rd1136, %rd1135, 42;
	mov.b64	{%r1507, %r1508}, %rd1128;
	mov.b64	%rd1137, {%r1507, %r1508};
	xor.b64  	%rd1138, %rd1137, %rd1136;
	shr.u64 	%rd1139, %rd1135, 22;
	mov.b64	{%r1509, %r1510}, %rd1131;
	mov.b64	%rd1140, {%r1509, %r1510};
	xor.b64  	%rd1141, %rd1140, %rd1139;
	shr.u64 	%rd1142, %rd999, 42;
	and.b64  	%rd1143, %rd1142, 56;
	add.s64 	%rd1144, %rd12, %rd1143;
	ld.local.u64 	%rd1145, [%rd1144];
	shl.b64 	%rd1146, %rd1145, 45;
	mov.b64	{%r1511, %r1512}, %rd1138;
	mov.b64	%rd1147, {%r1511, %r1512};
	xor.b64  	%rd1148, %rd1147, %rd1146;
	shr.u64 	%rd1149, %rd1145, 19;
	mov.b64	{%r1513, %r1514}, %rd1141;
	mov.b64	%rd1150, {%r1513, %r1514};
	xor.b64  	%rd1151, %rd1150, %rd1149;
	shr.u64 	%rd1152, %rd999, 45;
	and.b64  	%rd1153, %rd1152, 56;
	add.s64 	%rd1154, %rd12, %rd1153;
	ld.local.u64 	%rd1155, [%rd1154];
	shl.b64 	%rd1156, %rd1155, 48;
	mov.b64	{%r1515, %r1516}, %rd1148;
	mov.b64	%rd1157, {%r1515, %r1516};
	xor.b64  	%rd1158, %rd1157, %rd1156;
	shr.u64 	%rd1159, %rd1155, 16;
	mov.b64	{%r1517, %r1518}, %rd1151;
	mov.b64	%rd1160, {%r1517, %r1518};
	xor.b64  	%rd1161, %rd1160, %rd1159;
	shr.u64 	%rd1162, %rd999, 48;
	and.b64  	%rd1163, %rd1162, 56;
	add.s64 	%rd1164, %rd12, %rd1163;
	ld.local.u64 	%rd1165, [%rd1164];
	shl.b64 	%rd1166, %rd1165, 51;
	mov.b64	{%r1519, %r1520}, %rd1158;
	mov.b64	%rd1167, {%r1519, %r1520};
	xor.b64  	%rd1168, %rd1167, %rd1166;
	shr.u64 	%rd1169, %rd1165, 13;
	mov.b64	{%r1521, %r1522}, %rd1161;
	mov.b64	%rd1170, {%r1521, %r1522};
	xor.b64  	%rd1171, %rd1170, %rd1169;
	shr.u64 	%rd1172, %rd999, 51;
	and.b64  	%rd1173, %rd1172, 56;
	add.s64 	%rd1174, %rd12, %rd1173;
	ld.local.u64 	%rd1175, [%rd1174];
	shl.b64 	%rd1176, %rd1175, 54;
	mov.b64	{%r1523, %r1524}, %rd1168;
	mov.b64	%rd1177, {%r1523, %r1524};
	xor.b64  	%rd1178, %rd1177, %rd1176;
	shr.u64 	%rd1179, %rd1175, 10;
	mov.b64	{%r1525, %r1526}, %rd1171;
	mov.b64	%rd1180, {%r1525, %r1526};
	xor.b64  	%rd1181, %rd1180, %rd1179;
	shr.u64 	%rd1182, %rd999, 54;
	and.b64  	%rd1183, %rd1182, 56;
	add.s64 	%rd1184, %rd12, %rd1183;
	ld.local.u64 	%rd1185, [%rd1184];
	shl.b64 	%rd1186, %rd1185, 57;
	mov.b64	{%r1527, %r1528}, %rd1178;
	mov.b64	%rd1187, {%r1527, %r1528};
	xor.b64  	%rd1188, %rd1187, %rd1186;
	shr.u64 	%rd1189, %rd1185, 7;
	mov.b64	{%r1529, %r1530}, %rd1181;
	mov.b64	%rd1190, {%r1529, %r1530};
	xor.b64  	%rd1191, %rd1190, %rd1189;
	shr.u64 	%rd1192, %rd999, 57;
	and.b64  	%rd1193, %rd1192, 56;
	add.s64 	%rd1194, %rd12, %rd1193;
	ld.local.u64 	%rd1195, [%rd1194];
	shl.b64 	%rd1196, %rd1195, 60;
	mov.b64	{%r1531, %r1532}, %rd1188;
	mov.b64	%rd1197, {%r1531, %r1532};
	xor.b64  	%rd1198, %rd1197, %rd1196;
	shr.u64 	%rd1199, %rd1195, 4;
	mov.b64	{%r1533, %r1534}, %rd1191;
	mov.b64	%rd1200, {%r1533, %r1534};
	xor.b64  	%rd1201, %rd1200, %rd1199;
	shr.u64 	%rd1202, %rd999, 60;
	and.b64  	%rd1203, %rd1202, 8;
	add.s64 	%rd1204, %rd12, %rd1203;
	ld.local.u64 	%rd1205, [%rd1204];
	shl.b64 	%rd1206, %rd1205, 63;
	mov.b64	{%r1535, %r1536}, %rd1198;
	mov.b64	%rd1207, {%r1535, %r1536};
	xor.b64  	%rd1208, %rd1207, %rd1206;
	shr.u64 	%rd1209, %rd1205, 1;
	mov.b64	{%r1537, %r1538}, %rd1201;
	mov.b64	%rd1210, {%r1537, %r1538};
	xor.b64  	%rd1211, %rd1210, %rd1209;
	mov.b64	{%r6524, %r6525}, %rd1211;
	mov.b64	{%r6526, %r6527}, %rd1208;
	setp.gt.s32	%p21, %r165, -1;
	@%p21 bra 	BB0_35;

	shr.u32 	%r1539, %r162, 1;
	shl.b32 	%r1540, %r163, 31;
	shr.u32 	%r1541, %r163, 1;
	and.b32  	%r1542, %r1539, 1533916891;
	or.b32  	%r1543, %r1542, %r1540;
	xor.b32  	%r6524, %r6524, %r1543;
	and.b32  	%r1544, %r1541, 920350134;
	xor.b32  	%r6525, %r6525, %r1544;

BB0_35:
	and.b32  	%r1545, %r165, 1073741824;
	setp.eq.s32	%p22, %r1545, 0;
	@%p22 bra 	BB0_37;

	shr.u32 	%r1546, %r162, 2;
	shl.b32 	%r1547, %r163, 30;
	or.b32  	%r1548, %r1546, %r1547;
	shr.u32 	%r1549, %r163, 2;
	and.b32  	%r1550, %r1548, 1227133513;
	xor.b32  	%r6524, %r6524, %r1550;
	and.b32  	%r1551, %r1549, 306783378;
	xor.b32  	%r6525, %r6525, %r1551;

BB0_37:
	mov.u32 	%r6454, 0;
	xor.b32  	%r6730, %r6526, %r154;
	xor.b32  	%r6731, %r6527, %r155;
	xor.b32  	%r6732, %r6524, %r156;
	xor.b32  	%r6733, %r6525, %r157;
	mov.b64	%rd1212, {%r6732, %r6733};
	mov.b64	%rd1213, {%r6730, %r6731};
	cvt.u32.u64	%r1552, %rd1213;
	cvt.s32.s16 	%r1553, %r1552;
	mov.b64	%rd1214, {%r6723, %r6724};
	mov.b64	%rd1215, {%r6721, %r6722};
	cvt.u32.u64	%r1554, %rd1215;
	cvt.s32.s16 	%r1555, %r1554;
	mad.lo.s32 	%r1556, %r1553, %r1555, 16384;
	bfe.u32 	%r1557, %r1556, 15, 16;
	shr.s32 	%r1558, %r1552, 16;
	shr.s32 	%r1559, %r1554, 16;
	mad.lo.s32 	%r1560, %r1558, %r1559, 16384;
	mov.b64	%rd1216, {%r1557, %r6454};
	shl.b32 	%r1562, %r1560, 1;
	cvt.u64.u32	%rd1217, %r1562;
	and.b64  	%rd1218, %rd1216, 65535;
	shr.u64 	%rd1219, %rd1212, 16;
	shr.u64 	%rd1220, %rd1213, 16;
	cvt.u32.u64	%r1563, %rd1220;
	shr.s32 	%r1564, %r1563, 16;
	shr.u64 	%rd1221, %rd1214, 16;
	shr.u64 	%rd1222, %rd1215, 16;
	cvt.u32.u64	%r1565, %rd1222;
	shr.s32 	%r1566, %r1565, 16;
	mad.lo.s32 	%r1567, %r1564, %r1566, 16384;
	mul.wide.u32 	%rd1223, %r1567, 131072;
	shr.u64 	%rd1224, %rd1213, 32;
	cvt.u32.u64	%r1568, %rd1224;
	shr.s32 	%r1569, %r1568, 16;
	shr.u64 	%rd1225, %rd1215, 32;
	cvt.u32.u64	%r1570, %rd1225;
	shr.s32 	%r1571, %r1570, 16;
	mad.lo.s32 	%r1572, %r1569, %r1571, 16384;
	shr.u32 	%r1573, %r1572, 15;
	cvt.u64.u32	%rd1226, %r1573;
	shl.b64 	%rd1227, %rd1226, 48;
	and.b64  	%rd1228, %rd1223, 281470681743360;
	shl.b64 	%rd1229, %rd1212, 16;
	cvt.u32.u64	%r1574, %rd1229;
	shr.s32 	%r1575, %r1574, 16;
	shl.b64 	%rd1230, %rd1214, 16;
	cvt.u32.u64	%r1576, %rd1230;
	shr.s32 	%r1577, %r1576, 16;
	mad.lo.s32 	%r1578, %r1575, %r1577, 16384;
	shr.u32 	%r1579, %r1578, 15;
	cvt.u64.u32	%rd1231, %r1579;
	cvt.u32.u64	%r1580, %rd1212;
	shr.s32 	%r1581, %r1580, 16;
	cvt.u32.u64	%r1582, %rd1214;
	shr.s32 	%r1583, %r1582, 16;
	mad.lo.s32 	%r1584, %r1581, %r1583, 16384;
	and.b64  	%rd1232, %rd1231, 65535;
	and.b64  	%rd1233, %rd1217, 4294901760;
	or.b64  	%rd1234, %rd1233, %rd1218;
	or.b64  	%rd1235, %rd1234, %rd1228;
	shl.b32 	%r1585, %r1584, 1;
	cvt.u64.u32	%rd1236, %r1585;
	cvt.u32.u64	%r1586, %rd1219;
	shr.s32 	%r1587, %r1586, 16;
	cvt.u32.u64	%r1588, %rd1221;
	shr.s32 	%r1589, %r1588, 16;
	mad.lo.s32 	%r1590, %r1587, %r1589, 16384;
	mul.wide.u32 	%rd1237, %r1590, 131072;
	and.b64  	%rd1238, %rd1236, 4294901760;
	or.b64  	%rd1239, %rd1238, %rd1232;
	shr.u64 	%rd1240, %rd1212, 32;
	cvt.u32.u64	%r1591, %rd1240;
	shr.s32 	%r1592, %r1591, 16;
	shr.u64 	%rd1241, %rd1214, 32;
	cvt.u32.u64	%r1593, %rd1241;
	shr.s32 	%r1594, %r1593, 16;
	mad.lo.s32 	%r1595, %r1592, %r1594, 16384;
	shr.u32 	%r1596, %r1595, 15;
	cvt.u64.u32	%rd1242, %r1596;
	shl.b64 	%rd1243, %rd1242, 48;
	and.b64  	%rd1244, %rd1237, 281470681743360;
	or.b64  	%rd1245, %rd1244, %rd1243;
	or.b64  	%rd1246, %rd1245, %rd1239;
	or.b64  	%rd1247, %rd1235, %rd1227;
	mov.b64	{%r1597, %r1598}, %rd1246;
	mov.b64	{%r1599, %r1600}, %rd1247;
	xor.b32  	%r6502, %r1600, %r6722;
	xor.b32  	%r6501, %r1599, %r6721;
	xor.b32  	%r6504, %r1598, %r6724;
	xor.b32  	%r6503, %r1597, %r6723;
	cvt.u64.u32	%rd1248, %r6729;
	add.s64 	%rd1249, %rd1248, %rd9;
	shl.b64 	%rd1250, %rd1249, 1;
	mov.u64 	%rd1251, verus_gpu_hash$sharedrand;
	add.s64 	%rd1252, %rd1251, %rd1250;
	st.shared.u16 	[%rd1252], %rs236;
	mov.u64 	%rd1253, verus_gpu_hash$sharedrandex;
	add.s64 	%rd1254, %rd1253, %rd1250;
	st.shared.u16 	[%rd1254], %rs237;
	cvt.u64.u16	%rd1255, %rs236;
	add.s64 	%rd1256, %rd1255, %rd3;
	shl.b64 	%rd1257, %rd1256, 4;
	add.s64 	%rd1258, %rd203, %rd1257;
	st.global.v4.u32 	[%rd1258], {%r158, %r159, %r160, %r161};
	cvt.u64.u16	%rd1259, %rs237;
	add.s64 	%rd1260, %rd1259, %rd3;
	shl.b64 	%rd1261, %rd1260, 4;
	add.s64 	%rd1262, %rd203, %rd1261;
	st.global.v4.u32 	[%rd1262], {%r6501, %r6502, %r6503, %r6504};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p23, %r6729, 32;
	@%p23 bra 	BB0_134;

	mov.u32 	%r6455, 0;
	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p24, %r6729, 0;
	mov.u16 	%rs238, 0;
	@%p24 bra 	BB0_39;
	bra.uni 	BB0_40;

BB0_39:
	mov.u32 	%r6721, %r158;
	mov.u32 	%r6722, %r159;
	mov.u32 	%r6723, %r160;
	mov.u32 	%r6724, %r161;
	mov.u32 	%r6729, %r6455;
	bra.uni 	BB0_41;

BB0_40:
	cvt.u64.u32	%rd1263, %r6731;
	shr.u32 	%r1602, %r6730, 5;
	cvt.u64.u32	%rd1264, %r1602;
	cvt.u16.u32	%rs71, %r1602;
	and.b16  	%rs236, %rs71, 511;
	cvt.u16.u32	%rs72, %r6731;
	and.b16  	%rs237, %rs72, 511;
	cvt.u16.u64	%rs73, %rd6123;
	and.b16  	%rs238, %rs73, 28;
	and.b64  	%rd1265, %rd1264, 511;
	add.s64 	%rd1266, %rd1265, %rd3;
	shl.b64 	%rd1267, %rd1266, 4;
	add.s64 	%rd1268, %rd203, %rd1267;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd1268];
	and.b64  	%rd1269, %rd1263, 511;
	add.s64 	%rd1270, %rd1269, %rd3;
	shl.b64 	%rd1271, %rd1270, 4;
	add.s64 	%rd1272, %rd203, %rd1271;
	ld.global.v4.u32 	{%r6501, %r6502, %r6503, %r6504}, [%rd1272];
	mul.wide.u32 	%rd1273, %r6730, 16;
	and.b64  	%rd1274, %rd1273, 48;
	add.s64 	%rd6122, %rd1, %rd1274;

BB0_41:
	and.b16  	%rs74, %rs238, 255;
	setp.ne.s16	%p25, %rs74, 4;
	@%p25 bra 	BB0_54;

	ld.local.v4.u32 	{%r1611, %r1612, %r1613, %r1614}, [%rd6122];
	xor.b32  	%r223, %r6723, %r1613;
	xor.b32  	%r222, %r6722, %r1612;
	xor.b32  	%r221, %r6721, %r1611;
	xor.b32  	%r224, %r6724, %r1614;
	mov.b64	%rd1275, {%r223, %r224};
	mov.u64 	%rd1276, 0;
	st.local.v2.u64 	[%rd12], {%rd1276, %rd1275};
	shl.b64 	%rd1279, %rd1275, 1;
	xor.b64  	%rd1280, %rd1279, %rd1275;
	st.local.v2.u64 	[%rd15], {%rd1279, %rd1280};
	shl.b64 	%rd1281, %rd1275, 2;
	xor.b64  	%rd1282, %rd1281, %rd1275;
	st.local.v2.u64 	[%rd15+16], {%rd1281, %rd1282};
	shl.b64 	%rd1283, %rd1280, 1;
	xor.b64  	%rd1284, %rd1283, %rd1275;
	st.local.v2.u64 	[%rd15+32], {%rd1283, %rd1284};
	mov.b64	%rd1285, {%r221, %r222};
	shl.b64 	%rd1286, %rd1285, 3;
	and.b64  	%rd1287, %rd1286, 56;
	add.s64 	%rd1288, %rd12, %rd1287;
	and.b64  	%rd1289, %rd1285, 56;
	add.s64 	%rd1290, %rd12, %rd1289;
	ld.local.u64 	%rd1291, [%rd1290];
	shl.b64 	%rd1292, %rd1291, 3;
	ld.local.u64 	%rd1293, [%rd1288];
	xor.b64  	%rd1294, %rd1293, %rd1292;
	shr.u64 	%rd1295, %rd1291, 61;
	mov.u32 	%r1615, 0;
	mov.b64	%rd1296, {%r1615, %r1615};
	xor.b64  	%rd1297, %rd1296, %rd1295;
	shr.u64 	%rd1298, %rd1285, 3;
	and.b64  	%rd1299, %rd1298, 56;
	add.s64 	%rd1300, %rd12, %rd1299;
	ld.local.u64 	%rd1301, [%rd1300];
	shl.b64 	%rd1302, %rd1301, 6;
	mov.b64	{%r1616, %r1617}, %rd1294;
	mov.b64	%rd1303, {%r1616, %r1617};
	xor.b64  	%rd1304, %rd1303, %rd1302;
	shr.u64 	%rd1305, %rd1301, 58;
	mov.b64	{%r1618, %r1619}, %rd1297;
	mov.b64	%rd1306, {%r1618, %r1619};
	xor.b64  	%rd1307, %rd1306, %rd1305;
	shr.u64 	%rd1308, %rd1285, 6;
	and.b64  	%rd1309, %rd1308, 56;
	add.s64 	%rd1310, %rd12, %rd1309;
	ld.local.u64 	%rd1311, [%rd1310];
	shl.b64 	%rd1312, %rd1311, 9;
	mov.b64	{%r1620, %r1621}, %rd1304;
	mov.b64	%rd1313, {%r1620, %r1621};
	xor.b64  	%rd1314, %rd1313, %rd1312;
	shr.u64 	%rd1315, %rd1311, 55;
	mov.b64	{%r1622, %r1623}, %rd1307;
	mov.b64	%rd1316, {%r1622, %r1623};
	xor.b64  	%rd1317, %rd1316, %rd1315;
	shr.u64 	%rd1318, %rd1285, 9;
	and.b64  	%rd1319, %rd1318, 56;
	add.s64 	%rd1320, %rd12, %rd1319;
	ld.local.u64 	%rd1321, [%rd1320];
	shl.b64 	%rd1322, %rd1321, 12;
	mov.b64	{%r1624, %r1625}, %rd1314;
	mov.b64	%rd1323, {%r1624, %r1625};
	xor.b64  	%rd1324, %rd1323, %rd1322;
	shr.u64 	%rd1325, %rd1321, 52;
	mov.b64	{%r1626, %r1627}, %rd1317;
	mov.b64	%rd1326, {%r1626, %r1627};
	xor.b64  	%rd1327, %rd1326, %rd1325;
	shr.u64 	%rd1328, %rd1285, 12;
	and.b64  	%rd1329, %rd1328, 56;
	add.s64 	%rd1330, %rd12, %rd1329;
	ld.local.u64 	%rd1331, [%rd1330];
	shl.b64 	%rd1332, %rd1331, 15;
	mov.b64	{%r1628, %r1629}, %rd1324;
	mov.b64	%rd1333, {%r1628, %r1629};
	xor.b64  	%rd1334, %rd1333, %rd1332;
	shr.u64 	%rd1335, %rd1331, 49;
	mov.b64	{%r1630, %r1631}, %rd1327;
	mov.b64	%rd1336, {%r1630, %r1631};
	xor.b64  	%rd1337, %rd1336, %rd1335;
	shr.u64 	%rd1338, %rd1285, 15;
	and.b64  	%rd1339, %rd1338, 56;
	add.s64 	%rd1340, %rd12, %rd1339;
	ld.local.u64 	%rd1341, [%rd1340];
	shl.b64 	%rd1342, %rd1341, 18;
	mov.b64	{%r1632, %r1633}, %rd1334;
	mov.b64	%rd1343, {%r1632, %r1633};
	xor.b64  	%rd1344, %rd1343, %rd1342;
	shr.u64 	%rd1345, %rd1341, 46;
	mov.b64	{%r1634, %r1635}, %rd1337;
	mov.b64	%rd1346, {%r1634, %r1635};
	xor.b64  	%rd1347, %rd1346, %rd1345;
	shr.u64 	%rd1348, %rd1285, 18;
	and.b64  	%rd1349, %rd1348, 56;
	add.s64 	%rd1350, %rd12, %rd1349;
	ld.local.u64 	%rd1351, [%rd1350];
	shl.b64 	%rd1352, %rd1351, 21;
	mov.b64	{%r1636, %r1637}, %rd1344;
	mov.b64	%rd1353, {%r1636, %r1637};
	xor.b64  	%rd1354, %rd1353, %rd1352;
	shr.u64 	%rd1355, %rd1351, 43;
	mov.b64	{%r1638, %r1639}, %rd1347;
	mov.b64	%rd1356, {%r1638, %r1639};
	xor.b64  	%rd1357, %rd1356, %rd1355;
	shr.u64 	%rd1358, %rd1285, 21;
	and.b64  	%rd1359, %rd1358, 56;
	add.s64 	%rd1360, %rd12, %rd1359;
	ld.local.u64 	%rd1361, [%rd1360];
	shl.b64 	%rd1362, %rd1361, 24;
	mov.b64	{%r1640, %r1641}, %rd1354;
	mov.b64	%rd1363, {%r1640, %r1641};
	xor.b64  	%rd1364, %rd1363, %rd1362;
	shr.u64 	%rd1365, %rd1361, 40;
	mov.b64	{%r1642, %r1643}, %rd1357;
	mov.b64	%rd1366, {%r1642, %r1643};
	xor.b64  	%rd1367, %rd1366, %rd1365;
	shr.u64 	%rd1368, %rd1285, 24;
	and.b64  	%rd1369, %rd1368, 56;
	add.s64 	%rd1370, %rd12, %rd1369;
	ld.local.u64 	%rd1371, [%rd1370];
	shl.b64 	%rd1372, %rd1371, 27;
	mov.b64	{%r1644, %r1645}, %rd1364;
	mov.b64	%rd1373, {%r1644, %r1645};
	xor.b64  	%rd1374, %rd1373, %rd1372;
	shr.u64 	%rd1375, %rd1371, 37;
	mov.b64	{%r1646, %r1647}, %rd1367;
	mov.b64	%rd1376, {%r1646, %r1647};
	xor.b64  	%rd1377, %rd1376, %rd1375;
	shr.u64 	%rd1378, %rd1285, 27;
	and.b64  	%rd1379, %rd1378, 56;
	add.s64 	%rd1380, %rd12, %rd1379;
	ld.local.u64 	%rd1381, [%rd1380];
	shl.b64 	%rd1382, %rd1381, 30;
	mov.b64	{%r1648, %r1649}, %rd1374;
	mov.b64	%rd1383, {%r1648, %r1649};
	xor.b64  	%rd1384, %rd1383, %rd1382;
	shr.u64 	%rd1385, %rd1381, 34;
	mov.b64	{%r1650, %r1651}, %rd1377;
	mov.b64	%rd1386, {%r1650, %r1651};
	xor.b64  	%rd1387, %rd1386, %rd1385;
	shr.u64 	%rd1388, %rd1285, 30;
	and.b64  	%rd1389, %rd1388, 56;
	add.s64 	%rd1390, %rd12, %rd1389;
	ld.local.u64 	%rd1391, [%rd1390];
	shl.b64 	%rd1392, %rd1391, 33;
	mov.b64	{%r1652, %r1653}, %rd1384;
	mov.b64	%rd1393, {%r1652, %r1653};
	xor.b64  	%rd1394, %rd1393, %rd1392;
	shr.u64 	%rd1395, %rd1391, 31;
	mov.b64	{%r1654, %r1655}, %rd1387;
	mov.b64	%rd1396, {%r1654, %r1655};
	xor.b64  	%rd1397, %rd1396, %rd1395;
	shr.u64 	%rd1398, %rd1285, 33;
	and.b64  	%rd1399, %rd1398, 56;
	add.s64 	%rd1400, %rd12, %rd1399;
	ld.local.u64 	%rd1401, [%rd1400];
	shl.b64 	%rd1402, %rd1401, 36;
	mov.b64	{%r1656, %r1657}, %rd1394;
	mov.b64	%rd1403, {%r1656, %r1657};
	xor.b64  	%rd1404, %rd1403, %rd1402;
	shr.u64 	%rd1405, %rd1401, 28;
	mov.b64	{%r1658, %r1659}, %rd1397;
	mov.b64	%rd1406, {%r1658, %r1659};
	xor.b64  	%rd1407, %rd1406, %rd1405;
	shr.u64 	%rd1408, %rd1285, 36;
	and.b64  	%rd1409, %rd1408, 56;
	add.s64 	%rd1410, %rd12, %rd1409;
	ld.local.u64 	%rd1411, [%rd1410];
	shl.b64 	%rd1412, %rd1411, 39;
	mov.b64	{%r1660, %r1661}, %rd1404;
	mov.b64	%rd1413, {%r1660, %r1661};
	xor.b64  	%rd1414, %rd1413, %rd1412;
	shr.u64 	%rd1415, %rd1411, 25;
	mov.b64	{%r1662, %r1663}, %rd1407;
	mov.b64	%rd1416, {%r1662, %r1663};
	xor.b64  	%rd1417, %rd1416, %rd1415;
	shr.u64 	%rd1418, %rd1285, 39;
	and.b64  	%rd1419, %rd1418, 56;
	add.s64 	%rd1420, %rd12, %rd1419;
	ld.local.u64 	%rd1421, [%rd1420];
	shl.b64 	%rd1422, %rd1421, 42;
	mov.b64	{%r1664, %r1665}, %rd1414;
	mov.b64	%rd1423, {%r1664, %r1665};
	xor.b64  	%rd1424, %rd1423, %rd1422;
	shr.u64 	%rd1425, %rd1421, 22;
	mov.b64	{%r1666, %r1667}, %rd1417;
	mov.b64	%rd1426, {%r1666, %r1667};
	xor.b64  	%rd1427, %rd1426, %rd1425;
	shr.u64 	%rd1428, %rd1285, 42;
	and.b64  	%rd1429, %rd1428, 56;
	add.s64 	%rd1430, %rd12, %rd1429;
	ld.local.u64 	%rd1431, [%rd1430];
	shl.b64 	%rd1432, %rd1431, 45;
	mov.b64	{%r1668, %r1669}, %rd1424;
	mov.b64	%rd1433, {%r1668, %r1669};
	xor.b64  	%rd1434, %rd1433, %rd1432;
	shr.u64 	%rd1435, %rd1431, 19;
	mov.b64	{%r1670, %r1671}, %rd1427;
	mov.b64	%rd1436, {%r1670, %r1671};
	xor.b64  	%rd1437, %rd1436, %rd1435;
	shr.u64 	%rd1438, %rd1285, 45;
	and.b64  	%rd1439, %rd1438, 56;
	add.s64 	%rd1440, %rd12, %rd1439;
	ld.local.u64 	%rd1441, [%rd1440];
	shl.b64 	%rd1442, %rd1441, 48;
	mov.b64	{%r1672, %r1673}, %rd1434;
	mov.b64	%rd1443, {%r1672, %r1673};
	xor.b64  	%rd1444, %rd1443, %rd1442;
	shr.u64 	%rd1445, %rd1441, 16;
	mov.b64	{%r1674, %r1675}, %rd1437;
	mov.b64	%rd1446, {%r1674, %r1675};
	xor.b64  	%rd1447, %rd1446, %rd1445;
	shr.u64 	%rd1448, %rd1285, 48;
	and.b64  	%rd1449, %rd1448, 56;
	add.s64 	%rd1450, %rd12, %rd1449;
	ld.local.u64 	%rd1451, [%rd1450];
	shl.b64 	%rd1452, %rd1451, 51;
	mov.b64	{%r1676, %r1677}, %rd1444;
	mov.b64	%rd1453, {%r1676, %r1677};
	xor.b64  	%rd1454, %rd1453, %rd1452;
	shr.u64 	%rd1455, %rd1451, 13;
	mov.b64	{%r1678, %r1679}, %rd1447;
	mov.b64	%rd1456, {%r1678, %r1679};
	xor.b64  	%rd1457, %rd1456, %rd1455;
	shr.u64 	%rd1458, %rd1285, 51;
	and.b64  	%rd1459, %rd1458, 56;
	add.s64 	%rd1460, %rd12, %rd1459;
	ld.local.u64 	%rd1461, [%rd1460];
	shl.b64 	%rd1462, %rd1461, 54;
	mov.b64	{%r1680, %r1681}, %rd1454;
	mov.b64	%rd1463, {%r1680, %r1681};
	xor.b64  	%rd1464, %rd1463, %rd1462;
	shr.u64 	%rd1465, %rd1461, 10;
	mov.b64	{%r1682, %r1683}, %rd1457;
	mov.b64	%rd1466, {%r1682, %r1683};
	xor.b64  	%rd1467, %rd1466, %rd1465;
	shr.u64 	%rd1468, %rd1285, 54;
	and.b64  	%rd1469, %rd1468, 56;
	add.s64 	%rd1470, %rd12, %rd1469;
	ld.local.u64 	%rd1471, [%rd1470];
	shl.b64 	%rd1472, %rd1471, 57;
	mov.b64	{%r1684, %r1685}, %rd1464;
	mov.b64	%rd1473, {%r1684, %r1685};
	xor.b64  	%rd1474, %rd1473, %rd1472;
	shr.u64 	%rd1475, %rd1471, 7;
	mov.b64	{%r1686, %r1687}, %rd1467;
	mov.b64	%rd1476, {%r1686, %r1687};
	xor.b64  	%rd1477, %rd1476, %rd1475;
	shr.u64 	%rd1478, %rd1285, 57;
	and.b64  	%rd1479, %rd1478, 56;
	add.s64 	%rd1480, %rd12, %rd1479;
	ld.local.u64 	%rd1481, [%rd1480];
	shl.b64 	%rd1482, %rd1481, 60;
	mov.b64	{%r1688, %r1689}, %rd1474;
	mov.b64	%rd1483, {%r1688, %r1689};
	xor.b64  	%rd1484, %rd1483, %rd1482;
	shr.u64 	%rd1485, %rd1481, 4;
	mov.b64	{%r1690, %r1691}, %rd1477;
	mov.b64	%rd1486, {%r1690, %r1691};
	xor.b64  	%rd1487, %rd1486, %rd1485;
	shr.u64 	%rd1488, %rd1285, 60;
	and.b64  	%rd1489, %rd1488, 8;
	add.s64 	%rd1490, %rd12, %rd1489;
	ld.local.u64 	%rd1491, [%rd1490];
	shl.b64 	%rd1492, %rd1491, 63;
	mov.b64	{%r1692, %r1693}, %rd1484;
	mov.b64	%rd1493, {%r1692, %r1693};
	xor.b64  	%rd1494, %rd1493, %rd1492;
	shr.u64 	%rd1495, %rd1491, 1;
	mov.b64	{%r1694, %r1695}, %rd1487;
	mov.b64	%rd1496, {%r1694, %r1695};
	xor.b64  	%rd1497, %rd1496, %rd1495;
	mov.b64	{%r6545, %r6546}, %rd1497;
	mov.b64	{%r6547, %r6548}, %rd1494;
	setp.gt.s32	%p26, %r224, -1;
	@%p26 bra 	BB0_44;

	xor.b32  	%r6466, %r6722, %r1612;
	xor.b32  	%r6465, %r6721, %r1611;
	shr.u32 	%r1696, %r6465, 1;
	shl.b32 	%r1697, %r6466, 31;
	shr.u32 	%r1698, %r6466, 1;
	and.b32  	%r1699, %r1696, 1533916891;
	or.b32  	%r1700, %r1699, %r1697;
	xor.b32  	%r6545, %r6545, %r1700;
	and.b32  	%r1701, %r1698, 920350134;
	xor.b32  	%r6546, %r6546, %r1701;

BB0_44:
	and.b32  	%r1702, %r224, 1073741824;
	setp.eq.s32	%p27, %r1702, 0;
	@%p27 bra 	BB0_46;

	xor.b32  	%r6464, %r6722, %r1612;
	xor.b32  	%r6463, %r6721, %r1611;
	shr.u32 	%r1703, %r6463, 2;
	shl.b32 	%r1704, %r6464, 30;
	or.b32  	%r1705, %r1703, %r1704;
	shr.u32 	%r1706, %r6464, 2;
	and.b32  	%r1707, %r1705, 1227133513;
	xor.b32  	%r6545, %r6545, %r1707;
	and.b32  	%r1708, %r1706, 306783378;
	xor.b32  	%r6546, %r6546, %r1708;

BB0_46:
	mov.u64 	%rd6089, 0;
	mov.b64	%rd1498, {%r1613, %r1614};
	st.local.v2.u64 	[%rd12], {%rd6089, %rd1498};
	shl.b64 	%rd1502, %rd1498, 1;
	xor.b64  	%rd1503, %rd1502, %rd1498;
	st.local.v2.u64 	[%rd15], {%rd1502, %rd1503};
	shl.b64 	%rd1504, %rd1498, 2;
	xor.b64  	%rd1505, %rd1504, %rd1498;
	st.local.v2.u64 	[%rd15+16], {%rd1504, %rd1505};
	shl.b64 	%rd1506, %rd1503, 1;
	xor.b64  	%rd1507, %rd1506, %rd1498;
	st.local.v2.u64 	[%rd15+32], {%rd1506, %rd1507};
	mov.b64	%rd1508, {%r1611, %r1612};
	shl.b64 	%rd1509, %rd1508, 3;
	and.b64  	%rd1510, %rd1509, 56;
	add.s64 	%rd1511, %rd12, %rd1510;
	and.b64  	%rd1512, %rd1508, 56;
	add.s64 	%rd1513, %rd12, %rd1512;
	ld.local.u64 	%rd1514, [%rd1513];
	shl.b64 	%rd1515, %rd1514, 3;
	ld.local.u64 	%rd1516, [%rd1511];
	xor.b64  	%rd1517, %rd1516, %rd1515;
	shr.u64 	%rd1518, %rd1514, 61;
	mov.b64	%rd1519, {%r1615, %r1615};
	xor.b64  	%rd1520, %rd1519, %rd1518;
	shr.u64 	%rd1521, %rd1508, 3;
	and.b64  	%rd1522, %rd1521, 56;
	add.s64 	%rd1523, %rd12, %rd1522;
	ld.local.u64 	%rd1524, [%rd1523];
	shl.b64 	%rd1525, %rd1524, 6;
	mov.b64	{%r1710, %r1711}, %rd1517;
	mov.b64	%rd1526, {%r1710, %r1711};
	xor.b64  	%rd1527, %rd1526, %rd1525;
	shr.u64 	%rd1528, %rd1524, 58;
	mov.b64	{%r1712, %r1713}, %rd1520;
	mov.b64	%rd1529, {%r1712, %r1713};
	xor.b64  	%rd1530, %rd1529, %rd1528;
	shr.u64 	%rd1531, %rd1508, 6;
	and.b64  	%rd1532, %rd1531, 56;
	add.s64 	%rd1533, %rd12, %rd1532;
	ld.local.u64 	%rd1534, [%rd1533];
	shl.b64 	%rd1535, %rd1534, 9;
	mov.b64	{%r1714, %r1715}, %rd1527;
	mov.b64	%rd1536, {%r1714, %r1715};
	xor.b64  	%rd1537, %rd1536, %rd1535;
	shr.u64 	%rd1538, %rd1534, 55;
	mov.b64	{%r1716, %r1717}, %rd1530;
	mov.b64	%rd1539, {%r1716, %r1717};
	xor.b64  	%rd1540, %rd1539, %rd1538;
	shr.u64 	%rd1541, %rd1508, 9;
	and.b64  	%rd1542, %rd1541, 56;
	add.s64 	%rd1543, %rd12, %rd1542;
	ld.local.u64 	%rd1544, [%rd1543];
	shl.b64 	%rd1545, %rd1544, 12;
	mov.b64	{%r1718, %r1719}, %rd1537;
	mov.b64	%rd1546, {%r1718, %r1719};
	xor.b64  	%rd1547, %rd1546, %rd1545;
	shr.u64 	%rd1548, %rd1544, 52;
	mov.b64	{%r1720, %r1721}, %rd1540;
	mov.b64	%rd1549, {%r1720, %r1721};
	xor.b64  	%rd1550, %rd1549, %rd1548;
	shr.u64 	%rd1551, %rd1508, 12;
	and.b64  	%rd1552, %rd1551, 56;
	add.s64 	%rd1553, %rd12, %rd1552;
	ld.local.u64 	%rd1554, [%rd1553];
	shl.b64 	%rd1555, %rd1554, 15;
	mov.b64	{%r1722, %r1723}, %rd1547;
	mov.b64	%rd1556, {%r1722, %r1723};
	xor.b64  	%rd1557, %rd1556, %rd1555;
	shr.u64 	%rd1558, %rd1554, 49;
	mov.b64	{%r1724, %r1725}, %rd1550;
	mov.b64	%rd1559, {%r1724, %r1725};
	xor.b64  	%rd1560, %rd1559, %rd1558;
	shr.u64 	%rd1561, %rd1508, 15;
	and.b64  	%rd1562, %rd1561, 56;
	add.s64 	%rd1563, %rd12, %rd1562;
	ld.local.u64 	%rd1564, [%rd1563];
	shl.b64 	%rd1565, %rd1564, 18;
	mov.b64	{%r1726, %r1727}, %rd1557;
	mov.b64	%rd1566, {%r1726, %r1727};
	xor.b64  	%rd1567, %rd1566, %rd1565;
	shr.u64 	%rd1568, %rd1564, 46;
	mov.b64	{%r1728, %r1729}, %rd1560;
	mov.b64	%rd1569, {%r1728, %r1729};
	xor.b64  	%rd1570, %rd1569, %rd1568;
	shr.u64 	%rd1571, %rd1508, 18;
	and.b64  	%rd1572, %rd1571, 56;
	add.s64 	%rd1573, %rd12, %rd1572;
	ld.local.u64 	%rd1574, [%rd1573];
	shl.b64 	%rd1575, %rd1574, 21;
	mov.b64	{%r1730, %r1731}, %rd1567;
	mov.b64	%rd1576, {%r1730, %r1731};
	xor.b64  	%rd1577, %rd1576, %rd1575;
	shr.u64 	%rd1578, %rd1574, 43;
	mov.b64	{%r1732, %r1733}, %rd1570;
	mov.b64	%rd1579, {%r1732, %r1733};
	xor.b64  	%rd1580, %rd1579, %rd1578;
	shr.u64 	%rd1581, %rd1508, 21;
	and.b64  	%rd1582, %rd1581, 56;
	add.s64 	%rd1583, %rd12, %rd1582;
	ld.local.u64 	%rd1584, [%rd1583];
	shl.b64 	%rd1585, %rd1584, 24;
	mov.b64	{%r1734, %r1735}, %rd1577;
	mov.b64	%rd1586, {%r1734, %r1735};
	xor.b64  	%rd1587, %rd1586, %rd1585;
	shr.u64 	%rd1588, %rd1584, 40;
	mov.b64	{%r1736, %r1737}, %rd1580;
	mov.b64	%rd1589, {%r1736, %r1737};
	xor.b64  	%rd1590, %rd1589, %rd1588;
	shr.u64 	%rd1591, %rd1508, 24;
	and.b64  	%rd1592, %rd1591, 56;
	add.s64 	%rd1593, %rd12, %rd1592;
	ld.local.u64 	%rd1594, [%rd1593];
	shl.b64 	%rd1595, %rd1594, 27;
	mov.b64	{%r1738, %r1739}, %rd1587;
	mov.b64	%rd1596, {%r1738, %r1739};
	xor.b64  	%rd1597, %rd1596, %rd1595;
	shr.u64 	%rd1598, %rd1594, 37;
	mov.b64	{%r1740, %r1741}, %rd1590;
	mov.b64	%rd1599, {%r1740, %r1741};
	xor.b64  	%rd1600, %rd1599, %rd1598;
	shr.u64 	%rd1601, %rd1508, 27;
	and.b64  	%rd1602, %rd1601, 56;
	add.s64 	%rd1603, %rd12, %rd1602;
	ld.local.u64 	%rd1604, [%rd1603];
	shl.b64 	%rd1605, %rd1604, 30;
	mov.b64	{%r1742, %r1743}, %rd1597;
	mov.b64	%rd1606, {%r1742, %r1743};
	xor.b64  	%rd1607, %rd1606, %rd1605;
	shr.u64 	%rd1608, %rd1604, 34;
	mov.b64	{%r1744, %r1745}, %rd1600;
	mov.b64	%rd1609, {%r1744, %r1745};
	xor.b64  	%rd1610, %rd1609, %rd1608;
	shr.u64 	%rd1611, %rd1508, 30;
	and.b64  	%rd1612, %rd1611, 56;
	add.s64 	%rd1613, %rd12, %rd1612;
	ld.local.u64 	%rd1614, [%rd1613];
	shl.b64 	%rd1615, %rd1614, 33;
	mov.b64	{%r1746, %r1747}, %rd1607;
	mov.b64	%rd1616, {%r1746, %r1747};
	xor.b64  	%rd1617, %rd1616, %rd1615;
	shr.u64 	%rd1618, %rd1614, 31;
	mov.b64	{%r1748, %r1749}, %rd1610;
	mov.b64	%rd1619, {%r1748, %r1749};
	xor.b64  	%rd1620, %rd1619, %rd1618;
	shr.u64 	%rd1621, %rd1508, 33;
	and.b64  	%rd1622, %rd1621, 56;
	add.s64 	%rd1623, %rd12, %rd1622;
	ld.local.u64 	%rd1624, [%rd1623];
	shl.b64 	%rd1625, %rd1624, 36;
	mov.b64	{%r1750, %r1751}, %rd1617;
	mov.b64	%rd1626, {%r1750, %r1751};
	xor.b64  	%rd1627, %rd1626, %rd1625;
	shr.u64 	%rd1628, %rd1624, 28;
	mov.b64	{%r1752, %r1753}, %rd1620;
	mov.b64	%rd1629, {%r1752, %r1753};
	xor.b64  	%rd1630, %rd1629, %rd1628;
	shr.u64 	%rd1631, %rd1508, 36;
	and.b64  	%rd1632, %rd1631, 56;
	add.s64 	%rd1633, %rd12, %rd1632;
	ld.local.u64 	%rd1634, [%rd1633];
	shl.b64 	%rd1635, %rd1634, 39;
	mov.b64	{%r1754, %r1755}, %rd1627;
	mov.b64	%rd1636, {%r1754, %r1755};
	xor.b64  	%rd1637, %rd1636, %rd1635;
	shr.u64 	%rd1638, %rd1634, 25;
	mov.b64	{%r1756, %r1757}, %rd1630;
	mov.b64	%rd1639, {%r1756, %r1757};
	xor.b64  	%rd1640, %rd1639, %rd1638;
	shr.u64 	%rd1641, %rd1508, 39;
	and.b64  	%rd1642, %rd1641, 56;
	add.s64 	%rd1643, %rd12, %rd1642;
	ld.local.u64 	%rd1644, [%rd1643];
	shl.b64 	%rd1645, %rd1644, 42;
	mov.b64	{%r1758, %r1759}, %rd1637;
	mov.b64	%rd1646, {%r1758, %r1759};
	xor.b64  	%rd1647, %rd1646, %rd1645;
	shr.u64 	%rd1648, %rd1644, 22;
	mov.b64	{%r1760, %r1761}, %rd1640;
	mov.b64	%rd1649, {%r1760, %r1761};
	xor.b64  	%rd1650, %rd1649, %rd1648;
	shr.u64 	%rd1651, %rd1508, 42;
	and.b64  	%rd1652, %rd1651, 56;
	add.s64 	%rd1653, %rd12, %rd1652;
	ld.local.u64 	%rd1654, [%rd1653];
	shl.b64 	%rd1655, %rd1654, 45;
	mov.b64	{%r1762, %r1763}, %rd1647;
	mov.b64	%rd1656, {%r1762, %r1763};
	xor.b64  	%rd1657, %rd1656, %rd1655;
	shr.u64 	%rd1658, %rd1654, 19;
	mov.b64	{%r1764, %r1765}, %rd1650;
	mov.b64	%rd1659, {%r1764, %r1765};
	xor.b64  	%rd1660, %rd1659, %rd1658;
	shr.u64 	%rd1661, %rd1508, 45;
	and.b64  	%rd1662, %rd1661, 56;
	add.s64 	%rd1663, %rd12, %rd1662;
	ld.local.u64 	%rd1664, [%rd1663];
	shl.b64 	%rd1665, %rd1664, 48;
	mov.b64	{%r1766, %r1767}, %rd1657;
	mov.b64	%rd1666, {%r1766, %r1767};
	xor.b64  	%rd1667, %rd1666, %rd1665;
	shr.u64 	%rd1668, %rd1664, 16;
	mov.b64	{%r1768, %r1769}, %rd1660;
	mov.b64	%rd1669, {%r1768, %r1769};
	xor.b64  	%rd1670, %rd1669, %rd1668;
	shr.u64 	%rd1671, %rd1508, 48;
	and.b64  	%rd1672, %rd1671, 56;
	add.s64 	%rd1673, %rd12, %rd1672;
	ld.local.u64 	%rd1674, [%rd1673];
	shl.b64 	%rd1675, %rd1674, 51;
	mov.b64	{%r1770, %r1771}, %rd1667;
	mov.b64	%rd1676, {%r1770, %r1771};
	xor.b64  	%rd1677, %rd1676, %rd1675;
	shr.u64 	%rd1678, %rd1674, 13;
	mov.b64	{%r1772, %r1773}, %rd1670;
	mov.b64	%rd1679, {%r1772, %r1773};
	xor.b64  	%rd1680, %rd1679, %rd1678;
	shr.u64 	%rd1681, %rd1508, 51;
	and.b64  	%rd1682, %rd1681, 56;
	add.s64 	%rd1683, %rd12, %rd1682;
	ld.local.u64 	%rd1684, [%rd1683];
	shl.b64 	%rd1685, %rd1684, 54;
	mov.b64	{%r1774, %r1775}, %rd1677;
	mov.b64	%rd1686, {%r1774, %r1775};
	xor.b64  	%rd1687, %rd1686, %rd1685;
	shr.u64 	%rd1688, %rd1684, 10;
	mov.b64	{%r1776, %r1777}, %rd1680;
	mov.b64	%rd1689, {%r1776, %r1777};
	xor.b64  	%rd1690, %rd1689, %rd1688;
	shr.u64 	%rd1691, %rd1508, 54;
	and.b64  	%rd1692, %rd1691, 56;
	add.s64 	%rd1693, %rd12, %rd1692;
	ld.local.u64 	%rd1694, [%rd1693];
	shl.b64 	%rd1695, %rd1694, 57;
	mov.b64	{%r1778, %r1779}, %rd1687;
	mov.b64	%rd1696, {%r1778, %r1779};
	xor.b64  	%rd1697, %rd1696, %rd1695;
	shr.u64 	%rd1698, %rd1694, 7;
	mov.b64	{%r1780, %r1781}, %rd1690;
	mov.b64	%rd1699, {%r1780, %r1781};
	xor.b64  	%rd1700, %rd1699, %rd1698;
	shr.u64 	%rd1701, %rd1508, 57;
	and.b64  	%rd1702, %rd1701, 56;
	add.s64 	%rd1703, %rd12, %rd1702;
	ld.local.u64 	%rd1704, [%rd1703];
	shl.b64 	%rd1705, %rd1704, 60;
	mov.b64	{%r1782, %r1783}, %rd1697;
	mov.b64	%rd1706, {%r1782, %r1783};
	xor.b64  	%rd1707, %rd1706, %rd1705;
	shr.u64 	%rd1708, %rd1704, 4;
	mov.b64	{%r1784, %r1785}, %rd1700;
	mov.b64	%rd1709, {%r1784, %r1785};
	xor.b64  	%rd1710, %rd1709, %rd1708;
	shr.u64 	%rd1711, %rd1508, 60;
	and.b64  	%rd1712, %rd1711, 8;
	add.s64 	%rd1713, %rd12, %rd1712;
	ld.local.u64 	%rd1714, [%rd1713];
	shl.b64 	%rd1715, %rd1714, 63;
	mov.b64	{%r1786, %r1787}, %rd1707;
	mov.b64	%rd1716, {%r1786, %r1787};
	xor.b64  	%rd1717, %rd1716, %rd1715;
	shr.u64 	%rd1718, %rd1714, 1;
	mov.b64	{%r1788, %r1789}, %rd1710;
	mov.b64	%rd1719, {%r1788, %r1789};
	xor.b64  	%rd1720, %rd1719, %rd1718;
	mov.b64	{%r6553, %r6554}, %rd1720;
	mov.b64	{%r6555, %r6556}, %rd1717;
	setp.gt.s32	%p28, %r1614, -1;
	@%p28 bra 	BB0_48;

	shr.u32 	%r1790, %r1611, 1;
	shl.b32 	%r1791, %r1612, 31;
	shr.u32 	%r1792, %r1612, 1;
	and.b32  	%r1793, %r1790, 1533916891;
	or.b32  	%r1794, %r1793, %r1791;
	xor.b32  	%r6553, %r6553, %r1794;
	and.b32  	%r1795, %r1792, 920350134;
	xor.b32  	%r6554, %r6554, %r1795;

BB0_48:
	and.b32  	%r1796, %r1614, 1073741824;
	setp.eq.s32	%p29, %r1796, 0;
	@%p29 bra 	BB0_50;

	shr.u32 	%r1797, %r1611, 2;
	shl.b32 	%r1798, %r1612, 30;
	or.b32  	%r1799, %r1797, %r1798;
	shr.u32 	%r1800, %r1612, 2;
	and.b32  	%r1801, %r1799, 1227133513;
	xor.b32  	%r6553, %r6553, %r1801;
	and.b32  	%r1802, %r1800, 306783378;
	xor.b32  	%r6554, %r6554, %r1802;

BB0_50:
	xor.b32  	%r1803, %r6546, %r6733;
	xor.b32  	%r1804, %r6545, %r6732;
	xor.b32  	%r1805, %r6548, %r6731;
	xor.b32  	%r1806, %r6547, %r6730;
	xor.b32  	%r1807, %r6555, %r1806;
	xor.b32  	%r1808, %r6556, %r1805;
	xor.b32  	%r1809, %r6553, %r1804;
	xor.b32  	%r1810, %r6554, %r1803;
	mov.b64	%rd1721, {%r1809, %r1810};
	mov.b64	%rd1722, {%r1807, %r1808};
	cvt.u32.u64	%r1811, %rd1722;
	cvt.s32.s16 	%r1812, %r1811;
	mov.b64	%rd1723, {%r6723, %r6724};
	mov.b64	%rd1724, {%r6721, %r6722};
	cvt.u32.u64	%r1813, %rd1724;
	cvt.s32.s16 	%r1814, %r1813;
	mad.lo.s32 	%r1815, %r1812, %r1814, 16384;
	bfe.u32 	%r1816, %r1815, 15, 16;
	shr.s32 	%r1817, %r1811, 16;
	shr.s32 	%r1818, %r1813, 16;
	mad.lo.s32 	%r1819, %r1817, %r1818, 16384;
	mov.b64	%rd1725, {%r1816, %r1615};
	shl.b32 	%r1821, %r1819, 1;
	cvt.u64.u32	%rd1726, %r1821;
	and.b64  	%rd1727, %rd1725, 65535;
	shr.u64 	%rd1728, %rd1721, 16;
	shr.u64 	%rd1729, %rd1722, 16;
	cvt.u32.u64	%r1822, %rd1729;
	shr.s32 	%r1823, %r1822, 16;
	shr.u64 	%rd1730, %rd1723, 16;
	shr.u64 	%rd1731, %rd1724, 16;
	cvt.u32.u64	%r1824, %rd1731;
	shr.s32 	%r1825, %r1824, 16;
	mad.lo.s32 	%r1826, %r1823, %r1825, 16384;
	mul.wide.u32 	%rd1732, %r1826, 131072;
	shr.u64 	%rd1733, %rd1722, 32;
	cvt.u32.u64	%r1827, %rd1733;
	shr.s32 	%r1828, %r1827, 16;
	shr.u64 	%rd1734, %rd1724, 32;
	cvt.u32.u64	%r1829, %rd1734;
	shr.s32 	%r1830, %r1829, 16;
	mad.lo.s32 	%r1831, %r1828, %r1830, 16384;
	shr.u32 	%r1832, %r1831, 15;
	cvt.u64.u32	%rd1735, %r1832;
	shl.b64 	%rd1736, %rd1735, 48;
	and.b64  	%rd1737, %rd1732, 281470681743360;
	shl.b64 	%rd1738, %rd1721, 16;
	cvt.u32.u64	%r1833, %rd1738;
	shr.s32 	%r1834, %r1833, 16;
	shl.b64 	%rd1739, %rd1723, 16;
	cvt.u32.u64	%r1835, %rd1739;
	shr.s32 	%r1836, %r1835, 16;
	mad.lo.s32 	%r1837, %r1834, %r1836, 16384;
	shr.u32 	%r1838, %r1837, 15;
	cvt.u64.u32	%rd1740, %r1838;
	cvt.u32.u64	%r1839, %rd1721;
	shr.s32 	%r1840, %r1839, 16;
	cvt.u32.u64	%r1841, %rd1723;
	shr.s32 	%r1842, %r1841, 16;
	mad.lo.s32 	%r1843, %r1840, %r1842, 16384;
	and.b64  	%rd1741, %rd1740, 65535;
	and.b64  	%rd1742, %rd1726, 4294901760;
	or.b64  	%rd1743, %rd1742, %rd1727;
	or.b64  	%rd1744, %rd1743, %rd1737;
	shl.b32 	%r1844, %r1843, 1;
	cvt.u64.u32	%rd1745, %r1844;
	cvt.u32.u64	%r1845, %rd1728;
	shr.s32 	%r1846, %r1845, 16;
	cvt.u32.u64	%r1847, %rd1730;
	shr.s32 	%r1848, %r1847, 16;
	mad.lo.s32 	%r1849, %r1846, %r1848, 16384;
	mul.wide.u32 	%rd1746, %r1849, 131072;
	and.b64  	%rd1747, %rd1745, 4294901760;
	or.b64  	%rd1748, %rd1747, %rd1741;
	shr.u64 	%rd1749, %rd1721, 32;
	cvt.u32.u64	%r1850, %rd1749;
	shr.s32 	%r1851, %r1850, 16;
	shr.u64 	%rd1750, %rd1723, 32;
	cvt.u32.u64	%r1852, %rd1750;
	shr.s32 	%r1853, %r1852, 16;
	mad.lo.s32 	%r1854, %r1851, %r1853, 16384;
	shr.u32 	%r1855, %r1854, 15;
	cvt.u64.u32	%rd1751, %r1855;
	shl.b64 	%rd1752, %rd1751, 48;
	and.b64  	%rd1753, %rd1746, 281470681743360;
	or.b64  	%rd1754, %rd1753, %rd1752;
	or.b64  	%rd1755, %rd1754, %rd1748;
	or.b64  	%rd1756, %rd1744, %rd1736;
	mov.b64	{%r1856, %r1857}, %rd1755;
	mov.b64	{%r1858, %r1859}, %rd1756;
	xor.b32  	%r268, %r1859, %r6722;
	xor.b32  	%r267, %r1858, %r6721;
	xor.b32  	%r270, %r1857, %r6724;
	xor.b32  	%r269, %r1856, %r6723;
	shl.b64 	%rd1757, %rd6123, 1;
	and.b64  	%rd1758, %rd1757, 2;
	mov.u64 	%rd1759, 1;
	sub.s64 	%rd1760, %rd1759, %rd1758;
	shl.b64 	%rd1761, %rd1760, 4;
	add.s64 	%rd1762, %rd6122, %rd1761;
	ld.local.v4.u32 	{%r1860, %r1861, %r1862, %r1863}, [%rd1762];
	xor.b32  	%r1868, %r6503, %r1862;
	xor.b32  	%r1869, %r6504, %r1863;
	xor.b32  	%r1870, %r6501, %r1860;
	xor.b32  	%r1871, %r6502, %r1861;
	xor.b32  	%r6731, %r1871, %r1808;
	xor.b32  	%r6730, %r1870, %r1807;
	xor.b32  	%r6733, %r1869, %r1810;
	xor.b32  	%r6732, %r1868, %r1809;
	mov.b64	%rd1763, {%r6732, %r6733};
	mov.b64	%rd1764, {%r6730, %r6731};
	cvt.u32.u64	%r1872, %rd1764;
	cvt.s32.s16 	%r1873, %r1872;
	mov.b64	%rd1765, {%r6503, %r6504};
	mov.b64	%rd1766, {%r6501, %r6502};
	cvt.u32.u64	%r1874, %rd1766;
	cvt.s32.s16 	%r1875, %r1874;
	mad.lo.s32 	%r1876, %r1873, %r1875, 16384;
	bfe.u32 	%r1877, %r1876, 15, 16;
	shr.s32 	%r1878, %r1872, 16;
	shr.s32 	%r1879, %r1874, 16;
	mad.lo.s32 	%r1880, %r1878, %r1879, 16384;
	mov.b64	%rd1767, {%r1877, %r1615};
	shl.b32 	%r1881, %r1880, 1;
	cvt.u64.u32	%rd1768, %r1881;
	and.b64  	%rd1769, %rd1767, 65535;
	shr.u64 	%rd1770, %rd1763, 16;
	shr.u64 	%rd1771, %rd1764, 16;
	cvt.u32.u64	%r1882, %rd1771;
	shr.s32 	%r1883, %r1882, 16;
	shr.u64 	%rd1772, %rd1765, 16;
	shr.u64 	%rd1773, %rd1766, 16;
	cvt.u32.u64	%r1884, %rd1773;
	shr.s32 	%r1885, %r1884, 16;
	mad.lo.s32 	%r1886, %r1883, %r1885, 16384;
	mul.wide.u32 	%rd1774, %r1886, 131072;
	shr.u64 	%rd1775, %rd1764, 32;
	cvt.u32.u64	%r1887, %rd1775;
	shr.s32 	%r1888, %r1887, 16;
	shr.u64 	%rd1776, %rd1766, 32;
	cvt.u32.u64	%r1889, %rd1776;
	shr.s32 	%r1890, %r1889, 16;
	mad.lo.s32 	%r1891, %r1888, %r1890, 16384;
	shr.u32 	%r1892, %r1891, 15;
	cvt.u64.u32	%rd1777, %r1892;
	shl.b64 	%rd1778, %rd1777, 48;
	and.b64  	%rd1779, %rd1774, 281470681743360;
	shl.b64 	%rd1780, %rd1763, 16;
	cvt.u32.u64	%r1893, %rd1780;
	shr.s32 	%r1894, %r1893, 16;
	shl.b64 	%rd1781, %rd1765, 16;
	cvt.u32.u64	%r1895, %rd1781;
	shr.s32 	%r1896, %r1895, 16;
	mad.lo.s32 	%r1897, %r1894, %r1896, 16384;
	shr.u32 	%r1898, %r1897, 15;
	cvt.u64.u32	%rd1782, %r1898;
	cvt.u32.u64	%r1899, %rd1763;
	shr.s32 	%r1900, %r1899, 16;
	cvt.u32.u64	%r1901, %rd1765;
	shr.s32 	%r1902, %r1901, 16;
	mad.lo.s32 	%r1903, %r1900, %r1902, 16384;
	and.b64  	%rd1783, %rd1782, 65535;
	and.b64  	%rd1784, %rd1768, 4294901760;
	or.b64  	%rd1785, %rd1784, %rd1769;
	or.b64  	%rd1786, %rd1785, %rd1779;
	shl.b32 	%r1904, %r1903, 1;
	cvt.u64.u32	%rd1787, %r1904;
	cvt.u32.u64	%r1905, %rd1770;
	shr.s32 	%r1906, %r1905, 16;
	cvt.u32.u64	%r1907, %rd1772;
	shr.s32 	%r1908, %r1907, 16;
	mad.lo.s32 	%r1909, %r1906, %r1908, 16384;
	mul.wide.u32 	%rd1788, %r1909, 131072;
	and.b64  	%rd1789, %rd1787, 4294901760;
	or.b64  	%rd1790, %rd1789, %rd1783;
	shr.u64 	%rd1791, %rd1763, 32;
	cvt.u32.u64	%r1910, %rd1791;
	shr.s32 	%r1911, %r1910, 16;
	shr.u64 	%rd1792, %rd1765, 32;
	cvt.u32.u64	%r1912, %rd1792;
	shr.s32 	%r1913, %r1912, 16;
	mad.lo.s32 	%r1914, %r1911, %r1913, 16384;
	shr.u32 	%r1915, %r1914, 15;
	cvt.u64.u32	%rd1793, %r1915;
	shl.b64 	%rd1794, %rd1793, 48;
	and.b64  	%rd1795, %rd1788, 281470681743360;
	or.b64  	%rd1796, %rd1795, %rd1794;
	or.b64  	%rd1797, %rd1796, %rd1790;
	or.b64  	%rd1798, %rd1786, %rd1778;
	mov.b64	{%r1916, %r1917}, %rd1797;
	mov.b64	{%r1918, %r1919}, %rd1798;
	xor.b32  	%r6724, %r1917, %r6504;
	xor.b32  	%r6723, %r1916, %r6503;
	xor.b32  	%r6722, %r1919, %r6502;
	xor.b32  	%r6721, %r1918, %r6501;
	cvt.u64.u32	%rd1799, %r6729;
	add.s64 	%rd1800, %rd1799, %rd9;
	shl.b64 	%rd1801, %rd1800, 1;
	mov.u64 	%rd1802, verus_gpu_hash$sharedrand;
	add.s64 	%rd1803, %rd1802, %rd1801;
	st.shared.u16 	[%rd1803], %rs236;
	mov.u64 	%rd1804, verus_gpu_hash$sharedrandex;
	add.s64 	%rd1805, %rd1804, %rd1801;
	st.shared.u16 	[%rd1805], %rs237;
	cvt.u64.u16	%rd1806, %rs236;
	add.s64 	%rd1807, %rd1806, %rd3;
	shl.b64 	%rd1808, %rd1807, 4;
	add.s64 	%rd1809, %rd203, %rd1808;
	st.global.v4.u32 	[%rd1809], {%r6721, %r6722, %r6723, %r6724};
	cvt.u64.u16	%rd1810, %rs237;
	add.s64 	%rd1811, %rd1810, %rd3;
	shl.b64 	%rd1812, %rd1811, 4;
	add.s64 	%rd1813, %rd203, %rd1812;
	st.global.v4.u32 	[%rd1813], {%r267, %r268, %r269, %r270};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p30, %r6729, 32;
	@%p30 bra 	BB0_134;

	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p31, %r6729, 0;
	mov.u16 	%rs238, 4;
	@%p31 bra 	BB0_52;
	bra.uni 	BB0_53;

BB0_52:
	mov.u32 	%r6501, %r267;
	mov.u32 	%r6502, %r268;
	mov.u32 	%r6503, %r269;
	mov.u32 	%r6504, %r270;
	mov.u32 	%r6729, %r1615;
	bra.uni 	BB0_54;

BB0_53:
	cvt.u64.u32	%rd1814, %r6731;
	shr.u32 	%r1921, %r6730, 5;
	cvt.u64.u32	%rd1815, %r1921;
	cvt.u16.u32	%rs76, %r1921;
	and.b16  	%rs236, %rs76, 511;
	cvt.u16.u32	%rs77, %r6731;
	and.b16  	%rs237, %rs77, 511;
	cvt.u16.u64	%rs78, %rd6123;
	and.b16  	%rs238, %rs78, 28;
	and.b64  	%rd1816, %rd1815, 511;
	add.s64 	%rd1817, %rd1816, %rd3;
	shl.b64 	%rd1818, %rd1817, 4;
	add.s64 	%rd1819, %rd203, %rd1818;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd1819];
	and.b64  	%rd1820, %rd1814, 511;
	add.s64 	%rd1821, %rd1820, %rd3;
	shl.b64 	%rd1822, %rd1821, 4;
	add.s64 	%rd1823, %rd203, %rd1822;
	ld.global.v4.u32 	{%r6501, %r6502, %r6503, %r6504}, [%rd1823];
	mul.wide.u32 	%rd1824, %r6730, 16;
	and.b64  	%rd1825, %rd1824, 48;
	add.s64 	%rd6122, %rd1, %rd1825;

BB0_54:
	and.b16  	%rs79, %rs238, 255;
	setp.ne.s16	%p32, %rs79, 8;
	@%p32 bra 	BB0_67;

	ld.local.v4.u32 	{%r1930, %r1931, %r1932, %r1933}, [%rd6122];
	xor.b32  	%r1938, %r6504, %r1933;
	xor.b32  	%r1939, %r6503, %r1932;
	xor.b32  	%r1940, %r6501, %r1930;
	xor.b32  	%r1941, %r6502, %r1931;
	xor.b32  	%r302, %r1941, %r6731;
	xor.b32  	%r301, %r1940, %r6730;
	xor.b32  	%r303, %r1939, %r6732;
	xor.b32  	%r304, %r1938, %r6733;
	mov.b64	%rd1826, {%r303, %r304};
	mov.b64	%rd1827, {%r301, %r302};
	cvt.u32.u64	%r1942, %rd1827;
	cvt.s32.s16 	%r1943, %r1942;
	mov.b64	%rd1828, {%r6503, %r6504};
	mov.b64	%rd1829, {%r6501, %r6502};
	cvt.u32.u64	%r1944, %rd1829;
	cvt.s32.s16 	%r1945, %r1944;
	mad.lo.s32 	%r1946, %r1943, %r1945, 16384;
	bfe.u32 	%r1947, %r1946, 15, 16;
	shr.s32 	%r1948, %r1942, 16;
	shr.s32 	%r1949, %r1944, 16;
	mad.lo.s32 	%r1950, %r1948, %r1949, 16384;
	mov.u32 	%r1951, 0;
	mov.b64	%rd1830, {%r1947, %r1951};
	shl.b32 	%r1952, %r1950, 1;
	cvt.u64.u32	%rd1831, %r1952;
	and.b64  	%rd1832, %rd1830, 65535;
	shr.u64 	%rd1833, %rd1826, 16;
	shr.u64 	%rd1834, %rd1827, 16;
	cvt.u32.u64	%r1953, %rd1834;
	shr.s32 	%r1954, %r1953, 16;
	shr.u64 	%rd1835, %rd1828, 16;
	shr.u64 	%rd1836, %rd1829, 16;
	cvt.u32.u64	%r1955, %rd1836;
	shr.s32 	%r1956, %r1955, 16;
	mad.lo.s32 	%r1957, %r1954, %r1956, 16384;
	mul.wide.u32 	%rd1837, %r1957, 131072;
	shr.u64 	%rd1838, %rd1827, 32;
	cvt.u32.u64	%r1958, %rd1838;
	shr.s32 	%r1959, %r1958, 16;
	shr.u64 	%rd1839, %rd1829, 32;
	cvt.u32.u64	%r1960, %rd1839;
	shr.s32 	%r1961, %r1960, 16;
	mad.lo.s32 	%r1962, %r1959, %r1961, 16384;
	shr.u32 	%r1963, %r1962, 15;
	cvt.u64.u32	%rd1840, %r1963;
	shl.b64 	%rd1841, %rd1840, 48;
	and.b64  	%rd1842, %rd1837, 281470681743360;
	shl.b64 	%rd1843, %rd1826, 16;
	mov.u64 	%rd1844, 0;
	cvt.u32.u64	%r1964, %rd1843;
	shr.s32 	%r1965, %r1964, 16;
	shl.b64 	%rd1845, %rd1828, 16;
	cvt.u32.u64	%r1966, %rd1845;
	shr.s32 	%r1967, %r1966, 16;
	mad.lo.s32 	%r1968, %r1965, %r1967, 16384;
	shr.u32 	%r1969, %r1968, 15;
	cvt.u64.u32	%rd1846, %r1969;
	cvt.u32.u64	%r1970, %rd1826;
	shr.s32 	%r1971, %r1970, 16;
	cvt.u32.u64	%r1972, %rd1828;
	shr.s32 	%r1973, %r1972, 16;
	mad.lo.s32 	%r1974, %r1971, %r1973, 16384;
	and.b64  	%rd1847, %rd1846, 65535;
	and.b64  	%rd1848, %rd1831, 4294901760;
	or.b64  	%rd1849, %rd1848, %rd1832;
	or.b64  	%rd1850, %rd1849, %rd1842;
	shl.b32 	%r1975, %r1974, 1;
	cvt.u64.u32	%rd1851, %r1975;
	cvt.u32.u64	%r1976, %rd1833;
	shr.s32 	%r1977, %r1976, 16;
	cvt.u32.u64	%r1978, %rd1835;
	shr.s32 	%r1979, %r1978, 16;
	mad.lo.s32 	%r1980, %r1977, %r1979, 16384;
	mul.wide.u32 	%rd1852, %r1980, 131072;
	and.b64  	%rd1853, %rd1851, 4294901760;
	or.b64  	%rd1854, %rd1853, %rd1847;
	shr.u64 	%rd1855, %rd1826, 32;
	cvt.u32.u64	%r1981, %rd1855;
	shr.s32 	%r1982, %r1981, 16;
	shr.u64 	%rd1856, %rd1828, 32;
	cvt.u32.u64	%r1983, %rd1856;
	shr.s32 	%r1984, %r1983, 16;
	mad.lo.s32 	%r1985, %r1982, %r1984, 16384;
	shr.u32 	%r1986, %r1985, 15;
	cvt.u64.u32	%rd1857, %r1986;
	shl.b64 	%rd1858, %rd1857, 48;
	and.b64  	%rd1859, %rd1852, 281470681743360;
	or.b64  	%rd1860, %rd1859, %rd1858;
	or.b64  	%rd1861, %rd1860, %rd1854;
	or.b64  	%rd1862, %rd1850, %rd1841;
	mov.b64	{%r1987, %r1988}, %rd1861;
	mov.b64	{%r1989, %r1990}, %rd1862;
	xor.b32  	%r308, %r1988, %r6504;
	xor.b32  	%r307, %r1987, %r6503;
	xor.b32  	%r306, %r1990, %r6502;
	xor.b32  	%r305, %r1989, %r6501;
	shl.b64 	%rd1863, %rd6123, 1;
	and.b64  	%rd1864, %rd1863, 2;
	mov.u64 	%rd1865, 1;
	sub.s64 	%rd1866, %rd1865, %rd1864;
	shl.b64 	%rd1867, %rd1866, 4;
	add.s64 	%rd1868, %rd6122, %rd1867;
	ld.local.v4.u32 	{%r1991, %r1992, %r1993, %r1994}, [%rd1868];
	xor.b32  	%r316, %r6724, %r1994;
	xor.b32  	%r315, %r6723, %r1993;
	xor.b32  	%r314, %r6722, %r1992;
	xor.b32  	%r313, %r6721, %r1991;
	mov.b64	%rd1869, {%r315, %r316};
	st.local.v2.u64 	[%rd12], {%rd1844, %rd1869};
	shl.b64 	%rd1872, %rd1869, 1;
	xor.b64  	%rd1873, %rd1872, %rd1869;
	st.local.v2.u64 	[%rd15], {%rd1872, %rd1873};
	shl.b64 	%rd1874, %rd1869, 2;
	xor.b64  	%rd1875, %rd1874, %rd1869;
	st.local.v2.u64 	[%rd15+16], {%rd1874, %rd1875};
	shl.b64 	%rd1876, %rd1873, 1;
	xor.b64  	%rd1877, %rd1876, %rd1869;
	st.local.v2.u64 	[%rd15+32], {%rd1876, %rd1877};
	mov.b64	%rd1878, {%r313, %r314};
	shl.b64 	%rd1879, %rd1878, 3;
	and.b64  	%rd1880, %rd1879, 56;
	add.s64 	%rd1881, %rd12, %rd1880;
	and.b64  	%rd1882, %rd1878, 56;
	add.s64 	%rd1883, %rd12, %rd1882;
	ld.local.u64 	%rd1884, [%rd1883];
	shl.b64 	%rd1885, %rd1884, 3;
	ld.local.u64 	%rd1886, [%rd1881];
	xor.b64  	%rd1887, %rd1886, %rd1885;
	shr.u64 	%rd1888, %rd1884, 61;
	mov.b64	%rd1889, {%r1951, %r1951};
	xor.b64  	%rd1890, %rd1889, %rd1888;
	shr.u64 	%rd1891, %rd1878, 3;
	and.b64  	%rd1892, %rd1891, 56;
	add.s64 	%rd1893, %rd12, %rd1892;
	ld.local.u64 	%rd1894, [%rd1893];
	shl.b64 	%rd1895, %rd1894, 6;
	mov.b64	{%r1995, %r1996}, %rd1887;
	mov.b64	%rd1896, {%r1995, %r1996};
	xor.b64  	%rd1897, %rd1896, %rd1895;
	shr.u64 	%rd1898, %rd1894, 58;
	mov.b64	{%r1997, %r1998}, %rd1890;
	mov.b64	%rd1899, {%r1997, %r1998};
	xor.b64  	%rd1900, %rd1899, %rd1898;
	shr.u64 	%rd1901, %rd1878, 6;
	and.b64  	%rd1902, %rd1901, 56;
	add.s64 	%rd1903, %rd12, %rd1902;
	ld.local.u64 	%rd1904, [%rd1903];
	shl.b64 	%rd1905, %rd1904, 9;
	mov.b64	{%r1999, %r2000}, %rd1897;
	mov.b64	%rd1906, {%r1999, %r2000};
	xor.b64  	%rd1907, %rd1906, %rd1905;
	shr.u64 	%rd1908, %rd1904, 55;
	mov.b64	{%r2001, %r2002}, %rd1900;
	mov.b64	%rd1909, {%r2001, %r2002};
	xor.b64  	%rd1910, %rd1909, %rd1908;
	shr.u64 	%rd1911, %rd1878, 9;
	and.b64  	%rd1912, %rd1911, 56;
	add.s64 	%rd1913, %rd12, %rd1912;
	ld.local.u64 	%rd1914, [%rd1913];
	shl.b64 	%rd1915, %rd1914, 12;
	mov.b64	{%r2003, %r2004}, %rd1907;
	mov.b64	%rd1916, {%r2003, %r2004};
	xor.b64  	%rd1917, %rd1916, %rd1915;
	shr.u64 	%rd1918, %rd1914, 52;
	mov.b64	{%r2005, %r2006}, %rd1910;
	mov.b64	%rd1919, {%r2005, %r2006};
	xor.b64  	%rd1920, %rd1919, %rd1918;
	shr.u64 	%rd1921, %rd1878, 12;
	and.b64  	%rd1922, %rd1921, 56;
	add.s64 	%rd1923, %rd12, %rd1922;
	ld.local.u64 	%rd1924, [%rd1923];
	shl.b64 	%rd1925, %rd1924, 15;
	mov.b64	{%r2007, %r2008}, %rd1917;
	mov.b64	%rd1926, {%r2007, %r2008};
	xor.b64  	%rd1927, %rd1926, %rd1925;
	shr.u64 	%rd1928, %rd1924, 49;
	mov.b64	{%r2009, %r2010}, %rd1920;
	mov.b64	%rd1929, {%r2009, %r2010};
	xor.b64  	%rd1930, %rd1929, %rd1928;
	shr.u64 	%rd1931, %rd1878, 15;
	and.b64  	%rd1932, %rd1931, 56;
	add.s64 	%rd1933, %rd12, %rd1932;
	ld.local.u64 	%rd1934, [%rd1933];
	shl.b64 	%rd1935, %rd1934, 18;
	mov.b64	{%r2011, %r2012}, %rd1927;
	mov.b64	%rd1936, {%r2011, %r2012};
	xor.b64  	%rd1937, %rd1936, %rd1935;
	shr.u64 	%rd1938, %rd1934, 46;
	mov.b64	{%r2013, %r2014}, %rd1930;
	mov.b64	%rd1939, {%r2013, %r2014};
	xor.b64  	%rd1940, %rd1939, %rd1938;
	shr.u64 	%rd1941, %rd1878, 18;
	and.b64  	%rd1942, %rd1941, 56;
	add.s64 	%rd1943, %rd12, %rd1942;
	ld.local.u64 	%rd1944, [%rd1943];
	shl.b64 	%rd1945, %rd1944, 21;
	mov.b64	{%r2015, %r2016}, %rd1937;
	mov.b64	%rd1946, {%r2015, %r2016};
	xor.b64  	%rd1947, %rd1946, %rd1945;
	shr.u64 	%rd1948, %rd1944, 43;
	mov.b64	{%r2017, %r2018}, %rd1940;
	mov.b64	%rd1949, {%r2017, %r2018};
	xor.b64  	%rd1950, %rd1949, %rd1948;
	shr.u64 	%rd1951, %rd1878, 21;
	and.b64  	%rd1952, %rd1951, 56;
	add.s64 	%rd1953, %rd12, %rd1952;
	ld.local.u64 	%rd1954, [%rd1953];
	shl.b64 	%rd1955, %rd1954, 24;
	mov.b64	{%r2019, %r2020}, %rd1947;
	mov.b64	%rd1956, {%r2019, %r2020};
	xor.b64  	%rd1957, %rd1956, %rd1955;
	shr.u64 	%rd1958, %rd1954, 40;
	mov.b64	{%r2021, %r2022}, %rd1950;
	mov.b64	%rd1959, {%r2021, %r2022};
	xor.b64  	%rd1960, %rd1959, %rd1958;
	shr.u64 	%rd1961, %rd1878, 24;
	and.b64  	%rd1962, %rd1961, 56;
	add.s64 	%rd1963, %rd12, %rd1962;
	ld.local.u64 	%rd1964, [%rd1963];
	shl.b64 	%rd1965, %rd1964, 27;
	mov.b64	{%r2023, %r2024}, %rd1957;
	mov.b64	%rd1966, {%r2023, %r2024};
	xor.b64  	%rd1967, %rd1966, %rd1965;
	shr.u64 	%rd1968, %rd1964, 37;
	mov.b64	{%r2025, %r2026}, %rd1960;
	mov.b64	%rd1969, {%r2025, %r2026};
	xor.b64  	%rd1970, %rd1969, %rd1968;
	shr.u64 	%rd1971, %rd1878, 27;
	and.b64  	%rd1972, %rd1971, 56;
	add.s64 	%rd1973, %rd12, %rd1972;
	ld.local.u64 	%rd1974, [%rd1973];
	shl.b64 	%rd1975, %rd1974, 30;
	mov.b64	{%r2027, %r2028}, %rd1967;
	mov.b64	%rd1976, {%r2027, %r2028};
	xor.b64  	%rd1977, %rd1976, %rd1975;
	shr.u64 	%rd1978, %rd1974, 34;
	mov.b64	{%r2029, %r2030}, %rd1970;
	mov.b64	%rd1979, {%r2029, %r2030};
	xor.b64  	%rd1980, %rd1979, %rd1978;
	shr.u64 	%rd1981, %rd1878, 30;
	and.b64  	%rd1982, %rd1981, 56;
	add.s64 	%rd1983, %rd12, %rd1982;
	ld.local.u64 	%rd1984, [%rd1983];
	shl.b64 	%rd1985, %rd1984, 33;
	mov.b64	{%r2031, %r2032}, %rd1977;
	mov.b64	%rd1986, {%r2031, %r2032};
	xor.b64  	%rd1987, %rd1986, %rd1985;
	shr.u64 	%rd1988, %rd1984, 31;
	mov.b64	{%r2033, %r2034}, %rd1980;
	mov.b64	%rd1989, {%r2033, %r2034};
	xor.b64  	%rd1990, %rd1989, %rd1988;
	shr.u64 	%rd1991, %rd1878, 33;
	and.b64  	%rd1992, %rd1991, 56;
	add.s64 	%rd1993, %rd12, %rd1992;
	ld.local.u64 	%rd1994, [%rd1993];
	shl.b64 	%rd1995, %rd1994, 36;
	mov.b64	{%r2035, %r2036}, %rd1987;
	mov.b64	%rd1996, {%r2035, %r2036};
	xor.b64  	%rd1997, %rd1996, %rd1995;
	shr.u64 	%rd1998, %rd1994, 28;
	mov.b64	{%r2037, %r2038}, %rd1990;
	mov.b64	%rd1999, {%r2037, %r2038};
	xor.b64  	%rd2000, %rd1999, %rd1998;
	shr.u64 	%rd2001, %rd1878, 36;
	and.b64  	%rd2002, %rd2001, 56;
	add.s64 	%rd2003, %rd12, %rd2002;
	ld.local.u64 	%rd2004, [%rd2003];
	shl.b64 	%rd2005, %rd2004, 39;
	mov.b64	{%r2039, %r2040}, %rd1997;
	mov.b64	%rd2006, {%r2039, %r2040};
	xor.b64  	%rd2007, %rd2006, %rd2005;
	shr.u64 	%rd2008, %rd2004, 25;
	mov.b64	{%r2041, %r2042}, %rd2000;
	mov.b64	%rd2009, {%r2041, %r2042};
	xor.b64  	%rd2010, %rd2009, %rd2008;
	shr.u64 	%rd2011, %rd1878, 39;
	and.b64  	%rd2012, %rd2011, 56;
	add.s64 	%rd2013, %rd12, %rd2012;
	ld.local.u64 	%rd2014, [%rd2013];
	shl.b64 	%rd2015, %rd2014, 42;
	mov.b64	{%r2043, %r2044}, %rd2007;
	mov.b64	%rd2016, {%r2043, %r2044};
	xor.b64  	%rd2017, %rd2016, %rd2015;
	shr.u64 	%rd2018, %rd2014, 22;
	mov.b64	{%r2045, %r2046}, %rd2010;
	mov.b64	%rd2019, {%r2045, %r2046};
	xor.b64  	%rd2020, %rd2019, %rd2018;
	shr.u64 	%rd2021, %rd1878, 42;
	and.b64  	%rd2022, %rd2021, 56;
	add.s64 	%rd2023, %rd12, %rd2022;
	ld.local.u64 	%rd2024, [%rd2023];
	shl.b64 	%rd2025, %rd2024, 45;
	mov.b64	{%r2047, %r2048}, %rd2017;
	mov.b64	%rd2026, {%r2047, %r2048};
	xor.b64  	%rd2027, %rd2026, %rd2025;
	shr.u64 	%rd2028, %rd2024, 19;
	mov.b64	{%r2049, %r2050}, %rd2020;
	mov.b64	%rd2029, {%r2049, %r2050};
	xor.b64  	%rd2030, %rd2029, %rd2028;
	shr.u64 	%rd2031, %rd1878, 45;
	and.b64  	%rd2032, %rd2031, 56;
	add.s64 	%rd2033, %rd12, %rd2032;
	ld.local.u64 	%rd2034, [%rd2033];
	shl.b64 	%rd2035, %rd2034, 48;
	mov.b64	{%r2051, %r2052}, %rd2027;
	mov.b64	%rd2036, {%r2051, %r2052};
	xor.b64  	%rd2037, %rd2036, %rd2035;
	shr.u64 	%rd2038, %rd2034, 16;
	mov.b64	{%r2053, %r2054}, %rd2030;
	mov.b64	%rd2039, {%r2053, %r2054};
	xor.b64  	%rd2040, %rd2039, %rd2038;
	shr.u64 	%rd2041, %rd1878, 48;
	and.b64  	%rd2042, %rd2041, 56;
	add.s64 	%rd2043, %rd12, %rd2042;
	ld.local.u64 	%rd2044, [%rd2043];
	shl.b64 	%rd2045, %rd2044, 51;
	mov.b64	{%r2055, %r2056}, %rd2037;
	mov.b64	%rd2046, {%r2055, %r2056};
	xor.b64  	%rd2047, %rd2046, %rd2045;
	shr.u64 	%rd2048, %rd2044, 13;
	mov.b64	{%r2057, %r2058}, %rd2040;
	mov.b64	%rd2049, {%r2057, %r2058};
	xor.b64  	%rd2050, %rd2049, %rd2048;
	shr.u64 	%rd2051, %rd1878, 51;
	and.b64  	%rd2052, %rd2051, 56;
	add.s64 	%rd2053, %rd12, %rd2052;
	ld.local.u64 	%rd2054, [%rd2053];
	shl.b64 	%rd2055, %rd2054, 54;
	mov.b64	{%r2059, %r2060}, %rd2047;
	mov.b64	%rd2056, {%r2059, %r2060};
	xor.b64  	%rd2057, %rd2056, %rd2055;
	shr.u64 	%rd2058, %rd2054, 10;
	mov.b64	{%r2061, %r2062}, %rd2050;
	mov.b64	%rd2059, {%r2061, %r2062};
	xor.b64  	%rd2060, %rd2059, %rd2058;
	shr.u64 	%rd2061, %rd1878, 54;
	and.b64  	%rd2062, %rd2061, 56;
	add.s64 	%rd2063, %rd12, %rd2062;
	ld.local.u64 	%rd2064, [%rd2063];
	shl.b64 	%rd2065, %rd2064, 57;
	mov.b64	{%r2063, %r2064}, %rd2057;
	mov.b64	%rd2066, {%r2063, %r2064};
	xor.b64  	%rd2067, %rd2066, %rd2065;
	shr.u64 	%rd2068, %rd2064, 7;
	mov.b64	{%r2065, %r2066}, %rd2060;
	mov.b64	%rd2069, {%r2065, %r2066};
	xor.b64  	%rd2070, %rd2069, %rd2068;
	shr.u64 	%rd2071, %rd1878, 57;
	and.b64  	%rd2072, %rd2071, 56;
	add.s64 	%rd2073, %rd12, %rd2072;
	ld.local.u64 	%rd2074, [%rd2073];
	shl.b64 	%rd2075, %rd2074, 60;
	mov.b64	{%r2067, %r2068}, %rd2067;
	mov.b64	%rd2076, {%r2067, %r2068};
	xor.b64  	%rd2077, %rd2076, %rd2075;
	shr.u64 	%rd2078, %rd2074, 4;
	mov.b64	{%r2069, %r2070}, %rd2070;
	mov.b64	%rd2079, {%r2069, %r2070};
	xor.b64  	%rd2080, %rd2079, %rd2078;
	shr.u64 	%rd2081, %rd1878, 60;
	and.b64  	%rd2082, %rd2081, 8;
	add.s64 	%rd2083, %rd12, %rd2082;
	ld.local.u64 	%rd2084, [%rd2083];
	shl.b64 	%rd2085, %rd2084, 63;
	mov.b64	{%r2071, %r2072}, %rd2077;
	mov.b64	%rd2086, {%r2071, %r2072};
	xor.b64  	%rd2087, %rd2086, %rd2085;
	shr.u64 	%rd2088, %rd2084, 1;
	mov.b64	{%r2073, %r2074}, %rd2080;
	mov.b64	%rd2089, {%r2073, %r2074};
	xor.b64  	%rd2090, %rd2089, %rd2088;
	mov.b64	{%r6574, %r6575}, %rd2090;
	mov.b64	{%r6576, %r6577}, %rd2087;
	setp.gt.s32	%p33, %r316, -1;
	@%p33 bra 	BB0_57;

	xor.b32  	%r6447, %r6722, %r1992;
	xor.b32  	%r6446, %r6721, %r1991;
	shr.u32 	%r2075, %r6446, 1;
	shl.b32 	%r2076, %r6447, 31;
	shr.u32 	%r2077, %r6447, 1;
	and.b32  	%r2078, %r2075, 1533916891;
	or.b32  	%r2079, %r2078, %r2076;
	xor.b32  	%r6574, %r6574, %r2079;
	and.b32  	%r2080, %r2077, 920350134;
	xor.b32  	%r6575, %r6575, %r2080;

BB0_57:
	and.b32  	%r2081, %r316, 1073741824;
	setp.eq.s32	%p34, %r2081, 0;
	@%p34 bra 	BB0_59;

	xor.b32  	%r6445, %r6722, %r1992;
	xor.b32  	%r6444, %r6721, %r1991;
	shr.u32 	%r2082, %r6444, 2;
	shl.b32 	%r2083, %r6445, 30;
	or.b32  	%r2084, %r2082, %r2083;
	shr.u32 	%r2085, %r6445, 2;
	and.b32  	%r2086, %r2084, 1227133513;
	xor.b32  	%r6574, %r6574, %r2086;
	and.b32  	%r2087, %r2085, 306783378;
	xor.b32  	%r6575, %r6575, %r2087;

BB0_59:
	mov.u32 	%r6456, 0;
	mov.u64 	%rd6075, 0;
	xor.b32  	%r338, %r6576, %r301;
	xor.b32  	%r339, %r6577, %r302;
	xor.b32  	%r340, %r6574, %r303;
	xor.b32  	%r341, %r6575, %r304;
	mov.b64	%rd2091, {%r1993, %r1994};
	st.local.v2.u64 	[%rd12], {%rd6075, %rd2091};
	shl.b64 	%rd2095, %rd2091, 1;
	xor.b64  	%rd2096, %rd2095, %rd2091;
	st.local.v2.u64 	[%rd15], {%rd2095, %rd2096};
	shl.b64 	%rd2097, %rd2091, 2;
	xor.b64  	%rd2098, %rd2097, %rd2091;
	st.local.v2.u64 	[%rd15+16], {%rd2097, %rd2098};
	shl.b64 	%rd2099, %rd2096, 1;
	xor.b64  	%rd2100, %rd2099, %rd2091;
	st.local.v2.u64 	[%rd15+32], {%rd2099, %rd2100};
	mov.b64	%rd2101, {%r1991, %r1992};
	shl.b64 	%rd2102, %rd2101, 3;
	and.b64  	%rd2103, %rd2102, 56;
	add.s64 	%rd2104, %rd12, %rd2103;
	and.b64  	%rd2105, %rd2101, 56;
	add.s64 	%rd2106, %rd12, %rd2105;
	ld.local.u64 	%rd2107, [%rd2106];
	shl.b64 	%rd2108, %rd2107, 3;
	ld.local.u64 	%rd2109, [%rd2104];
	xor.b64  	%rd2110, %rd2109, %rd2108;
	shr.u64 	%rd2111, %rd2107, 61;
	mov.b64	%rd2112, {%r6456, %r6456};
	xor.b64  	%rd2113, %rd2112, %rd2111;
	shr.u64 	%rd2114, %rd2101, 3;
	and.b64  	%rd2115, %rd2114, 56;
	add.s64 	%rd2116, %rd12, %rd2115;
	ld.local.u64 	%rd2117, [%rd2116];
	shl.b64 	%rd2118, %rd2117, 6;
	mov.b64	{%r2089, %r2090}, %rd2110;
	mov.b64	%rd2119, {%r2089, %r2090};
	xor.b64  	%rd2120, %rd2119, %rd2118;
	shr.u64 	%rd2121, %rd2117, 58;
	mov.b64	{%r2091, %r2092}, %rd2113;
	mov.b64	%rd2122, {%r2091, %r2092};
	xor.b64  	%rd2123, %rd2122, %rd2121;
	shr.u64 	%rd2124, %rd2101, 6;
	and.b64  	%rd2125, %rd2124, 56;
	add.s64 	%rd2126, %rd12, %rd2125;
	ld.local.u64 	%rd2127, [%rd2126];
	shl.b64 	%rd2128, %rd2127, 9;
	mov.b64	{%r2093, %r2094}, %rd2120;
	mov.b64	%rd2129, {%r2093, %r2094};
	xor.b64  	%rd2130, %rd2129, %rd2128;
	shr.u64 	%rd2131, %rd2127, 55;
	mov.b64	{%r2095, %r2096}, %rd2123;
	mov.b64	%rd2132, {%r2095, %r2096};
	xor.b64  	%rd2133, %rd2132, %rd2131;
	shr.u64 	%rd2134, %rd2101, 9;
	and.b64  	%rd2135, %rd2134, 56;
	add.s64 	%rd2136, %rd12, %rd2135;
	ld.local.u64 	%rd2137, [%rd2136];
	shl.b64 	%rd2138, %rd2137, 12;
	mov.b64	{%r2097, %r2098}, %rd2130;
	mov.b64	%rd2139, {%r2097, %r2098};
	xor.b64  	%rd2140, %rd2139, %rd2138;
	shr.u64 	%rd2141, %rd2137, 52;
	mov.b64	{%r2099, %r2100}, %rd2133;
	mov.b64	%rd2142, {%r2099, %r2100};
	xor.b64  	%rd2143, %rd2142, %rd2141;
	shr.u64 	%rd2144, %rd2101, 12;
	and.b64  	%rd2145, %rd2144, 56;
	add.s64 	%rd2146, %rd12, %rd2145;
	ld.local.u64 	%rd2147, [%rd2146];
	shl.b64 	%rd2148, %rd2147, 15;
	mov.b64	{%r2101, %r2102}, %rd2140;
	mov.b64	%rd2149, {%r2101, %r2102};
	xor.b64  	%rd2150, %rd2149, %rd2148;
	shr.u64 	%rd2151, %rd2147, 49;
	mov.b64	{%r2103, %r2104}, %rd2143;
	mov.b64	%rd2152, {%r2103, %r2104};
	xor.b64  	%rd2153, %rd2152, %rd2151;
	shr.u64 	%rd2154, %rd2101, 15;
	and.b64  	%rd2155, %rd2154, 56;
	add.s64 	%rd2156, %rd12, %rd2155;
	ld.local.u64 	%rd2157, [%rd2156];
	shl.b64 	%rd2158, %rd2157, 18;
	mov.b64	{%r2105, %r2106}, %rd2150;
	mov.b64	%rd2159, {%r2105, %r2106};
	xor.b64  	%rd2160, %rd2159, %rd2158;
	shr.u64 	%rd2161, %rd2157, 46;
	mov.b64	{%r2107, %r2108}, %rd2153;
	mov.b64	%rd2162, {%r2107, %r2108};
	xor.b64  	%rd2163, %rd2162, %rd2161;
	shr.u64 	%rd2164, %rd2101, 18;
	and.b64  	%rd2165, %rd2164, 56;
	add.s64 	%rd2166, %rd12, %rd2165;
	ld.local.u64 	%rd2167, [%rd2166];
	shl.b64 	%rd2168, %rd2167, 21;
	mov.b64	{%r2109, %r2110}, %rd2160;
	mov.b64	%rd2169, {%r2109, %r2110};
	xor.b64  	%rd2170, %rd2169, %rd2168;
	shr.u64 	%rd2171, %rd2167, 43;
	mov.b64	{%r2111, %r2112}, %rd2163;
	mov.b64	%rd2172, {%r2111, %r2112};
	xor.b64  	%rd2173, %rd2172, %rd2171;
	shr.u64 	%rd2174, %rd2101, 21;
	and.b64  	%rd2175, %rd2174, 56;
	add.s64 	%rd2176, %rd12, %rd2175;
	ld.local.u64 	%rd2177, [%rd2176];
	shl.b64 	%rd2178, %rd2177, 24;
	mov.b64	{%r2113, %r2114}, %rd2170;
	mov.b64	%rd2179, {%r2113, %r2114};
	xor.b64  	%rd2180, %rd2179, %rd2178;
	shr.u64 	%rd2181, %rd2177, 40;
	mov.b64	{%r2115, %r2116}, %rd2173;
	mov.b64	%rd2182, {%r2115, %r2116};
	xor.b64  	%rd2183, %rd2182, %rd2181;
	shr.u64 	%rd2184, %rd2101, 24;
	and.b64  	%rd2185, %rd2184, 56;
	add.s64 	%rd2186, %rd12, %rd2185;
	ld.local.u64 	%rd2187, [%rd2186];
	shl.b64 	%rd2188, %rd2187, 27;
	mov.b64	{%r2117, %r2118}, %rd2180;
	mov.b64	%rd2189, {%r2117, %r2118};
	xor.b64  	%rd2190, %rd2189, %rd2188;
	shr.u64 	%rd2191, %rd2187, 37;
	mov.b64	{%r2119, %r2120}, %rd2183;
	mov.b64	%rd2192, {%r2119, %r2120};
	xor.b64  	%rd2193, %rd2192, %rd2191;
	shr.u64 	%rd2194, %rd2101, 27;
	and.b64  	%rd2195, %rd2194, 56;
	add.s64 	%rd2196, %rd12, %rd2195;
	ld.local.u64 	%rd2197, [%rd2196];
	shl.b64 	%rd2198, %rd2197, 30;
	mov.b64	{%r2121, %r2122}, %rd2190;
	mov.b64	%rd2199, {%r2121, %r2122};
	xor.b64  	%rd2200, %rd2199, %rd2198;
	shr.u64 	%rd2201, %rd2197, 34;
	mov.b64	{%r2123, %r2124}, %rd2193;
	mov.b64	%rd2202, {%r2123, %r2124};
	xor.b64  	%rd2203, %rd2202, %rd2201;
	shr.u64 	%rd2204, %rd2101, 30;
	and.b64  	%rd2205, %rd2204, 56;
	add.s64 	%rd2206, %rd12, %rd2205;
	ld.local.u64 	%rd2207, [%rd2206];
	shl.b64 	%rd2208, %rd2207, 33;
	mov.b64	{%r2125, %r2126}, %rd2200;
	mov.b64	%rd2209, {%r2125, %r2126};
	xor.b64  	%rd2210, %rd2209, %rd2208;
	shr.u64 	%rd2211, %rd2207, 31;
	mov.b64	{%r2127, %r2128}, %rd2203;
	mov.b64	%rd2212, {%r2127, %r2128};
	xor.b64  	%rd2213, %rd2212, %rd2211;
	shr.u64 	%rd2214, %rd2101, 33;
	and.b64  	%rd2215, %rd2214, 56;
	add.s64 	%rd2216, %rd12, %rd2215;
	ld.local.u64 	%rd2217, [%rd2216];
	shl.b64 	%rd2218, %rd2217, 36;
	mov.b64	{%r2129, %r2130}, %rd2210;
	mov.b64	%rd2219, {%r2129, %r2130};
	xor.b64  	%rd2220, %rd2219, %rd2218;
	shr.u64 	%rd2221, %rd2217, 28;
	mov.b64	{%r2131, %r2132}, %rd2213;
	mov.b64	%rd2222, {%r2131, %r2132};
	xor.b64  	%rd2223, %rd2222, %rd2221;
	shr.u64 	%rd2224, %rd2101, 36;
	and.b64  	%rd2225, %rd2224, 56;
	add.s64 	%rd2226, %rd12, %rd2225;
	ld.local.u64 	%rd2227, [%rd2226];
	shl.b64 	%rd2228, %rd2227, 39;
	mov.b64	{%r2133, %r2134}, %rd2220;
	mov.b64	%rd2229, {%r2133, %r2134};
	xor.b64  	%rd2230, %rd2229, %rd2228;
	shr.u64 	%rd2231, %rd2227, 25;
	mov.b64	{%r2135, %r2136}, %rd2223;
	mov.b64	%rd2232, {%r2135, %r2136};
	xor.b64  	%rd2233, %rd2232, %rd2231;
	shr.u64 	%rd2234, %rd2101, 39;
	and.b64  	%rd2235, %rd2234, 56;
	add.s64 	%rd2236, %rd12, %rd2235;
	ld.local.u64 	%rd2237, [%rd2236];
	shl.b64 	%rd2238, %rd2237, 42;
	mov.b64	{%r2137, %r2138}, %rd2230;
	mov.b64	%rd2239, {%r2137, %r2138};
	xor.b64  	%rd2240, %rd2239, %rd2238;
	shr.u64 	%rd2241, %rd2237, 22;
	mov.b64	{%r2139, %r2140}, %rd2233;
	mov.b64	%rd2242, {%r2139, %r2140};
	xor.b64  	%rd2243, %rd2242, %rd2241;
	shr.u64 	%rd2244, %rd2101, 42;
	and.b64  	%rd2245, %rd2244, 56;
	add.s64 	%rd2246, %rd12, %rd2245;
	ld.local.u64 	%rd2247, [%rd2246];
	shl.b64 	%rd2248, %rd2247, 45;
	mov.b64	{%r2141, %r2142}, %rd2240;
	mov.b64	%rd2249, {%r2141, %r2142};
	xor.b64  	%rd2250, %rd2249, %rd2248;
	shr.u64 	%rd2251, %rd2247, 19;
	mov.b64	{%r2143, %r2144}, %rd2243;
	mov.b64	%rd2252, {%r2143, %r2144};
	xor.b64  	%rd2253, %rd2252, %rd2251;
	shr.u64 	%rd2254, %rd2101, 45;
	and.b64  	%rd2255, %rd2254, 56;
	add.s64 	%rd2256, %rd12, %rd2255;
	ld.local.u64 	%rd2257, [%rd2256];
	shl.b64 	%rd2258, %rd2257, 48;
	mov.b64	{%r2145, %r2146}, %rd2250;
	mov.b64	%rd2259, {%r2145, %r2146};
	xor.b64  	%rd2260, %rd2259, %rd2258;
	shr.u64 	%rd2261, %rd2257, 16;
	mov.b64	{%r2147, %r2148}, %rd2253;
	mov.b64	%rd2262, {%r2147, %r2148};
	xor.b64  	%rd2263, %rd2262, %rd2261;
	shr.u64 	%rd2264, %rd2101, 48;
	and.b64  	%rd2265, %rd2264, 56;
	add.s64 	%rd2266, %rd12, %rd2265;
	ld.local.u64 	%rd2267, [%rd2266];
	shl.b64 	%rd2268, %rd2267, 51;
	mov.b64	{%r2149, %r2150}, %rd2260;
	mov.b64	%rd2269, {%r2149, %r2150};
	xor.b64  	%rd2270, %rd2269, %rd2268;
	shr.u64 	%rd2271, %rd2267, 13;
	mov.b64	{%r2151, %r2152}, %rd2263;
	mov.b64	%rd2272, {%r2151, %r2152};
	xor.b64  	%rd2273, %rd2272, %rd2271;
	shr.u64 	%rd2274, %rd2101, 51;
	and.b64  	%rd2275, %rd2274, 56;
	add.s64 	%rd2276, %rd12, %rd2275;
	ld.local.u64 	%rd2277, [%rd2276];
	shl.b64 	%rd2278, %rd2277, 54;
	mov.b64	{%r2153, %r2154}, %rd2270;
	mov.b64	%rd2279, {%r2153, %r2154};
	xor.b64  	%rd2280, %rd2279, %rd2278;
	shr.u64 	%rd2281, %rd2277, 10;
	mov.b64	{%r2155, %r2156}, %rd2273;
	mov.b64	%rd2282, {%r2155, %r2156};
	xor.b64  	%rd2283, %rd2282, %rd2281;
	shr.u64 	%rd2284, %rd2101, 54;
	and.b64  	%rd2285, %rd2284, 56;
	add.s64 	%rd2286, %rd12, %rd2285;
	ld.local.u64 	%rd2287, [%rd2286];
	shl.b64 	%rd2288, %rd2287, 57;
	mov.b64	{%r2157, %r2158}, %rd2280;
	mov.b64	%rd2289, {%r2157, %r2158};
	xor.b64  	%rd2290, %rd2289, %rd2288;
	shr.u64 	%rd2291, %rd2287, 7;
	mov.b64	{%r2159, %r2160}, %rd2283;
	mov.b64	%rd2292, {%r2159, %r2160};
	xor.b64  	%rd2293, %rd2292, %rd2291;
	shr.u64 	%rd2294, %rd2101, 57;
	and.b64  	%rd2295, %rd2294, 56;
	add.s64 	%rd2296, %rd12, %rd2295;
	ld.local.u64 	%rd2297, [%rd2296];
	shl.b64 	%rd2298, %rd2297, 60;
	mov.b64	{%r2161, %r2162}, %rd2290;
	mov.b64	%rd2299, {%r2161, %r2162};
	xor.b64  	%rd2300, %rd2299, %rd2298;
	shr.u64 	%rd2301, %rd2297, 4;
	mov.b64	{%r2163, %r2164}, %rd2293;
	mov.b64	%rd2302, {%r2163, %r2164};
	xor.b64  	%rd2303, %rd2302, %rd2301;
	shr.u64 	%rd2304, %rd2101, 60;
	and.b64  	%rd2305, %rd2304, 8;
	add.s64 	%rd2306, %rd12, %rd2305;
	ld.local.u64 	%rd2307, [%rd2306];
	shl.b64 	%rd2308, %rd2307, 63;
	mov.b64	{%r2165, %r2166}, %rd2300;
	mov.b64	%rd2309, {%r2165, %r2166};
	xor.b64  	%rd2310, %rd2309, %rd2308;
	shr.u64 	%rd2311, %rd2307, 1;
	mov.b64	{%r2167, %r2168}, %rd2303;
	mov.b64	%rd2312, {%r2167, %r2168};
	xor.b64  	%rd2313, %rd2312, %rd2311;
	mov.b64	{%r6582, %r6583}, %rd2313;
	mov.b64	{%r6584, %r6585}, %rd2310;
	setp.gt.s32	%p35, %r1994, -1;
	@%p35 bra 	BB0_61;

	shr.u32 	%r2169, %r1991, 1;
	shl.b32 	%r2170, %r1992, 31;
	shr.u32 	%r2171, %r1992, 1;
	and.b32  	%r2172, %r2169, 1533916891;
	or.b32  	%r2173, %r2172, %r2170;
	xor.b32  	%r6582, %r6582, %r2173;
	and.b32  	%r2174, %r2171, 920350134;
	xor.b32  	%r6583, %r6583, %r2174;

BB0_61:
	and.b32  	%r2175, %r1994, 1073741824;
	setp.eq.s32	%p36, %r2175, 0;
	@%p36 bra 	BB0_63;

	shr.u32 	%r2176, %r1991, 2;
	shl.b32 	%r2177, %r1992, 30;
	or.b32  	%r2178, %r2176, %r2177;
	shr.u32 	%r2179, %r1992, 2;
	and.b32  	%r2180, %r2178, 1227133513;
	xor.b32  	%r6582, %r6582, %r2180;
	and.b32  	%r2181, %r2179, 306783378;
	xor.b32  	%r6583, %r6583, %r2181;

BB0_63:
	mov.u32 	%r6457, 0;
	xor.b32  	%r6730, %r6584, %r338;
	xor.b32  	%r6731, %r6585, %r339;
	xor.b32  	%r6732, %r6582, %r340;
	xor.b32  	%r6733, %r6583, %r341;
	mov.b64	%rd2314, {%r6732, %r6733};
	mov.b64	%rd2315, {%r6730, %r6731};
	cvt.u32.u64	%r2182, %rd2315;
	cvt.s32.s16 	%r2183, %r2182;
	mov.b64	%rd2316, {%r6723, %r6724};
	mov.b64	%rd2317, {%r6721, %r6722};
	cvt.u32.u64	%r2184, %rd2317;
	cvt.s32.s16 	%r2185, %r2184;
	mad.lo.s32 	%r2186, %r2183, %r2185, 16384;
	bfe.u32 	%r2187, %r2186, 15, 16;
	shr.s32 	%r2188, %r2182, 16;
	shr.s32 	%r2189, %r2184, 16;
	mad.lo.s32 	%r2190, %r2188, %r2189, 16384;
	mov.b64	%rd2318, {%r2187, %r6457};
	shl.b32 	%r2192, %r2190, 1;
	cvt.u64.u32	%rd2319, %r2192;
	and.b64  	%rd2320, %rd2318, 65535;
	shr.u64 	%rd2321, %rd2314, 16;
	shr.u64 	%rd2322, %rd2315, 16;
	cvt.u32.u64	%r2193, %rd2322;
	shr.s32 	%r2194, %r2193, 16;
	shr.u64 	%rd2323, %rd2316, 16;
	shr.u64 	%rd2324, %rd2317, 16;
	cvt.u32.u64	%r2195, %rd2324;
	shr.s32 	%r2196, %r2195, 16;
	mad.lo.s32 	%r2197, %r2194, %r2196, 16384;
	mul.wide.u32 	%rd2325, %r2197, 131072;
	shr.u64 	%rd2326, %rd2315, 32;
	cvt.u32.u64	%r2198, %rd2326;
	shr.s32 	%r2199, %r2198, 16;
	shr.u64 	%rd2327, %rd2317, 32;
	cvt.u32.u64	%r2200, %rd2327;
	shr.s32 	%r2201, %r2200, 16;
	mad.lo.s32 	%r2202, %r2199, %r2201, 16384;
	shr.u32 	%r2203, %r2202, 15;
	cvt.u64.u32	%rd2328, %r2203;
	shl.b64 	%rd2329, %rd2328, 48;
	and.b64  	%rd2330, %rd2325, 281470681743360;
	shl.b64 	%rd2331, %rd2314, 16;
	cvt.u32.u64	%r2204, %rd2331;
	shr.s32 	%r2205, %r2204, 16;
	shl.b64 	%rd2332, %rd2316, 16;
	cvt.u32.u64	%r2206, %rd2332;
	shr.s32 	%r2207, %r2206, 16;
	mad.lo.s32 	%r2208, %r2205, %r2207, 16384;
	shr.u32 	%r2209, %r2208, 15;
	cvt.u64.u32	%rd2333, %r2209;
	cvt.u32.u64	%r2210, %rd2314;
	shr.s32 	%r2211, %r2210, 16;
	cvt.u32.u64	%r2212, %rd2316;
	shr.s32 	%r2213, %r2212, 16;
	mad.lo.s32 	%r2214, %r2211, %r2213, 16384;
	and.b64  	%rd2334, %rd2333, 65535;
	and.b64  	%rd2335, %rd2319, 4294901760;
	or.b64  	%rd2336, %rd2335, %rd2320;
	or.b64  	%rd2337, %rd2336, %rd2330;
	shl.b32 	%r2215, %r2214, 1;
	cvt.u64.u32	%rd2338, %r2215;
	cvt.u32.u64	%r2216, %rd2321;
	shr.s32 	%r2217, %r2216, 16;
	cvt.u32.u64	%r2218, %rd2323;
	shr.s32 	%r2219, %r2218, 16;
	mad.lo.s32 	%r2220, %r2217, %r2219, 16384;
	mul.wide.u32 	%rd2339, %r2220, 131072;
	and.b64  	%rd2340, %rd2338, 4294901760;
	or.b64  	%rd2341, %rd2340, %rd2334;
	shr.u64 	%rd2342, %rd2314, 32;
	cvt.u32.u64	%r2221, %rd2342;
	shr.s32 	%r2222, %r2221, 16;
	shr.u64 	%rd2343, %rd2316, 32;
	cvt.u32.u64	%r2223, %rd2343;
	shr.s32 	%r2224, %r2223, 16;
	mad.lo.s32 	%r2225, %r2222, %r2224, 16384;
	shr.u32 	%r2226, %r2225, 15;
	cvt.u64.u32	%rd2344, %r2226;
	shl.b64 	%rd2345, %rd2344, 48;
	and.b64  	%rd2346, %rd2339, 281470681743360;
	or.b64  	%rd2347, %rd2346, %rd2345;
	or.b64  	%rd2348, %rd2347, %rd2341;
	or.b64  	%rd2349, %rd2337, %rd2329;
	mov.b64	{%r2227, %r2228}, %rd2348;
	mov.b64	{%r2229, %r2230}, %rd2349;
	xor.b32  	%r6502, %r2230, %r6722;
	xor.b32  	%r6501, %r2229, %r6721;
	xor.b32  	%r6504, %r2228, %r6724;
	xor.b32  	%r6503, %r2227, %r6723;
	cvt.u64.u32	%rd2350, %r6729;
	add.s64 	%rd2351, %rd2350, %rd9;
	shl.b64 	%rd2352, %rd2351, 1;
	mov.u64 	%rd2353, verus_gpu_hash$sharedrand;
	add.s64 	%rd2354, %rd2353, %rd2352;
	st.shared.u16 	[%rd2354], %rs236;
	mov.u64 	%rd2355, verus_gpu_hash$sharedrandex;
	add.s64 	%rd2356, %rd2355, %rd2352;
	st.shared.u16 	[%rd2356], %rs237;
	cvt.u64.u16	%rd2357, %rs236;
	add.s64 	%rd2358, %rd2357, %rd3;
	shl.b64 	%rd2359, %rd2358, 4;
	add.s64 	%rd2360, %rd203, %rd2359;
	st.global.v4.u32 	[%rd2360], {%r305, %r306, %r307, %r308};
	cvt.u64.u16	%rd2361, %rs237;
	add.s64 	%rd2362, %rd2361, %rd3;
	shl.b64 	%rd2363, %rd2362, 4;
	add.s64 	%rd2364, %rd203, %rd2363;
	st.global.v4.u32 	[%rd2364], {%r6501, %r6502, %r6503, %r6504};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p37, %r6729, 32;
	@%p37 bra 	BB0_134;

	mov.u32 	%r6458, 0;
	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p38, %r6729, 0;
	mov.u16 	%rs238, 8;
	@%p38 bra 	BB0_65;
	bra.uni 	BB0_66;

BB0_65:
	mov.u32 	%r6721, %r305;
	mov.u32 	%r6722, %r306;
	mov.u32 	%r6723, %r307;
	mov.u32 	%r6724, %r308;
	mov.u32 	%r6729, %r6458;
	bra.uni 	BB0_67;

BB0_66:
	cvt.u64.u32	%rd2365, %r6731;
	shr.u32 	%r2232, %r6730, 5;
	cvt.u64.u32	%rd2366, %r2232;
	cvt.u16.u32	%rs81, %r2232;
	and.b16  	%rs236, %rs81, 511;
	cvt.u16.u32	%rs82, %r6731;
	and.b16  	%rs237, %rs82, 511;
	cvt.u16.u64	%rs83, %rd6123;
	and.b16  	%rs238, %rs83, 28;
	and.b64  	%rd2367, %rd2366, 511;
	add.s64 	%rd2368, %rd2367, %rd3;
	shl.b64 	%rd2369, %rd2368, 4;
	add.s64 	%rd2370, %rd203, %rd2369;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd2370];
	and.b64  	%rd2371, %rd2365, 511;
	add.s64 	%rd2372, %rd2371, %rd3;
	shl.b64 	%rd2373, %rd2372, 4;
	add.s64 	%rd2374, %rd203, %rd2373;
	ld.global.v4.u32 	{%r6501, %r6502, %r6503, %r6504}, [%rd2374];
	mul.wide.u32 	%rd2375, %r6730, 16;
	and.b64  	%rd2376, %rd2375, 48;
	add.s64 	%rd6122, %rd1, %rd2376;

BB0_67:
	and.b16  	%rs84, %rs238, 255;
	setp.ne.s16	%p39, %rs84, 12;
	@%p39 bra 	BB0_68;

	mov.b64	{%r2241, %r2242}, %rd6123;
	and.b64  	%rd2377, %rd6123, 1;
	mov.u64 	%rd2378, 1;
	shl.b64 	%rd2379, %rd2377, 1;
	sub.s64 	%rd2380, %rd2378, %rd2379;
	shl.b64 	%rd2381, %rd2380, 4;
	add.s64 	%rd2382, %rd6122, %rd2381;
	ld.local.v4.u32 	{%r2243, %r2244, %r2245, %r2246}, [%rd2382];
	xor.b32  	%r2251, %r6721, %r2243;
	xor.b32  	%r2252, %r6722, %r2244;
	xor.b32  	%r2253, %r6723, %r2245;
	xor.b32  	%r2254, %r6724, %r2246;
	xor.b32  	%r6733, %r2254, %r6733;
	xor.b32  	%r6732, %r2253, %r6732;
	xor.b32  	%r6731, %r2252, %r6731;
	xor.b32  	%r393, %r2251, %r6730;
	mov.b64	%rd43, {%r393, %r6731};
	cvt.s64.s32	%rd44, %r2241;
	or.b64  	%rd2383, %rd43, %rd44;
	and.b64  	%rd2384, %rd2383, -4294967296;
	setp.eq.s64	%p40, %rd2384, 0;
	@%p40 bra 	BB0_71;
	bra.uni 	BB0_70;

BB0_71:
	cvt.u32.u64	%r2255, %rd44;
	cvt.u32.u64	%r2256, %rd43;
	rem.u32 	%r2257, %r2256, %r2255;
	cvt.u64.u32	%rd6106, %r2257;
	bra.uni 	BB0_72;

BB0_68:
	mov.u32 	%r6625, %r6501;
	mov.u32 	%r6626, %r6502;
	mov.u32 	%r6627, %r6503;
	mov.u32 	%r6628, %r6504;
	bra.uni 	BB0_86;

BB0_70:
	rem.s64 	%rd6106, %rd43, %rd44;

BB0_72:
	cvt.u32.u64	%r2258, %rd6106;
	xor.b32  	%r6730, %r2258, %r393;
	mov.b64	%rd2385, {%r6732, %r6733};
	mov.b64	%rd2386, {%r6730, %r6731};
	cvt.u32.u64	%r2259, %rd2386;
	cvt.s32.s16 	%r2260, %r2259;
	mov.b64	%rd2387, {%r6723, %r6724};
	mov.b64	%rd2388, {%r6721, %r6722};
	cvt.u32.u64	%r2261, %rd2388;
	cvt.s32.s16 	%r2262, %r2261;
	mad.lo.s32 	%r2263, %r2260, %r2262, 16384;
	bfe.u32 	%r2264, %r2263, 15, 16;
	shr.s32 	%r2265, %r2259, 16;
	shr.s32 	%r2266, %r2261, 16;
	mad.lo.s32 	%r2267, %r2265, %r2266, 16384;
	mov.u32 	%r2268, 0;
	mov.b64	%rd2389, {%r2264, %r2268};
	shl.b32 	%r2269, %r2267, 1;
	cvt.u64.u32	%rd2390, %r2269;
	and.b64  	%rd2391, %rd2389, 65535;
	shr.u64 	%rd2392, %rd2385, 16;
	shr.u64 	%rd2393, %rd2386, 16;
	cvt.u32.u64	%r2270, %rd2393;
	shr.s32 	%r2271, %r2270, 16;
	shr.u64 	%rd2394, %rd2387, 16;
	shr.u64 	%rd2395, %rd2388, 16;
	cvt.u32.u64	%r2272, %rd2395;
	shr.s32 	%r2273, %r2272, 16;
	mad.lo.s32 	%r2274, %r2271, %r2273, 16384;
	mul.wide.u32 	%rd2396, %r2274, 131072;
	shr.u64 	%rd2397, %rd2386, 32;
	cvt.u32.u64	%r2275, %rd2397;
	shr.s32 	%r2276, %r2275, 16;
	shr.u64 	%rd2398, %rd2388, 32;
	cvt.u32.u64	%r2277, %rd2398;
	shr.s32 	%r2278, %r2277, 16;
	mad.lo.s32 	%r2279, %r2276, %r2278, 16384;
	shr.u32 	%r2280, %r2279, 15;
	cvt.u64.u32	%rd2399, %r2280;
	shl.b64 	%rd2400, %rd2399, 48;
	and.b64  	%rd2401, %rd2396, 281470681743360;
	shl.b64 	%rd2402, %rd2385, 16;
	cvt.u32.u64	%r2281, %rd2402;
	shr.s32 	%r2282, %r2281, 16;
	shl.b64 	%rd2403, %rd2387, 16;
	cvt.u32.u64	%r2283, %rd2403;
	shr.s32 	%r2284, %r2283, 16;
	mad.lo.s32 	%r2285, %r2282, %r2284, 16384;
	shr.u32 	%r2286, %r2285, 15;
	cvt.u64.u32	%rd2404, %r2286;
	cvt.u32.u64	%r2287, %rd2385;
	shr.s32 	%r2288, %r2287, 16;
	cvt.u32.u64	%r2289, %rd2387;
	shr.s32 	%r2290, %r2289, 16;
	mad.lo.s32 	%r2291, %r2288, %r2290, 16384;
	and.b64  	%rd2405, %rd2404, 65535;
	and.b64  	%rd2406, %rd2390, 4294901760;
	or.b64  	%rd2407, %rd2406, %rd2391;
	or.b64  	%rd2408, %rd2407, %rd2401;
	shl.b32 	%r2292, %r2291, 1;
	cvt.u64.u32	%rd2409, %r2292;
	cvt.u32.u64	%r2293, %rd2392;
	shr.s32 	%r2294, %r2293, 16;
	cvt.u32.u64	%r2295, %rd2394;
	shr.s32 	%r2296, %r2295, 16;
	mad.lo.s32 	%r2297, %r2294, %r2296, 16384;
	mul.wide.u32 	%rd2410, %r2297, 131072;
	and.b64  	%rd2411, %rd2409, 4294901760;
	or.b64  	%rd2412, %rd2411, %rd2405;
	shr.u64 	%rd2413, %rd2385, 32;
	cvt.u32.u64	%r2298, %rd2413;
	shr.s32 	%r2299, %r2298, 16;
	shr.u64 	%rd2414, %rd2387, 32;
	cvt.u32.u64	%r2300, %rd2414;
	shr.s32 	%r2301, %r2300, 16;
	mad.lo.s32 	%r2302, %r2299, %r2301, 16384;
	shr.u32 	%r2303, %r2302, 15;
	cvt.u64.u32	%rd2415, %r2303;
	shl.b64 	%rd2416, %rd2415, 48;
	and.b64  	%rd2417, %rd2410, 281470681743360;
	or.b64  	%rd2418, %rd2417, %rd2416;
	or.b64  	%rd2419, %rd2418, %rd2412;
	or.b64  	%rd2420, %rd2408, %rd2400;
	mov.b64	{%r2304, %r2305}, %rd2419;
	mov.b64	{%r2306, %r2307}, %rd2420;
	xor.b32  	%r6628, %r2305, %r6724;
	xor.b32  	%r6627, %r2304, %r6723;
	xor.b32  	%r6626, %r2307, %r6722;
	xor.b32  	%r6625, %r2306, %r6721;
	and.b64  	%rd2421, %rd43, 1;
	setp.eq.b64	%p41, %rd2421, 1;
	@!%p41 bra 	BB0_82;
	bra.uni 	BB0_73;

BB0_73:
	ld.local.v4.u32 	{%r2308, %r2309, %r2310, %r2311}, [%rd6122];
	xor.b32  	%r411, %r6503, %r2310;
	xor.b32  	%r410, %r6502, %r2309;
	xor.b32  	%r409, %r6501, %r2308;
	xor.b32  	%r412, %r6504, %r2311;
	mov.b64	%rd2422, {%r411, %r412};
	mov.u64 	%rd2423, 0;
	st.local.v2.u64 	[%rd12], {%rd2423, %rd2422};
	shl.b64 	%rd2426, %rd2422, 1;
	xor.b64  	%rd2427, %rd2426, %rd2422;
	st.local.v2.u64 	[%rd15], {%rd2426, %rd2427};
	shl.b64 	%rd2428, %rd2422, 2;
	xor.b64  	%rd2429, %rd2428, %rd2422;
	st.local.v2.u64 	[%rd15+16], {%rd2428, %rd2429};
	shl.b64 	%rd2430, %rd2427, 1;
	xor.b64  	%rd2431, %rd2430, %rd2422;
	st.local.v2.u64 	[%rd15+32], {%rd2430, %rd2431};
	mov.b64	%rd2432, {%r409, %r410};
	shl.b64 	%rd2433, %rd2432, 3;
	and.b64  	%rd2434, %rd2433, 56;
	add.s64 	%rd2435, %rd12, %rd2434;
	and.b64  	%rd2436, %rd2432, 56;
	add.s64 	%rd2437, %rd12, %rd2436;
	ld.local.u64 	%rd2438, [%rd2437];
	shl.b64 	%rd2439, %rd2438, 3;
	ld.local.u64 	%rd2440, [%rd2435];
	xor.b64  	%rd2441, %rd2440, %rd2439;
	shr.u64 	%rd2442, %rd2438, 61;
	mov.b64	%rd2443, {%r2268, %r2268};
	xor.b64  	%rd2444, %rd2443, %rd2442;
	shr.u64 	%rd2445, %rd2432, 3;
	and.b64  	%rd2446, %rd2445, 56;
	add.s64 	%rd2447, %rd12, %rd2446;
	ld.local.u64 	%rd2448, [%rd2447];
	shl.b64 	%rd2449, %rd2448, 6;
	mov.b64	{%r2313, %r2314}, %rd2441;
	mov.b64	%rd2450, {%r2313, %r2314};
	xor.b64  	%rd2451, %rd2450, %rd2449;
	shr.u64 	%rd2452, %rd2448, 58;
	mov.b64	{%r2315, %r2316}, %rd2444;
	mov.b64	%rd2453, {%r2315, %r2316};
	xor.b64  	%rd2454, %rd2453, %rd2452;
	shr.u64 	%rd2455, %rd2432, 6;
	and.b64  	%rd2456, %rd2455, 56;
	add.s64 	%rd2457, %rd12, %rd2456;
	ld.local.u64 	%rd2458, [%rd2457];
	shl.b64 	%rd2459, %rd2458, 9;
	mov.b64	{%r2317, %r2318}, %rd2451;
	mov.b64	%rd2460, {%r2317, %r2318};
	xor.b64  	%rd2461, %rd2460, %rd2459;
	shr.u64 	%rd2462, %rd2458, 55;
	mov.b64	{%r2319, %r2320}, %rd2454;
	mov.b64	%rd2463, {%r2319, %r2320};
	xor.b64  	%rd2464, %rd2463, %rd2462;
	shr.u64 	%rd2465, %rd2432, 9;
	and.b64  	%rd2466, %rd2465, 56;
	add.s64 	%rd2467, %rd12, %rd2466;
	ld.local.u64 	%rd2468, [%rd2467];
	shl.b64 	%rd2469, %rd2468, 12;
	mov.b64	{%r2321, %r2322}, %rd2461;
	mov.b64	%rd2470, {%r2321, %r2322};
	xor.b64  	%rd2471, %rd2470, %rd2469;
	shr.u64 	%rd2472, %rd2468, 52;
	mov.b64	{%r2323, %r2324}, %rd2464;
	mov.b64	%rd2473, {%r2323, %r2324};
	xor.b64  	%rd2474, %rd2473, %rd2472;
	shr.u64 	%rd2475, %rd2432, 12;
	and.b64  	%rd2476, %rd2475, 56;
	add.s64 	%rd2477, %rd12, %rd2476;
	ld.local.u64 	%rd2478, [%rd2477];
	shl.b64 	%rd2479, %rd2478, 15;
	mov.b64	{%r2325, %r2326}, %rd2471;
	mov.b64	%rd2480, {%r2325, %r2326};
	xor.b64  	%rd2481, %rd2480, %rd2479;
	shr.u64 	%rd2482, %rd2478, 49;
	mov.b64	{%r2327, %r2328}, %rd2474;
	mov.b64	%rd2483, {%r2327, %r2328};
	xor.b64  	%rd2484, %rd2483, %rd2482;
	shr.u64 	%rd2485, %rd2432, 15;
	and.b64  	%rd2486, %rd2485, 56;
	add.s64 	%rd2487, %rd12, %rd2486;
	ld.local.u64 	%rd2488, [%rd2487];
	shl.b64 	%rd2489, %rd2488, 18;
	mov.b64	{%r2329, %r2330}, %rd2481;
	mov.b64	%rd2490, {%r2329, %r2330};
	xor.b64  	%rd2491, %rd2490, %rd2489;
	shr.u64 	%rd2492, %rd2488, 46;
	mov.b64	{%r2331, %r2332}, %rd2484;
	mov.b64	%rd2493, {%r2331, %r2332};
	xor.b64  	%rd2494, %rd2493, %rd2492;
	shr.u64 	%rd2495, %rd2432, 18;
	and.b64  	%rd2496, %rd2495, 56;
	add.s64 	%rd2497, %rd12, %rd2496;
	ld.local.u64 	%rd2498, [%rd2497];
	shl.b64 	%rd2499, %rd2498, 21;
	mov.b64	{%r2333, %r2334}, %rd2491;
	mov.b64	%rd2500, {%r2333, %r2334};
	xor.b64  	%rd2501, %rd2500, %rd2499;
	shr.u64 	%rd2502, %rd2498, 43;
	mov.b64	{%r2335, %r2336}, %rd2494;
	mov.b64	%rd2503, {%r2335, %r2336};
	xor.b64  	%rd2504, %rd2503, %rd2502;
	shr.u64 	%rd2505, %rd2432, 21;
	and.b64  	%rd2506, %rd2505, 56;
	add.s64 	%rd2507, %rd12, %rd2506;
	ld.local.u64 	%rd2508, [%rd2507];
	shl.b64 	%rd2509, %rd2508, 24;
	mov.b64	{%r2337, %r2338}, %rd2501;
	mov.b64	%rd2510, {%r2337, %r2338};
	xor.b64  	%rd2511, %rd2510, %rd2509;
	shr.u64 	%rd2512, %rd2508, 40;
	mov.b64	{%r2339, %r2340}, %rd2504;
	mov.b64	%rd2513, {%r2339, %r2340};
	xor.b64  	%rd2514, %rd2513, %rd2512;
	shr.u64 	%rd2515, %rd2432, 24;
	and.b64  	%rd2516, %rd2515, 56;
	add.s64 	%rd2517, %rd12, %rd2516;
	ld.local.u64 	%rd2518, [%rd2517];
	shl.b64 	%rd2519, %rd2518, 27;
	mov.b64	{%r2341, %r2342}, %rd2511;
	mov.b64	%rd2520, {%r2341, %r2342};
	xor.b64  	%rd2521, %rd2520, %rd2519;
	shr.u64 	%rd2522, %rd2518, 37;
	mov.b64	{%r2343, %r2344}, %rd2514;
	mov.b64	%rd2523, {%r2343, %r2344};
	xor.b64  	%rd2524, %rd2523, %rd2522;
	shr.u64 	%rd2525, %rd2432, 27;
	and.b64  	%rd2526, %rd2525, 56;
	add.s64 	%rd2527, %rd12, %rd2526;
	ld.local.u64 	%rd2528, [%rd2527];
	shl.b64 	%rd2529, %rd2528, 30;
	mov.b64	{%r2345, %r2346}, %rd2521;
	mov.b64	%rd2530, {%r2345, %r2346};
	xor.b64  	%rd2531, %rd2530, %rd2529;
	shr.u64 	%rd2532, %rd2528, 34;
	mov.b64	{%r2347, %r2348}, %rd2524;
	mov.b64	%rd2533, {%r2347, %r2348};
	xor.b64  	%rd2534, %rd2533, %rd2532;
	shr.u64 	%rd2535, %rd2432, 30;
	and.b64  	%rd2536, %rd2535, 56;
	add.s64 	%rd2537, %rd12, %rd2536;
	ld.local.u64 	%rd2538, [%rd2537];
	shl.b64 	%rd2539, %rd2538, 33;
	mov.b64	{%r2349, %r2350}, %rd2531;
	mov.b64	%rd2540, {%r2349, %r2350};
	xor.b64  	%rd2541, %rd2540, %rd2539;
	shr.u64 	%rd2542, %rd2538, 31;
	mov.b64	{%r2351, %r2352}, %rd2534;
	mov.b64	%rd2543, {%r2351, %r2352};
	xor.b64  	%rd2544, %rd2543, %rd2542;
	shr.u64 	%rd2545, %rd2432, 33;
	and.b64  	%rd2546, %rd2545, 56;
	add.s64 	%rd2547, %rd12, %rd2546;
	ld.local.u64 	%rd2548, [%rd2547];
	shl.b64 	%rd2549, %rd2548, 36;
	mov.b64	{%r2353, %r2354}, %rd2541;
	mov.b64	%rd2550, {%r2353, %r2354};
	xor.b64  	%rd2551, %rd2550, %rd2549;
	shr.u64 	%rd2552, %rd2548, 28;
	mov.b64	{%r2355, %r2356}, %rd2544;
	mov.b64	%rd2553, {%r2355, %r2356};
	xor.b64  	%rd2554, %rd2553, %rd2552;
	shr.u64 	%rd2555, %rd2432, 36;
	and.b64  	%rd2556, %rd2555, 56;
	add.s64 	%rd2557, %rd12, %rd2556;
	ld.local.u64 	%rd2558, [%rd2557];
	shl.b64 	%rd2559, %rd2558, 39;
	mov.b64	{%r2357, %r2358}, %rd2551;
	mov.b64	%rd2560, {%r2357, %r2358};
	xor.b64  	%rd2561, %rd2560, %rd2559;
	shr.u64 	%rd2562, %rd2558, 25;
	mov.b64	{%r2359, %r2360}, %rd2554;
	mov.b64	%rd2563, {%r2359, %r2360};
	xor.b64  	%rd2564, %rd2563, %rd2562;
	shr.u64 	%rd2565, %rd2432, 39;
	and.b64  	%rd2566, %rd2565, 56;
	add.s64 	%rd2567, %rd12, %rd2566;
	ld.local.u64 	%rd2568, [%rd2567];
	shl.b64 	%rd2569, %rd2568, 42;
	mov.b64	{%r2361, %r2362}, %rd2561;
	mov.b64	%rd2570, {%r2361, %r2362};
	xor.b64  	%rd2571, %rd2570, %rd2569;
	shr.u64 	%rd2572, %rd2568, 22;
	mov.b64	{%r2363, %r2364}, %rd2564;
	mov.b64	%rd2573, {%r2363, %r2364};
	xor.b64  	%rd2574, %rd2573, %rd2572;
	shr.u64 	%rd2575, %rd2432, 42;
	and.b64  	%rd2576, %rd2575, 56;
	add.s64 	%rd2577, %rd12, %rd2576;
	ld.local.u64 	%rd2578, [%rd2577];
	shl.b64 	%rd2579, %rd2578, 45;
	mov.b64	{%r2365, %r2366}, %rd2571;
	mov.b64	%rd2580, {%r2365, %r2366};
	xor.b64  	%rd2581, %rd2580, %rd2579;
	shr.u64 	%rd2582, %rd2578, 19;
	mov.b64	{%r2367, %r2368}, %rd2574;
	mov.b64	%rd2583, {%r2367, %r2368};
	xor.b64  	%rd2584, %rd2583, %rd2582;
	shr.u64 	%rd2585, %rd2432, 45;
	and.b64  	%rd2586, %rd2585, 56;
	add.s64 	%rd2587, %rd12, %rd2586;
	ld.local.u64 	%rd2588, [%rd2587];
	shl.b64 	%rd2589, %rd2588, 48;
	mov.b64	{%r2369, %r2370}, %rd2581;
	mov.b64	%rd2590, {%r2369, %r2370};
	xor.b64  	%rd2591, %rd2590, %rd2589;
	shr.u64 	%rd2592, %rd2588, 16;
	mov.b64	{%r2371, %r2372}, %rd2584;
	mov.b64	%rd2593, {%r2371, %r2372};
	xor.b64  	%rd2594, %rd2593, %rd2592;
	shr.u64 	%rd2595, %rd2432, 48;
	and.b64  	%rd2596, %rd2595, 56;
	add.s64 	%rd2597, %rd12, %rd2596;
	ld.local.u64 	%rd2598, [%rd2597];
	shl.b64 	%rd2599, %rd2598, 51;
	mov.b64	{%r2373, %r2374}, %rd2591;
	mov.b64	%rd2600, {%r2373, %r2374};
	xor.b64  	%rd2601, %rd2600, %rd2599;
	shr.u64 	%rd2602, %rd2598, 13;
	mov.b64	{%r2375, %r2376}, %rd2594;
	mov.b64	%rd2603, {%r2375, %r2376};
	xor.b64  	%rd2604, %rd2603, %rd2602;
	shr.u64 	%rd2605, %rd2432, 51;
	and.b64  	%rd2606, %rd2605, 56;
	add.s64 	%rd2607, %rd12, %rd2606;
	ld.local.u64 	%rd2608, [%rd2607];
	shl.b64 	%rd2609, %rd2608, 54;
	mov.b64	{%r2377, %r2378}, %rd2601;
	mov.b64	%rd2610, {%r2377, %r2378};
	xor.b64  	%rd2611, %rd2610, %rd2609;
	shr.u64 	%rd2612, %rd2608, 10;
	mov.b64	{%r2379, %r2380}, %rd2604;
	mov.b64	%rd2613, {%r2379, %r2380};
	xor.b64  	%rd2614, %rd2613, %rd2612;
	shr.u64 	%rd2615, %rd2432, 54;
	and.b64  	%rd2616, %rd2615, 56;
	add.s64 	%rd2617, %rd12, %rd2616;
	ld.local.u64 	%rd2618, [%rd2617];
	shl.b64 	%rd2619, %rd2618, 57;
	mov.b64	{%r2381, %r2382}, %rd2611;
	mov.b64	%rd2620, {%r2381, %r2382};
	xor.b64  	%rd2621, %rd2620, %rd2619;
	shr.u64 	%rd2622, %rd2618, 7;
	mov.b64	{%r2383, %r2384}, %rd2614;
	mov.b64	%rd2623, {%r2383, %r2384};
	xor.b64  	%rd2624, %rd2623, %rd2622;
	shr.u64 	%rd2625, %rd2432, 57;
	and.b64  	%rd2626, %rd2625, 56;
	add.s64 	%rd2627, %rd12, %rd2626;
	ld.local.u64 	%rd2628, [%rd2627];
	shl.b64 	%rd2629, %rd2628, 60;
	mov.b64	{%r2385, %r2386}, %rd2621;
	mov.b64	%rd2630, {%r2385, %r2386};
	xor.b64  	%rd2631, %rd2630, %rd2629;
	shr.u64 	%rd2632, %rd2628, 4;
	mov.b64	{%r2387, %r2388}, %rd2624;
	mov.b64	%rd2633, {%r2387, %r2388};
	xor.b64  	%rd2634, %rd2633, %rd2632;
	shr.u64 	%rd2635, %rd2432, 60;
	and.b64  	%rd2636, %rd2635, 8;
	add.s64 	%rd2637, %rd12, %rd2636;
	ld.local.u64 	%rd2638, [%rd2637];
	shl.b64 	%rd2639, %rd2638, 63;
	mov.b64	{%r2389, %r2390}, %rd2631;
	mov.b64	%rd2640, {%r2389, %r2390};
	xor.b64  	%rd2641, %rd2640, %rd2639;
	shr.u64 	%rd2642, %rd2638, 1;
	mov.b64	{%r2391, %r2392}, %rd2634;
	mov.b64	%rd2643, {%r2391, %r2392};
	xor.b64  	%rd2644, %rd2643, %rd2642;
	mov.b64	{%r6603, %r6604}, %rd2644;
	mov.b64	{%r6605, %r6606}, %rd2641;
	setp.gt.s32	%p42, %r412, -1;
	@%p42 bra 	BB0_75;

	xor.b32  	%r6451, %r6502, %r2309;
	xor.b32  	%r6450, %r6501, %r2308;
	shr.u32 	%r2393, %r6450, 1;
	shl.b32 	%r2394, %r6451, 31;
	shr.u32 	%r2395, %r6451, 1;
	and.b32  	%r2396, %r2393, 1533916891;
	or.b32  	%r2397, %r2396, %r2394;
	xor.b32  	%r6603, %r6603, %r2397;
	and.b32  	%r2398, %r2395, 920350134;
	xor.b32  	%r6604, %r6604, %r2398;

BB0_75:
	and.b32  	%r2399, %r412, 1073741824;
	setp.eq.s32	%p43, %r2399, 0;
	@%p43 bra 	BB0_77;

	xor.b32  	%r6449, %r6502, %r2309;
	xor.b32  	%r6448, %r6501, %r2308;
	shr.u32 	%r2400, %r6448, 2;
	shl.b32 	%r2401, %r6449, 30;
	or.b32  	%r2402, %r2400, %r2401;
	shr.u32 	%r2403, %r6449, 2;
	and.b32  	%r2404, %r2402, 1227133513;
	xor.b32  	%r6603, %r6603, %r2404;
	and.b32  	%r2405, %r2403, 306783378;
	xor.b32  	%r6604, %r6604, %r2405;

BB0_77:
	mov.u32 	%r6459, 0;
	mov.u64 	%rd6076, 0;
	xor.b32  	%r434, %r6605, %r6730;
	xor.b32  	%r435, %r6606, %r6731;
	xor.b32  	%r436, %r6603, %r6732;
	xor.b32  	%r437, %r6604, %r6733;
	mov.b64	%rd2645, {%r2310, %r2311};
	st.local.v2.u64 	[%rd12], {%rd6076, %rd2645};
	shl.b64 	%rd2649, %rd2645, 1;
	xor.b64  	%rd2650, %rd2649, %rd2645;
	st.local.v2.u64 	[%rd15], {%rd2649, %rd2650};
	shl.b64 	%rd2651, %rd2645, 2;
	xor.b64  	%rd2652, %rd2651, %rd2645;
	st.local.v2.u64 	[%rd15+16], {%rd2651, %rd2652};
	shl.b64 	%rd2653, %rd2650, 1;
	xor.b64  	%rd2654, %rd2653, %rd2645;
	st.local.v2.u64 	[%rd15+32], {%rd2653, %rd2654};
	mov.b64	%rd2655, {%r2308, %r2309};
	shl.b64 	%rd2656, %rd2655, 3;
	and.b64  	%rd2657, %rd2656, 56;
	add.s64 	%rd2658, %rd12, %rd2657;
	and.b64  	%rd2659, %rd2655, 56;
	add.s64 	%rd2660, %rd12, %rd2659;
	ld.local.u64 	%rd2661, [%rd2660];
	shl.b64 	%rd2662, %rd2661, 3;
	ld.local.u64 	%rd2663, [%rd2658];
	xor.b64  	%rd2664, %rd2663, %rd2662;
	shr.u64 	%rd2665, %rd2661, 61;
	mov.b64	%rd2666, {%r6459, %r6459};
	xor.b64  	%rd2667, %rd2666, %rd2665;
	shr.u64 	%rd2668, %rd2655, 3;
	and.b64  	%rd2669, %rd2668, 56;
	add.s64 	%rd2670, %rd12, %rd2669;
	ld.local.u64 	%rd2671, [%rd2670];
	shl.b64 	%rd2672, %rd2671, 6;
	mov.b64	{%r2407, %r2408}, %rd2664;
	mov.b64	%rd2673, {%r2407, %r2408};
	xor.b64  	%rd2674, %rd2673, %rd2672;
	shr.u64 	%rd2675, %rd2671, 58;
	mov.b64	{%r2409, %r2410}, %rd2667;
	mov.b64	%rd2676, {%r2409, %r2410};
	xor.b64  	%rd2677, %rd2676, %rd2675;
	shr.u64 	%rd2678, %rd2655, 6;
	and.b64  	%rd2679, %rd2678, 56;
	add.s64 	%rd2680, %rd12, %rd2679;
	ld.local.u64 	%rd2681, [%rd2680];
	shl.b64 	%rd2682, %rd2681, 9;
	mov.b64	{%r2411, %r2412}, %rd2674;
	mov.b64	%rd2683, {%r2411, %r2412};
	xor.b64  	%rd2684, %rd2683, %rd2682;
	shr.u64 	%rd2685, %rd2681, 55;
	mov.b64	{%r2413, %r2414}, %rd2677;
	mov.b64	%rd2686, {%r2413, %r2414};
	xor.b64  	%rd2687, %rd2686, %rd2685;
	shr.u64 	%rd2688, %rd2655, 9;
	and.b64  	%rd2689, %rd2688, 56;
	add.s64 	%rd2690, %rd12, %rd2689;
	ld.local.u64 	%rd2691, [%rd2690];
	shl.b64 	%rd2692, %rd2691, 12;
	mov.b64	{%r2415, %r2416}, %rd2684;
	mov.b64	%rd2693, {%r2415, %r2416};
	xor.b64  	%rd2694, %rd2693, %rd2692;
	shr.u64 	%rd2695, %rd2691, 52;
	mov.b64	{%r2417, %r2418}, %rd2687;
	mov.b64	%rd2696, {%r2417, %r2418};
	xor.b64  	%rd2697, %rd2696, %rd2695;
	shr.u64 	%rd2698, %rd2655, 12;
	and.b64  	%rd2699, %rd2698, 56;
	add.s64 	%rd2700, %rd12, %rd2699;
	ld.local.u64 	%rd2701, [%rd2700];
	shl.b64 	%rd2702, %rd2701, 15;
	mov.b64	{%r2419, %r2420}, %rd2694;
	mov.b64	%rd2703, {%r2419, %r2420};
	xor.b64  	%rd2704, %rd2703, %rd2702;
	shr.u64 	%rd2705, %rd2701, 49;
	mov.b64	{%r2421, %r2422}, %rd2697;
	mov.b64	%rd2706, {%r2421, %r2422};
	xor.b64  	%rd2707, %rd2706, %rd2705;
	shr.u64 	%rd2708, %rd2655, 15;
	and.b64  	%rd2709, %rd2708, 56;
	add.s64 	%rd2710, %rd12, %rd2709;
	ld.local.u64 	%rd2711, [%rd2710];
	shl.b64 	%rd2712, %rd2711, 18;
	mov.b64	{%r2423, %r2424}, %rd2704;
	mov.b64	%rd2713, {%r2423, %r2424};
	xor.b64  	%rd2714, %rd2713, %rd2712;
	shr.u64 	%rd2715, %rd2711, 46;
	mov.b64	{%r2425, %r2426}, %rd2707;
	mov.b64	%rd2716, {%r2425, %r2426};
	xor.b64  	%rd2717, %rd2716, %rd2715;
	shr.u64 	%rd2718, %rd2655, 18;
	and.b64  	%rd2719, %rd2718, 56;
	add.s64 	%rd2720, %rd12, %rd2719;
	ld.local.u64 	%rd2721, [%rd2720];
	shl.b64 	%rd2722, %rd2721, 21;
	mov.b64	{%r2427, %r2428}, %rd2714;
	mov.b64	%rd2723, {%r2427, %r2428};
	xor.b64  	%rd2724, %rd2723, %rd2722;
	shr.u64 	%rd2725, %rd2721, 43;
	mov.b64	{%r2429, %r2430}, %rd2717;
	mov.b64	%rd2726, {%r2429, %r2430};
	xor.b64  	%rd2727, %rd2726, %rd2725;
	shr.u64 	%rd2728, %rd2655, 21;
	and.b64  	%rd2729, %rd2728, 56;
	add.s64 	%rd2730, %rd12, %rd2729;
	ld.local.u64 	%rd2731, [%rd2730];
	shl.b64 	%rd2732, %rd2731, 24;
	mov.b64	{%r2431, %r2432}, %rd2724;
	mov.b64	%rd2733, {%r2431, %r2432};
	xor.b64  	%rd2734, %rd2733, %rd2732;
	shr.u64 	%rd2735, %rd2731, 40;
	mov.b64	{%r2433, %r2434}, %rd2727;
	mov.b64	%rd2736, {%r2433, %r2434};
	xor.b64  	%rd2737, %rd2736, %rd2735;
	shr.u64 	%rd2738, %rd2655, 24;
	and.b64  	%rd2739, %rd2738, 56;
	add.s64 	%rd2740, %rd12, %rd2739;
	ld.local.u64 	%rd2741, [%rd2740];
	shl.b64 	%rd2742, %rd2741, 27;
	mov.b64	{%r2435, %r2436}, %rd2734;
	mov.b64	%rd2743, {%r2435, %r2436};
	xor.b64  	%rd2744, %rd2743, %rd2742;
	shr.u64 	%rd2745, %rd2741, 37;
	mov.b64	{%r2437, %r2438}, %rd2737;
	mov.b64	%rd2746, {%r2437, %r2438};
	xor.b64  	%rd2747, %rd2746, %rd2745;
	shr.u64 	%rd2748, %rd2655, 27;
	and.b64  	%rd2749, %rd2748, 56;
	add.s64 	%rd2750, %rd12, %rd2749;
	ld.local.u64 	%rd2751, [%rd2750];
	shl.b64 	%rd2752, %rd2751, 30;
	mov.b64	{%r2439, %r2440}, %rd2744;
	mov.b64	%rd2753, {%r2439, %r2440};
	xor.b64  	%rd2754, %rd2753, %rd2752;
	shr.u64 	%rd2755, %rd2751, 34;
	mov.b64	{%r2441, %r2442}, %rd2747;
	mov.b64	%rd2756, {%r2441, %r2442};
	xor.b64  	%rd2757, %rd2756, %rd2755;
	shr.u64 	%rd2758, %rd2655, 30;
	and.b64  	%rd2759, %rd2758, 56;
	add.s64 	%rd2760, %rd12, %rd2759;
	ld.local.u64 	%rd2761, [%rd2760];
	shl.b64 	%rd2762, %rd2761, 33;
	mov.b64	{%r2443, %r2444}, %rd2754;
	mov.b64	%rd2763, {%r2443, %r2444};
	xor.b64  	%rd2764, %rd2763, %rd2762;
	shr.u64 	%rd2765, %rd2761, 31;
	mov.b64	{%r2445, %r2446}, %rd2757;
	mov.b64	%rd2766, {%r2445, %r2446};
	xor.b64  	%rd2767, %rd2766, %rd2765;
	shr.u64 	%rd2768, %rd2655, 33;
	and.b64  	%rd2769, %rd2768, 56;
	add.s64 	%rd2770, %rd12, %rd2769;
	ld.local.u64 	%rd2771, [%rd2770];
	shl.b64 	%rd2772, %rd2771, 36;
	mov.b64	{%r2447, %r2448}, %rd2764;
	mov.b64	%rd2773, {%r2447, %r2448};
	xor.b64  	%rd2774, %rd2773, %rd2772;
	shr.u64 	%rd2775, %rd2771, 28;
	mov.b64	{%r2449, %r2450}, %rd2767;
	mov.b64	%rd2776, {%r2449, %r2450};
	xor.b64  	%rd2777, %rd2776, %rd2775;
	shr.u64 	%rd2778, %rd2655, 36;
	and.b64  	%rd2779, %rd2778, 56;
	add.s64 	%rd2780, %rd12, %rd2779;
	ld.local.u64 	%rd2781, [%rd2780];
	shl.b64 	%rd2782, %rd2781, 39;
	mov.b64	{%r2451, %r2452}, %rd2774;
	mov.b64	%rd2783, {%r2451, %r2452};
	xor.b64  	%rd2784, %rd2783, %rd2782;
	shr.u64 	%rd2785, %rd2781, 25;
	mov.b64	{%r2453, %r2454}, %rd2777;
	mov.b64	%rd2786, {%r2453, %r2454};
	xor.b64  	%rd2787, %rd2786, %rd2785;
	shr.u64 	%rd2788, %rd2655, 39;
	and.b64  	%rd2789, %rd2788, 56;
	add.s64 	%rd2790, %rd12, %rd2789;
	ld.local.u64 	%rd2791, [%rd2790];
	shl.b64 	%rd2792, %rd2791, 42;
	mov.b64	{%r2455, %r2456}, %rd2784;
	mov.b64	%rd2793, {%r2455, %r2456};
	xor.b64  	%rd2794, %rd2793, %rd2792;
	shr.u64 	%rd2795, %rd2791, 22;
	mov.b64	{%r2457, %r2458}, %rd2787;
	mov.b64	%rd2796, {%r2457, %r2458};
	xor.b64  	%rd2797, %rd2796, %rd2795;
	shr.u64 	%rd2798, %rd2655, 42;
	and.b64  	%rd2799, %rd2798, 56;
	add.s64 	%rd2800, %rd12, %rd2799;
	ld.local.u64 	%rd2801, [%rd2800];
	shl.b64 	%rd2802, %rd2801, 45;
	mov.b64	{%r2459, %r2460}, %rd2794;
	mov.b64	%rd2803, {%r2459, %r2460};
	xor.b64  	%rd2804, %rd2803, %rd2802;
	shr.u64 	%rd2805, %rd2801, 19;
	mov.b64	{%r2461, %r2462}, %rd2797;
	mov.b64	%rd2806, {%r2461, %r2462};
	xor.b64  	%rd2807, %rd2806, %rd2805;
	shr.u64 	%rd2808, %rd2655, 45;
	and.b64  	%rd2809, %rd2808, 56;
	add.s64 	%rd2810, %rd12, %rd2809;
	ld.local.u64 	%rd2811, [%rd2810];
	shl.b64 	%rd2812, %rd2811, 48;
	mov.b64	{%r2463, %r2464}, %rd2804;
	mov.b64	%rd2813, {%r2463, %r2464};
	xor.b64  	%rd2814, %rd2813, %rd2812;
	shr.u64 	%rd2815, %rd2811, 16;
	mov.b64	{%r2465, %r2466}, %rd2807;
	mov.b64	%rd2816, {%r2465, %r2466};
	xor.b64  	%rd2817, %rd2816, %rd2815;
	shr.u64 	%rd2818, %rd2655, 48;
	and.b64  	%rd2819, %rd2818, 56;
	add.s64 	%rd2820, %rd12, %rd2819;
	ld.local.u64 	%rd2821, [%rd2820];
	shl.b64 	%rd2822, %rd2821, 51;
	mov.b64	{%r2467, %r2468}, %rd2814;
	mov.b64	%rd2823, {%r2467, %r2468};
	xor.b64  	%rd2824, %rd2823, %rd2822;
	shr.u64 	%rd2825, %rd2821, 13;
	mov.b64	{%r2469, %r2470}, %rd2817;
	mov.b64	%rd2826, {%r2469, %r2470};
	xor.b64  	%rd2827, %rd2826, %rd2825;
	shr.u64 	%rd2828, %rd2655, 51;
	and.b64  	%rd2829, %rd2828, 56;
	add.s64 	%rd2830, %rd12, %rd2829;
	ld.local.u64 	%rd2831, [%rd2830];
	shl.b64 	%rd2832, %rd2831, 54;
	mov.b64	{%r2471, %r2472}, %rd2824;
	mov.b64	%rd2833, {%r2471, %r2472};
	xor.b64  	%rd2834, %rd2833, %rd2832;
	shr.u64 	%rd2835, %rd2831, 10;
	mov.b64	{%r2473, %r2474}, %rd2827;
	mov.b64	%rd2836, {%r2473, %r2474};
	xor.b64  	%rd2837, %rd2836, %rd2835;
	shr.u64 	%rd2838, %rd2655, 54;
	and.b64  	%rd2839, %rd2838, 56;
	add.s64 	%rd2840, %rd12, %rd2839;
	ld.local.u64 	%rd2841, [%rd2840];
	shl.b64 	%rd2842, %rd2841, 57;
	mov.b64	{%r2475, %r2476}, %rd2834;
	mov.b64	%rd2843, {%r2475, %r2476};
	xor.b64  	%rd2844, %rd2843, %rd2842;
	shr.u64 	%rd2845, %rd2841, 7;
	mov.b64	{%r2477, %r2478}, %rd2837;
	mov.b64	%rd2846, {%r2477, %r2478};
	xor.b64  	%rd2847, %rd2846, %rd2845;
	shr.u64 	%rd2848, %rd2655, 57;
	and.b64  	%rd2849, %rd2848, 56;
	add.s64 	%rd2850, %rd12, %rd2849;
	ld.local.u64 	%rd2851, [%rd2850];
	shl.b64 	%rd2852, %rd2851, 60;
	mov.b64	{%r2479, %r2480}, %rd2844;
	mov.b64	%rd2853, {%r2479, %r2480};
	xor.b64  	%rd2854, %rd2853, %rd2852;
	shr.u64 	%rd2855, %rd2851, 4;
	mov.b64	{%r2481, %r2482}, %rd2847;
	mov.b64	%rd2856, {%r2481, %r2482};
	xor.b64  	%rd2857, %rd2856, %rd2855;
	shr.u64 	%rd2858, %rd2655, 60;
	and.b64  	%rd2859, %rd2858, 8;
	add.s64 	%rd2860, %rd12, %rd2859;
	ld.local.u64 	%rd2861, [%rd2860];
	shl.b64 	%rd2862, %rd2861, 63;
	mov.b64	{%r2483, %r2484}, %rd2854;
	mov.b64	%rd2863, {%r2483, %r2484};
	xor.b64  	%rd2864, %rd2863, %rd2862;
	shr.u64 	%rd2865, %rd2861, 1;
	mov.b64	{%r2485, %r2486}, %rd2857;
	mov.b64	%rd2866, {%r2485, %r2486};
	xor.b64  	%rd2867, %rd2866, %rd2865;
	mov.b64	{%r6611, %r6612}, %rd2867;
	mov.b64	{%r6613, %r6614}, %rd2864;
	setp.gt.s32	%p44, %r2311, -1;
	@%p44 bra 	BB0_79;

	shr.u32 	%r2487, %r2308, 1;
	shl.b32 	%r2488, %r2309, 31;
	shr.u32 	%r2489, %r2309, 1;
	and.b32  	%r2490, %r2487, 1533916891;
	or.b32  	%r2491, %r2490, %r2488;
	xor.b32  	%r6611, %r6611, %r2491;
	and.b32  	%r2492, %r2489, 920350134;
	xor.b32  	%r6612, %r6612, %r2492;

BB0_79:
	and.b32  	%r2493, %r2311, 1073741824;
	setp.eq.s32	%p45, %r2493, 0;
	@%p45 bra 	BB0_81;

	shr.u32 	%r2494, %r2308, 2;
	shl.b32 	%r2495, %r2309, 30;
	or.b32  	%r2496, %r2494, %r2495;
	shr.u32 	%r2497, %r2309, 2;
	and.b32  	%r2498, %r2496, 1227133513;
	xor.b32  	%r6611, %r6611, %r2498;
	and.b32  	%r2499, %r2497, 306783378;
	xor.b32  	%r6612, %r6612, %r2499;

BB0_81:
	mov.u32 	%r6460, 0;
	xor.b32  	%r6730, %r6613, %r434;
	xor.b32  	%r6731, %r6614, %r435;
	xor.b32  	%r6732, %r6611, %r436;
	xor.b32  	%r6733, %r6612, %r437;
	mov.b64	%rd2868, {%r6732, %r6733};
	mov.b64	%rd2869, {%r6730, %r6731};
	cvt.u32.u64	%r2500, %rd2869;
	cvt.s32.s16 	%r2501, %r2500;
	mov.b64	%rd2870, {%r6503, %r6504};
	mov.b64	%rd2871, {%r6501, %r6502};
	cvt.u32.u64	%r2502, %rd2871;
	cvt.s32.s16 	%r2503, %r2502;
	mad.lo.s32 	%r2504, %r2501, %r2503, 16384;
	bfe.u32 	%r2505, %r2504, 15, 16;
	shr.s32 	%r2506, %r2500, 16;
	shr.s32 	%r2507, %r2502, 16;
	mad.lo.s32 	%r2508, %r2506, %r2507, 16384;
	mov.b64	%rd2872, {%r2505, %r6460};
	shl.b32 	%r2510, %r2508, 1;
	cvt.u64.u32	%rd2873, %r2510;
	and.b64  	%rd2874, %rd2872, 65535;
	shr.u64 	%rd2875, %rd2868, 16;
	shr.u64 	%rd2876, %rd2869, 16;
	cvt.u32.u64	%r2511, %rd2876;
	shr.s32 	%r2512, %r2511, 16;
	shr.u64 	%rd2877, %rd2870, 16;
	shr.u64 	%rd2878, %rd2871, 16;
	cvt.u32.u64	%r2513, %rd2878;
	shr.s32 	%r2514, %r2513, 16;
	mad.lo.s32 	%r2515, %r2512, %r2514, 16384;
	mul.wide.u32 	%rd2879, %r2515, 131072;
	shr.u64 	%rd2880, %rd2869, 32;
	cvt.u32.u64	%r2516, %rd2880;
	shr.s32 	%r2517, %r2516, 16;
	shr.u64 	%rd2881, %rd2871, 32;
	cvt.u32.u64	%r2518, %rd2881;
	shr.s32 	%r2519, %r2518, 16;
	mad.lo.s32 	%r2520, %r2517, %r2519, 16384;
	shr.u32 	%r2521, %r2520, 15;
	cvt.u64.u32	%rd2882, %r2521;
	shl.b64 	%rd2883, %rd2882, 48;
	and.b64  	%rd2884, %rd2879, 281470681743360;
	shl.b64 	%rd2885, %rd2868, 16;
	cvt.u32.u64	%r2522, %rd2885;
	shr.s32 	%r2523, %r2522, 16;
	shl.b64 	%rd2886, %rd2870, 16;
	cvt.u32.u64	%r2524, %rd2886;
	shr.s32 	%r2525, %r2524, 16;
	mad.lo.s32 	%r2526, %r2523, %r2525, 16384;
	shr.u32 	%r2527, %r2526, 15;
	cvt.u64.u32	%rd2887, %r2527;
	cvt.u32.u64	%r2528, %rd2868;
	shr.s32 	%r2529, %r2528, 16;
	cvt.u32.u64	%r2530, %rd2870;
	shr.s32 	%r2531, %r2530, 16;
	mad.lo.s32 	%r2532, %r2529, %r2531, 16384;
	and.b64  	%rd2888, %rd2887, 65535;
	and.b64  	%rd2889, %rd2873, 4294901760;
	or.b64  	%rd2890, %rd2889, %rd2874;
	or.b64  	%rd2891, %rd2890, %rd2884;
	shl.b32 	%r2533, %r2532, 1;
	cvt.u64.u32	%rd2892, %r2533;
	cvt.u32.u64	%r2534, %rd2875;
	shr.s32 	%r2535, %r2534, 16;
	cvt.u32.u64	%r2536, %rd2877;
	shr.s32 	%r2537, %r2536, 16;
	mad.lo.s32 	%r2538, %r2535, %r2537, 16384;
	mul.wide.u32 	%rd2893, %r2538, 131072;
	and.b64  	%rd2894, %rd2892, 4294901760;
	or.b64  	%rd2895, %rd2894, %rd2888;
	shr.u64 	%rd2896, %rd2868, 32;
	cvt.u32.u64	%r2539, %rd2896;
	shr.s32 	%r2540, %r2539, 16;
	shr.u64 	%rd2897, %rd2870, 32;
	cvt.u32.u64	%r2541, %rd2897;
	shr.s32 	%r2542, %r2541, 16;
	mad.lo.s32 	%r2543, %r2540, %r2542, 16384;
	shr.u32 	%r2544, %r2543, 15;
	cvt.u64.u32	%rd2898, %r2544;
	shl.b64 	%rd2899, %rd2898, 48;
	and.b64  	%rd2900, %rd2893, 281470681743360;
	or.b64  	%rd2901, %rd2900, %rd2899;
	or.b64  	%rd2902, %rd2901, %rd2895;
	or.b64  	%rd2903, %rd2891, %rd2883;
	mov.b64	{%r2545, %r2546}, %rd2902;
	mov.b64	{%r2547, %r2548}, %rd2903;
	xor.b32  	%r6504, %r2546, %r6504;
	xor.b32  	%r6503, %r2545, %r6503;
	xor.b32  	%r6502, %r2548, %r6502;
	xor.b32  	%r6501, %r2547, %r6501;

BB0_82:
	cvt.u64.u32	%rd2904, %r6729;
	add.s64 	%rd2905, %rd2904, %rd9;
	shl.b64 	%rd2906, %rd2905, 1;
	mov.u64 	%rd2907, verus_gpu_hash$sharedrand;
	add.s64 	%rd2908, %rd2907, %rd2906;
	st.shared.u16 	[%rd2908], %rs236;
	mov.u64 	%rd2909, verus_gpu_hash$sharedrandex;
	add.s64 	%rd2910, %rd2909, %rd2906;
	st.shared.u16 	[%rd2910], %rs237;
	cvt.u64.u16	%rd2911, %rs236;
	add.s64 	%rd2912, %rd2911, %rd3;
	shl.b64 	%rd2913, %rd2912, 4;
	add.s64 	%rd2914, %rd203, %rd2913;
	st.global.v4.u32 	[%rd2914], {%r6501, %r6502, %r6503, %r6504};
	cvt.u64.u16	%rd2915, %rs237;
	add.s64 	%rd2916, %rd2915, %rd3;
	shl.b64 	%rd2917, %rd2916, 4;
	add.s64 	%rd2918, %rd203, %rd2917;
	st.global.v4.u32 	[%rd2918], {%r6625, %r6626, %r6627, %r6628};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p46, %r6729, 32;
	@%p46 bra 	BB0_134;

	mov.u32 	%r6461, 0;
	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p47, %r6729, 0;
	mov.u16 	%rs238, 12;
	@%p47 bra 	BB0_84;
	bra.uni 	BB0_85;

BB0_84:
	mov.u32 	%r6721, %r6501;
	mov.u32 	%r6722, %r6502;
	mov.u32 	%r6723, %r6503;
	mov.u32 	%r6724, %r6504;
	mov.u32 	%r6729, %r6461;
	bra.uni 	BB0_86;

BB0_85:
	cvt.u64.u32	%rd2919, %r6731;
	shr.u32 	%r2550, %r6730, 5;
	cvt.u64.u32	%rd2920, %r2550;
	cvt.u16.u32	%rs86, %r2550;
	and.b16  	%rs236, %rs86, 511;
	cvt.u16.u32	%rs87, %r6731;
	and.b16  	%rs237, %rs87, 511;
	cvt.u16.u64	%rs88, %rd6123;
	and.b16  	%rs238, %rs88, 28;
	and.b64  	%rd2921, %rd2920, 511;
	add.s64 	%rd2922, %rd2921, %rd3;
	shl.b64 	%rd2923, %rd2922, 4;
	add.s64 	%rd2924, %rd203, %rd2923;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd2924];
	and.b64  	%rd2925, %rd2919, 511;
	add.s64 	%rd2926, %rd2925, %rd3;
	shl.b64 	%rd2927, %rd2926, 4;
	add.s64 	%rd2928, %rd203, %rd2927;
	ld.global.v4.u32 	{%r6625, %r6626, %r6627, %r6628}, [%rd2928];
	mul.wide.u32 	%rd2929, %r6730, 16;
	and.b64  	%rd2930, %rd2929, 48;
	add.s64 	%rd6122, %rd1, %rd2930;

BB0_86:
	and.b16  	%rs89, %rs238, 255;
	setp.ne.s16	%p48, %rs89, 16;
	@%p48 bra 	BB0_87;

	cvt.u64.u16	%rd2931, %rs236;
	add.s64 	%rd2932, %rd2931, %rd3;
	shl.b64 	%rd2933, %rd2932, 4;
	add.s64 	%rd2934, %rd203, %rd2933;
	and.b64  	%rd2935, %rd6123, 1;
	mov.u64 	%rd2936, 1;
	shl.b64 	%rd2937, %rd2935, 1;
	sub.s64 	%rd2938, %rd2936, %rd2937;
	shl.b64 	%rd2939, %rd2938, 4;
	add.s64 	%rd2940, %rd6122, %rd2939;
	ld.local.v4.u32 	{%r2559, %r2560, %r2561, %r2562}, [%rd2940];
	and.b32  	%r2564, %r2559, 255;
	mul.wide.u32 	%rd2941, %r2564, 4;
	mov.u64 	%rd2942, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd2943, %rd2942, %rd2941;
	and.b32  	%r2566, %r2560, 255;
	mul.wide.u32 	%rd2944, %r2566, 4;
	add.s64 	%rd2945, %rd2942, %rd2944;
	and.b32  	%r2568, %r2561, 255;
	mul.wide.u32 	%rd2946, %r2568, 4;
	add.s64 	%rd2947, %rd2942, %rd2946;
	and.b32  	%r2570, %r2562, 255;
	mul.wide.u32 	%rd2948, %r2570, 4;
	add.s64 	%rd2949, %rd2942, %rd2948;
	bfe.u32 	%r2571, %r2560, 8, 8;
	mul.wide.u32 	%rd2950, %r2571, 4;
	add.s64 	%rd2951, %rd2950, %rd2942;
	ld.shared.u32 	%r2572, [%rd2951+1024];
	ld.shared.u32 	%r2573, [%rd2943];
	xor.b32  	%r2574, %r2572, %r2573;
	bfe.u32 	%r2575, %r2561, 8, 8;
	mul.wide.u32 	%rd2952, %r2575, 4;
	add.s64 	%rd2953, %rd2952, %rd2942;
	ld.shared.u32 	%r2576, [%rd2953+1024];
	ld.shared.u32 	%r2577, [%rd2945];
	xor.b32  	%r2578, %r2576, %r2577;
	bfe.u32 	%r2579, %r2562, 8, 8;
	mul.wide.u32 	%rd2954, %r2579, 4;
	add.s64 	%rd2955, %rd2954, %rd2942;
	ld.shared.u32 	%r2580, [%rd2955+1024];
	ld.shared.u32 	%r2581, [%rd2947];
	xor.b32  	%r2582, %r2580, %r2581;
	bfe.u32 	%r2583, %r2559, 8, 8;
	mul.wide.u32 	%rd2956, %r2583, 4;
	add.s64 	%rd2957, %rd2956, %rd2942;
	ld.shared.u32 	%r2584, [%rd2957+1024];
	ld.shared.u32 	%r2585, [%rd2949];
	xor.b32  	%r2586, %r2584, %r2585;
	bfe.u32 	%r2587, %r2561, 16, 8;
	mul.wide.u32 	%rd2958, %r2587, 4;
	add.s64 	%rd2959, %rd2958, %rd2942;
	ld.shared.u32 	%r2588, [%rd2959+2048];
	xor.b32  	%r2589, %r2574, %r2588;
	shr.u32 	%r2590, %r2561, 24;
	bfe.u32 	%r2591, %r2562, 16, 8;
	mul.wide.u32 	%rd2960, %r2591, 4;
	add.s64 	%rd2961, %rd2960, %rd2942;
	ld.shared.u32 	%r2592, [%rd2961+2048];
	xor.b32  	%r2593, %r2578, %r2592;
	shr.u32 	%r2594, %r2562, 24;
	bfe.u32 	%r2595, %r2559, 16, 8;
	mul.wide.u32 	%rd2962, %r2595, 4;
	add.s64 	%rd2963, %rd2962, %rd2942;
	ld.shared.u32 	%r2596, [%rd2963+2048];
	xor.b32  	%r2597, %r2582, %r2596;
	shr.u32 	%r2598, %r2559, 24;
	bfe.u32 	%r2599, %r2560, 16, 8;
	mul.wide.u32 	%rd2964, %r2599, 4;
	add.s64 	%rd2965, %rd2964, %rd2942;
	ld.shared.u32 	%r2600, [%rd2965+2048];
	xor.b32  	%r2601, %r2586, %r2600;
	shr.u32 	%r2602, %r2560, 24;
	mul.wide.u32 	%rd2966, %r2594, 4;
	add.s64 	%rd2967, %rd2966, %rd2942;
	ld.shared.u32 	%r2603, [%rd2967+3072];
	mul.wide.u32 	%rd2968, %r2598, 4;
	add.s64 	%rd2969, %rd2968, %rd2942;
	ld.shared.u32 	%r2604, [%rd2969+3072];
	mul.wide.u32 	%rd2970, %r2602, 4;
	add.s64 	%rd2971, %rd2970, %rd2942;
	ld.shared.u32 	%r2605, [%rd2971+3072];
	mul.wide.u32 	%rd2972, %r2590, 4;
	add.s64 	%rd2973, %rd2972, %rd2942;
	ld.shared.u32 	%r2606, [%rd2973+3072];
	xor.b32  	%r2607, %r2601, %r2606;
	xor.b32  	%r2608, %r2597, %r2605;
	xor.b32  	%r2609, %r2593, %r2604;
	xor.b32  	%r2610, %r2589, %r2603;
	ld.global.v4.u32 	{%r2611, %r2612, %r2613, %r2614}, [%rd2934];
	ld.local.v4.u32 	{%r2619, %r2620, %r2621, %r2622}, [%rd6122];
	and.b32  	%r2624, %r2619, 255;
	mul.wide.u32 	%rd2974, %r2624, 4;
	add.s64 	%rd2975, %rd2942, %rd2974;
	and.b32  	%r2626, %r2620, 255;
	mul.wide.u32 	%rd2976, %r2626, 4;
	add.s64 	%rd2977, %rd2942, %rd2976;
	and.b32  	%r2628, %r2621, 255;
	mul.wide.u32 	%rd2978, %r2628, 4;
	add.s64 	%rd2979, %rd2942, %rd2978;
	and.b32  	%r2630, %r2622, 255;
	mul.wide.u32 	%rd2980, %r2630, 4;
	add.s64 	%rd2981, %rd2942, %rd2980;
	bfe.u32 	%r2631, %r2620, 8, 8;
	mul.wide.u32 	%rd2982, %r2631, 4;
	add.s64 	%rd2983, %rd2982, %rd2942;
	ld.shared.u32 	%r2632, [%rd2983+1024];
	ld.shared.u32 	%r2633, [%rd2975];
	xor.b32  	%r2634, %r2632, %r2633;
	bfe.u32 	%r2635, %r2621, 8, 8;
	mul.wide.u32 	%rd2984, %r2635, 4;
	add.s64 	%rd2985, %rd2984, %rd2942;
	ld.shared.u32 	%r2636, [%rd2985+1024];
	ld.shared.u32 	%r2637, [%rd2977];
	xor.b32  	%r2638, %r2636, %r2637;
	bfe.u32 	%r2639, %r2622, 8, 8;
	mul.wide.u32 	%rd2986, %r2639, 4;
	add.s64 	%rd2987, %rd2986, %rd2942;
	ld.shared.u32 	%r2640, [%rd2987+1024];
	ld.shared.u32 	%r2641, [%rd2979];
	xor.b32  	%r2642, %r2640, %r2641;
	bfe.u32 	%r2643, %r2619, 8, 8;
	mul.wide.u32 	%rd2988, %r2643, 4;
	add.s64 	%rd2989, %rd2988, %rd2942;
	ld.shared.u32 	%r2644, [%rd2989+1024];
	ld.shared.u32 	%r2645, [%rd2981];
	xor.b32  	%r2646, %r2644, %r2645;
	bfe.u32 	%r2647, %r2621, 16, 8;
	mul.wide.u32 	%rd2990, %r2647, 4;
	add.s64 	%rd2991, %rd2990, %rd2942;
	ld.shared.u32 	%r2648, [%rd2991+2048];
	xor.b32  	%r2649, %r2634, %r2648;
	shr.u32 	%r2650, %r2621, 24;
	bfe.u32 	%r2651, %r2622, 16, 8;
	mul.wide.u32 	%rd2992, %r2651, 4;
	add.s64 	%rd2993, %rd2992, %rd2942;
	ld.shared.u32 	%r2652, [%rd2993+2048];
	xor.b32  	%r2653, %r2638, %r2652;
	shr.u32 	%r2654, %r2622, 24;
	bfe.u32 	%r2655, %r2619, 16, 8;
	mul.wide.u32 	%rd2994, %r2655, 4;
	add.s64 	%rd2995, %rd2994, %rd2942;
	ld.shared.u32 	%r2656, [%rd2995+2048];
	xor.b32  	%r2657, %r2642, %r2656;
	shr.u32 	%r2658, %r2619, 24;
	bfe.u32 	%r2659, %r2620, 16, 8;
	mul.wide.u32 	%rd2996, %r2659, 4;
	add.s64 	%rd2997, %rd2996, %rd2942;
	ld.shared.u32 	%r2660, [%rd2997+2048];
	xor.b32  	%r2661, %r2646, %r2660;
	shr.u32 	%r2662, %r2620, 24;
	mul.wide.u32 	%rd2998, %r2654, 4;
	add.s64 	%rd2999, %rd2998, %rd2942;
	ld.shared.u32 	%r2663, [%rd2999+3072];
	mul.wide.u32 	%rd3000, %r2658, 4;
	add.s64 	%rd3001, %rd3000, %rd2942;
	ld.shared.u32 	%r2664, [%rd3001+3072];
	mul.wide.u32 	%rd3002, %r2662, 4;
	add.s64 	%rd3003, %rd3002, %rd2942;
	ld.shared.u32 	%r2665, [%rd3003+3072];
	mul.wide.u32 	%rd3004, %r2650, 4;
	add.s64 	%rd3005, %rd3004, %rd2942;
	ld.shared.u32 	%r2666, [%rd3005+3072];
	xor.b32  	%r2667, %r2661, %r2666;
	xor.b32  	%r2668, %r2657, %r2665;
	xor.b32  	%r2669, %r2653, %r2664;
	xor.b32  	%r2670, %r2649, %r2663;
	ld.global.v4.u32 	{%r2671, %r2672, %r2673, %r2674}, [%rd2934+16];
	xor.b32  	%r2679, %r2610, %r2611;
	and.b32  	%r2680, %r2679, 255;
	mul.wide.u32 	%rd3006, %r2680, 4;
	add.s64 	%rd3007, %rd2942, %rd3006;
	xor.b32  	%r2681, %r2609, %r2612;
	and.b32  	%r2682, %r2681, 255;
	mul.wide.u32 	%rd3008, %r2682, 4;
	add.s64 	%rd3009, %rd2942, %rd3008;
	xor.b32  	%r2683, %r2608, %r2613;
	and.b32  	%r2684, %r2683, 255;
	mul.wide.u32 	%rd3010, %r2684, 4;
	add.s64 	%rd3011, %rd2942, %rd3010;
	xor.b32  	%r2685, %r2607, %r2614;
	and.b32  	%r2686, %r2685, 255;
	mul.wide.u32 	%rd3012, %r2686, 4;
	add.s64 	%rd3013, %rd2942, %rd3012;
	bfe.u32 	%r2687, %r2681, 8, 8;
	mul.wide.u32 	%rd3014, %r2687, 4;
	add.s64 	%rd3015, %rd3014, %rd2942;
	ld.shared.u32 	%r2688, [%rd3015+1024];
	ld.shared.u32 	%r2689, [%rd3007];
	xor.b32  	%r2690, %r2688, %r2689;
	bfe.u32 	%r2691, %r2683, 8, 8;
	mul.wide.u32 	%rd3016, %r2691, 4;
	add.s64 	%rd3017, %rd3016, %rd2942;
	ld.shared.u32 	%r2692, [%rd3017+1024];
	ld.shared.u32 	%r2693, [%rd3009];
	xor.b32  	%r2694, %r2692, %r2693;
	bfe.u32 	%r2695, %r2685, 8, 8;
	mul.wide.u32 	%rd3018, %r2695, 4;
	add.s64 	%rd3019, %rd3018, %rd2942;
	ld.shared.u32 	%r2696, [%rd3019+1024];
	ld.shared.u32 	%r2697, [%rd3011];
	xor.b32  	%r2698, %r2696, %r2697;
	bfe.u32 	%r2699, %r2679, 8, 8;
	mul.wide.u32 	%rd3020, %r2699, 4;
	add.s64 	%rd3021, %rd3020, %rd2942;
	ld.shared.u32 	%r2700, [%rd3021+1024];
	ld.shared.u32 	%r2701, [%rd3013];
	xor.b32  	%r2702, %r2700, %r2701;
	bfe.u32 	%r2703, %r2683, 16, 8;
	mul.wide.u32 	%rd3022, %r2703, 4;
	add.s64 	%rd3023, %rd3022, %rd2942;
	ld.shared.u32 	%r2704, [%rd3023+2048];
	xor.b32  	%r2705, %r2690, %r2704;
	shr.u32 	%r2706, %r2683, 24;
	bfe.u32 	%r2707, %r2685, 16, 8;
	mul.wide.u32 	%rd3024, %r2707, 4;
	add.s64 	%rd3025, %rd3024, %rd2942;
	ld.shared.u32 	%r2708, [%rd3025+2048];
	xor.b32  	%r2709, %r2694, %r2708;
	shr.u32 	%r2710, %r2685, 24;
	bfe.u32 	%r2711, %r2679, 16, 8;
	mul.wide.u32 	%rd3026, %r2711, 4;
	add.s64 	%rd3027, %rd3026, %rd2942;
	ld.shared.u32 	%r2712, [%rd3027+2048];
	xor.b32  	%r2713, %r2698, %r2712;
	shr.u32 	%r2714, %r2679, 24;
	bfe.u32 	%r2715, %r2681, 16, 8;
	mul.wide.u32 	%rd3028, %r2715, 4;
	add.s64 	%rd3029, %rd3028, %rd2942;
	ld.shared.u32 	%r2716, [%rd3029+2048];
	xor.b32  	%r2717, %r2702, %r2716;
	shr.u32 	%r2718, %r2681, 24;
	mul.wide.u32 	%rd3030, %r2710, 4;
	add.s64 	%rd3031, %rd3030, %rd2942;
	ld.shared.u32 	%r2719, [%rd3031+3072];
	mul.wide.u32 	%rd3032, %r2714, 4;
	add.s64 	%rd3033, %rd3032, %rd2942;
	ld.shared.u32 	%r2720, [%rd3033+3072];
	mul.wide.u32 	%rd3034, %r2718, 4;
	add.s64 	%rd3035, %rd3034, %rd2942;
	ld.shared.u32 	%r2721, [%rd3035+3072];
	mul.wide.u32 	%rd3036, %r2706, 4;
	add.s64 	%rd3037, %rd3036, %rd2942;
	ld.shared.u32 	%r2722, [%rd3037+3072];
	xor.b32  	%r2723, %r2717, %r2722;
	xor.b32  	%r2724, %r2713, %r2721;
	xor.b32  	%r2725, %r2709, %r2720;
	xor.b32  	%r2726, %r2705, %r2719;
	ld.global.v4.u32 	{%r2727, %r2728, %r2729, %r2730}, [%rd2934+32];
	xor.b32  	%r2735, %r2670, %r2671;
	and.b32  	%r2736, %r2735, 255;
	mul.wide.u32 	%rd3038, %r2736, 4;
	add.s64 	%rd3039, %rd2942, %rd3038;
	xor.b32  	%r2737, %r2669, %r2672;
	and.b32  	%r2738, %r2737, 255;
	mul.wide.u32 	%rd3040, %r2738, 4;
	add.s64 	%rd3041, %rd2942, %rd3040;
	xor.b32  	%r2739, %r2668, %r2673;
	and.b32  	%r2740, %r2739, 255;
	mul.wide.u32 	%rd3042, %r2740, 4;
	add.s64 	%rd3043, %rd2942, %rd3042;
	xor.b32  	%r2741, %r2667, %r2674;
	and.b32  	%r2742, %r2741, 255;
	mul.wide.u32 	%rd3044, %r2742, 4;
	add.s64 	%rd3045, %rd2942, %rd3044;
	bfe.u32 	%r2743, %r2737, 8, 8;
	mul.wide.u32 	%rd3046, %r2743, 4;
	add.s64 	%rd3047, %rd3046, %rd2942;
	ld.shared.u32 	%r2744, [%rd3047+1024];
	ld.shared.u32 	%r2745, [%rd3039];
	xor.b32  	%r2746, %r2744, %r2745;
	bfe.u32 	%r2747, %r2739, 8, 8;
	mul.wide.u32 	%rd3048, %r2747, 4;
	add.s64 	%rd3049, %rd3048, %rd2942;
	ld.shared.u32 	%r2748, [%rd3049+1024];
	ld.shared.u32 	%r2749, [%rd3041];
	xor.b32  	%r2750, %r2748, %r2749;
	bfe.u32 	%r2751, %r2741, 8, 8;
	mul.wide.u32 	%rd3050, %r2751, 4;
	add.s64 	%rd3051, %rd3050, %rd2942;
	ld.shared.u32 	%r2752, [%rd3051+1024];
	ld.shared.u32 	%r2753, [%rd3043];
	xor.b32  	%r2754, %r2752, %r2753;
	bfe.u32 	%r2755, %r2735, 8, 8;
	mul.wide.u32 	%rd3052, %r2755, 4;
	add.s64 	%rd3053, %rd3052, %rd2942;
	ld.shared.u32 	%r2756, [%rd3053+1024];
	ld.shared.u32 	%r2757, [%rd3045];
	xor.b32  	%r2758, %r2756, %r2757;
	bfe.u32 	%r2759, %r2739, 16, 8;
	mul.wide.u32 	%rd3054, %r2759, 4;
	add.s64 	%rd3055, %rd3054, %rd2942;
	ld.shared.u32 	%r2760, [%rd3055+2048];
	xor.b32  	%r2761, %r2746, %r2760;
	shr.u32 	%r2762, %r2739, 24;
	bfe.u32 	%r2763, %r2741, 16, 8;
	mul.wide.u32 	%rd3056, %r2763, 4;
	add.s64 	%rd3057, %rd3056, %rd2942;
	ld.shared.u32 	%r2764, [%rd3057+2048];
	xor.b32  	%r2765, %r2750, %r2764;
	shr.u32 	%r2766, %r2741, 24;
	bfe.u32 	%r2767, %r2735, 16, 8;
	mul.wide.u32 	%rd3058, %r2767, 4;
	add.s64 	%rd3059, %rd3058, %rd2942;
	ld.shared.u32 	%r2768, [%rd3059+2048];
	xor.b32  	%r2769, %r2754, %r2768;
	shr.u32 	%r2770, %r2735, 24;
	bfe.u32 	%r2771, %r2737, 16, 8;
	mul.wide.u32 	%rd3060, %r2771, 4;
	add.s64 	%rd3061, %rd3060, %rd2942;
	ld.shared.u32 	%r2772, [%rd3061+2048];
	xor.b32  	%r2773, %r2758, %r2772;
	shr.u32 	%r2774, %r2737, 24;
	mul.wide.u32 	%rd3062, %r2766, 4;
	add.s64 	%rd3063, %rd3062, %rd2942;
	ld.shared.u32 	%r2775, [%rd3063+3072];
	mul.wide.u32 	%rd3064, %r2770, 4;
	add.s64 	%rd3065, %rd3064, %rd2942;
	ld.shared.u32 	%r2776, [%rd3065+3072];
	mul.wide.u32 	%rd3066, %r2774, 4;
	add.s64 	%rd3067, %rd3066, %rd2942;
	ld.shared.u32 	%r2777, [%rd3067+3072];
	mul.wide.u32 	%rd3068, %r2762, 4;
	add.s64 	%rd3069, %rd3068, %rd2942;
	ld.shared.u32 	%r2778, [%rd3069+3072];
	xor.b32  	%r2779, %r2773, %r2778;
	xor.b32  	%r2780, %r2769, %r2777;
	xor.b32  	%r2781, %r2765, %r2776;
	xor.b32  	%r2782, %r2761, %r2775;
	ld.global.v4.u32 	{%r2783, %r2784, %r2785, %r2786}, [%rd2934+48];
	xor.b32  	%r2791, %r2726, %r2727;
	and.b32  	%r2792, %r2791, 255;
	mul.wide.u32 	%rd3070, %r2792, 4;
	add.s64 	%rd3071, %rd2942, %rd3070;
	xor.b32  	%r2793, %r2782, %r2783;
	and.b32  	%r2794, %r2793, 255;
	mul.wide.u32 	%rd3072, %r2794, 4;
	add.s64 	%rd3073, %rd2942, %rd3072;
	xor.b32  	%r2795, %r2725, %r2728;
	and.b32  	%r2796, %r2795, 255;
	mul.wide.u32 	%rd3074, %r2796, 4;
	add.s64 	%rd3075, %rd2942, %rd3074;
	xor.b32  	%r2797, %r2781, %r2784;
	and.b32  	%r2798, %r2797, 255;
	mul.wide.u32 	%rd3076, %r2798, 4;
	add.s64 	%rd3077, %rd2942, %rd3076;
	bfe.u32 	%r2799, %r2793, 8, 8;
	mul.wide.u32 	%rd3078, %r2799, 4;
	add.s64 	%rd3079, %rd3078, %rd2942;
	ld.shared.u32 	%r2800, [%rd3079+1024];
	ld.shared.u32 	%r2801, [%rd3071];
	xor.b32  	%r2802, %r2800, %r2801;
	bfe.u32 	%r2803, %r2795, 8, 8;
	mul.wide.u32 	%rd3080, %r2803, 4;
	add.s64 	%rd3081, %rd3080, %rd2942;
	ld.shared.u32 	%r2804, [%rd3081+1024];
	ld.shared.u32 	%r2805, [%rd3073];
	xor.b32  	%r2806, %r2804, %r2805;
	bfe.u32 	%r2807, %r2797, 8, 8;
	mul.wide.u32 	%rd3082, %r2807, 4;
	add.s64 	%rd3083, %rd3082, %rd2942;
	ld.shared.u32 	%r2808, [%rd3083+1024];
	ld.shared.u32 	%r2809, [%rd3075];
	xor.b32  	%r2810, %r2808, %r2809;
	bfe.u32 	%r2811, %r2791, 8, 8;
	mul.wide.u32 	%rd3084, %r2811, 4;
	add.s64 	%rd3085, %rd3084, %rd2942;
	ld.shared.u32 	%r2812, [%rd3085+1024];
	ld.shared.u32 	%r2813, [%rd3077];
	xor.b32  	%r2814, %r2812, %r2813;
	bfe.u32 	%r2815, %r2795, 16, 8;
	mul.wide.u32 	%rd3086, %r2815, 4;
	add.s64 	%rd3087, %rd3086, %rd2942;
	ld.shared.u32 	%r2816, [%rd3087+2048];
	xor.b32  	%r2817, %r2802, %r2816;
	shr.u32 	%r2818, %r2795, 24;
	bfe.u32 	%r2819, %r2797, 16, 8;
	mul.wide.u32 	%rd3088, %r2819, 4;
	add.s64 	%rd3089, %rd3088, %rd2942;
	ld.shared.u32 	%r2820, [%rd3089+2048];
	xor.b32  	%r2821, %r2806, %r2820;
	shr.u32 	%r2822, %r2797, 24;
	bfe.u32 	%r2823, %r2791, 16, 8;
	mul.wide.u32 	%rd3090, %r2823, 4;
	add.s64 	%rd3091, %rd3090, %rd2942;
	ld.shared.u32 	%r2824, [%rd3091+2048];
	xor.b32  	%r2825, %r2810, %r2824;
	shr.u32 	%r2826, %r2791, 24;
	bfe.u32 	%r2827, %r2793, 16, 8;
	mul.wide.u32 	%rd3092, %r2827, 4;
	add.s64 	%rd3093, %rd3092, %rd2942;
	ld.shared.u32 	%r2828, [%rd3093+2048];
	xor.b32  	%r2829, %r2814, %r2828;
	shr.u32 	%r2830, %r2793, 24;
	mul.wide.u32 	%rd3094, %r2822, 4;
	add.s64 	%rd3095, %rd3094, %rd2942;
	ld.shared.u32 	%r2831, [%rd3095+3072];
	mul.wide.u32 	%rd3096, %r2826, 4;
	add.s64 	%rd3097, %rd3096, %rd2942;
	ld.shared.u32 	%r2832, [%rd3097+3072];
	mul.wide.u32 	%rd3098, %r2830, 4;
	add.s64 	%rd3099, %rd3098, %rd2942;
	ld.shared.u32 	%r2833, [%rd3099+3072];
	mul.wide.u32 	%rd3100, %r2818, 4;
	add.s64 	%rd3101, %rd3100, %rd2942;
	ld.shared.u32 	%r2834, [%rd3101+3072];
	xor.b32  	%r2835, %r2829, %r2834;
	xor.b32  	%r2836, %r2825, %r2833;
	xor.b32  	%r2837, %r2821, %r2832;
	xor.b32  	%r2838, %r2817, %r2831;
	ld.global.v4.u32 	{%r2839, %r2840, %r2841, %r2842}, [%rd2934+64];
	xor.b32  	%r2847, %r2724, %r2729;
	and.b32  	%r2848, %r2847, 255;
	mul.wide.u32 	%rd3102, %r2848, 4;
	add.s64 	%rd3103, %rd2942, %rd3102;
	xor.b32  	%r2849, %r2780, %r2785;
	and.b32  	%r2850, %r2849, 255;
	mul.wide.u32 	%rd3104, %r2850, 4;
	add.s64 	%rd3105, %rd2942, %rd3104;
	xor.b32  	%r2851, %r2723, %r2730;
	and.b32  	%r2852, %r2851, 255;
	mul.wide.u32 	%rd3106, %r2852, 4;
	add.s64 	%rd3107, %rd2942, %rd3106;
	xor.b32  	%r2853, %r2779, %r2786;
	and.b32  	%r2854, %r2853, 255;
	mul.wide.u32 	%rd3108, %r2854, 4;
	add.s64 	%rd3109, %rd2942, %rd3108;
	bfe.u32 	%r2855, %r2849, 8, 8;
	mul.wide.u32 	%rd3110, %r2855, 4;
	add.s64 	%rd3111, %rd3110, %rd2942;
	ld.shared.u32 	%r2856, [%rd3111+1024];
	ld.shared.u32 	%r2857, [%rd3103];
	xor.b32  	%r2858, %r2856, %r2857;
	bfe.u32 	%r2859, %r2851, 8, 8;
	mul.wide.u32 	%rd3112, %r2859, 4;
	add.s64 	%rd3113, %rd3112, %rd2942;
	ld.shared.u32 	%r2860, [%rd3113+1024];
	ld.shared.u32 	%r2861, [%rd3105];
	xor.b32  	%r2862, %r2860, %r2861;
	bfe.u32 	%r2863, %r2853, 8, 8;
	mul.wide.u32 	%rd3114, %r2863, 4;
	add.s64 	%rd3115, %rd3114, %rd2942;
	ld.shared.u32 	%r2864, [%rd3115+1024];
	ld.shared.u32 	%r2865, [%rd3107];
	xor.b32  	%r2866, %r2864, %r2865;
	bfe.u32 	%r2867, %r2847, 8, 8;
	mul.wide.u32 	%rd3116, %r2867, 4;
	add.s64 	%rd3117, %rd3116, %rd2942;
	ld.shared.u32 	%r2868, [%rd3117+1024];
	ld.shared.u32 	%r2869, [%rd3109];
	xor.b32  	%r2870, %r2868, %r2869;
	bfe.u32 	%r2871, %r2851, 16, 8;
	mul.wide.u32 	%rd3118, %r2871, 4;
	add.s64 	%rd3119, %rd3118, %rd2942;
	ld.shared.u32 	%r2872, [%rd3119+2048];
	xor.b32  	%r2873, %r2858, %r2872;
	shr.u32 	%r2874, %r2851, 24;
	bfe.u32 	%r2875, %r2853, 16, 8;
	mul.wide.u32 	%rd3120, %r2875, 4;
	add.s64 	%rd3121, %rd3120, %rd2942;
	ld.shared.u32 	%r2876, [%rd3121+2048];
	xor.b32  	%r2877, %r2862, %r2876;
	shr.u32 	%r2878, %r2853, 24;
	bfe.u32 	%r2879, %r2847, 16, 8;
	mul.wide.u32 	%rd3122, %r2879, 4;
	add.s64 	%rd3123, %rd3122, %rd2942;
	ld.shared.u32 	%r2880, [%rd3123+2048];
	xor.b32  	%r2881, %r2866, %r2880;
	shr.u32 	%r2882, %r2847, 24;
	bfe.u32 	%r2883, %r2849, 16, 8;
	mul.wide.u32 	%rd3124, %r2883, 4;
	add.s64 	%rd3125, %rd3124, %rd2942;
	ld.shared.u32 	%r2884, [%rd3125+2048];
	xor.b32  	%r2885, %r2870, %r2884;
	shr.u32 	%r2886, %r2849, 24;
	mul.wide.u32 	%rd3126, %r2878, 4;
	add.s64 	%rd3127, %rd3126, %rd2942;
	ld.shared.u32 	%r2887, [%rd3127+3072];
	mul.wide.u32 	%rd3128, %r2882, 4;
	add.s64 	%rd3129, %rd3128, %rd2942;
	ld.shared.u32 	%r2888, [%rd3129+3072];
	mul.wide.u32 	%rd3130, %r2886, 4;
	add.s64 	%rd3131, %rd3130, %rd2942;
	ld.shared.u32 	%r2889, [%rd3131+3072];
	mul.wide.u32 	%rd3132, %r2874, 4;
	add.s64 	%rd3133, %rd3132, %rd2942;
	ld.shared.u32 	%r2890, [%rd3133+3072];
	xor.b32  	%r2891, %r2885, %r2890;
	xor.b32  	%r2892, %r2881, %r2889;
	xor.b32  	%r2893, %r2877, %r2888;
	xor.b32  	%r2894, %r2873, %r2887;
	ld.global.v4.u32 	{%r2895, %r2896, %r2897, %r2898}, [%rd2934+80];
	xor.b32  	%r2903, %r2838, %r2839;
	and.b32  	%r2904, %r2903, 255;
	mul.wide.u32 	%rd3134, %r2904, 4;
	add.s64 	%rd3135, %rd2942, %rd3134;
	xor.b32  	%r2905, %r2837, %r2840;
	and.b32  	%r2906, %r2905, 255;
	mul.wide.u32 	%rd3136, %r2906, 4;
	add.s64 	%rd3137, %rd2942, %rd3136;
	xor.b32  	%r2907, %r2836, %r2841;
	and.b32  	%r2908, %r2907, 255;
	mul.wide.u32 	%rd3138, %r2908, 4;
	add.s64 	%rd3139, %rd2942, %rd3138;
	xor.b32  	%r2909, %r2835, %r2842;
	and.b32  	%r2910, %r2909, 255;
	mul.wide.u32 	%rd3140, %r2910, 4;
	add.s64 	%rd3141, %rd2942, %rd3140;
	bfe.u32 	%r2911, %r2905, 8, 8;
	mul.wide.u32 	%rd3142, %r2911, 4;
	add.s64 	%rd3143, %rd3142, %rd2942;
	ld.shared.u32 	%r2912, [%rd3143+1024];
	ld.shared.u32 	%r2913, [%rd3135];
	xor.b32  	%r2914, %r2912, %r2913;
	bfe.u32 	%r2915, %r2907, 8, 8;
	mul.wide.u32 	%rd3144, %r2915, 4;
	add.s64 	%rd3145, %rd3144, %rd2942;
	ld.shared.u32 	%r2916, [%rd3145+1024];
	ld.shared.u32 	%r2917, [%rd3137];
	xor.b32  	%r2918, %r2916, %r2917;
	bfe.u32 	%r2919, %r2909, 8, 8;
	mul.wide.u32 	%rd3146, %r2919, 4;
	add.s64 	%rd3147, %rd3146, %rd2942;
	ld.shared.u32 	%r2920, [%rd3147+1024];
	ld.shared.u32 	%r2921, [%rd3139];
	xor.b32  	%r2922, %r2920, %r2921;
	bfe.u32 	%r2923, %r2903, 8, 8;
	mul.wide.u32 	%rd3148, %r2923, 4;
	add.s64 	%rd3149, %rd3148, %rd2942;
	ld.shared.u32 	%r2924, [%rd3149+1024];
	ld.shared.u32 	%r2925, [%rd3141];
	xor.b32  	%r2926, %r2924, %r2925;
	bfe.u32 	%r2927, %r2907, 16, 8;
	mul.wide.u32 	%rd3150, %r2927, 4;
	add.s64 	%rd3151, %rd3150, %rd2942;
	ld.shared.u32 	%r2928, [%rd3151+2048];
	xor.b32  	%r2929, %r2914, %r2928;
	shr.u32 	%r2930, %r2907, 24;
	bfe.u32 	%r2931, %r2909, 16, 8;
	mul.wide.u32 	%rd3152, %r2931, 4;
	add.s64 	%rd3153, %rd3152, %rd2942;
	ld.shared.u32 	%r2932, [%rd3153+2048];
	xor.b32  	%r2933, %r2918, %r2932;
	shr.u32 	%r2934, %r2909, 24;
	bfe.u32 	%r2935, %r2903, 16, 8;
	mul.wide.u32 	%rd3154, %r2935, 4;
	add.s64 	%rd3155, %rd3154, %rd2942;
	ld.shared.u32 	%r2936, [%rd3155+2048];
	xor.b32  	%r2937, %r2922, %r2936;
	shr.u32 	%r2938, %r2903, 24;
	bfe.u32 	%r2939, %r2905, 16, 8;
	mul.wide.u32 	%rd3156, %r2939, 4;
	add.s64 	%rd3157, %rd3156, %rd2942;
	ld.shared.u32 	%r2940, [%rd3157+2048];
	xor.b32  	%r2941, %r2926, %r2940;
	shr.u32 	%r2942, %r2905, 24;
	mul.wide.u32 	%rd3158, %r2934, 4;
	add.s64 	%rd3159, %rd3158, %rd2942;
	ld.shared.u32 	%r2943, [%rd3159+3072];
	mul.wide.u32 	%rd3160, %r2938, 4;
	add.s64 	%rd3161, %rd3160, %rd2942;
	ld.shared.u32 	%r2944, [%rd3161+3072];
	mul.wide.u32 	%rd3162, %r2942, 4;
	add.s64 	%rd3163, %rd3162, %rd2942;
	ld.shared.u32 	%r2945, [%rd3163+3072];
	mul.wide.u32 	%rd3164, %r2930, 4;
	add.s64 	%rd3165, %rd3164, %rd2942;
	ld.shared.u32 	%r2946, [%rd3165+3072];
	xor.b32  	%r2947, %r2941, %r2946;
	xor.b32  	%r2948, %r2937, %r2945;
	xor.b32  	%r2949, %r2933, %r2944;
	xor.b32  	%r2950, %r2929, %r2943;
	ld.global.v4.u32 	{%r2951, %r2952, %r2953, %r2954}, [%rd2934+96];
	xor.b32  	%r2959, %r2894, %r2895;
	and.b32  	%r2960, %r2959, 255;
	mul.wide.u32 	%rd3166, %r2960, 4;
	add.s64 	%rd3167, %rd2942, %rd3166;
	xor.b32  	%r2961, %r2893, %r2896;
	and.b32  	%r2962, %r2961, 255;
	mul.wide.u32 	%rd3168, %r2962, 4;
	add.s64 	%rd3169, %rd2942, %rd3168;
	xor.b32  	%r2963, %r2892, %r2897;
	and.b32  	%r2964, %r2963, 255;
	mul.wide.u32 	%rd3170, %r2964, 4;
	add.s64 	%rd3171, %rd2942, %rd3170;
	xor.b32  	%r2965, %r2891, %r2898;
	and.b32  	%r2966, %r2965, 255;
	mul.wide.u32 	%rd3172, %r2966, 4;
	add.s64 	%rd3173, %rd2942, %rd3172;
	bfe.u32 	%r2967, %r2961, 8, 8;
	mul.wide.u32 	%rd3174, %r2967, 4;
	add.s64 	%rd3175, %rd3174, %rd2942;
	ld.shared.u32 	%r2968, [%rd3175+1024];
	ld.shared.u32 	%r2969, [%rd3167];
	xor.b32  	%r2970, %r2968, %r2969;
	bfe.u32 	%r2971, %r2963, 8, 8;
	mul.wide.u32 	%rd3176, %r2971, 4;
	add.s64 	%rd3177, %rd3176, %rd2942;
	ld.shared.u32 	%r2972, [%rd3177+1024];
	ld.shared.u32 	%r2973, [%rd3169];
	xor.b32  	%r2974, %r2972, %r2973;
	bfe.u32 	%r2975, %r2965, 8, 8;
	mul.wide.u32 	%rd3178, %r2975, 4;
	add.s64 	%rd3179, %rd3178, %rd2942;
	ld.shared.u32 	%r2976, [%rd3179+1024];
	ld.shared.u32 	%r2977, [%rd3171];
	xor.b32  	%r2978, %r2976, %r2977;
	bfe.u32 	%r2979, %r2959, 8, 8;
	mul.wide.u32 	%rd3180, %r2979, 4;
	add.s64 	%rd3181, %rd3180, %rd2942;
	ld.shared.u32 	%r2980, [%rd3181+1024];
	ld.shared.u32 	%r2981, [%rd3173];
	xor.b32  	%r2982, %r2980, %r2981;
	bfe.u32 	%r2983, %r2963, 16, 8;
	mul.wide.u32 	%rd3182, %r2983, 4;
	add.s64 	%rd3183, %rd3182, %rd2942;
	ld.shared.u32 	%r2984, [%rd3183+2048];
	xor.b32  	%r2985, %r2970, %r2984;
	shr.u32 	%r2986, %r2963, 24;
	bfe.u32 	%r2987, %r2965, 16, 8;
	mul.wide.u32 	%rd3184, %r2987, 4;
	add.s64 	%rd3185, %rd3184, %rd2942;
	ld.shared.u32 	%r2988, [%rd3185+2048];
	xor.b32  	%r2989, %r2974, %r2988;
	shr.u32 	%r2990, %r2965, 24;
	bfe.u32 	%r2991, %r2959, 16, 8;
	mul.wide.u32 	%rd3186, %r2991, 4;
	add.s64 	%rd3187, %rd3186, %rd2942;
	ld.shared.u32 	%r2992, [%rd3187+2048];
	xor.b32  	%r2993, %r2978, %r2992;
	shr.u32 	%r2994, %r2959, 24;
	bfe.u32 	%r2995, %r2961, 16, 8;
	mul.wide.u32 	%rd3188, %r2995, 4;
	add.s64 	%rd3189, %rd3188, %rd2942;
	ld.shared.u32 	%r2996, [%rd3189+2048];
	xor.b32  	%r2997, %r2982, %r2996;
	shr.u32 	%r2998, %r2961, 24;
	mul.wide.u32 	%rd3190, %r2990, 4;
	add.s64 	%rd3191, %rd3190, %rd2942;
	ld.shared.u32 	%r2999, [%rd3191+3072];
	mul.wide.u32 	%rd3192, %r2994, 4;
	add.s64 	%rd3193, %rd3192, %rd2942;
	ld.shared.u32 	%r3000, [%rd3193+3072];
	mul.wide.u32 	%rd3194, %r2998, 4;
	add.s64 	%rd3195, %rd3194, %rd2942;
	ld.shared.u32 	%r3001, [%rd3195+3072];
	mul.wide.u32 	%rd3196, %r2986, 4;
	add.s64 	%rd3197, %rd3196, %rd2942;
	ld.shared.u32 	%r3002, [%rd3197+3072];
	xor.b32  	%r3003, %r2997, %r3002;
	xor.b32  	%r3004, %r2993, %r3001;
	xor.b32  	%r3005, %r2989, %r3000;
	xor.b32  	%r3006, %r2985, %r2999;
	ld.global.v4.u32 	{%r3007, %r3008, %r3009, %r3010}, [%rd2934+112];
	xor.b32  	%r3015, %r2950, %r2951;
	and.b32  	%r3016, %r3015, 255;
	mul.wide.u32 	%rd3198, %r3016, 4;
	add.s64 	%rd3199, %rd2942, %rd3198;
	xor.b32  	%r3017, %r3006, %r3007;
	and.b32  	%r3018, %r3017, 255;
	mul.wide.u32 	%rd3200, %r3018, 4;
	add.s64 	%rd3201, %rd2942, %rd3200;
	xor.b32  	%r3019, %r2949, %r2952;
	and.b32  	%r3020, %r3019, 255;
	mul.wide.u32 	%rd3202, %r3020, 4;
	add.s64 	%rd3203, %rd2942, %rd3202;
	xor.b32  	%r3021, %r3005, %r3008;
	and.b32  	%r3022, %r3021, 255;
	mul.wide.u32 	%rd3204, %r3022, 4;
	add.s64 	%rd3205, %rd2942, %rd3204;
	bfe.u32 	%r3023, %r3017, 8, 8;
	mul.wide.u32 	%rd3206, %r3023, 4;
	add.s64 	%rd3207, %rd3206, %rd2942;
	ld.shared.u32 	%r3024, [%rd3207+1024];
	ld.shared.u32 	%r3025, [%rd3199];
	xor.b32  	%r3026, %r3024, %r3025;
	bfe.u32 	%r3027, %r3019, 8, 8;
	mul.wide.u32 	%rd3208, %r3027, 4;
	add.s64 	%rd3209, %rd3208, %rd2942;
	ld.shared.u32 	%r3028, [%rd3209+1024];
	ld.shared.u32 	%r3029, [%rd3201];
	xor.b32  	%r3030, %r3028, %r3029;
	bfe.u32 	%r3031, %r3021, 8, 8;
	mul.wide.u32 	%rd3210, %r3031, 4;
	add.s64 	%rd3211, %rd3210, %rd2942;
	ld.shared.u32 	%r3032, [%rd3211+1024];
	ld.shared.u32 	%r3033, [%rd3203];
	xor.b32  	%r3034, %r3032, %r3033;
	bfe.u32 	%r3035, %r3015, 8, 8;
	mul.wide.u32 	%rd3212, %r3035, 4;
	add.s64 	%rd3213, %rd3212, %rd2942;
	ld.shared.u32 	%r3036, [%rd3213+1024];
	ld.shared.u32 	%r3037, [%rd3205];
	xor.b32  	%r3038, %r3036, %r3037;
	bfe.u32 	%r3039, %r3019, 16, 8;
	mul.wide.u32 	%rd3214, %r3039, 4;
	add.s64 	%rd3215, %rd3214, %rd2942;
	ld.shared.u32 	%r3040, [%rd3215+2048];
	xor.b32  	%r3041, %r3026, %r3040;
	shr.u32 	%r3042, %r3019, 24;
	bfe.u32 	%r3043, %r3021, 16, 8;
	mul.wide.u32 	%rd3216, %r3043, 4;
	add.s64 	%rd3217, %rd3216, %rd2942;
	ld.shared.u32 	%r3044, [%rd3217+2048];
	xor.b32  	%r3045, %r3030, %r3044;
	shr.u32 	%r3046, %r3021, 24;
	bfe.u32 	%r3047, %r3015, 16, 8;
	mul.wide.u32 	%rd3218, %r3047, 4;
	add.s64 	%rd3219, %rd3218, %rd2942;
	ld.shared.u32 	%r3048, [%rd3219+2048];
	xor.b32  	%r3049, %r3034, %r3048;
	shr.u32 	%r3050, %r3015, 24;
	bfe.u32 	%r3051, %r3017, 16, 8;
	mul.wide.u32 	%rd3220, %r3051, 4;
	add.s64 	%rd3221, %rd3220, %rd2942;
	ld.shared.u32 	%r3052, [%rd3221+2048];
	xor.b32  	%r3053, %r3038, %r3052;
	shr.u32 	%r3054, %r3017, 24;
	mul.wide.u32 	%rd3222, %r3046, 4;
	add.s64 	%rd3223, %rd3222, %rd2942;
	ld.shared.u32 	%r3055, [%rd3223+3072];
	mul.wide.u32 	%rd3224, %r3050, 4;
	add.s64 	%rd3225, %rd3224, %rd2942;
	ld.shared.u32 	%r3056, [%rd3225+3072];
	mul.wide.u32 	%rd3226, %r3054, 4;
	add.s64 	%rd3227, %rd3226, %rd2942;
	ld.shared.u32 	%r3057, [%rd3227+3072];
	mul.wide.u32 	%rd3228, %r3042, 4;
	add.s64 	%rd3229, %rd3228, %rd2942;
	ld.shared.u32 	%r3058, [%rd3229+3072];
	xor.b32  	%r3059, %r3053, %r3058;
	xor.b32  	%r3060, %r3049, %r3057;
	xor.b32  	%r3061, %r3045, %r3056;
	xor.b32  	%r3062, %r3041, %r3055;
	ld.global.v4.u32 	{%r3063, %r3064, %r3065, %r3066}, [%rd2934+128];
	xor.b32  	%r3071, %r2948, %r2953;
	and.b32  	%r3072, %r3071, 255;
	mul.wide.u32 	%rd3230, %r3072, 4;
	add.s64 	%rd3231, %rd2942, %rd3230;
	xor.b32  	%r3073, %r3004, %r3009;
	and.b32  	%r3074, %r3073, 255;
	mul.wide.u32 	%rd3232, %r3074, 4;
	add.s64 	%rd3233, %rd2942, %rd3232;
	xor.b32  	%r3075, %r2947, %r2954;
	and.b32  	%r3076, %r3075, 255;
	mul.wide.u32 	%rd3234, %r3076, 4;
	add.s64 	%rd3235, %rd2942, %rd3234;
	xor.b32  	%r3077, %r3003, %r3010;
	and.b32  	%r3078, %r3077, 255;
	mul.wide.u32 	%rd3236, %r3078, 4;
	add.s64 	%rd3237, %rd2942, %rd3236;
	bfe.u32 	%r3079, %r3073, 8, 8;
	mul.wide.u32 	%rd3238, %r3079, 4;
	add.s64 	%rd3239, %rd3238, %rd2942;
	ld.shared.u32 	%r3080, [%rd3239+1024];
	ld.shared.u32 	%r3081, [%rd3231];
	xor.b32  	%r3082, %r3080, %r3081;
	bfe.u32 	%r3083, %r3075, 8, 8;
	mul.wide.u32 	%rd3240, %r3083, 4;
	add.s64 	%rd3241, %rd3240, %rd2942;
	ld.shared.u32 	%r3084, [%rd3241+1024];
	ld.shared.u32 	%r3085, [%rd3233];
	xor.b32  	%r3086, %r3084, %r3085;
	bfe.u32 	%r3087, %r3077, 8, 8;
	mul.wide.u32 	%rd3242, %r3087, 4;
	add.s64 	%rd3243, %rd3242, %rd2942;
	ld.shared.u32 	%r3088, [%rd3243+1024];
	ld.shared.u32 	%r3089, [%rd3235];
	xor.b32  	%r3090, %r3088, %r3089;
	bfe.u32 	%r3091, %r3071, 8, 8;
	mul.wide.u32 	%rd3244, %r3091, 4;
	add.s64 	%rd3245, %rd3244, %rd2942;
	ld.shared.u32 	%r3092, [%rd3245+1024];
	ld.shared.u32 	%r3093, [%rd3237];
	xor.b32  	%r3094, %r3092, %r3093;
	bfe.u32 	%r3095, %r3075, 16, 8;
	mul.wide.u32 	%rd3246, %r3095, 4;
	add.s64 	%rd3247, %rd3246, %rd2942;
	ld.shared.u32 	%r3096, [%rd3247+2048];
	xor.b32  	%r3097, %r3082, %r3096;
	shr.u32 	%r3098, %r3075, 24;
	bfe.u32 	%r3099, %r3077, 16, 8;
	mul.wide.u32 	%rd3248, %r3099, 4;
	add.s64 	%rd3249, %rd3248, %rd2942;
	ld.shared.u32 	%r3100, [%rd3249+2048];
	xor.b32  	%r3101, %r3086, %r3100;
	shr.u32 	%r3102, %r3077, 24;
	bfe.u32 	%r3103, %r3071, 16, 8;
	mul.wide.u32 	%rd3250, %r3103, 4;
	add.s64 	%rd3251, %rd3250, %rd2942;
	ld.shared.u32 	%r3104, [%rd3251+2048];
	xor.b32  	%r3105, %r3090, %r3104;
	shr.u32 	%r3106, %r3071, 24;
	bfe.u32 	%r3107, %r3073, 16, 8;
	mul.wide.u32 	%rd3252, %r3107, 4;
	add.s64 	%rd3253, %rd3252, %rd2942;
	ld.shared.u32 	%r3108, [%rd3253+2048];
	xor.b32  	%r3109, %r3094, %r3108;
	shr.u32 	%r3110, %r3073, 24;
	mul.wide.u32 	%rd3254, %r3102, 4;
	add.s64 	%rd3255, %rd3254, %rd2942;
	ld.shared.u32 	%r3111, [%rd3255+3072];
	mul.wide.u32 	%rd3256, %r3106, 4;
	add.s64 	%rd3257, %rd3256, %rd2942;
	ld.shared.u32 	%r3112, [%rd3257+3072];
	mul.wide.u32 	%rd3258, %r3110, 4;
	add.s64 	%rd3259, %rd3258, %rd2942;
	ld.shared.u32 	%r3113, [%rd3259+3072];
	mul.wide.u32 	%rd3260, %r3098, 4;
	add.s64 	%rd3261, %rd3260, %rd2942;
	ld.shared.u32 	%r3114, [%rd3261+3072];
	xor.b32  	%r3115, %r3109, %r3114;
	xor.b32  	%r3116, %r3105, %r3113;
	xor.b32  	%r3117, %r3101, %r3112;
	xor.b32  	%r3118, %r3097, %r3111;
	ld.global.v4.u32 	{%r3119, %r3120, %r3121, %r3122}, [%rd2934+144];
	xor.b32  	%r3127, %r3062, %r3063;
	and.b32  	%r3128, %r3127, 255;
	mul.wide.u32 	%rd3262, %r3128, 4;
	add.s64 	%rd3263, %rd2942, %rd3262;
	xor.b32  	%r3129, %r3061, %r3064;
	and.b32  	%r3130, %r3129, 255;
	mul.wide.u32 	%rd3264, %r3130, 4;
	add.s64 	%rd3265, %rd2942, %rd3264;
	xor.b32  	%r3131, %r3060, %r3065;
	and.b32  	%r3132, %r3131, 255;
	mul.wide.u32 	%rd3266, %r3132, 4;
	add.s64 	%rd3267, %rd2942, %rd3266;
	xor.b32  	%r3133, %r3059, %r3066;
	and.b32  	%r3134, %r3133, 255;
	mul.wide.u32 	%rd3268, %r3134, 4;
	add.s64 	%rd3269, %rd2942, %rd3268;
	bfe.u32 	%r3135, %r3129, 8, 8;
	mul.wide.u32 	%rd3270, %r3135, 4;
	add.s64 	%rd3271, %rd3270, %rd2942;
	ld.shared.u32 	%r3136, [%rd3271+1024];
	ld.shared.u32 	%r3137, [%rd3263];
	xor.b32  	%r3138, %r3136, %r3137;
	bfe.u32 	%r3139, %r3131, 8, 8;
	mul.wide.u32 	%rd3272, %r3139, 4;
	add.s64 	%rd3273, %rd3272, %rd2942;
	ld.shared.u32 	%r3140, [%rd3273+1024];
	ld.shared.u32 	%r3141, [%rd3265];
	xor.b32  	%r3142, %r3140, %r3141;
	bfe.u32 	%r3143, %r3133, 8, 8;
	mul.wide.u32 	%rd3274, %r3143, 4;
	add.s64 	%rd3275, %rd3274, %rd2942;
	ld.shared.u32 	%r3144, [%rd3275+1024];
	ld.shared.u32 	%r3145, [%rd3267];
	xor.b32  	%r3146, %r3144, %r3145;
	bfe.u32 	%r3147, %r3127, 8, 8;
	mul.wide.u32 	%rd3276, %r3147, 4;
	add.s64 	%rd3277, %rd3276, %rd2942;
	ld.shared.u32 	%r3148, [%rd3277+1024];
	ld.shared.u32 	%r3149, [%rd3269];
	xor.b32  	%r3150, %r3148, %r3149;
	bfe.u32 	%r3151, %r3131, 16, 8;
	mul.wide.u32 	%rd3278, %r3151, 4;
	add.s64 	%rd3279, %rd3278, %rd2942;
	ld.shared.u32 	%r3152, [%rd3279+2048];
	xor.b32  	%r3153, %r3138, %r3152;
	shr.u32 	%r3154, %r3131, 24;
	bfe.u32 	%r3155, %r3133, 16, 8;
	mul.wide.u32 	%rd3280, %r3155, 4;
	add.s64 	%rd3281, %rd3280, %rd2942;
	ld.shared.u32 	%r3156, [%rd3281+2048];
	xor.b32  	%r3157, %r3142, %r3156;
	shr.u32 	%r3158, %r3133, 24;
	bfe.u32 	%r3159, %r3127, 16, 8;
	mul.wide.u32 	%rd3282, %r3159, 4;
	add.s64 	%rd3283, %rd3282, %rd2942;
	ld.shared.u32 	%r3160, [%rd3283+2048];
	xor.b32  	%r3161, %r3146, %r3160;
	shr.u32 	%r3162, %r3127, 24;
	bfe.u32 	%r3163, %r3129, 16, 8;
	mul.wide.u32 	%rd3284, %r3163, 4;
	add.s64 	%rd3285, %rd3284, %rd2942;
	ld.shared.u32 	%r3164, [%rd3285+2048];
	xor.b32  	%r3165, %r3150, %r3164;
	shr.u32 	%r3166, %r3129, 24;
	mul.wide.u32 	%rd3286, %r3158, 4;
	add.s64 	%rd3287, %rd3286, %rd2942;
	ld.shared.u32 	%r3167, [%rd3287+3072];
	mul.wide.u32 	%rd3288, %r3162, 4;
	add.s64 	%rd3289, %rd3288, %rd2942;
	ld.shared.u32 	%r3168, [%rd3289+3072];
	mul.wide.u32 	%rd3290, %r3166, 4;
	add.s64 	%rd3291, %rd3290, %rd2942;
	ld.shared.u32 	%r3169, [%rd3291+3072];
	mul.wide.u32 	%rd3292, %r3154, 4;
	add.s64 	%rd3293, %rd3292, %rd2942;
	ld.shared.u32 	%r3170, [%rd3293+3072];
	xor.b32  	%r3171, %r3165, %r3170;
	xor.b32  	%r3172, %r3157, %r3168;
	xor.b32  	%r3173, %r3153, %r3167;
	xor.b32  	%r3174, %r3161, %r3169;
	ld.global.v4.u32 	{%r3175, %r3176, %r3177, %r3178}, [%rd2934+160];
	xor.b32  	%r3183, %r3174, %r3177;
	xor.b32  	%r3184, %r3173, %r3175;
	xor.b32  	%r3185, %r3118, %r3119;
	and.b32  	%r3186, %r3185, 255;
	mul.wide.u32 	%rd3294, %r3186, 4;
	add.s64 	%rd3295, %rd2942, %rd3294;
	xor.b32  	%r3187, %r3117, %r3120;
	and.b32  	%r3188, %r3187, 255;
	mul.wide.u32 	%rd3296, %r3188, 4;
	add.s64 	%rd3297, %rd2942, %rd3296;
	xor.b32  	%r3189, %r3116, %r3121;
	and.b32  	%r3190, %r3189, 255;
	mul.wide.u32 	%rd3298, %r3190, 4;
	add.s64 	%rd3299, %rd2942, %rd3298;
	xor.b32  	%r3191, %r3115, %r3122;
	and.b32  	%r3192, %r3191, 255;
	mul.wide.u32 	%rd3300, %r3192, 4;
	add.s64 	%rd3301, %rd2942, %rd3300;
	bfe.u32 	%r3193, %r3187, 8, 8;
	mul.wide.u32 	%rd3302, %r3193, 4;
	add.s64 	%rd3303, %rd3302, %rd2942;
	ld.shared.u32 	%r3194, [%rd3303+1024];
	ld.shared.u32 	%r3195, [%rd3295];
	xor.b32  	%r3196, %r3194, %r3195;
	bfe.u32 	%r3197, %r3189, 8, 8;
	mul.wide.u32 	%rd3304, %r3197, 4;
	add.s64 	%rd3305, %rd3304, %rd2942;
	ld.shared.u32 	%r3198, [%rd3305+1024];
	ld.shared.u32 	%r3199, [%rd3297];
	xor.b32  	%r3200, %r3198, %r3199;
	bfe.u32 	%r3201, %r3191, 8, 8;
	mul.wide.u32 	%rd3306, %r3201, 4;
	add.s64 	%rd3307, %rd3306, %rd2942;
	ld.shared.u32 	%r3202, [%rd3307+1024];
	ld.shared.u32 	%r3203, [%rd3299];
	xor.b32  	%r3204, %r3202, %r3203;
	bfe.u32 	%r3205, %r3185, 8, 8;
	mul.wide.u32 	%rd3308, %r3205, 4;
	add.s64 	%rd3309, %rd3308, %rd2942;
	ld.shared.u32 	%r3206, [%rd3309+1024];
	ld.shared.u32 	%r3207, [%rd3301];
	xor.b32  	%r3208, %r3206, %r3207;
	bfe.u32 	%r3209, %r3189, 16, 8;
	mul.wide.u32 	%rd3310, %r3209, 4;
	add.s64 	%rd3311, %rd3310, %rd2942;
	ld.shared.u32 	%r3210, [%rd3311+2048];
	xor.b32  	%r3211, %r3196, %r3210;
	shr.u32 	%r3212, %r3189, 24;
	bfe.u32 	%r3213, %r3191, 16, 8;
	mul.wide.u32 	%rd3312, %r3213, 4;
	add.s64 	%rd3313, %rd3312, %rd2942;
	ld.shared.u32 	%r3214, [%rd3313+2048];
	xor.b32  	%r3215, %r3200, %r3214;
	shr.u32 	%r3216, %r3191, 24;
	bfe.u32 	%r3217, %r3185, 16, 8;
	mul.wide.u32 	%rd3314, %r3217, 4;
	add.s64 	%rd3315, %rd3314, %rd2942;
	ld.shared.u32 	%r3218, [%rd3315+2048];
	xor.b32  	%r3219, %r3204, %r3218;
	shr.u32 	%r3220, %r3185, 24;
	bfe.u32 	%r3221, %r3187, 16, 8;
	mul.wide.u32 	%rd3316, %r3221, 4;
	add.s64 	%rd3317, %rd3316, %rd2942;
	ld.shared.u32 	%r3222, [%rd3317+2048];
	xor.b32  	%r3223, %r3208, %r3222;
	shr.u32 	%r3224, %r3187, 24;
	mul.wide.u32 	%rd3318, %r3216, 4;
	add.s64 	%rd3319, %rd3318, %rd2942;
	ld.shared.u32 	%r3225, [%rd3319+3072];
	mul.wide.u32 	%rd3320, %r3220, 4;
	add.s64 	%rd3321, %rd3320, %rd2942;
	ld.shared.u32 	%r3226, [%rd3321+3072];
	mul.wide.u32 	%rd3322, %r3224, 4;
	add.s64 	%rd3323, %rd3322, %rd2942;
	ld.shared.u32 	%r3227, [%rd3323+3072];
	mul.wide.u32 	%rd3324, %r3212, 4;
	add.s64 	%rd3325, %rd3324, %rd2942;
	ld.shared.u32 	%r3228, [%rd3325+3072];
	xor.b32  	%r3229, %r3215, %r3226;
	xor.b32  	%r3230, %r3211, %r3225;
	xor.b32  	%r3231, %r3219, %r3227;
	xor.b32  	%r3232, %r3223, %r3228;
	ld.global.v4.u32 	{%r3233, %r3234, %r3235, %r3236}, [%rd2934+176];
	xor.b32  	%r3241, %r3232, %r3236;
	xor.b32  	%r3242, %r3231, %r3235;
	xor.b32  	%r3243, %r3230, %r3233;
	xor.b32  	%r3244, %r3229, %r3234;
	xor.b32  	%r3245, %r3172, %r3176;
	xor.b32  	%r3246, %r3171, %r3178;
	xor.b32  	%r3247, %r3245, %r6732;
	xor.b32  	%r3248, %r3244, %r6733;
	xor.b32  	%r3249, %r3184, %r6730;
	xor.b32  	%r3250, %r3243, %r6731;
	xor.b32  	%r6731, %r3242, %r3250;
	xor.b32  	%r6730, %r3183, %r3249;
	xor.b32  	%r6733, %r3241, %r3248;
	xor.b32  	%r6732, %r3246, %r3247;
	mov.b64	%rd3326, {%r6732, %r6733};
	mov.b64	%rd3327, {%r6730, %r6731};
	cvt.u32.u64	%r3251, %rd3327;
	cvt.s32.s16 	%r3252, %r3251;
	mov.b64	%rd3328, {%r6723, %r6724};
	mov.b64	%rd3329, {%r6721, %r6722};
	cvt.u32.u64	%r3253, %rd3329;
	cvt.s32.s16 	%r3254, %r3253;
	mad.lo.s32 	%r3255, %r3252, %r3254, 16384;
	bfe.u32 	%r3256, %r3255, 15, 16;
	shr.s32 	%r3257, %r3251, 16;
	shr.s32 	%r3258, %r3253, 16;
	mad.lo.s32 	%r3259, %r3257, %r3258, 16384;
	mov.u32 	%r3260, 0;
	mov.b64	%rd3330, {%r3256, %r3260};
	shl.b32 	%r3261, %r3259, 1;
	cvt.u64.u32	%rd3331, %r3261;
	and.b64  	%rd3332, %rd3330, 65535;
	shr.u64 	%rd3333, %rd3326, 16;
	shr.u64 	%rd3334, %rd3327, 16;
	cvt.u32.u64	%r3262, %rd3334;
	shr.s32 	%r3263, %r3262, 16;
	shr.u64 	%rd3335, %rd3328, 16;
	shr.u64 	%rd3336, %rd3329, 16;
	cvt.u32.u64	%r3264, %rd3336;
	shr.s32 	%r3265, %r3264, 16;
	mad.lo.s32 	%r3266, %r3263, %r3265, 16384;
	mul.wide.u32 	%rd3337, %r3266, 131072;
	shr.u64 	%rd3338, %rd3327, 32;
	cvt.u32.u64	%r3267, %rd3338;
	shr.s32 	%r3268, %r3267, 16;
	shr.u64 	%rd3339, %rd3329, 32;
	cvt.u32.u64	%r3269, %rd3339;
	shr.s32 	%r3270, %r3269, 16;
	mad.lo.s32 	%r3271, %r3268, %r3270, 16384;
	shr.u32 	%r3272, %r3271, 15;
	cvt.u64.u32	%rd3340, %r3272;
	shl.b64 	%rd3341, %rd3340, 48;
	and.b64  	%rd3342, %rd3337, 281470681743360;
	shl.b64 	%rd3343, %rd3326, 16;
	cvt.u32.u64	%r3273, %rd3343;
	shr.s32 	%r3274, %r3273, 16;
	shl.b64 	%rd3344, %rd3328, 16;
	cvt.u32.u64	%r3275, %rd3344;
	shr.s32 	%r3276, %r3275, 16;
	mad.lo.s32 	%r3277, %r3274, %r3276, 16384;
	shr.u32 	%r3278, %r3277, 15;
	cvt.u64.u32	%rd3345, %r3278;
	cvt.u32.u64	%r3279, %rd3326;
	shr.s32 	%r3280, %r3279, 16;
	cvt.u32.u64	%r3281, %rd3328;
	shr.s32 	%r3282, %r3281, 16;
	mad.lo.s32 	%r3283, %r3280, %r3282, 16384;
	and.b64  	%rd3346, %rd3345, 65535;
	and.b64  	%rd3347, %rd3331, 4294901760;
	or.b64  	%rd3348, %rd3347, %rd3332;
	or.b64  	%rd3349, %rd3348, %rd3342;
	shl.b32 	%r3284, %r3283, 1;
	cvt.u64.u32	%rd3350, %r3284;
	cvt.u32.u64	%r3285, %rd3333;
	shr.s32 	%r3286, %r3285, 16;
	cvt.u32.u64	%r3287, %rd3335;
	shr.s32 	%r3288, %r3287, 16;
	mad.lo.s32 	%r3289, %r3286, %r3288, 16384;
	mul.wide.u32 	%rd3351, %r3289, 131072;
	and.b64  	%rd3352, %rd3350, 4294901760;
	or.b64  	%rd3353, %rd3352, %rd3346;
	shr.u64 	%rd3354, %rd3326, 32;
	cvt.u32.u64	%r3290, %rd3354;
	shr.s32 	%r3291, %r3290, 16;
	shr.u64 	%rd3355, %rd3328, 32;
	cvt.u32.u64	%r3292, %rd3355;
	shr.s32 	%r3293, %r3292, 16;
	mad.lo.s32 	%r3294, %r3291, %r3293, 16384;
	shr.u32 	%r3295, %r3294, 15;
	cvt.u64.u32	%rd3356, %r3295;
	shl.b64 	%rd3357, %rd3356, 48;
	and.b64  	%rd3358, %rd3351, 281470681743360;
	or.b64  	%rd3359, %rd3358, %rd3357;
	or.b64  	%rd3360, %rd3359, %rd3353;
	or.b64  	%rd3361, %rd3349, %rd3341;
	mov.b64	{%r3296, %r3297}, %rd3360;
	mov.b64	{%r3298, %r3299}, %rd3361;
	xor.b32  	%r6641, %r6724, %r3297;
	xor.b32  	%r6640, %r6723, %r3296;
	xor.b32  	%r6639, %r6722, %r3299;
	xor.b32  	%r6638, %r6721, %r3298;
	cvt.u64.u32	%rd3362, %r6729;
	add.s64 	%rd3363, %rd3362, %rd9;
	shl.b64 	%rd3364, %rd3363, 1;
	mov.u64 	%rd3365, verus_gpu_hash$sharedrand;
	add.s64 	%rd3366, %rd3365, %rd3364;
	st.shared.u16 	[%rd3366], %rs236;
	mov.u64 	%rd3367, verus_gpu_hash$sharedrandex;
	add.s64 	%rd3368, %rd3367, %rd3364;
	st.shared.u16 	[%rd3368], %rs237;
	st.global.v4.u32 	[%rd2934], {%r6625, %r6626, %r6627, %r6628};
	cvt.u64.u16	%rd3369, %rs237;
	add.s64 	%rd3370, %rd3369, %rd3;
	shl.b64 	%rd3371, %rd3370, 4;
	add.s64 	%rd3372, %rd203, %rd3371;
	st.global.v4.u32 	[%rd3372], {%r6638, %r6639, %r6640, %r6641};
	add.s32 	%r505, %r6729, 1;
	setp.eq.s32	%p49, %r505, 32;
	@%p49 bra 	BB0_134;

	add.s32 	%r6452, %r6729, 1;
	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p50, %r6452, 0;
	mov.u16 	%rs238, 16;
	@%p50 bra 	BB0_90;
	bra.uni 	BB0_91;

BB0_90:
	mov.u32 	%r6721, %r6625;
	mov.u32 	%r6722, %r6626;
	mov.u32 	%r6723, %r6627;
	mov.u32 	%r6724, %r6628;
	mov.u32 	%r6729, %r3260;
	bra.uni 	BB0_92;

BB0_87:
	mov.u32 	%r6638, %r6625;
	mov.u32 	%r6639, %r6626;
	mov.u32 	%r6640, %r6627;
	mov.u32 	%r6641, %r6628;
	bra.uni 	BB0_92;

BB0_91:
	add.s32 	%r6729, %r6729, 1;
	cvt.u64.u32	%rd3373, %r6731;
	shr.u32 	%r3301, %r6730, 5;
	cvt.u64.u32	%rd3374, %r3301;
	cvt.u16.u32	%rs91, %r3301;
	and.b16  	%rs236, %rs91, 511;
	cvt.u16.u32	%rs92, %r6731;
	and.b16  	%rs237, %rs92, 511;
	cvt.u16.u64	%rs93, %rd6123;
	and.b16  	%rs238, %rs93, 28;
	and.b64  	%rd3375, %rd3374, 511;
	add.s64 	%rd3376, %rd3375, %rd3;
	shl.b64 	%rd3377, %rd3376, 4;
	add.s64 	%rd3378, %rd203, %rd3377;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd3378];
	and.b64  	%rd3379, %rd3373, 511;
	add.s64 	%rd3380, %rd3379, %rd3;
	shl.b64 	%rd3381, %rd3380, 4;
	add.s64 	%rd3382, %rd203, %rd3381;
	ld.global.v4.u32 	{%r6638, %r6639, %r6640, %r6641}, [%rd3382];
	mul.wide.u32 	%rd3383, %r6730, 16;
	and.b64  	%rd3384, %rd3383, 48;
	add.s64 	%rd6122, %rd1, %rd3384;

BB0_92:
	and.b16  	%rs94, %rs238, 255;
	setp.ne.s16	%p51, %rs94, 20;
	@%p51 bra 	BB0_93;

	and.b64  	%rd3386, %rd6123, 1;
	mov.u64 	%rd3387, 1;
	shl.b64 	%rd3388, %rd3386, 1;
	sub.s64 	%rd3389, %rd3387, %rd3388;
	shl.b64 	%rd3390, %rd3389, 4;
	add.s64 	%rd56, %rd6122, %rd3390;
	shr.u64 	%rd6111, %rd6123, 61;
	cvt.u64.u16	%rd3391, %rs236;
	add.s64 	%rd3392, %rd3391, %rd3;
	shl.b64 	%rd3393, %rd3392, 4;
	add.s64 	%rd6112, %rd203, %rd3393;
	mov.u64 	%rd6113, 0;

BB0_95:
	cvt.u32.u64	%r3310, %rd6111;
	and.b32  	%r3311, %r3310, 63;
	mov.u64 	%rd3394, 268435456;
	shl.b64 	%rd3395, %rd3394, %r3311;
	and.b64  	%rd3396, %rd3395, %rd6123;
	setp.gt.u64	%p52, %rd3396, 268435455;
	ld.global.v4.u32 	{%r3312, %r3313, %r3314, %r3315}, [%rd6112];
	and.b64  	%rd62, %rd6111, 1;
	@%p52 bra 	BB0_97;
	bra.uni 	BB0_96;

BB0_97:
	setp.eq.s64	%p54, %rd62, 0;
	selp.b64	%rd3576, %rd56, %rd6122, %p54;
	ld.local.v4.u32 	{%r3537, %r3538, %r3539, %r3540}, [%rd3576];
	xor.b32  	%r542, %r3315, %r3540;
	xor.b32  	%r541, %r3314, %r3539;
	xor.b32  	%r540, %r3313, %r3538;
	xor.b32  	%r539, %r3312, %r3537;
	mov.b64	%rd3577, {%r541, %r542};
	mov.u64 	%rd3578, 0;
	st.local.v2.u64 	[%rd12], {%rd3578, %rd3577};
	shl.b64 	%rd3581, %rd3577, 1;
	xor.b64  	%rd3582, %rd3581, %rd3577;
	st.local.v2.u64 	[%rd15], {%rd3581, %rd3582};
	shl.b64 	%rd3583, %rd3577, 2;
	xor.b64  	%rd3584, %rd3583, %rd3577;
	st.local.v2.u64 	[%rd15+16], {%rd3583, %rd3584};
	shl.b64 	%rd3585, %rd3582, 1;
	xor.b64  	%rd3586, %rd3585, %rd3577;
	st.local.v2.u64 	[%rd15+32], {%rd3585, %rd3586};
	mov.b64	%rd3587, {%r539, %r540};
	shl.b64 	%rd3588, %rd3587, 3;
	and.b64  	%rd3589, %rd3588, 56;
	add.s64 	%rd3590, %rd12, %rd3589;
	and.b64  	%rd3591, %rd3587, 56;
	add.s64 	%rd3592, %rd12, %rd3591;
	ld.local.u64 	%rd3593, [%rd3592];
	shl.b64 	%rd3594, %rd3593, 3;
	ld.local.u64 	%rd3595, [%rd3590];
	xor.b64  	%rd3596, %rd3595, %rd3594;
	shr.u64 	%rd3597, %rd3593, 61;
	mov.u32 	%r3545, 0;
	mov.b64	%rd3598, {%r3545, %r3545};
	xor.b64  	%rd3599, %rd3598, %rd3597;
	shr.u64 	%rd3600, %rd3587, 3;
	and.b64  	%rd3601, %rd3600, 56;
	add.s64 	%rd3602, %rd12, %rd3601;
	ld.local.u64 	%rd3603, [%rd3602];
	shl.b64 	%rd3604, %rd3603, 6;
	mov.b64	{%r3546, %r3547}, %rd3596;
	mov.b64	%rd3605, {%r3546, %r3547};
	xor.b64  	%rd3606, %rd3605, %rd3604;
	shr.u64 	%rd3607, %rd3603, 58;
	mov.b64	{%r3548, %r3549}, %rd3599;
	mov.b64	%rd3608, {%r3548, %r3549};
	xor.b64  	%rd3609, %rd3608, %rd3607;
	shr.u64 	%rd3610, %rd3587, 6;
	and.b64  	%rd3611, %rd3610, 56;
	add.s64 	%rd3612, %rd12, %rd3611;
	ld.local.u64 	%rd3613, [%rd3612];
	shl.b64 	%rd3614, %rd3613, 9;
	mov.b64	{%r3550, %r3551}, %rd3606;
	mov.b64	%rd3615, {%r3550, %r3551};
	xor.b64  	%rd3616, %rd3615, %rd3614;
	shr.u64 	%rd3617, %rd3613, 55;
	mov.b64	{%r3552, %r3553}, %rd3609;
	mov.b64	%rd3618, {%r3552, %r3553};
	xor.b64  	%rd3619, %rd3618, %rd3617;
	shr.u64 	%rd3620, %rd3587, 9;
	and.b64  	%rd3621, %rd3620, 56;
	add.s64 	%rd3622, %rd12, %rd3621;
	ld.local.u64 	%rd3623, [%rd3622];
	shl.b64 	%rd3624, %rd3623, 12;
	mov.b64	{%r3554, %r3555}, %rd3616;
	mov.b64	%rd3625, {%r3554, %r3555};
	xor.b64  	%rd3626, %rd3625, %rd3624;
	shr.u64 	%rd3627, %rd3623, 52;
	mov.b64	{%r3556, %r3557}, %rd3619;
	mov.b64	%rd3628, {%r3556, %r3557};
	xor.b64  	%rd3629, %rd3628, %rd3627;
	shr.u64 	%rd3630, %rd3587, 12;
	and.b64  	%rd3631, %rd3630, 56;
	add.s64 	%rd3632, %rd12, %rd3631;
	ld.local.u64 	%rd3633, [%rd3632];
	shl.b64 	%rd3634, %rd3633, 15;
	mov.b64	{%r3558, %r3559}, %rd3626;
	mov.b64	%rd3635, {%r3558, %r3559};
	xor.b64  	%rd3636, %rd3635, %rd3634;
	shr.u64 	%rd3637, %rd3633, 49;
	mov.b64	{%r3560, %r3561}, %rd3629;
	mov.b64	%rd3638, {%r3560, %r3561};
	xor.b64  	%rd3639, %rd3638, %rd3637;
	shr.u64 	%rd3640, %rd3587, 15;
	and.b64  	%rd3641, %rd3640, 56;
	add.s64 	%rd3642, %rd12, %rd3641;
	ld.local.u64 	%rd3643, [%rd3642];
	shl.b64 	%rd3644, %rd3643, 18;
	mov.b64	{%r3562, %r3563}, %rd3636;
	mov.b64	%rd3645, {%r3562, %r3563};
	xor.b64  	%rd3646, %rd3645, %rd3644;
	shr.u64 	%rd3647, %rd3643, 46;
	mov.b64	{%r3564, %r3565}, %rd3639;
	mov.b64	%rd3648, {%r3564, %r3565};
	xor.b64  	%rd3649, %rd3648, %rd3647;
	shr.u64 	%rd3650, %rd3587, 18;
	and.b64  	%rd3651, %rd3650, 56;
	add.s64 	%rd3652, %rd12, %rd3651;
	ld.local.u64 	%rd3653, [%rd3652];
	shl.b64 	%rd3654, %rd3653, 21;
	mov.b64	{%r3566, %r3567}, %rd3646;
	mov.b64	%rd3655, {%r3566, %r3567};
	xor.b64  	%rd3656, %rd3655, %rd3654;
	shr.u64 	%rd3657, %rd3653, 43;
	mov.b64	{%r3568, %r3569}, %rd3649;
	mov.b64	%rd3658, {%r3568, %r3569};
	xor.b64  	%rd3659, %rd3658, %rd3657;
	shr.u64 	%rd3660, %rd3587, 21;
	and.b64  	%rd3661, %rd3660, 56;
	add.s64 	%rd3662, %rd12, %rd3661;
	ld.local.u64 	%rd3663, [%rd3662];
	shl.b64 	%rd3664, %rd3663, 24;
	mov.b64	{%r3570, %r3571}, %rd3656;
	mov.b64	%rd3665, {%r3570, %r3571};
	xor.b64  	%rd3666, %rd3665, %rd3664;
	shr.u64 	%rd3667, %rd3663, 40;
	mov.b64	{%r3572, %r3573}, %rd3659;
	mov.b64	%rd3668, {%r3572, %r3573};
	xor.b64  	%rd3669, %rd3668, %rd3667;
	shr.u64 	%rd3670, %rd3587, 24;
	and.b64  	%rd3671, %rd3670, 56;
	add.s64 	%rd3672, %rd12, %rd3671;
	ld.local.u64 	%rd3673, [%rd3672];
	shl.b64 	%rd3674, %rd3673, 27;
	mov.b64	{%r3574, %r3575}, %rd3666;
	mov.b64	%rd3675, {%r3574, %r3575};
	xor.b64  	%rd3676, %rd3675, %rd3674;
	shr.u64 	%rd3677, %rd3673, 37;
	mov.b64	{%r3576, %r3577}, %rd3669;
	mov.b64	%rd3678, {%r3576, %r3577};
	xor.b64  	%rd3679, %rd3678, %rd3677;
	shr.u64 	%rd3680, %rd3587, 27;
	and.b64  	%rd3681, %rd3680, 56;
	add.s64 	%rd3682, %rd12, %rd3681;
	ld.local.u64 	%rd3683, [%rd3682];
	shl.b64 	%rd3684, %rd3683, 30;
	mov.b64	{%r3578, %r3579}, %rd3676;
	mov.b64	%rd3685, {%r3578, %r3579};
	xor.b64  	%rd3686, %rd3685, %rd3684;
	shr.u64 	%rd3687, %rd3683, 34;
	mov.b64	{%r3580, %r3581}, %rd3679;
	mov.b64	%rd3688, {%r3580, %r3581};
	xor.b64  	%rd3689, %rd3688, %rd3687;
	shr.u64 	%rd3690, %rd3587, 30;
	and.b64  	%rd3691, %rd3690, 56;
	add.s64 	%rd3692, %rd12, %rd3691;
	ld.local.u64 	%rd3693, [%rd3692];
	shl.b64 	%rd3694, %rd3693, 33;
	mov.b64	{%r3582, %r3583}, %rd3686;
	mov.b64	%rd3695, {%r3582, %r3583};
	xor.b64  	%rd3696, %rd3695, %rd3694;
	shr.u64 	%rd3697, %rd3693, 31;
	mov.b64	{%r3584, %r3585}, %rd3689;
	mov.b64	%rd3698, {%r3584, %r3585};
	xor.b64  	%rd3699, %rd3698, %rd3697;
	shr.u64 	%rd3700, %rd3587, 33;
	and.b64  	%rd3701, %rd3700, 56;
	add.s64 	%rd3702, %rd12, %rd3701;
	ld.local.u64 	%rd3703, [%rd3702];
	shl.b64 	%rd3704, %rd3703, 36;
	mov.b64	{%r3586, %r3587}, %rd3696;
	mov.b64	%rd3705, {%r3586, %r3587};
	xor.b64  	%rd3706, %rd3705, %rd3704;
	shr.u64 	%rd3707, %rd3703, 28;
	mov.b64	{%r3588, %r3589}, %rd3699;
	mov.b64	%rd3708, {%r3588, %r3589};
	xor.b64  	%rd3709, %rd3708, %rd3707;
	shr.u64 	%rd3710, %rd3587, 36;
	and.b64  	%rd3711, %rd3710, 56;
	add.s64 	%rd3712, %rd12, %rd3711;
	ld.local.u64 	%rd3713, [%rd3712];
	shl.b64 	%rd3714, %rd3713, 39;
	mov.b64	{%r3590, %r3591}, %rd3706;
	mov.b64	%rd3715, {%r3590, %r3591};
	xor.b64  	%rd3716, %rd3715, %rd3714;
	shr.u64 	%rd3717, %rd3713, 25;
	mov.b64	{%r3592, %r3593}, %rd3709;
	mov.b64	%rd3718, {%r3592, %r3593};
	xor.b64  	%rd3719, %rd3718, %rd3717;
	shr.u64 	%rd3720, %rd3587, 39;
	and.b64  	%rd3721, %rd3720, 56;
	add.s64 	%rd3722, %rd12, %rd3721;
	ld.local.u64 	%rd3723, [%rd3722];
	shl.b64 	%rd3724, %rd3723, 42;
	mov.b64	{%r3594, %r3595}, %rd3716;
	mov.b64	%rd3725, {%r3594, %r3595};
	xor.b64  	%rd3726, %rd3725, %rd3724;
	shr.u64 	%rd3727, %rd3723, 22;
	mov.b64	{%r3596, %r3597}, %rd3719;
	mov.b64	%rd3728, {%r3596, %r3597};
	xor.b64  	%rd3729, %rd3728, %rd3727;
	shr.u64 	%rd3730, %rd3587, 42;
	and.b64  	%rd3731, %rd3730, 56;
	add.s64 	%rd3732, %rd12, %rd3731;
	ld.local.u64 	%rd3733, [%rd3732];
	shl.b64 	%rd3734, %rd3733, 45;
	mov.b64	{%r3598, %r3599}, %rd3726;
	mov.b64	%rd3735, {%r3598, %r3599};
	xor.b64  	%rd3736, %rd3735, %rd3734;
	shr.u64 	%rd3737, %rd3733, 19;
	mov.b64	{%r3600, %r3601}, %rd3729;
	mov.b64	%rd3738, {%r3600, %r3601};
	xor.b64  	%rd3739, %rd3738, %rd3737;
	shr.u64 	%rd3740, %rd3587, 45;
	and.b64  	%rd3741, %rd3740, 56;
	add.s64 	%rd3742, %rd12, %rd3741;
	ld.local.u64 	%rd3743, [%rd3742];
	shl.b64 	%rd3744, %rd3743, 48;
	mov.b64	{%r3602, %r3603}, %rd3736;
	mov.b64	%rd3745, {%r3602, %r3603};
	xor.b64  	%rd3746, %rd3745, %rd3744;
	shr.u64 	%rd3747, %rd3743, 16;
	mov.b64	{%r3604, %r3605}, %rd3739;
	mov.b64	%rd3748, {%r3604, %r3605};
	xor.b64  	%rd3749, %rd3748, %rd3747;
	shr.u64 	%rd3750, %rd3587, 48;
	and.b64  	%rd3751, %rd3750, 56;
	add.s64 	%rd3752, %rd12, %rd3751;
	ld.local.u64 	%rd3753, [%rd3752];
	shl.b64 	%rd3754, %rd3753, 51;
	mov.b64	{%r3606, %r3607}, %rd3746;
	mov.b64	%rd3755, {%r3606, %r3607};
	xor.b64  	%rd3756, %rd3755, %rd3754;
	shr.u64 	%rd3757, %rd3753, 13;
	mov.b64	{%r3608, %r3609}, %rd3749;
	mov.b64	%rd3758, {%r3608, %r3609};
	xor.b64  	%rd3759, %rd3758, %rd3757;
	shr.u64 	%rd3760, %rd3587, 51;
	and.b64  	%rd3761, %rd3760, 56;
	add.s64 	%rd3762, %rd12, %rd3761;
	ld.local.u64 	%rd3763, [%rd3762];
	shl.b64 	%rd3764, %rd3763, 54;
	mov.b64	{%r3610, %r3611}, %rd3756;
	mov.b64	%rd3765, {%r3610, %r3611};
	xor.b64  	%rd3766, %rd3765, %rd3764;
	shr.u64 	%rd3767, %rd3763, 10;
	mov.b64	{%r3612, %r3613}, %rd3759;
	mov.b64	%rd3768, {%r3612, %r3613};
	xor.b64  	%rd3769, %rd3768, %rd3767;
	shr.u64 	%rd3770, %rd3587, 54;
	and.b64  	%rd3771, %rd3770, 56;
	add.s64 	%rd3772, %rd12, %rd3771;
	ld.local.u64 	%rd3773, [%rd3772];
	shl.b64 	%rd3774, %rd3773, 57;
	mov.b64	{%r3614, %r3615}, %rd3766;
	mov.b64	%rd3775, {%r3614, %r3615};
	xor.b64  	%rd3776, %rd3775, %rd3774;
	shr.u64 	%rd3777, %rd3773, 7;
	mov.b64	{%r3616, %r3617}, %rd3769;
	mov.b64	%rd3778, {%r3616, %r3617};
	xor.b64  	%rd3779, %rd3778, %rd3777;
	shr.u64 	%rd3780, %rd3587, 57;
	and.b64  	%rd3781, %rd3780, 56;
	add.s64 	%rd3782, %rd12, %rd3781;
	ld.local.u64 	%rd3783, [%rd3782];
	shl.b64 	%rd3784, %rd3783, 60;
	mov.b64	{%r3618, %r3619}, %rd3776;
	mov.b64	%rd3785, {%r3618, %r3619};
	xor.b64  	%rd3786, %rd3785, %rd3784;
	shr.u64 	%rd3787, %rd3783, 4;
	mov.b64	{%r3620, %r3621}, %rd3779;
	mov.b64	%rd3788, {%r3620, %r3621};
	xor.b64  	%rd3789, %rd3788, %rd3787;
	shr.u64 	%rd3790, %rd3587, 60;
	and.b64  	%rd3791, %rd3790, 8;
	add.s64 	%rd3792, %rd12, %rd3791;
	ld.local.u64 	%rd3793, [%rd3792];
	shl.b64 	%rd3794, %rd3793, 63;
	mov.b64	{%r3622, %r3623}, %rd3786;
	mov.b64	%rd3795, {%r3622, %r3623};
	xor.b64  	%rd3796, %rd3795, %rd3794;
	shr.u64 	%rd3797, %rd3793, 1;
	mov.b64	{%r3624, %r3625}, %rd3789;
	mov.b64	%rd3798, {%r3624, %r3625};
	xor.b64  	%rd3799, %rd3798, %rd3797;
	mov.b64	{%r6657, %r6658}, %rd3799;
	mov.b64	{%r6659, %r6660}, %rd3796;
	setp.gt.s32	%p55, %r542, -1;
	@%p55 bra 	BB0_99;

	shr.u32 	%r3626, %r539, 1;
	shl.b32 	%r3627, %r540, 31;
	shr.u32 	%r3628, %r540, 1;
	and.b32  	%r3629, %r3626, 1533916891;
	or.b32  	%r3630, %r3629, %r3627;
	xor.b32  	%r6657, %r6657, %r3630;
	and.b32  	%r3631, %r3628, 920350134;
	xor.b32  	%r6658, %r6658, %r3631;

BB0_99:
	and.b32  	%r3632, %r542, 1073741824;
	setp.eq.s32	%p56, %r3632, 0;
	@%p56 bra 	BB0_101;

	shr.u32 	%r3633, %r539, 2;
	shl.b32 	%r3634, %r540, 30;
	or.b32  	%r3635, %r3633, %r3634;
	shr.u32 	%r3636, %r540, 2;
	and.b32  	%r3637, %r3635, 1227133513;
	xor.b32  	%r6657, %r6657, %r3637;
	and.b32  	%r3638, %r3636, 306783378;
	xor.b32  	%r6658, %r6658, %r3638;

BB0_101:
	xor.b32  	%r6730, %r6659, %r6730;
	xor.b32  	%r6731, %r6660, %r6731;
	xor.b32  	%r6732, %r6657, %r6732;
	xor.b32  	%r6733, %r6658, %r6733;
	bra.uni 	BB0_102;

BB0_96:
	setp.eq.s64	%p53, %rd62, 0;
	selp.b64	%rd3397, %rd6122, %rd56, %p53;
	add.s64 	%rd63, %rd6113, 1;
	shl.b64 	%rd3398, %rd6113, 6;
	add.s64 	%rd3399, %rd3398, %rd6112;
	and.b32  	%r3316, %r3312, 255;
	mul.wide.u32 	%rd3400, %r3316, 4;
	mov.u64 	%rd3401, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd3402, %rd3401, %rd3400;
	shr.u32 	%r3317, %r3312, 8;
	mul.wide.u32 	%rd3403, %r3313, 4;
	and.b64  	%rd3404, %rd3403, 1020;
	add.s64 	%rd3405, %rd3401, %rd3404;
	shr.u32 	%r3318, %r3313, 8;
	mul.wide.u32 	%rd3406, %r3314, 4;
	and.b64  	%rd3407, %rd3406, 1020;
	add.s64 	%rd3408, %rd3401, %rd3407;
	shr.u32 	%r3319, %r3314, 8;
	mul.wide.u32 	%rd3409, %r3315, 4;
	and.b64  	%rd3410, %rd3409, 1020;
	add.s64 	%rd3411, %rd3401, %rd3410;
	shr.u32 	%r3320, %r3315, 8;
	mul.wide.u32 	%rd3412, %r3318, 4;
	and.b64  	%rd3413, %rd3412, 1020;
	add.s64 	%rd3414, %rd3413, %rd3401;
	ld.shared.u32 	%r3321, [%rd3414+1024];
	ld.shared.u32 	%r3322, [%rd3402];
	xor.b32  	%r3323, %r3321, %r3322;
	shr.u32 	%r3324, %r3313, 16;
	mul.wide.u32 	%rd3415, %r3319, 4;
	and.b64  	%rd3416, %rd3415, 1020;
	add.s64 	%rd3417, %rd3416, %rd3401;
	ld.shared.u32 	%r3325, [%rd3417+1024];
	ld.shared.u32 	%r3326, [%rd3405];
	xor.b32  	%r3327, %r3325, %r3326;
	shr.u32 	%r3328, %r3314, 16;
	mul.wide.u32 	%rd3418, %r3320, 4;
	and.b64  	%rd3419, %rd3418, 1020;
	add.s64 	%rd3420, %rd3419, %rd3401;
	ld.shared.u32 	%r3329, [%rd3420+1024];
	ld.shared.u32 	%r3330, [%rd3408];
	xor.b32  	%r3331, %r3329, %r3330;
	shr.u32 	%r3332, %r3315, 16;
	mul.wide.u32 	%rd3421, %r3317, 4;
	and.b64  	%rd3422, %rd3421, 1020;
	add.s64 	%rd3423, %rd3422, %rd3401;
	ld.shared.u32 	%r3333, [%rd3423+1024];
	ld.shared.u32 	%r3334, [%rd3411];
	xor.b32  	%r3335, %r3333, %r3334;
	shr.u32 	%r3336, %r3312, 16;
	mul.wide.u32 	%rd3424, %r3328, 4;
	and.b64  	%rd3425, %rd3424, 1020;
	add.s64 	%rd3426, %rd3425, %rd3401;
	ld.shared.u32 	%r3337, [%rd3426+2048];
	xor.b32  	%r3338, %r3323, %r3337;
	shr.u32 	%r3339, %r3314, 24;
	mul.wide.u32 	%rd3427, %r3332, 4;
	and.b64  	%rd3428, %rd3427, 1020;
	add.s64 	%rd3429, %rd3428, %rd3401;
	ld.shared.u32 	%r3340, [%rd3429+2048];
	xor.b32  	%r3341, %r3327, %r3340;
	shr.u32 	%r3342, %r3315, 24;
	mul.wide.u32 	%rd3430, %r3336, 4;
	and.b64  	%rd3431, %rd3430, 1020;
	add.s64 	%rd3432, %rd3431, %rd3401;
	ld.shared.u32 	%r3343, [%rd3432+2048];
	xor.b32  	%r3344, %r3331, %r3343;
	shr.u32 	%r3345, %r3312, 24;
	mul.wide.u32 	%rd3433, %r3324, 4;
	and.b64  	%rd3434, %rd3433, 1020;
	add.s64 	%rd3435, %rd3434, %rd3401;
	ld.shared.u32 	%r3346, [%rd3435+2048];
	xor.b32  	%r3347, %r3335, %r3346;
	shr.u32 	%r3348, %r3313, 24;
	mul.wide.u32 	%rd3436, %r3342, 4;
	add.s64 	%rd3437, %rd3436, %rd3401;
	ld.shared.u32 	%r3349, [%rd3437+3072];
	mul.wide.u32 	%rd3438, %r3345, 4;
	add.s64 	%rd3439, %rd3438, %rd3401;
	ld.shared.u32 	%r3350, [%rd3439+3072];
	mul.wide.u32 	%rd3440, %r3348, 4;
	add.s64 	%rd3441, %rd3440, %rd3401;
	ld.shared.u32 	%r3351, [%rd3441+3072];
	mul.wide.u32 	%rd3442, %r3339, 4;
	add.s64 	%rd3443, %rd3442, %rd3401;
	ld.shared.u32 	%r3352, [%rd3443+3072];
	xor.b32  	%r3353, %r3347, %r3352;
	xor.b32  	%r3354, %r3344, %r3351;
	xor.b32  	%r3355, %r3341, %r3350;
	xor.b32  	%r3356, %r3338, %r3349;
	ld.global.v4.u32 	{%r3357, %r3358, %r3359, %r3360}, [%rd3399+16];
	ld.local.v4.u32 	{%r3365, %r3366, %r3367, %r3368}, [%rd3397];
	mul.wide.u32 	%rd3444, %r3365, 4;
	and.b64  	%rd3445, %rd3444, 1020;
	add.s64 	%rd3446, %rd3401, %rd3445;
	shr.u32 	%r3370, %r3365, 8;
	mul.wide.u32 	%rd3447, %r3366, 4;
	and.b64  	%rd3448, %rd3447, 1020;
	add.s64 	%rd3449, %rd3401, %rd3448;
	shr.u32 	%r3372, %r3366, 8;
	mul.wide.u32 	%rd3450, %r3367, 4;
	and.b64  	%rd3451, %rd3450, 1020;
	add.s64 	%rd3452, %rd3401, %rd3451;
	shr.u32 	%r3374, %r3367, 8;
	mul.wide.u32 	%rd3453, %r3368, 4;
	and.b64  	%rd3454, %rd3453, 1020;
	add.s64 	%rd3455, %rd3401, %rd3454;
	shr.u32 	%r3376, %r3368, 8;
	mul.wide.u32 	%rd3456, %r3372, 4;
	and.b64  	%rd3457, %rd3456, 1020;
	add.s64 	%rd3458, %rd3457, %rd3401;
	ld.shared.u32 	%r3377, [%rd3458+1024];
	ld.shared.u32 	%r3378, [%rd3446];
	xor.b32  	%r3379, %r3377, %r3378;
	shr.u32 	%r3380, %r3366, 16;
	mul.wide.u32 	%rd3459, %r3374, 4;
	and.b64  	%rd3460, %rd3459, 1020;
	add.s64 	%rd3461, %rd3460, %rd3401;
	ld.shared.u32 	%r3381, [%rd3461+1024];
	ld.shared.u32 	%r3382, [%rd3449];
	xor.b32  	%r3383, %r3381, %r3382;
	shr.u32 	%r3384, %r3367, 16;
	mul.wide.u32 	%rd3462, %r3376, 4;
	and.b64  	%rd3463, %rd3462, 1020;
	add.s64 	%rd3464, %rd3463, %rd3401;
	ld.shared.u32 	%r3385, [%rd3464+1024];
	ld.shared.u32 	%r3386, [%rd3452];
	xor.b32  	%r3387, %r3385, %r3386;
	shr.u32 	%r3388, %r3368, 16;
	mul.wide.u32 	%rd3465, %r3370, 4;
	and.b64  	%rd3466, %rd3465, 1020;
	add.s64 	%rd3467, %rd3466, %rd3401;
	ld.shared.u32 	%r3389, [%rd3467+1024];
	ld.shared.u32 	%r3390, [%rd3455];
	xor.b32  	%r3391, %r3389, %r3390;
	shr.u32 	%r3392, %r3365, 16;
	mul.wide.u32 	%rd3468, %r3384, 4;
	and.b64  	%rd3469, %rd3468, 1020;
	add.s64 	%rd3470, %rd3469, %rd3401;
	ld.shared.u32 	%r3393, [%rd3470+2048];
	xor.b32  	%r3394, %r3379, %r3393;
	shr.u32 	%r3395, %r3367, 24;
	mul.wide.u32 	%rd3471, %r3388, 4;
	and.b64  	%rd3472, %rd3471, 1020;
	add.s64 	%rd3473, %rd3472, %rd3401;
	ld.shared.u32 	%r3396, [%rd3473+2048];
	xor.b32  	%r3397, %r3383, %r3396;
	shr.u32 	%r3398, %r3368, 24;
	mul.wide.u32 	%rd3474, %r3392, 4;
	and.b64  	%rd3475, %rd3474, 1020;
	add.s64 	%rd3476, %rd3475, %rd3401;
	ld.shared.u32 	%r3399, [%rd3476+2048];
	xor.b32  	%r3400, %r3387, %r3399;
	shr.u32 	%r3401, %r3365, 24;
	mul.wide.u32 	%rd3477, %r3380, 4;
	and.b64  	%rd3478, %rd3477, 1020;
	add.s64 	%rd3479, %rd3478, %rd3401;
	ld.shared.u32 	%r3402, [%rd3479+2048];
	xor.b32  	%r3403, %r3391, %r3402;
	shr.u32 	%r3404, %r3366, 24;
	mul.wide.u32 	%rd3480, %r3398, 4;
	add.s64 	%rd3481, %rd3480, %rd3401;
	ld.shared.u32 	%r3405, [%rd3481+3072];
	mul.wide.u32 	%rd3482, %r3401, 4;
	add.s64 	%rd3483, %rd3482, %rd3401;
	ld.shared.u32 	%r3406, [%rd3483+3072];
	mul.wide.u32 	%rd3484, %r3404, 4;
	add.s64 	%rd3485, %rd3484, %rd3401;
	ld.shared.u32 	%r3407, [%rd3485+3072];
	mul.wide.u32 	%rd3486, %r3395, 4;
	add.s64 	%rd3487, %rd3486, %rd3401;
	ld.shared.u32 	%r3408, [%rd3487+3072];
	xor.b32  	%r3409, %r3403, %r3408;
	xor.b32  	%r3410, %r3400, %r3407;
	xor.b32  	%r3411, %r3397, %r3406;
	xor.b32  	%r3412, %r3394, %r3405;
	ld.global.v4.u32 	{%r3413, %r3414, %r3415, %r3416}, [%rd3399+32];
	xor.b32  	%r3421, %r3356, %r3357;
	mul.wide.u32 	%rd3488, %r3421, 4;
	and.b64  	%rd3489, %rd3488, 1020;
	add.s64 	%rd3490, %rd3401, %rd3489;
	shr.u32 	%r3422, %r3421, 8;
	xor.b32  	%r3423, %r3355, %r3358;
	mul.wide.u32 	%rd3491, %r3423, 4;
	and.b64  	%rd3492, %rd3491, 1020;
	add.s64 	%rd3493, %rd3401, %rd3492;
	shr.u32 	%r3424, %r3423, 8;
	xor.b32  	%r3425, %r3354, %r3359;
	mul.wide.u32 	%rd3494, %r3425, 4;
	and.b64  	%rd3495, %rd3494, 1020;
	add.s64 	%rd3496, %rd3401, %rd3495;
	shr.u32 	%r3426, %r3425, 8;
	xor.b32  	%r3427, %r3353, %r3360;
	mul.wide.u32 	%rd3497, %r3427, 4;
	and.b64  	%rd3498, %rd3497, 1020;
	add.s64 	%rd3499, %rd3401, %rd3498;
	shr.u32 	%r3428, %r3427, 8;
	mul.wide.u32 	%rd3500, %r3424, 4;
	and.b64  	%rd3501, %rd3500, 1020;
	add.s64 	%rd3502, %rd3501, %rd3401;
	ld.shared.u32 	%r3429, [%rd3502+1024];
	ld.shared.u32 	%r3430, [%rd3490];
	xor.b32  	%r3431, %r3429, %r3430;
	shr.u32 	%r3432, %r3423, 16;
	mul.wide.u32 	%rd3503, %r3426, 4;
	and.b64  	%rd3504, %rd3503, 1020;
	add.s64 	%rd3505, %rd3504, %rd3401;
	ld.shared.u32 	%r3433, [%rd3505+1024];
	ld.shared.u32 	%r3434, [%rd3493];
	xor.b32  	%r3435, %r3433, %r3434;
	shr.u32 	%r3436, %r3425, 16;
	mul.wide.u32 	%rd3506, %r3428, 4;
	and.b64  	%rd3507, %rd3506, 1020;
	add.s64 	%rd3508, %rd3507, %rd3401;
	ld.shared.u32 	%r3437, [%rd3508+1024];
	ld.shared.u32 	%r3438, [%rd3496];
	xor.b32  	%r3439, %r3437, %r3438;
	shr.u32 	%r3440, %r3427, 16;
	mul.wide.u32 	%rd3509, %r3422, 4;
	and.b64  	%rd3510, %rd3509, 1020;
	add.s64 	%rd3511, %rd3510, %rd3401;
	ld.shared.u32 	%r3441, [%rd3511+1024];
	ld.shared.u32 	%r3442, [%rd3499];
	xor.b32  	%r3443, %r3441, %r3442;
	shr.u32 	%r3444, %r3421, 16;
	mul.wide.u32 	%rd3512, %r3436, 4;
	and.b64  	%rd3513, %rd3512, 1020;
	add.s64 	%rd3514, %rd3513, %rd3401;
	ld.shared.u32 	%r3445, [%rd3514+2048];
	xor.b32  	%r3446, %r3431, %r3445;
	shr.u32 	%r3447, %r3425, 24;
	mul.wide.u32 	%rd3515, %r3440, 4;
	and.b64  	%rd3516, %rd3515, 1020;
	add.s64 	%rd3517, %rd3516, %rd3401;
	ld.shared.u32 	%r3448, [%rd3517+2048];
	xor.b32  	%r3449, %r3435, %r3448;
	shr.u32 	%r3450, %r3427, 24;
	mul.wide.u32 	%rd3518, %r3444, 4;
	and.b64  	%rd3519, %rd3518, 1020;
	add.s64 	%rd3520, %rd3519, %rd3401;
	ld.shared.u32 	%r3451, [%rd3520+2048];
	xor.b32  	%r3452, %r3439, %r3451;
	shr.u32 	%r3453, %r3421, 24;
	mul.wide.u32 	%rd3521, %r3432, 4;
	and.b64  	%rd3522, %rd3521, 1020;
	add.s64 	%rd3523, %rd3522, %rd3401;
	ld.shared.u32 	%r3454, [%rd3523+2048];
	xor.b32  	%r3455, %r3443, %r3454;
	shr.u32 	%r3456, %r3423, 24;
	mul.wide.u32 	%rd3524, %r3450, 4;
	add.s64 	%rd3525, %rd3524, %rd3401;
	ld.shared.u32 	%r3457, [%rd3525+3072];
	mul.wide.u32 	%rd3526, %r3453, 4;
	add.s64 	%rd3527, %rd3526, %rd3401;
	ld.shared.u32 	%r3458, [%rd3527+3072];
	mul.wide.u32 	%rd3528, %r3456, 4;
	add.s64 	%rd3529, %rd3528, %rd3401;
	ld.shared.u32 	%r3459, [%rd3529+3072];
	mul.wide.u32 	%rd3530, %r3447, 4;
	add.s64 	%rd3531, %rd3530, %rd3401;
	ld.shared.u32 	%r3460, [%rd3531+3072];
	xor.b32  	%r3461, %r3455, %r3460;
	xor.b32  	%r3462, %r3449, %r3458;
	xor.b32  	%r3463, %r3446, %r3457;
	xor.b32  	%r3464, %r3452, %r3459;
	ld.global.v4.u32 	{%r3465, %r3466, %r3467, %r3468}, [%rd3399+48];
	xor.b32  	%r3473, %r3464, %r3467;
	xor.b32  	%r3474, %r3463, %r3465;
	xor.b32  	%r3475, %r3412, %r3413;
	mul.wide.u32 	%rd3532, %r3475, 4;
	and.b64  	%rd3533, %rd3532, 1020;
	add.s64 	%rd3534, %rd3401, %rd3533;
	shr.u32 	%r3476, %r3475, 8;
	xor.b32  	%r3477, %r3411, %r3414;
	mul.wide.u32 	%rd3535, %r3477, 4;
	and.b64  	%rd3536, %rd3535, 1020;
	add.s64 	%rd3537, %rd3401, %rd3536;
	shr.u32 	%r3478, %r3477, 8;
	xor.b32  	%r3479, %r3410, %r3415;
	mul.wide.u32 	%rd3538, %r3479, 4;
	and.b64  	%rd3539, %rd3538, 1020;
	add.s64 	%rd3540, %rd3401, %rd3539;
	shr.u32 	%r3480, %r3479, 8;
	xor.b32  	%r3481, %r3409, %r3416;
	mul.wide.u32 	%rd3541, %r3481, 4;
	and.b64  	%rd3542, %rd3541, 1020;
	add.s64 	%rd3543, %rd3401, %rd3542;
	shr.u32 	%r3482, %r3481, 8;
	mul.wide.u32 	%rd3544, %r3478, 4;
	and.b64  	%rd3545, %rd3544, 1020;
	add.s64 	%rd3546, %rd3545, %rd3401;
	ld.shared.u32 	%r3483, [%rd3546+1024];
	ld.shared.u32 	%r3484, [%rd3534];
	xor.b32  	%r3485, %r3483, %r3484;
	shr.u32 	%r3486, %r3477, 16;
	mul.wide.u32 	%rd3547, %r3480, 4;
	and.b64  	%rd3548, %rd3547, 1020;
	add.s64 	%rd3549, %rd3548, %rd3401;
	ld.shared.u32 	%r3487, [%rd3549+1024];
	ld.shared.u32 	%r3488, [%rd3537];
	xor.b32  	%r3489, %r3487, %r3488;
	shr.u32 	%r3490, %r3479, 16;
	mul.wide.u32 	%rd3550, %r3482, 4;
	and.b64  	%rd3551, %rd3550, 1020;
	add.s64 	%rd3552, %rd3551, %rd3401;
	ld.shared.u32 	%r3491, [%rd3552+1024];
	ld.shared.u32 	%r3492, [%rd3540];
	xor.b32  	%r3493, %r3491, %r3492;
	shr.u32 	%r3494, %r3481, 16;
	mul.wide.u32 	%rd3553, %r3476, 4;
	and.b64  	%rd3554, %rd3553, 1020;
	add.s64 	%rd3555, %rd3554, %rd3401;
	ld.shared.u32 	%r3495, [%rd3555+1024];
	ld.shared.u32 	%r3496, [%rd3543];
	xor.b32  	%r3497, %r3495, %r3496;
	shr.u32 	%r3498, %r3475, 16;
	mul.wide.u32 	%rd3556, %r3490, 4;
	and.b64  	%rd3557, %rd3556, 1020;
	add.s64 	%rd3558, %rd3557, %rd3401;
	ld.shared.u32 	%r3499, [%rd3558+2048];
	xor.b32  	%r3500, %r3485, %r3499;
	shr.u32 	%r3501, %r3479, 24;
	mul.wide.u32 	%rd3559, %r3494, 4;
	and.b64  	%rd3560, %rd3559, 1020;
	add.s64 	%rd3561, %rd3560, %rd3401;
	ld.shared.u32 	%r3502, [%rd3561+2048];
	xor.b32  	%r3503, %r3489, %r3502;
	shr.u32 	%r3504, %r3481, 24;
	mul.wide.u32 	%rd3562, %r3498, 4;
	and.b64  	%rd3563, %rd3562, 1020;
	add.s64 	%rd3564, %rd3563, %rd3401;
	ld.shared.u32 	%r3505, [%rd3564+2048];
	xor.b32  	%r3506, %r3493, %r3505;
	shr.u32 	%r3507, %r3475, 24;
	mul.wide.u32 	%rd3565, %r3486, 4;
	and.b64  	%rd3566, %rd3565, 1020;
	add.s64 	%rd3567, %rd3566, %rd3401;
	ld.shared.u32 	%r3508, [%rd3567+2048];
	xor.b32  	%r3509, %r3497, %r3508;
	shr.u32 	%r3510, %r3477, 24;
	mul.wide.u32 	%rd3568, %r3504, 4;
	add.s64 	%rd3569, %rd3568, %rd3401;
	ld.shared.u32 	%r3511, [%rd3569+3072];
	mul.wide.u32 	%rd3570, %r3507, 4;
	add.s64 	%rd3571, %rd3570, %rd3401;
	ld.shared.u32 	%r3512, [%rd3571+3072];
	mul.wide.u32 	%rd3572, %r3510, 4;
	add.s64 	%rd3573, %rd3572, %rd3401;
	ld.shared.u32 	%r3513, [%rd3573+3072];
	mul.wide.u32 	%rd3574, %r3501, 4;
	add.s64 	%rd3575, %rd3574, %rd3401;
	ld.shared.u32 	%r3514, [%rd3575+3072];
	xor.b32  	%r3515, %r3500, %r3511;
	xor.b32  	%r3516, %r3503, %r3512;
	xor.b32  	%r3517, %r3509, %r3514;
	xor.b32  	%r3518, %r3506, %r3513;
	ld.global.v4.u32 	{%r3519, %r3520, %r3521, %r3522}, [%rd3399+64];
	xor.b32  	%r3527, %r3518, %r3521;
	xor.b32  	%r3528, %r3517, %r3522;
	xor.b32  	%r3529, %r3516, %r3520;
	xor.b32  	%r3530, %r3515, %r3519;
	xor.b32  	%r3531, %r3462, %r3466;
	xor.b32  	%r3532, %r3461, %r3468;
	xor.b32  	%r3533, %r3530, %r6731;
	xor.b32  	%r3534, %r3529, %r6733;
	xor.b32  	%r3535, %r3531, %r6732;
	xor.b32  	%r3536, %r3474, %r6730;
	xor.b32  	%r6733, %r3528, %r3534;
	xor.b32  	%r6731, %r3527, %r3533;
	xor.b32  	%r6730, %r3473, %r3536;
	xor.b32  	%r6732, %r3532, %r3535;
	mov.u64 	%rd6113, %rd63;

BB0_102:
	add.s64 	%rd6112, %rd6112, 16;
	add.s64 	%rd66, %rd6111, -1;
	setp.ne.s64	%p57, %rd6111, 0;
	mov.u64 	%rd6111, %rd66;
	@%p57 bra 	BB0_95;

	cvt.u64.u16	%rd6080, %rs236;
	add.s64 	%rd6079, %rd6080, %rd3;
	shl.b64 	%rd6078, %rd6079, 4;
	add.s64 	%rd6077, %rd203, %rd6078;
	mov.b64	%rd3800, {%r6732, %r6733};
	mov.b64	%rd3801, {%r6730, %r6731};
	cvt.u32.u64	%r3639, %rd3801;
	cvt.s32.s16 	%r3640, %r3639;
	mov.b64	%rd3802, {%r6723, %r6724};
	mov.b64	%rd3803, {%r6721, %r6722};
	cvt.u32.u64	%r3641, %rd3803;
	cvt.s32.s16 	%r3642, %r3641;
	mad.lo.s32 	%r3643, %r3640, %r3642, 16384;
	bfe.u32 	%r3644, %r3643, 15, 16;
	shr.s32 	%r3645, %r3639, 16;
	shr.s32 	%r3646, %r3641, 16;
	mad.lo.s32 	%r3647, %r3645, %r3646, 16384;
	mov.u32 	%r3648, 0;
	mov.b64	%rd3804, {%r3644, %r3648};
	shl.b32 	%r3649, %r3647, 1;
	cvt.u64.u32	%rd3805, %r3649;
	and.b64  	%rd3806, %rd3804, 65535;
	shr.u64 	%rd3807, %rd3800, 16;
	shr.u64 	%rd3808, %rd3801, 16;
	cvt.u32.u64	%r3650, %rd3808;
	shr.s32 	%r3651, %r3650, 16;
	shr.u64 	%rd3809, %rd3802, 16;
	shr.u64 	%rd3810, %rd3803, 16;
	cvt.u32.u64	%r3652, %rd3810;
	shr.s32 	%r3653, %r3652, 16;
	mad.lo.s32 	%r3654, %r3651, %r3653, 16384;
	mul.wide.u32 	%rd3811, %r3654, 131072;
	shr.u64 	%rd3812, %rd3801, 32;
	cvt.u32.u64	%r3655, %rd3812;
	shr.s32 	%r3656, %r3655, 16;
	shr.u64 	%rd3813, %rd3803, 32;
	cvt.u32.u64	%r3657, %rd3813;
	shr.s32 	%r3658, %r3657, 16;
	mad.lo.s32 	%r3659, %r3656, %r3658, 16384;
	shr.u32 	%r3660, %r3659, 15;
	cvt.u64.u32	%rd3814, %r3660;
	shl.b64 	%rd3815, %rd3814, 48;
	and.b64  	%rd3816, %rd3811, 281470681743360;
	shl.b64 	%rd3817, %rd3800, 16;
	cvt.u32.u64	%r3661, %rd3817;
	shr.s32 	%r3662, %r3661, 16;
	shl.b64 	%rd3818, %rd3802, 16;
	cvt.u32.u64	%r3663, %rd3818;
	shr.s32 	%r3664, %r3663, 16;
	mad.lo.s32 	%r3665, %r3662, %r3664, 16384;
	shr.u32 	%r3666, %r3665, 15;
	cvt.u64.u32	%rd3819, %r3666;
	cvt.u32.u64	%r3667, %rd3800;
	shr.s32 	%r3668, %r3667, 16;
	cvt.u32.u64	%r3669, %rd3802;
	shr.s32 	%r3670, %r3669, 16;
	mad.lo.s32 	%r3671, %r3668, %r3670, 16384;
	and.b64  	%rd3820, %rd3819, 65535;
	and.b64  	%rd3821, %rd3805, 4294901760;
	or.b64  	%rd3822, %rd3821, %rd3806;
	or.b64  	%rd3823, %rd3822, %rd3816;
	shl.b32 	%r3672, %r3671, 1;
	cvt.u64.u32	%rd3824, %r3672;
	cvt.u32.u64	%r3673, %rd3807;
	shr.s32 	%r3674, %r3673, 16;
	cvt.u32.u64	%r3675, %rd3809;
	shr.s32 	%r3676, %r3675, 16;
	mad.lo.s32 	%r3677, %r3674, %r3676, 16384;
	mul.wide.u32 	%rd3825, %r3677, 131072;
	and.b64  	%rd3826, %rd3824, 4294901760;
	or.b64  	%rd3827, %rd3826, %rd3820;
	shr.u64 	%rd3828, %rd3800, 32;
	cvt.u32.u64	%r3678, %rd3828;
	shr.s32 	%r3679, %r3678, 16;
	shr.u64 	%rd3829, %rd3802, 32;
	cvt.u32.u64	%r3680, %rd3829;
	shr.s32 	%r3681, %r3680, 16;
	mad.lo.s32 	%r3682, %r3679, %r3681, 16384;
	shr.u32 	%r3683, %r3682, 15;
	cvt.u64.u32	%rd3830, %r3683;
	shl.b64 	%rd3831, %rd3830, 48;
	and.b64  	%rd3832, %rd3825, 281470681743360;
	or.b64  	%rd3833, %rd3832, %rd3831;
	or.b64  	%rd3834, %rd3833, %rd3827;
	or.b64  	%rd3835, %rd3823, %rd3815;
	mov.b64	{%r3684, %r3685}, %rd3834;
	mov.b64	{%r3686, %r3687}, %rd3835;
	xor.b32  	%r6668, %r6722, %r3687;
	xor.b32  	%r6667, %r6721, %r3686;
	xor.b32  	%r6670, %r6724, %r3685;
	xor.b32  	%r6669, %r6723, %r3684;
	cvt.u64.u32	%rd3836, %r6729;
	add.s64 	%rd3837, %rd3836, %rd9;
	shl.b64 	%rd3838, %rd3837, 1;
	mov.u64 	%rd3839, verus_gpu_hash$sharedrand;
	add.s64 	%rd3840, %rd3839, %rd3838;
	st.shared.u16 	[%rd3840], %rs236;
	mov.u64 	%rd3841, verus_gpu_hash$sharedrandex;
	add.s64 	%rd3842, %rd3841, %rd3838;
	st.shared.u16 	[%rd3842], %rs237;
	st.global.v4.u32 	[%rd6077], {%r6638, %r6639, %r6640, %r6641};
	cvt.u64.u16	%rd3843, %rs237;
	add.s64 	%rd3844, %rd3843, %rd3;
	shl.b64 	%rd3845, %rd3844, 4;
	add.s64 	%rd3846, %rd203, %rd3845;
	st.global.v4.u32 	[%rd3846], {%r6667, %r6668, %r6669, %r6670};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p58, %r6729, 32;
	@%p58 bra 	BB0_134;

	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p59, %r6729, 0;
	mov.u16 	%rs238, 20;
	@%p59 bra 	BB0_105;
	bra.uni 	BB0_106;

BB0_105:
	mov.u32 	%r6721, %r6638;
	mov.u32 	%r6722, %r6639;
	mov.u32 	%r6723, %r6640;
	mov.u32 	%r6724, %r6641;
	mov.u32 	%r6729, %r3648;
	bra.uni 	BB0_107;

BB0_93:
	mov.u32 	%r6667, %r6638;
	mov.u32 	%r6668, %r6639;
	mov.u32 	%r6669, %r6640;
	mov.u32 	%r6670, %r6641;
	bra.uni 	BB0_107;

BB0_106:
	cvt.u64.u32	%rd3847, %r6731;
	shr.u32 	%r3689, %r6730, 5;
	cvt.u64.u32	%rd3848, %r3689;
	cvt.u16.u32	%rs96, %r3689;
	and.b16  	%rs236, %rs96, 511;
	cvt.u16.u32	%rs97, %r6731;
	and.b16  	%rs237, %rs97, 511;
	cvt.u16.u64	%rs98, %rd6123;
	and.b16  	%rs238, %rs98, 28;
	and.b64  	%rd3849, %rd3848, 511;
	add.s64 	%rd3850, %rd3849, %rd3;
	shl.b64 	%rd3851, %rd3850, 4;
	add.s64 	%rd3852, %rd203, %rd3851;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd3852];
	and.b64  	%rd3853, %rd3847, 511;
	add.s64 	%rd3854, %rd3853, %rd3;
	shl.b64 	%rd3855, %rd3854, 4;
	add.s64 	%rd3856, %rd203, %rd3855;
	ld.global.v4.u32 	{%r6667, %r6668, %r6669, %r6670}, [%rd3856];
	mul.wide.u32 	%rd3857, %r6730, 16;
	and.b64  	%rd3858, %rd3857, 48;
	add.s64 	%rd6122, %rd1, %rd3858;

BB0_107:
	and.b16  	%rs99, %rs238, 255;
	setp.ne.s16	%p60, %rs99, 24;
	@%p60 bra 	BB0_124;

	and.b64  	%rd3859, %rd6123, 1;
	mov.u64 	%rd3860, 1;
	shl.b64 	%rd3861, %rd3859, 1;
	sub.s64 	%rd3862, %rd3860, %rd3861;
	shl.b64 	%rd3863, %rd3862, 4;
	add.s64 	%rd71, %rd6122, %rd3863;
	shr.u64 	%rd6117, %rd6123, 61;
	cvt.u64.u16	%rd3864, %rs236;
	add.s64 	%rd3865, %rd3864, %rd3;
	shl.b64 	%rd3866, %rd3865, 4;
	add.s64 	%rd6118, %rd203, %rd3866;

BB0_109:
	cvt.u32.u64	%r3698, %rd6117;
	and.b32  	%r3699, %r3698, 63;
	mov.u64 	%rd3867, 268435456;
	shl.b64 	%rd3868, %rd3867, %r3699;
	and.b64  	%rd3869, %rd3868, %rd6123;
	setp.eq.s64	%p61, %rd3869, 0;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd6118];
	and.b64  	%rd77, %rd6117, 1;
	@%p61 bra 	BB0_114;
	bra.uni 	BB0_110;

BB0_114:
	setp.eq.s64	%p64, %rd77, 0;
	selp.b64	%rd3873, %rd6122, %rd71, %p64;
	ld.local.v4.u32 	{%r3715, %r3716, %r3717, %r3718}, [%rd3873];
	xor.b32  	%r613, %r6724, %r3718;
	xor.b32  	%r612, %r6723, %r3717;
	xor.b32  	%r611, %r6722, %r3716;
	xor.b32  	%r610, %r6721, %r3715;
	mov.b64	%rd3874, {%r612, %r613};
	mov.u64 	%rd3875, 0;
	st.local.v2.u64 	[%rd12], {%rd3875, %rd3874};
	shl.b64 	%rd3878, %rd3874, 1;
	xor.b64  	%rd3879, %rd3878, %rd3874;
	st.local.v2.u64 	[%rd15], {%rd3878, %rd3879};
	shl.b64 	%rd3880, %rd3874, 2;
	xor.b64  	%rd3881, %rd3880, %rd3874;
	st.local.v2.u64 	[%rd15+16], {%rd3880, %rd3881};
	shl.b64 	%rd3882, %rd3879, 1;
	xor.b64  	%rd3883, %rd3882, %rd3874;
	st.local.v2.u64 	[%rd15+32], {%rd3882, %rd3883};
	mov.b64	%rd3884, {%r610, %r611};
	shl.b64 	%rd3885, %rd3884, 3;
	and.b64  	%rd3886, %rd3885, 56;
	add.s64 	%rd3887, %rd12, %rd3886;
	and.b64  	%rd3888, %rd3884, 56;
	add.s64 	%rd3889, %rd12, %rd3888;
	ld.local.u64 	%rd3890, [%rd3889];
	shl.b64 	%rd3891, %rd3890, 3;
	ld.local.u64 	%rd3892, [%rd3887];
	xor.b64  	%rd3893, %rd3892, %rd3891;
	shr.u64 	%rd3894, %rd3890, 61;
	mov.u32 	%r3723, 0;
	mov.b64	%rd3895, {%r3723, %r3723};
	xor.b64  	%rd3896, %rd3895, %rd3894;
	shr.u64 	%rd3897, %rd3884, 3;
	and.b64  	%rd3898, %rd3897, 56;
	add.s64 	%rd3899, %rd12, %rd3898;
	ld.local.u64 	%rd3900, [%rd3899];
	shl.b64 	%rd3901, %rd3900, 6;
	mov.b64	{%r3724, %r3725}, %rd3893;
	mov.b64	%rd3902, {%r3724, %r3725};
	xor.b64  	%rd3903, %rd3902, %rd3901;
	shr.u64 	%rd3904, %rd3900, 58;
	mov.b64	{%r3726, %r3727}, %rd3896;
	mov.b64	%rd3905, {%r3726, %r3727};
	xor.b64  	%rd3906, %rd3905, %rd3904;
	shr.u64 	%rd3907, %rd3884, 6;
	and.b64  	%rd3908, %rd3907, 56;
	add.s64 	%rd3909, %rd12, %rd3908;
	ld.local.u64 	%rd3910, [%rd3909];
	shl.b64 	%rd3911, %rd3910, 9;
	mov.b64	{%r3728, %r3729}, %rd3903;
	mov.b64	%rd3912, {%r3728, %r3729};
	xor.b64  	%rd3913, %rd3912, %rd3911;
	shr.u64 	%rd3914, %rd3910, 55;
	mov.b64	{%r3730, %r3731}, %rd3906;
	mov.b64	%rd3915, {%r3730, %r3731};
	xor.b64  	%rd3916, %rd3915, %rd3914;
	shr.u64 	%rd3917, %rd3884, 9;
	and.b64  	%rd3918, %rd3917, 56;
	add.s64 	%rd3919, %rd12, %rd3918;
	ld.local.u64 	%rd3920, [%rd3919];
	shl.b64 	%rd3921, %rd3920, 12;
	mov.b64	{%r3732, %r3733}, %rd3913;
	mov.b64	%rd3922, {%r3732, %r3733};
	xor.b64  	%rd3923, %rd3922, %rd3921;
	shr.u64 	%rd3924, %rd3920, 52;
	mov.b64	{%r3734, %r3735}, %rd3916;
	mov.b64	%rd3925, {%r3734, %r3735};
	xor.b64  	%rd3926, %rd3925, %rd3924;
	shr.u64 	%rd3927, %rd3884, 12;
	and.b64  	%rd3928, %rd3927, 56;
	add.s64 	%rd3929, %rd12, %rd3928;
	ld.local.u64 	%rd3930, [%rd3929];
	shl.b64 	%rd3931, %rd3930, 15;
	mov.b64	{%r3736, %r3737}, %rd3923;
	mov.b64	%rd3932, {%r3736, %r3737};
	xor.b64  	%rd3933, %rd3932, %rd3931;
	shr.u64 	%rd3934, %rd3930, 49;
	mov.b64	{%r3738, %r3739}, %rd3926;
	mov.b64	%rd3935, {%r3738, %r3739};
	xor.b64  	%rd3936, %rd3935, %rd3934;
	shr.u64 	%rd3937, %rd3884, 15;
	and.b64  	%rd3938, %rd3937, 56;
	add.s64 	%rd3939, %rd12, %rd3938;
	ld.local.u64 	%rd3940, [%rd3939];
	shl.b64 	%rd3941, %rd3940, 18;
	mov.b64	{%r3740, %r3741}, %rd3933;
	mov.b64	%rd3942, {%r3740, %r3741};
	xor.b64  	%rd3943, %rd3942, %rd3941;
	shr.u64 	%rd3944, %rd3940, 46;
	mov.b64	{%r3742, %r3743}, %rd3936;
	mov.b64	%rd3945, {%r3742, %r3743};
	xor.b64  	%rd3946, %rd3945, %rd3944;
	shr.u64 	%rd3947, %rd3884, 18;
	and.b64  	%rd3948, %rd3947, 56;
	add.s64 	%rd3949, %rd12, %rd3948;
	ld.local.u64 	%rd3950, [%rd3949];
	shl.b64 	%rd3951, %rd3950, 21;
	mov.b64	{%r3744, %r3745}, %rd3943;
	mov.b64	%rd3952, {%r3744, %r3745};
	xor.b64  	%rd3953, %rd3952, %rd3951;
	shr.u64 	%rd3954, %rd3950, 43;
	mov.b64	{%r3746, %r3747}, %rd3946;
	mov.b64	%rd3955, {%r3746, %r3747};
	xor.b64  	%rd3956, %rd3955, %rd3954;
	shr.u64 	%rd3957, %rd3884, 21;
	and.b64  	%rd3958, %rd3957, 56;
	add.s64 	%rd3959, %rd12, %rd3958;
	ld.local.u64 	%rd3960, [%rd3959];
	shl.b64 	%rd3961, %rd3960, 24;
	mov.b64	{%r3748, %r3749}, %rd3953;
	mov.b64	%rd3962, {%r3748, %r3749};
	xor.b64  	%rd3963, %rd3962, %rd3961;
	shr.u64 	%rd3964, %rd3960, 40;
	mov.b64	{%r3750, %r3751}, %rd3956;
	mov.b64	%rd3965, {%r3750, %r3751};
	xor.b64  	%rd3966, %rd3965, %rd3964;
	shr.u64 	%rd3967, %rd3884, 24;
	and.b64  	%rd3968, %rd3967, 56;
	add.s64 	%rd3969, %rd12, %rd3968;
	ld.local.u64 	%rd3970, [%rd3969];
	shl.b64 	%rd3971, %rd3970, 27;
	mov.b64	{%r3752, %r3753}, %rd3963;
	mov.b64	%rd3972, {%r3752, %r3753};
	xor.b64  	%rd3973, %rd3972, %rd3971;
	shr.u64 	%rd3974, %rd3970, 37;
	mov.b64	{%r3754, %r3755}, %rd3966;
	mov.b64	%rd3975, {%r3754, %r3755};
	xor.b64  	%rd3976, %rd3975, %rd3974;
	shr.u64 	%rd3977, %rd3884, 27;
	and.b64  	%rd3978, %rd3977, 56;
	add.s64 	%rd3979, %rd12, %rd3978;
	ld.local.u64 	%rd3980, [%rd3979];
	shl.b64 	%rd3981, %rd3980, 30;
	mov.b64	{%r3756, %r3757}, %rd3973;
	mov.b64	%rd3982, {%r3756, %r3757};
	xor.b64  	%rd3983, %rd3982, %rd3981;
	shr.u64 	%rd3984, %rd3980, 34;
	mov.b64	{%r3758, %r3759}, %rd3976;
	mov.b64	%rd3985, {%r3758, %r3759};
	xor.b64  	%rd3986, %rd3985, %rd3984;
	shr.u64 	%rd3987, %rd3884, 30;
	and.b64  	%rd3988, %rd3987, 56;
	add.s64 	%rd3989, %rd12, %rd3988;
	ld.local.u64 	%rd3990, [%rd3989];
	shl.b64 	%rd3991, %rd3990, 33;
	mov.b64	{%r3760, %r3761}, %rd3983;
	mov.b64	%rd3992, {%r3760, %r3761};
	xor.b64  	%rd3993, %rd3992, %rd3991;
	shr.u64 	%rd3994, %rd3990, 31;
	mov.b64	{%r3762, %r3763}, %rd3986;
	mov.b64	%rd3995, {%r3762, %r3763};
	xor.b64  	%rd3996, %rd3995, %rd3994;
	shr.u64 	%rd3997, %rd3884, 33;
	and.b64  	%rd3998, %rd3997, 56;
	add.s64 	%rd3999, %rd12, %rd3998;
	ld.local.u64 	%rd4000, [%rd3999];
	shl.b64 	%rd4001, %rd4000, 36;
	mov.b64	{%r3764, %r3765}, %rd3993;
	mov.b64	%rd4002, {%r3764, %r3765};
	xor.b64  	%rd4003, %rd4002, %rd4001;
	shr.u64 	%rd4004, %rd4000, 28;
	mov.b64	{%r3766, %r3767}, %rd3996;
	mov.b64	%rd4005, {%r3766, %r3767};
	xor.b64  	%rd4006, %rd4005, %rd4004;
	shr.u64 	%rd4007, %rd3884, 36;
	and.b64  	%rd4008, %rd4007, 56;
	add.s64 	%rd4009, %rd12, %rd4008;
	ld.local.u64 	%rd4010, [%rd4009];
	shl.b64 	%rd4011, %rd4010, 39;
	mov.b64	{%r3768, %r3769}, %rd4003;
	mov.b64	%rd4012, {%r3768, %r3769};
	xor.b64  	%rd4013, %rd4012, %rd4011;
	shr.u64 	%rd4014, %rd4010, 25;
	mov.b64	{%r3770, %r3771}, %rd4006;
	mov.b64	%rd4015, {%r3770, %r3771};
	xor.b64  	%rd4016, %rd4015, %rd4014;
	shr.u64 	%rd4017, %rd3884, 39;
	and.b64  	%rd4018, %rd4017, 56;
	add.s64 	%rd4019, %rd12, %rd4018;
	ld.local.u64 	%rd4020, [%rd4019];
	shl.b64 	%rd4021, %rd4020, 42;
	mov.b64	{%r3772, %r3773}, %rd4013;
	mov.b64	%rd4022, {%r3772, %r3773};
	xor.b64  	%rd4023, %rd4022, %rd4021;
	shr.u64 	%rd4024, %rd4020, 22;
	mov.b64	{%r3774, %r3775}, %rd4016;
	mov.b64	%rd4025, {%r3774, %r3775};
	xor.b64  	%rd4026, %rd4025, %rd4024;
	shr.u64 	%rd4027, %rd3884, 42;
	and.b64  	%rd4028, %rd4027, 56;
	add.s64 	%rd4029, %rd12, %rd4028;
	ld.local.u64 	%rd4030, [%rd4029];
	shl.b64 	%rd4031, %rd4030, 45;
	mov.b64	{%r3776, %r3777}, %rd4023;
	mov.b64	%rd4032, {%r3776, %r3777};
	xor.b64  	%rd4033, %rd4032, %rd4031;
	shr.u64 	%rd4034, %rd4030, 19;
	mov.b64	{%r3778, %r3779}, %rd4026;
	mov.b64	%rd4035, {%r3778, %r3779};
	xor.b64  	%rd4036, %rd4035, %rd4034;
	shr.u64 	%rd4037, %rd3884, 45;
	and.b64  	%rd4038, %rd4037, 56;
	add.s64 	%rd4039, %rd12, %rd4038;
	ld.local.u64 	%rd4040, [%rd4039];
	shl.b64 	%rd4041, %rd4040, 48;
	mov.b64	{%r3780, %r3781}, %rd4033;
	mov.b64	%rd4042, {%r3780, %r3781};
	xor.b64  	%rd4043, %rd4042, %rd4041;
	shr.u64 	%rd4044, %rd4040, 16;
	mov.b64	{%r3782, %r3783}, %rd4036;
	mov.b64	%rd4045, {%r3782, %r3783};
	xor.b64  	%rd4046, %rd4045, %rd4044;
	shr.u64 	%rd4047, %rd3884, 48;
	and.b64  	%rd4048, %rd4047, 56;
	add.s64 	%rd4049, %rd12, %rd4048;
	ld.local.u64 	%rd4050, [%rd4049];
	shl.b64 	%rd4051, %rd4050, 51;
	mov.b64	{%r3784, %r3785}, %rd4043;
	mov.b64	%rd4052, {%r3784, %r3785};
	xor.b64  	%rd4053, %rd4052, %rd4051;
	shr.u64 	%rd4054, %rd4050, 13;
	mov.b64	{%r3786, %r3787}, %rd4046;
	mov.b64	%rd4055, {%r3786, %r3787};
	xor.b64  	%rd4056, %rd4055, %rd4054;
	shr.u64 	%rd4057, %rd3884, 51;
	and.b64  	%rd4058, %rd4057, 56;
	add.s64 	%rd4059, %rd12, %rd4058;
	ld.local.u64 	%rd4060, [%rd4059];
	shl.b64 	%rd4061, %rd4060, 54;
	mov.b64	{%r3788, %r3789}, %rd4053;
	mov.b64	%rd4062, {%r3788, %r3789};
	xor.b64  	%rd4063, %rd4062, %rd4061;
	shr.u64 	%rd4064, %rd4060, 10;
	mov.b64	{%r3790, %r3791}, %rd4056;
	mov.b64	%rd4065, {%r3790, %r3791};
	xor.b64  	%rd4066, %rd4065, %rd4064;
	shr.u64 	%rd4067, %rd3884, 54;
	and.b64  	%rd4068, %rd4067, 56;
	add.s64 	%rd4069, %rd12, %rd4068;
	ld.local.u64 	%rd4070, [%rd4069];
	shl.b64 	%rd4071, %rd4070, 57;
	mov.b64	{%r3792, %r3793}, %rd4063;
	mov.b64	%rd4072, {%r3792, %r3793};
	xor.b64  	%rd4073, %rd4072, %rd4071;
	shr.u64 	%rd4074, %rd4070, 7;
	mov.b64	{%r3794, %r3795}, %rd4066;
	mov.b64	%rd4075, {%r3794, %r3795};
	xor.b64  	%rd4076, %rd4075, %rd4074;
	shr.u64 	%rd4077, %rd3884, 57;
	and.b64  	%rd4078, %rd4077, 56;
	add.s64 	%rd4079, %rd12, %rd4078;
	ld.local.u64 	%rd4080, [%rd4079];
	shl.b64 	%rd4081, %rd4080, 60;
	mov.b64	{%r3796, %r3797}, %rd4073;
	mov.b64	%rd4082, {%r3796, %r3797};
	xor.b64  	%rd4083, %rd4082, %rd4081;
	shr.u64 	%rd4084, %rd4080, 4;
	mov.b64	{%r3798, %r3799}, %rd4076;
	mov.b64	%rd4085, {%r3798, %r3799};
	xor.b64  	%rd4086, %rd4085, %rd4084;
	shr.u64 	%rd4087, %rd3884, 60;
	and.b64  	%rd4088, %rd4087, 8;
	add.s64 	%rd4089, %rd12, %rd4088;
	ld.local.u64 	%rd4090, [%rd4089];
	shl.b64 	%rd4091, %rd4090, 63;
	mov.b64	{%r3800, %r3801}, %rd4083;
	mov.b64	%rd4092, {%r3800, %r3801};
	xor.b64  	%rd4093, %rd4092, %rd4091;
	shr.u64 	%rd4094, %rd4090, 1;
	mov.b64	{%r3802, %r3803}, %rd4086;
	mov.b64	%rd4095, {%r3802, %r3803};
	xor.b64  	%rd4096, %rd4095, %rd4094;
	mov.b64	{%r6686, %r6687}, %rd4096;
	mov.b64	{%r6688, %r6689}, %rd4093;
	setp.gt.s32	%p65, %r613, -1;
	@%p65 bra 	BB0_116;

	shr.u32 	%r3804, %r610, 1;
	shl.b32 	%r3805, %r611, 31;
	shr.u32 	%r3806, %r611, 1;
	and.b32  	%r3807, %r3804, 1533916891;
	or.b32  	%r3808, %r3807, %r3805;
	xor.b32  	%r6686, %r6686, %r3808;
	and.b32  	%r3809, %r3806, 920350134;
	xor.b32  	%r6687, %r6687, %r3809;

BB0_116:
	and.b32  	%r3810, %r613, 1073741824;
	setp.eq.s32	%p66, %r3810, 0;
	@%p66 bra 	BB0_118;

	shr.u32 	%r3811, %r610, 2;
	shl.b32 	%r3812, %r611, 30;
	or.b32  	%r3813, %r3811, %r3812;
	shr.u32 	%r3814, %r611, 2;
	and.b32  	%r3815, %r3813, 1227133513;
	xor.b32  	%r6686, %r6686, %r3815;
	and.b32  	%r3816, %r3814, 306783378;
	xor.b32  	%r6687, %r6687, %r3816;

BB0_118:
	mov.u32 	%r6434, 0;
	mov.b64	%rd4097, {%r6732, %r6733};
	mov.b64	%rd4098, {%r6730, %r6731};
	cvt.u32.u64	%r3817, %rd4098;
	cvt.s32.s16 	%r3818, %r3817;
	mov.b64	%rd4099, {%r6686, %r6687};
	mov.b64	%rd4100, {%r6688, %r6689};
	cvt.u32.u64	%r3819, %rd4100;
	cvt.s32.s16 	%r3820, %r3819;
	mad.lo.s32 	%r3821, %r3820, %r3818, 16384;
	bfe.u32 	%r3822, %r3821, 15, 16;
	shr.s32 	%r3823, %r3817, 16;
	shr.s32 	%r3824, %r3819, 16;
	mad.lo.s32 	%r3825, %r3824, %r3823, 16384;
	mov.b64	%rd4101, {%r3822, %r6434};
	shl.b32 	%r3827, %r3825, 1;
	cvt.u64.u32	%rd4102, %r3827;
	and.b64  	%rd4103, %rd4101, 65535;
	shr.u64 	%rd4104, %rd4097, 16;
	shr.u64 	%rd4105, %rd4098, 16;
	cvt.u32.u64	%r3828, %rd4105;
	shr.s32 	%r3829, %r3828, 16;
	shr.u64 	%rd4106, %rd4099, 16;
	shr.u64 	%rd4107, %rd4100, 16;
	cvt.u32.u64	%r3830, %rd4107;
	shr.s32 	%r3831, %r3830, 16;
	mad.lo.s32 	%r3832, %r3831, %r3829, 16384;
	mul.wide.u32 	%rd4108, %r3832, 131072;
	shr.u64 	%rd4109, %rd4098, 32;
	cvt.u32.u64	%r3833, %rd4109;
	shr.s32 	%r3834, %r3833, 16;
	shr.u64 	%rd4110, %rd4100, 32;
	cvt.u32.u64	%r3835, %rd4110;
	shr.s32 	%r3836, %r3835, 16;
	mad.lo.s32 	%r3837, %r3836, %r3834, 16384;
	shr.u32 	%r3838, %r3837, 15;
	cvt.u64.u32	%rd4111, %r3838;
	shl.b64 	%rd4112, %rd4111, 48;
	and.b64  	%rd4113, %rd4108, 281470681743360;
	shl.b64 	%rd4114, %rd4097, 16;
	cvt.u32.u64	%r3839, %rd4114;
	shr.s32 	%r3840, %r3839, 16;
	shl.b64 	%rd4115, %rd4099, 16;
	cvt.u32.u64	%r3841, %rd4115;
	shr.s32 	%r3842, %r3841, 16;
	mad.lo.s32 	%r3843, %r3842, %r3840, 16384;
	shr.u32 	%r3844, %r3843, 15;
	cvt.u64.u32	%rd4116, %r3844;
	cvt.u32.u64	%r3845, %rd4097;
	shr.s32 	%r3846, %r3845, 16;
	cvt.u32.u64	%r3847, %rd4099;
	shr.s32 	%r3848, %r3847, 16;
	mad.lo.s32 	%r3849, %r3848, %r3846, 16384;
	and.b64  	%rd4117, %rd4116, 65535;
	and.b64  	%rd4118, %rd4102, 4294901760;
	or.b64  	%rd4119, %rd4118, %rd4103;
	or.b64  	%rd4120, %rd4119, %rd4113;
	shl.b32 	%r3850, %r3849, 1;
	cvt.u64.u32	%rd4121, %r3850;
	cvt.u32.u64	%r3851, %rd4104;
	shr.s32 	%r3852, %r3851, 16;
	cvt.u32.u64	%r3853, %rd4106;
	shr.s32 	%r3854, %r3853, 16;
	mad.lo.s32 	%r3855, %r3854, %r3852, 16384;
	mul.wide.u32 	%rd4122, %r3855, 131072;
	and.b64  	%rd4123, %rd4121, 4294901760;
	or.b64  	%rd4124, %rd4123, %rd4117;
	shr.u64 	%rd4125, %rd4097, 32;
	cvt.u32.u64	%r3856, %rd4125;
	shr.s32 	%r3857, %r3856, 16;
	shr.u64 	%rd4126, %rd4099, 32;
	cvt.u32.u64	%r3858, %rd4126;
	shr.s32 	%r3859, %r3858, 16;
	mad.lo.s32 	%r3860, %r3859, %r3857, 16384;
	shr.u32 	%r3861, %r3860, 15;
	cvt.u64.u32	%rd4127, %r3861;
	shl.b64 	%rd4128, %rd4127, 48;
	and.b64  	%rd4129, %rd4122, 281470681743360;
	or.b64  	%rd4130, %rd4129, %rd4128;
	or.b64  	%rd4131, %rd4130, %rd4124;
	or.b64  	%rd4132, %rd4120, %rd4112;
	mov.b64	{%r6694, %r6695}, %rd4131;
	mov.b64	{%r6692, %r6693}, %rd4132;
	bra.uni 	BB0_119;

BB0_110:
	cvt.s64.s32 	%rd6081, %rd6123;
	setp.eq.s64	%p62, %rd77, 0;
	selp.b64	%rd3870, %rd71, %rd6122, %p62;
	ld.local.v4.u32 	{%r3704, %r3705, %r3706, %r3707}, [%rd3870];
	xor.b32  	%r3710, %r6722, %r3705;
	xor.b32  	%r3711, %r6721, %r3704;
	mov.b64	%rd78, {%r3711, %r3710};
	or.b64  	%rd3871, %rd78, %rd6081;
	and.b64  	%rd3872, %rd3871, -4294967296;
	setp.eq.s64	%p63, %rd3872, 0;
	@%p63 bra 	BB0_112;
	bra.uni 	BB0_111;

BB0_112:
	cvt.s64.s32 	%rd6083, %rd6123;
	cvt.u32.u64	%r3712, %rd6083;
	cvt.u32.u64	%r3713, %rd78;
	rem.u32 	%r3714, %r3713, %r3712;
	cvt.u64.u32	%rd6119, %r3714;
	bra.uni 	BB0_113;

BB0_111:
	cvt.s64.s32 	%rd6082, %rd6123;
	rem.s64 	%rd6119, %rd78, %rd6082;

BB0_113:
	cvt.u32.u64	%r6692, %rd6119;
	mov.u32 	%r6693, 0;
	mov.u32 	%r6694, %r6693;
	mov.u32 	%r6695, %r6693;

BB0_119:
	xor.b32  	%r6730, %r6692, %r6730;
	xor.b32  	%r6731, %r6693, %r6731;
	xor.b32  	%r6732, %r6694, %r6732;
	xor.b32  	%r6733, %r6695, %r6733;
	add.s64 	%rd6118, %rd6118, 16;
	add.s64 	%rd83, %rd6117, -1;
	setp.ne.s64	%p67, %rd6117, 0;
	mov.u64 	%rd6117, %rd83;
	@%p67 bra 	BB0_109;

	cvt.u64.u16	%rd6087, %rs236;
	add.s64 	%rd6086, %rd6087, %rd3;
	shl.b64 	%rd6085, %rd6086, 4;
	add.s64 	%rd6084, %rd203, %rd6085;
	xor.b32  	%r6667, %r6667, %r6730;
	xor.b32  	%r6668, %r6668, %r6731;
	xor.b32  	%r6669, %r6669, %r6732;
	xor.b32  	%r6670, %r6670, %r6733;
	cvt.u64.u32	%rd4133, %r6729;
	add.s64 	%rd4134, %rd4133, %rd9;
	shl.b64 	%rd4135, %rd4134, 1;
	mov.u64 	%rd4136, verus_gpu_hash$sharedrand;
	add.s64 	%rd4137, %rd4136, %rd4135;
	st.shared.u16 	[%rd4137], %rs236;
	mov.u64 	%rd4138, verus_gpu_hash$sharedrandex;
	add.s64 	%rd4139, %rd4138, %rd4135;
	st.shared.u16 	[%rd4139], %rs237;
	st.global.v4.u32 	[%rd6084], {%r6721, %r6722, %r6723, %r6724};
	cvt.u64.u16	%rd4140, %rs237;
	add.s64 	%rd4141, %rd4140, %rd3;
	shl.b64 	%rd4142, %rd4141, 4;
	add.s64 	%rd4143, %rd203, %rd4142;
	st.global.v4.u32 	[%rd4143], {%r6667, %r6668, %r6669, %r6670};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p68, %r6729, 32;
	@%p68 bra 	BB0_134;

	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p69, %r6729, 0;
	mov.u32 	%r3862, 0;
	mov.u16 	%rs238, 24;
	@%p69 bra 	BB0_122;
	bra.uni 	BB0_123;

BB0_122:
	mov.u32 	%r6729, %r3862;
	bra.uni 	BB0_124;

BB0_123:
	cvt.u64.u32	%rd4144, %r6731;
	shr.u32 	%r3863, %r6730, 5;
	cvt.u64.u32	%rd4145, %r3863;
	cvt.u16.u32	%rs101, %r3863;
	and.b16  	%rs236, %rs101, 511;
	cvt.u16.u32	%rs102, %r6731;
	and.b16  	%rs237, %rs102, 511;
	cvt.u16.u64	%rs103, %rd6123;
	and.b16  	%rs238, %rs103, 28;
	and.b64  	%rd4146, %rd4145, 511;
	add.s64 	%rd4147, %rd4146, %rd3;
	shl.b64 	%rd4148, %rd4147, 4;
	add.s64 	%rd4149, %rd203, %rd4148;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd4149];
	and.b64  	%rd4150, %rd4144, 511;
	add.s64 	%rd4151, %rd4150, %rd3;
	shl.b64 	%rd4152, %rd4151, 4;
	add.s64 	%rd4153, %rd203, %rd4152;
	ld.global.v4.u32 	{%r6667, %r6668, %r6669, %r6670}, [%rd4153];
	mul.wide.u32 	%rd4154, %r6730, 16;
	and.b64  	%rd4155, %rd4154, 48;
	add.s64 	%rd6122, %rd1, %rd4155;

BB0_124:
	and.b16  	%rs104, %rs238, 255;
	setp.ne.s16	%p70, %rs104, 28;
	@%p70 bra 	BB0_133;

	ld.local.v4.u32 	{%r3872, %r3873, %r3874, %r3875}, [%rd6122];
	xor.b32  	%r675, %r3874, %r6669;
	xor.b32  	%r674, %r3873, %r6668;
	xor.b32  	%r673, %r3872, %r6667;
	xor.b32  	%r676, %r3875, %r6670;
	mov.b64	%rd4156, {%r675, %r676};
	mov.u64 	%rd4157, 0;
	st.local.v2.u64 	[%rd12], {%rd4157, %rd4156};
	shl.b64 	%rd4160, %rd4156, 1;
	xor.b64  	%rd4161, %rd4160, %rd4156;
	st.local.v2.u64 	[%rd15], {%rd4160, %rd4161};
	shl.b64 	%rd4162, %rd4156, 2;
	xor.b64  	%rd4163, %rd4162, %rd4156;
	st.local.v2.u64 	[%rd15+16], {%rd4162, %rd4163};
	shl.b64 	%rd4164, %rd4161, 1;
	xor.b64  	%rd4165, %rd4164, %rd4156;
	st.local.v2.u64 	[%rd15+32], {%rd4164, %rd4165};
	mov.b64	%rd4166, {%r673, %r674};
	shl.b64 	%rd4167, %rd4166, 3;
	and.b64  	%rd4168, %rd4167, 56;
	add.s64 	%rd4169, %rd12, %rd4168;
	and.b64  	%rd4170, %rd4166, 56;
	add.s64 	%rd4171, %rd12, %rd4170;
	ld.local.u64 	%rd4172, [%rd4171];
	shl.b64 	%rd4173, %rd4172, 3;
	ld.local.u64 	%rd4174, [%rd4169];
	xor.b64  	%rd4175, %rd4174, %rd4173;
	shr.u64 	%rd4176, %rd4172, 61;
	mov.u32 	%r3880, 0;
	mov.b64	%rd4177, {%r3880, %r3880};
	xor.b64  	%rd4178, %rd4177, %rd4176;
	shr.u64 	%rd4179, %rd4166, 3;
	and.b64  	%rd4180, %rd4179, 56;
	add.s64 	%rd4181, %rd12, %rd4180;
	ld.local.u64 	%rd4182, [%rd4181];
	shl.b64 	%rd4183, %rd4182, 6;
	mov.b64	{%r3881, %r3882}, %rd4175;
	mov.b64	%rd4184, {%r3881, %r3882};
	xor.b64  	%rd4185, %rd4184, %rd4183;
	shr.u64 	%rd4186, %rd4182, 58;
	mov.b64	{%r3883, %r3884}, %rd4178;
	mov.b64	%rd4187, {%r3883, %r3884};
	xor.b64  	%rd4188, %rd4187, %rd4186;
	shr.u64 	%rd4189, %rd4166, 6;
	and.b64  	%rd4190, %rd4189, 56;
	add.s64 	%rd4191, %rd12, %rd4190;
	ld.local.u64 	%rd4192, [%rd4191];
	shl.b64 	%rd4193, %rd4192, 9;
	mov.b64	{%r3885, %r3886}, %rd4185;
	mov.b64	%rd4194, {%r3885, %r3886};
	xor.b64  	%rd4195, %rd4194, %rd4193;
	shr.u64 	%rd4196, %rd4192, 55;
	mov.b64	{%r3887, %r3888}, %rd4188;
	mov.b64	%rd4197, {%r3887, %r3888};
	xor.b64  	%rd4198, %rd4197, %rd4196;
	shr.u64 	%rd4199, %rd4166, 9;
	and.b64  	%rd4200, %rd4199, 56;
	add.s64 	%rd4201, %rd12, %rd4200;
	ld.local.u64 	%rd4202, [%rd4201];
	shl.b64 	%rd4203, %rd4202, 12;
	mov.b64	{%r3889, %r3890}, %rd4195;
	mov.b64	%rd4204, {%r3889, %r3890};
	xor.b64  	%rd4205, %rd4204, %rd4203;
	shr.u64 	%rd4206, %rd4202, 52;
	mov.b64	{%r3891, %r3892}, %rd4198;
	mov.b64	%rd4207, {%r3891, %r3892};
	xor.b64  	%rd4208, %rd4207, %rd4206;
	shr.u64 	%rd4209, %rd4166, 12;
	and.b64  	%rd4210, %rd4209, 56;
	add.s64 	%rd4211, %rd12, %rd4210;
	ld.local.u64 	%rd4212, [%rd4211];
	shl.b64 	%rd4213, %rd4212, 15;
	mov.b64	{%r3893, %r3894}, %rd4205;
	mov.b64	%rd4214, {%r3893, %r3894};
	xor.b64  	%rd4215, %rd4214, %rd4213;
	shr.u64 	%rd4216, %rd4212, 49;
	mov.b64	{%r3895, %r3896}, %rd4208;
	mov.b64	%rd4217, {%r3895, %r3896};
	xor.b64  	%rd4218, %rd4217, %rd4216;
	shr.u64 	%rd4219, %rd4166, 15;
	and.b64  	%rd4220, %rd4219, 56;
	add.s64 	%rd4221, %rd12, %rd4220;
	ld.local.u64 	%rd4222, [%rd4221];
	shl.b64 	%rd4223, %rd4222, 18;
	mov.b64	{%r3897, %r3898}, %rd4215;
	mov.b64	%rd4224, {%r3897, %r3898};
	xor.b64  	%rd4225, %rd4224, %rd4223;
	shr.u64 	%rd4226, %rd4222, 46;
	mov.b64	{%r3899, %r3900}, %rd4218;
	mov.b64	%rd4227, {%r3899, %r3900};
	xor.b64  	%rd4228, %rd4227, %rd4226;
	shr.u64 	%rd4229, %rd4166, 18;
	and.b64  	%rd4230, %rd4229, 56;
	add.s64 	%rd4231, %rd12, %rd4230;
	ld.local.u64 	%rd4232, [%rd4231];
	shl.b64 	%rd4233, %rd4232, 21;
	mov.b64	{%r3901, %r3902}, %rd4225;
	mov.b64	%rd4234, {%r3901, %r3902};
	xor.b64  	%rd4235, %rd4234, %rd4233;
	shr.u64 	%rd4236, %rd4232, 43;
	mov.b64	{%r3903, %r3904}, %rd4228;
	mov.b64	%rd4237, {%r3903, %r3904};
	xor.b64  	%rd4238, %rd4237, %rd4236;
	shr.u64 	%rd4239, %rd4166, 21;
	and.b64  	%rd4240, %rd4239, 56;
	add.s64 	%rd4241, %rd12, %rd4240;
	ld.local.u64 	%rd4242, [%rd4241];
	shl.b64 	%rd4243, %rd4242, 24;
	mov.b64	{%r3905, %r3906}, %rd4235;
	mov.b64	%rd4244, {%r3905, %r3906};
	xor.b64  	%rd4245, %rd4244, %rd4243;
	shr.u64 	%rd4246, %rd4242, 40;
	mov.b64	{%r3907, %r3908}, %rd4238;
	mov.b64	%rd4247, {%r3907, %r3908};
	xor.b64  	%rd4248, %rd4247, %rd4246;
	shr.u64 	%rd4249, %rd4166, 24;
	and.b64  	%rd4250, %rd4249, 56;
	add.s64 	%rd4251, %rd12, %rd4250;
	ld.local.u64 	%rd4252, [%rd4251];
	shl.b64 	%rd4253, %rd4252, 27;
	mov.b64	{%r3909, %r3910}, %rd4245;
	mov.b64	%rd4254, {%r3909, %r3910};
	xor.b64  	%rd4255, %rd4254, %rd4253;
	shr.u64 	%rd4256, %rd4252, 37;
	mov.b64	{%r3911, %r3912}, %rd4248;
	mov.b64	%rd4257, {%r3911, %r3912};
	xor.b64  	%rd4258, %rd4257, %rd4256;
	shr.u64 	%rd4259, %rd4166, 27;
	and.b64  	%rd4260, %rd4259, 56;
	add.s64 	%rd4261, %rd12, %rd4260;
	ld.local.u64 	%rd4262, [%rd4261];
	shl.b64 	%rd4263, %rd4262, 30;
	mov.b64	{%r3913, %r3914}, %rd4255;
	mov.b64	%rd4264, {%r3913, %r3914};
	xor.b64  	%rd4265, %rd4264, %rd4263;
	shr.u64 	%rd4266, %rd4262, 34;
	mov.b64	{%r3915, %r3916}, %rd4258;
	mov.b64	%rd4267, {%r3915, %r3916};
	xor.b64  	%rd4268, %rd4267, %rd4266;
	shr.u64 	%rd4269, %rd4166, 30;
	and.b64  	%rd4270, %rd4269, 56;
	add.s64 	%rd4271, %rd12, %rd4270;
	ld.local.u64 	%rd4272, [%rd4271];
	shl.b64 	%rd4273, %rd4272, 33;
	mov.b64	{%r3917, %r3918}, %rd4265;
	mov.b64	%rd4274, {%r3917, %r3918};
	xor.b64  	%rd4275, %rd4274, %rd4273;
	shr.u64 	%rd4276, %rd4272, 31;
	mov.b64	{%r3919, %r3920}, %rd4268;
	mov.b64	%rd4277, {%r3919, %r3920};
	xor.b64  	%rd4278, %rd4277, %rd4276;
	shr.u64 	%rd4279, %rd4166, 33;
	and.b64  	%rd4280, %rd4279, 56;
	add.s64 	%rd4281, %rd12, %rd4280;
	ld.local.u64 	%rd4282, [%rd4281];
	shl.b64 	%rd4283, %rd4282, 36;
	mov.b64	{%r3921, %r3922}, %rd4275;
	mov.b64	%rd4284, {%r3921, %r3922};
	xor.b64  	%rd4285, %rd4284, %rd4283;
	shr.u64 	%rd4286, %rd4282, 28;
	mov.b64	{%r3923, %r3924}, %rd4278;
	mov.b64	%rd4287, {%r3923, %r3924};
	xor.b64  	%rd4288, %rd4287, %rd4286;
	shr.u64 	%rd4289, %rd4166, 36;
	and.b64  	%rd4290, %rd4289, 56;
	add.s64 	%rd4291, %rd12, %rd4290;
	ld.local.u64 	%rd4292, [%rd4291];
	shl.b64 	%rd4293, %rd4292, 39;
	mov.b64	{%r3925, %r3926}, %rd4285;
	mov.b64	%rd4294, {%r3925, %r3926};
	xor.b64  	%rd4295, %rd4294, %rd4293;
	shr.u64 	%rd4296, %rd4292, 25;
	mov.b64	{%r3927, %r3928}, %rd4288;
	mov.b64	%rd4297, {%r3927, %r3928};
	xor.b64  	%rd4298, %rd4297, %rd4296;
	shr.u64 	%rd4299, %rd4166, 39;
	and.b64  	%rd4300, %rd4299, 56;
	add.s64 	%rd4301, %rd12, %rd4300;
	ld.local.u64 	%rd4302, [%rd4301];
	shl.b64 	%rd4303, %rd4302, 42;
	mov.b64	{%r3929, %r3930}, %rd4295;
	mov.b64	%rd4304, {%r3929, %r3930};
	xor.b64  	%rd4305, %rd4304, %rd4303;
	shr.u64 	%rd4306, %rd4302, 22;
	mov.b64	{%r3931, %r3932}, %rd4298;
	mov.b64	%rd4307, {%r3931, %r3932};
	xor.b64  	%rd4308, %rd4307, %rd4306;
	shr.u64 	%rd4309, %rd4166, 42;
	and.b64  	%rd4310, %rd4309, 56;
	add.s64 	%rd4311, %rd12, %rd4310;
	ld.local.u64 	%rd4312, [%rd4311];
	shl.b64 	%rd4313, %rd4312, 45;
	mov.b64	{%r3933, %r3934}, %rd4305;
	mov.b64	%rd4314, {%r3933, %r3934};
	xor.b64  	%rd4315, %rd4314, %rd4313;
	shr.u64 	%rd4316, %rd4312, 19;
	mov.b64	{%r3935, %r3936}, %rd4308;
	mov.b64	%rd4317, {%r3935, %r3936};
	xor.b64  	%rd4318, %rd4317, %rd4316;
	shr.u64 	%rd4319, %rd4166, 45;
	and.b64  	%rd4320, %rd4319, 56;
	add.s64 	%rd4321, %rd12, %rd4320;
	ld.local.u64 	%rd4322, [%rd4321];
	shl.b64 	%rd4323, %rd4322, 48;
	mov.b64	{%r3937, %r3938}, %rd4315;
	mov.b64	%rd4324, {%r3937, %r3938};
	xor.b64  	%rd4325, %rd4324, %rd4323;
	shr.u64 	%rd4326, %rd4322, 16;
	mov.b64	{%r3939, %r3940}, %rd4318;
	mov.b64	%rd4327, {%r3939, %r3940};
	xor.b64  	%rd4328, %rd4327, %rd4326;
	shr.u64 	%rd4329, %rd4166, 48;
	and.b64  	%rd4330, %rd4329, 56;
	add.s64 	%rd4331, %rd12, %rd4330;
	ld.local.u64 	%rd4332, [%rd4331];
	shl.b64 	%rd4333, %rd4332, 51;
	mov.b64	{%r3941, %r3942}, %rd4325;
	mov.b64	%rd4334, {%r3941, %r3942};
	xor.b64  	%rd4335, %rd4334, %rd4333;
	shr.u64 	%rd4336, %rd4332, 13;
	mov.b64	{%r3943, %r3944}, %rd4328;
	mov.b64	%rd4337, {%r3943, %r3944};
	xor.b64  	%rd4338, %rd4337, %rd4336;
	shr.u64 	%rd4339, %rd4166, 51;
	and.b64  	%rd4340, %rd4339, 56;
	add.s64 	%rd4341, %rd12, %rd4340;
	ld.local.u64 	%rd4342, [%rd4341];
	shl.b64 	%rd4343, %rd4342, 54;
	mov.b64	{%r3945, %r3946}, %rd4335;
	mov.b64	%rd4344, {%r3945, %r3946};
	xor.b64  	%rd4345, %rd4344, %rd4343;
	shr.u64 	%rd4346, %rd4342, 10;
	mov.b64	{%r3947, %r3948}, %rd4338;
	mov.b64	%rd4347, {%r3947, %r3948};
	xor.b64  	%rd4348, %rd4347, %rd4346;
	shr.u64 	%rd4349, %rd4166, 54;
	and.b64  	%rd4350, %rd4349, 56;
	add.s64 	%rd4351, %rd12, %rd4350;
	ld.local.u64 	%rd4352, [%rd4351];
	shl.b64 	%rd4353, %rd4352, 57;
	mov.b64	{%r3949, %r3950}, %rd4345;
	mov.b64	%rd4354, {%r3949, %r3950};
	xor.b64  	%rd4355, %rd4354, %rd4353;
	shr.u64 	%rd4356, %rd4352, 7;
	mov.b64	{%r3951, %r3952}, %rd4348;
	mov.b64	%rd4357, {%r3951, %r3952};
	xor.b64  	%rd4358, %rd4357, %rd4356;
	shr.u64 	%rd4359, %rd4166, 57;
	and.b64  	%rd4360, %rd4359, 56;
	add.s64 	%rd4361, %rd12, %rd4360;
	ld.local.u64 	%rd4362, [%rd4361];
	shl.b64 	%rd4363, %rd4362, 60;
	mov.b64	{%r3953, %r3954}, %rd4355;
	mov.b64	%rd4364, {%r3953, %r3954};
	xor.b64  	%rd4365, %rd4364, %rd4363;
	shr.u64 	%rd4366, %rd4362, 4;
	mov.b64	{%r3955, %r3956}, %rd4358;
	mov.b64	%rd4367, {%r3955, %r3956};
	xor.b64  	%rd4368, %rd4367, %rd4366;
	shr.u64 	%rd4369, %rd4166, 60;
	and.b64  	%rd4370, %rd4369, 8;
	add.s64 	%rd4371, %rd12, %rd4370;
	ld.local.u64 	%rd4372, [%rd4371];
	shl.b64 	%rd4373, %rd4372, 63;
	mov.b64	{%r3957, %r3958}, %rd4365;
	mov.b64	%rd4374, {%r3957, %r3958};
	xor.b64  	%rd4375, %rd4374, %rd4373;
	shr.u64 	%rd4376, %rd4372, 1;
	mov.b64	{%r3959, %r3960}, %rd4368;
	mov.b64	%rd4377, {%r3959, %r3960};
	xor.b64  	%rd4378, %rd4377, %rd4376;
	mov.b64	{%r6711, %r6712}, %rd4378;
	mov.b64	{%r6713, %r6714}, %rd4375;
	setp.gt.s32	%p71, %r676, -1;
	@%p71 bra 	BB0_127;

	shr.u32 	%r3961, %r673, 1;
	shl.b32 	%r3962, %r674, 31;
	shr.u32 	%r3963, %r674, 1;
	and.b32  	%r3964, %r3961, 1533916891;
	or.b32  	%r3965, %r3964, %r3962;
	xor.b32  	%r6711, %r6711, %r3965;
	and.b32  	%r3966, %r3963, 920350134;
	xor.b32  	%r6712, %r6712, %r3966;

BB0_127:
	and.b32  	%r3967, %r676, 1073741824;
	setp.eq.s32	%p72, %r3967, 0;
	@%p72 bra 	BB0_129;

	shr.u32 	%r3968, %r673, 2;
	shl.b32 	%r3969, %r674, 30;
	or.b32  	%r3970, %r3968, %r3969;
	shr.u32 	%r3971, %r674, 2;
	and.b32  	%r3972, %r3970, 1227133513;
	xor.b32  	%r6711, %r6711, %r3972;
	and.b32  	%r3973, %r3971, 306783378;
	xor.b32  	%r6712, %r6712, %r3973;

BB0_129:
	xor.b32  	%r3974, %r6713, %r6730;
	xor.b32  	%r3975, %r6714, %r6731;
	xor.b32  	%r3976, %r6711, %r6732;
	xor.b32  	%r3977, %r6712, %r6733;
	mov.b64	%rd4379, {%r3976, %r3977};
	mov.b64	%rd4380, {%r3974, %r3975};
	cvt.u32.u64	%r3978, %rd4380;
	cvt.s32.s16 	%r3979, %r3978;
	mov.b64	%rd4381, {%r6669, %r6670};
	mov.b64	%rd4382, {%r6667, %r6668};
	cvt.u32.u64	%r3980, %rd4382;
	cvt.s32.s16 	%r3981, %r3980;
	mad.lo.s32 	%r3982, %r3979, %r3981, 16384;
	bfe.u32 	%r3983, %r3982, 15, 16;
	shr.s32 	%r3984, %r3978, 16;
	shr.s32 	%r3985, %r3980, 16;
	mad.lo.s32 	%r3986, %r3984, %r3985, 16384;
	mov.b64	%rd4383, {%r3983, %r3880};
	shl.b32 	%r3988, %r3986, 1;
	cvt.u64.u32	%rd4384, %r3988;
	and.b64  	%rd4385, %rd4383, 65535;
	shr.u64 	%rd4386, %rd4379, 16;
	shr.u64 	%rd4387, %rd4380, 16;
	cvt.u32.u64	%r3989, %rd4387;
	shr.s32 	%r3990, %r3989, 16;
	shr.u64 	%rd4388, %rd4381, 16;
	shr.u64 	%rd4389, %rd4382, 16;
	cvt.u32.u64	%r3991, %rd4389;
	shr.s32 	%r3992, %r3991, 16;
	mad.lo.s32 	%r3993, %r3990, %r3992, 16384;
	mul.wide.u32 	%rd4390, %r3993, 131072;
	shr.u64 	%rd4391, %rd4380, 32;
	cvt.u32.u64	%r3994, %rd4391;
	shr.s32 	%r3995, %r3994, 16;
	shr.u64 	%rd4392, %rd4382, 32;
	cvt.u32.u64	%r3996, %rd4392;
	shr.s32 	%r3997, %r3996, 16;
	mad.lo.s32 	%r3998, %r3995, %r3997, 16384;
	shr.u32 	%r3999, %r3998, 15;
	cvt.u64.u32	%rd4393, %r3999;
	shl.b64 	%rd4394, %rd4393, 48;
	and.b64  	%rd4395, %rd4390, 281470681743360;
	shl.b64 	%rd4396, %rd4379, 16;
	cvt.u32.u64	%r4000, %rd4396;
	shr.s32 	%r4001, %r4000, 16;
	shl.b64 	%rd4397, %rd4381, 16;
	cvt.u32.u64	%r4002, %rd4397;
	shr.s32 	%r4003, %r4002, 16;
	mad.lo.s32 	%r4004, %r4001, %r4003, 16384;
	shr.u32 	%r4005, %r4004, 15;
	cvt.u64.u32	%rd4398, %r4005;
	cvt.u32.u64	%r4006, %rd4379;
	shr.s32 	%r4007, %r4006, 16;
	cvt.u32.u64	%r4008, %rd4381;
	shr.s32 	%r4009, %r4008, 16;
	mad.lo.s32 	%r4010, %r4007, %r4009, 16384;
	and.b64  	%rd4399, %rd4398, 65535;
	and.b64  	%rd4400, %rd4384, 4294901760;
	or.b64  	%rd4401, %rd4400, %rd4385;
	or.b64  	%rd4402, %rd4401, %rd4395;
	shl.b32 	%r4011, %r4010, 1;
	cvt.u64.u32	%rd4403, %r4011;
	cvt.u32.u64	%r4012, %rd4386;
	shr.s32 	%r4013, %r4012, 16;
	cvt.u32.u64	%r4014, %rd4388;
	shr.s32 	%r4015, %r4014, 16;
	mad.lo.s32 	%r4016, %r4013, %r4015, 16384;
	mul.wide.u32 	%rd4404, %r4016, 131072;
	and.b64  	%rd4405, %rd4403, 4294901760;
	or.b64  	%rd4406, %rd4405, %rd4399;
	shr.u64 	%rd4407, %rd4379, 32;
	cvt.u32.u64	%r4017, %rd4407;
	shr.s32 	%r4018, %r4017, 16;
	shr.u64 	%rd4408, %rd4381, 32;
	cvt.u32.u64	%r4019, %rd4408;
	shr.s32 	%r4020, %r4019, 16;
	mad.lo.s32 	%r4021, %r4018, %r4020, 16384;
	shr.u32 	%r4022, %r4021, 15;
	cvt.u64.u32	%rd4409, %r4022;
	shl.b64 	%rd4410, %rd4409, 48;
	and.b64  	%rd4411, %rd4404, 281470681743360;
	or.b64  	%rd4412, %rd4411, %rd4410;
	or.b64  	%rd4413, %rd4412, %rd4406;
	or.b64  	%rd4414, %rd4402, %rd4394;
	mov.b64	{%r4023, %r4024}, %rd4413;
	mov.b64	{%r4025, %r4026}, %rd4414;
	xor.b32  	%r699, %r4026, %r6668;
	xor.b32  	%r698, %r4025, %r6667;
	xor.b32  	%r701, %r4024, %r6670;
	xor.b32  	%r700, %r4023, %r6669;
	xor.b32  	%r6730, %r6721, %r3974;
	xor.b32  	%r6731, %r6722, %r3975;
	xor.b32  	%r6732, %r6723, %r3976;
	xor.b32  	%r6733, %r6724, %r3977;
	mov.b64	%rd4415, {%r6732, %r6733};
	mov.b64	%rd4416, {%r6730, %r6731};
	cvt.u32.u64	%r4027, %rd4416;
	cvt.s32.s16 	%r4028, %r4027;
	mov.b64	%rd4417, {%r6723, %r6724};
	mov.b64	%rd4418, {%r6721, %r6722};
	cvt.u32.u64	%r4029, %rd4418;
	cvt.s32.s16 	%r4030, %r4029;
	mad.lo.s32 	%r4031, %r4028, %r4030, 16384;
	bfe.u32 	%r4032, %r4031, 15, 16;
	shr.s32 	%r4033, %r4027, 16;
	shr.s32 	%r4034, %r4029, 16;
	mad.lo.s32 	%r4035, %r4033, %r4034, 16384;
	mov.b64	%rd4419, {%r4032, %r3880};
	shl.b32 	%r4036, %r4035, 1;
	cvt.u64.u32	%rd4420, %r4036;
	and.b64  	%rd4421, %rd4419, 65535;
	shr.u64 	%rd4422, %rd4415, 16;
	shr.u64 	%rd4423, %rd4416, 16;
	cvt.u32.u64	%r4037, %rd4423;
	shr.s32 	%r4038, %r4037, 16;
	shr.u64 	%rd4424, %rd4417, 16;
	shr.u64 	%rd4425, %rd4418, 16;
	cvt.u32.u64	%r4039, %rd4425;
	shr.s32 	%r4040, %r4039, 16;
	mad.lo.s32 	%r4041, %r4038, %r4040, 16384;
	mul.wide.u32 	%rd4426, %r4041, 131072;
	shr.u64 	%rd4427, %rd4416, 32;
	cvt.u32.u64	%r4042, %rd4427;
	shr.s32 	%r4043, %r4042, 16;
	shr.u64 	%rd4428, %rd4418, 32;
	cvt.u32.u64	%r4044, %rd4428;
	shr.s32 	%r4045, %r4044, 16;
	mad.lo.s32 	%r4046, %r4043, %r4045, 16384;
	shr.u32 	%r4047, %r4046, 15;
	cvt.u64.u32	%rd4429, %r4047;
	shl.b64 	%rd4430, %rd4429, 48;
	and.b64  	%rd4431, %rd4426, 281470681743360;
	shl.b64 	%rd4432, %rd4415, 16;
	cvt.u32.u64	%r4048, %rd4432;
	shr.s32 	%r4049, %r4048, 16;
	shl.b64 	%rd4433, %rd4417, 16;
	cvt.u32.u64	%r4050, %rd4433;
	shr.s32 	%r4051, %r4050, 16;
	mad.lo.s32 	%r4052, %r4049, %r4051, 16384;
	shr.u32 	%r4053, %r4052, 15;
	cvt.u64.u32	%rd4434, %r4053;
	cvt.u32.u64	%r4054, %rd4415;
	shr.s32 	%r4055, %r4054, 16;
	cvt.u32.u64	%r4056, %rd4417;
	shr.s32 	%r4057, %r4056, 16;
	mad.lo.s32 	%r4058, %r4055, %r4057, 16384;
	and.b64  	%rd4435, %rd4434, 65535;
	and.b64  	%rd4436, %rd4420, 4294901760;
	or.b64  	%rd4437, %rd4436, %rd4421;
	or.b64  	%rd4438, %rd4437, %rd4431;
	shl.b32 	%r4059, %r4058, 1;
	cvt.u64.u32	%rd4439, %r4059;
	cvt.u32.u64	%r4060, %rd4422;
	shr.s32 	%r4061, %r4060, 16;
	cvt.u32.u64	%r4062, %rd4424;
	shr.s32 	%r4063, %r4062, 16;
	mad.lo.s32 	%r4064, %r4061, %r4063, 16384;
	mul.wide.u32 	%rd4440, %r4064, 131072;
	and.b64  	%rd4441, %rd4439, 4294901760;
	or.b64  	%rd4442, %rd4441, %rd4435;
	shr.u64 	%rd4443, %rd4415, 32;
	cvt.u32.u64	%r4065, %rd4443;
	shr.s32 	%r4066, %r4065, 16;
	shr.u64 	%rd4444, %rd4417, 32;
	cvt.u32.u64	%r4067, %rd4444;
	shr.s32 	%r4068, %r4067, 16;
	mad.lo.s32 	%r4069, %r4066, %r4068, 16384;
	shr.u32 	%r4070, %r4069, 15;
	cvt.u64.u32	%rd4445, %r4070;
	shl.b64 	%rd4446, %rd4445, 48;
	and.b64  	%rd4447, %rd4440, 281470681743360;
	or.b64  	%rd4448, %rd4447, %rd4446;
	or.b64  	%rd4449, %rd4448, %rd4442;
	or.b64  	%rd4450, %rd4438, %rd4430;
	mov.b64	{%r4071, %r4072}, %rd4449;
	mov.b64	{%r4073, %r4074}, %rd4450;
	xor.b32  	%r6668, %r4074, %r6722;
	xor.b32  	%r6667, %r4073, %r6721;
	xor.b32  	%r6670, %r4072, %r6724;
	xor.b32  	%r6669, %r4071, %r6723;
	cvt.u64.u32	%rd4451, %r6729;
	add.s64 	%rd4452, %rd4451, %rd9;
	shl.b64 	%rd4453, %rd4452, 1;
	mov.u64 	%rd4454, verus_gpu_hash$sharedrand;
	add.s64 	%rd4455, %rd4454, %rd4453;
	st.shared.u16 	[%rd4455], %rs236;
	mov.u64 	%rd4456, verus_gpu_hash$sharedrandex;
	add.s64 	%rd4457, %rd4456, %rd4453;
	st.shared.u16 	[%rd4457], %rs237;
	cvt.u64.u16	%rd4458, %rs236;
	add.s64 	%rd4459, %rd4458, %rd3;
	shl.b64 	%rd4460, %rd4459, 4;
	add.s64 	%rd4461, %rd203, %rd4460;
	st.global.v4.u32 	[%rd4461], {%r698, %r699, %r700, %r701};
	cvt.u64.u16	%rd4462, %rs237;
	add.s64 	%rd4463, %rd4462, %rd3;
	shl.b64 	%rd4464, %rd4463, 4;
	add.s64 	%rd4465, %rd203, %rd4464;
	st.global.v4.u32 	[%rd4465], {%r6667, %r6668, %r6669, %r6670};
	add.s32 	%r6729, %r6729, 1;
	setp.eq.s32	%p73, %r6729, 32;
	@%p73 bra 	BB0_134;

	mov.b64	%rd6123, {%r6730, %r6731};
	setp.eq.s32	%p74, %r6729, 0;
	mov.u16 	%rs238, 28;
	@%p74 bra 	BB0_131;
	bra.uni 	BB0_132;

BB0_131:
	mov.u32 	%r6721, %r698;
	mov.u32 	%r6722, %r699;
	mov.u32 	%r6723, %r700;
	mov.u32 	%r6724, %r701;
	mov.u32 	%r6729, %r3880;
	bra.uni 	BB0_133;

BB0_132:
	cvt.u64.u32	%rd4466, %r6731;
	shr.u32 	%r4076, %r6730, 5;
	cvt.u64.u32	%rd4467, %r4076;
	cvt.u16.u32	%rs106, %r4076;
	and.b16  	%rs236, %rs106, 511;
	cvt.u16.u32	%rs107, %r6731;
	and.b16  	%rs237, %rs107, 511;
	cvt.u16.u64	%rs108, %rd6123;
	and.b16  	%rs238, %rs108, 28;
	and.b64  	%rd4468, %rd4467, 511;
	add.s64 	%rd4469, %rd4468, %rd3;
	shl.b64 	%rd4470, %rd4469, 4;
	add.s64 	%rd4471, %rd203, %rd4470;
	ld.global.v4.u32 	{%r6721, %r6722, %r6723, %r6724}, [%rd4471];
	and.b64  	%rd4472, %rd4466, 511;
	add.s64 	%rd4473, %rd4472, %rd3;
	shl.b64 	%rd4474, %rd4473, 4;
	add.s64 	%rd4475, %rd203, %rd4474;
	ld.global.v4.u32 	{%r6667, %r6668, %r6669, %r6670}, [%rd4475];
	mul.wide.u32 	%rd4476, %r6730, 16;
	and.b64  	%rd4477, %rd4476, 48;
	add.s64 	%rd6122, %rd1, %rd4477;

BB0_133:
	setp.ne.s32	%p75, %r6729, 32;
	@%p75 bra 	BB0_13;

BB0_134:
	st.local.v4.u32 	[%rd1], {%r744, %r745, %r746, %r747};
	st.local.v4.u32 	[%rd1+16], {%r748, %r749, %r750, %r751};
	mov.b64	%rd4480, {%r6732, %r6733};
	mov.u16 	%rs109, 16730;
	mov.u16 	%rs110, 30572;
	mov.u16 	%rs111, 11574;
	mov.u16 	%rs112, 6912;
	st.local.v4.u16 	[%rd12], {%rs112, %rs111, %rs110, %rs109};
	and.b64  	%rd4483, %rd4480, 7;
	add.s64 	%rd4484, %rd12, %rd4483;
	ld.local.u8 	%rd4485, [%rd4484];
	bfe.u64 	%rd4486, %rd4480, 3, 3;
	add.s64 	%rd4487, %rd12, %rd4486;
	ld.local.u8 	%r4085, [%rd4487];
	mul.wide.u32 	%rd4488, %r4085, 8;
	mov.b64	{%r4086, %r4087}, %rd4485;
	mov.b64	%rd4489, {%r4086, %r4087};
	xor.b64  	%rd4490, %rd4489, %rd4488;
	bfe.u64 	%rd4491, %rd4480, 6, 3;
	add.s64 	%rd4492, %rd12, %rd4491;
	ld.local.u8 	%r4088, [%rd4492];
	mul.wide.u32 	%rd4493, %r4088, 64;
	mov.b64	{%r4089, %r4090}, %rd4490;
	mov.b64	%rd4494, {%r4089, %r4090};
	xor.b64  	%rd4495, %rd4494, %rd4493;
	bfe.u64 	%rd4496, %rd4480, 9, 3;
	add.s64 	%rd4497, %rd12, %rd4496;
	ld.local.u8 	%r4091, [%rd4497];
	mul.wide.u32 	%rd4498, %r4091, 512;
	mov.b64	{%r4092, %r4093}, %rd4495;
	mov.b64	%rd4499, {%r4092, %r4093};
	xor.b64  	%rd4500, %rd4499, %rd4498;
	bfe.u64 	%rd4501, %rd4480, 12, 3;
	add.s64 	%rd4502, %rd12, %rd4501;
	ld.local.u8 	%r4094, [%rd4502];
	mul.wide.u32 	%rd4503, %r4094, 4096;
	mov.b64	{%r4095, %r4096}, %rd4500;
	mov.b64	%rd4504, {%r4095, %r4096};
	xor.b64  	%rd4505, %rd4504, %rd4503;
	bfe.u64 	%rd4506, %rd4480, 15, 3;
	add.s64 	%rd4507, %rd12, %rd4506;
	ld.local.u8 	%r4097, [%rd4507];
	mul.wide.u32 	%rd4508, %r4097, 32768;
	mov.b64	{%r4098, %r4099}, %rd4505;
	mov.b64	%rd4509, {%r4098, %r4099};
	xor.b64  	%rd4510, %rd4509, %rd4508;
	bfe.u64 	%rd4511, %rd4480, 18, 3;
	add.s64 	%rd4512, %rd12, %rd4511;
	ld.local.u8 	%r4100, [%rd4512];
	mul.wide.u32 	%rd4513, %r4100, 262144;
	mov.b64	{%r4101, %r4102}, %rd4510;
	mov.b64	%rd4514, {%r4101, %r4102};
	xor.b64  	%rd4515, %rd4514, %rd4513;
	bfe.u64 	%rd4516, %rd4480, 21, 3;
	add.s64 	%rd4517, %rd12, %rd4516;
	ld.local.u8 	%r4103, [%rd4517];
	mul.wide.u32 	%rd4518, %r4103, 2097152;
	mov.b64	{%r4104, %r4105}, %rd4515;
	mov.b64	%rd4519, {%r4104, %r4105};
	xor.b64  	%rd4520, %rd4519, %rd4518;
	bfe.u64 	%rd4521, %rd4480, 24, 3;
	add.s64 	%rd4522, %rd12, %rd4521;
	ld.local.u8 	%r4106, [%rd4522];
	mul.wide.u32 	%rd4523, %r4106, 16777216;
	mov.b64	{%r4107, %r4108}, %rd4520;
	mov.b64	%rd4524, {%r4107, %r4108};
	xor.b64  	%rd4525, %rd4524, %rd4523;
	bfe.u64 	%rd4526, %rd4480, 27, 3;
	add.s64 	%rd4527, %rd12, %rd4526;
	ld.local.u8 	%r4109, [%rd4527];
	mul.wide.u32 	%rd4528, %r4109, 134217728;
	mov.b64	{%r4110, %r4111}, %rd4525;
	mov.b64	%rd4529, {%r4110, %r4111};
	xor.b64  	%rd4530, %rd4529, %rd4528;
	bfe.u64 	%rd4531, %rd4480, 30, 3;
	add.s64 	%rd4532, %rd12, %rd4531;
	ld.local.u8 	%r4112, [%rd4532];
	mul.wide.u32 	%rd4533, %r4112, 1073741824;
	mov.b64	{%r4113, %r4114}, %rd4530;
	mov.b64	%rd4534, {%r4113, %r4114};
	xor.b64  	%rd4535, %rd4534, %rd4533;
	bfe.u64 	%rd4536, %rd4480, 33, 3;
	add.s64 	%rd4537, %rd12, %rd4536;
	ld.local.u8 	%rd4538, [%rd4537];
	shl.b64 	%rd4539, %rd4538, 33;
	mov.b64	{%r4115, %r4116}, %rd4535;
	mov.b64	%rd4540, {%r4115, %r4116};
	xor.b64  	%rd4541, %rd4540, %rd4539;
	bfe.u64 	%rd4542, %rd4480, 36, 3;
	add.s64 	%rd4543, %rd12, %rd4542;
	ld.local.u8 	%rd4544, [%rd4543];
	shl.b64 	%rd4545, %rd4544, 36;
	mov.b64	{%r4117, %r4118}, %rd4541;
	mov.b64	%rd4546, {%r4117, %r4118};
	xor.b64  	%rd4547, %rd4546, %rd4545;
	bfe.u64 	%rd4548, %rd4480, 39, 3;
	add.s64 	%rd4549, %rd12, %rd4548;
	ld.local.u8 	%rd4550, [%rd4549];
	shl.b64 	%rd4551, %rd4550, 39;
	mov.b64	{%r4119, %r4120}, %rd4547;
	mov.b64	%rd4552, {%r4119, %r4120};
	xor.b64  	%rd4553, %rd4552, %rd4551;
	bfe.u64 	%rd4554, %rd4480, 42, 3;
	add.s64 	%rd4555, %rd12, %rd4554;
	ld.local.u8 	%rd4556, [%rd4555];
	shl.b64 	%rd4557, %rd4556, 42;
	mov.b64	{%r4121, %r4122}, %rd4553;
	mov.b64	%rd4558, {%r4121, %r4122};
	xor.b64  	%rd4559, %rd4558, %rd4557;
	bfe.u64 	%rd4560, %rd4480, 45, 3;
	add.s64 	%rd4561, %rd12, %rd4560;
	ld.local.u8 	%rd4562, [%rd4561];
	shl.b64 	%rd4563, %rd4562, 45;
	mov.b64	{%r4123, %r4124}, %rd4559;
	mov.b64	%rd4564, {%r4123, %r4124};
	xor.b64  	%rd4565, %rd4564, %rd4563;
	bfe.u64 	%rd4566, %rd4480, 48, 3;
	add.s64 	%rd4567, %rd12, %rd4566;
	ld.local.u8 	%rd4568, [%rd4567];
	shl.b64 	%rd4569, %rd4568, 48;
	mov.b64	{%r4125, %r4126}, %rd4565;
	mov.b64	%rd4570, {%r4125, %r4126};
	xor.b64  	%rd4571, %rd4570, %rd4569;
	bfe.u64 	%rd4572, %rd4480, 51, 3;
	add.s64 	%rd4573, %rd12, %rd4572;
	ld.local.u8 	%rd4574, [%rd4573];
	shl.b64 	%rd4575, %rd4574, 51;
	mov.b64	{%r4127, %r4128}, %rd4571;
	mov.b64	%rd4576, {%r4127, %r4128};
	xor.b64  	%rd4577, %rd4576, %rd4575;
	bfe.u64 	%rd4578, %rd4480, 54, 3;
	add.s64 	%rd4579, %rd12, %rd4578;
	ld.local.u8 	%rd4580, [%rd4579];
	shl.b64 	%rd4581, %rd4580, 54;
	mov.b64	{%r4129, %r4130}, %rd4577;
	mov.b64	%rd4582, {%r4129, %r4130};
	xor.b64  	%rd4583, %rd4582, %rd4581;
	bfe.u64 	%rd4584, %rd4480, 57, 3;
	add.s64 	%rd4585, %rd12, %rd4584;
	ld.local.u8 	%rd4586, [%rd4585];
	shl.b64 	%rd4587, %rd4586, 57;
	mov.b64	{%r4131, %r4132}, %rd4583;
	mov.b64	%rd4588, {%r4131, %r4132};
	xor.b64  	%rd4589, %rd4588, %rd4587;
	shr.u64 	%rd4590, %rd4586, 7;
	mov.u32 	%r4133, 0;
	mov.b64	%rd94, {%r4133, %r4133};
	xor.b64  	%rd4591, %rd94, %rd4590;
	bfe.u64 	%rd4592, %rd4480, 60, 3;
	add.s64 	%rd4593, %rd12, %rd4592;
	ld.local.u8 	%rd4594, [%rd4593];
	shl.b64 	%rd4595, %rd4594, 60;
	mov.b64	{%r4134, %r4135}, %rd4589;
	mov.b64	%rd4596, {%r4134, %r4135};
	xor.b64  	%rd4597, %rd4596, %rd4595;
	shr.u64 	%rd4598, %rd4594, 4;
	mov.b64	{%r4136, %r4137}, %rd4591;
	mov.b64	%rd4599, {%r4136, %r4137};
	xor.b64  	%rd4600, %rd4599, %rd4598;
	shr.u64 	%rd4601, %rd4480, 63;
	add.s64 	%rd4602, %rd12, %rd4601;
	ld.local.u8 	%rd4603, [%rd4602];
	shl.b64 	%rd4604, %rd4603, 63;
	mov.b64	{%r4138, %r4139}, %rd4597;
	mov.b64	%rd4605, {%r4138, %r4139};
	xor.b64  	%rd4606, %rd4605, %rd4604;
	shr.u64 	%rd4607, %rd4603, 1;
	mov.b64	{%r4140, %r4141}, %rd4600;
	mov.b64	%rd4608, {%r4140, %r4141};
	xor.b64  	%rd4609, %rd4608, %rd4607;
	mov.b64	{%r736, %r737}, %rd4606;
	mov.b64	{%r738, %r739}, %rd4609;
	mov.u32 	%r4142, -1719488588;
	mov.u32 	%r4143, -168901672;
	mov.u32 	%r4144, 1096447852;
	mov.u32 	%r4145, 758520576;
	st.local.v4.u32 	[%rd12], {%r4145, %r4144, %r4143, %r4142};
	mov.b64	%rd93, {%r738, %r739};
	cvt.u16.u64	%rs113, %rd93;
	cvt.s16.s8 	%rs114, %rs113;
	setp.lt.s16	%p76, %rs114, 0;
	mov.u64 	%rd6133, 0;
	mov.u64 	%rd6124, %rd6133;
	@%p76 bra 	BB0_136;

	and.b64  	%rd4611, %rd93, 15;
	add.s64 	%rd4612, %rd12, %rd4611;
	ld.local.u8 	%rd6124, [%rd4612];

BB0_136:
	shr.u64 	%rd4615, %rd93, 8;
	shl.b64 	%rd4616, %rd94, 56;
	or.b64  	%rd99, %rd4615, %rd4616;
	cvt.u16.u64	%rs115, %rd4615;
	cvt.s16.s8 	%rs116, %rs115;
	setp.lt.s16	%p77, %rs116, 0;
	mov.u64 	%rd6125, %rd6133;
	@%p77 bra 	BB0_138;

	and.b64  	%rd4617, %rd99, 15;
	add.s64 	%rd4618, %rd12, %rd4617;
	ld.local.u8 	%rd4619, [%rd4618];
	shl.b64 	%rd6125, %rd4619, 8;

BB0_138:
	shr.u64 	%rd4622, %rd93, 16;
	shl.b64 	%rd4623, %rd94, 48;
	or.b64  	%rd105, %rd4622, %rd4623;
	cvt.u16.u64	%rs117, %rd4622;
	cvt.s16.s8 	%rs118, %rs117;
	setp.lt.s16	%p78, %rs118, 0;
	mov.u64 	%rd6126, %rd6133;
	@%p78 bra 	BB0_140;

	and.b64  	%rd4624, %rd105, 15;
	add.s64 	%rd4625, %rd12, %rd4624;
	ld.local.u8 	%rd4626, [%rd4625];
	shl.b64 	%rd6126, %rd4626, 16;

BB0_140:
	or.b64  	%rd4630, %rd6125, %rd6124;
	or.b64  	%rd111, %rd4630, %rd6126;
	shr.u64 	%rd4631, %rd93, 24;
	shl.b64 	%rd4632, %rd94, 40;
	or.b64  	%rd113, %rd4631, %rd4632;
	cvt.u16.u64	%rs119, %rd4631;
	cvt.s16.s8 	%rs120, %rs119;
	setp.lt.s16	%p79, %rs120, 0;
	mov.u64 	%rd6127, %rd6133;
	@%p79 bra 	BB0_142;

	and.b64  	%rd4633, %rd113, 15;
	add.s64 	%rd4634, %rd12, %rd4633;
	ld.local.u8 	%rd4635, [%rd4634];
	shl.b64 	%rd6127, %rd4635, 24;

BB0_142:
	shr.u64 	%rd4638, %rd93, 32;
	shl.b64 	%rd4639, %rd94, 32;
	or.b64  	%rd119, %rd4638, %rd4639;
	cvt.u16.u64	%rs121, %rd4638;
	cvt.s16.s8 	%rs122, %rs121;
	setp.lt.s16	%p80, %rs122, 0;
	mov.u64 	%rd6128, %rd6133;
	@%p80 bra 	BB0_144;

	and.b64  	%rd4640, %rd119, 15;
	add.s64 	%rd4641, %rd12, %rd4640;
	ld.local.u8 	%rd4642, [%rd4641];
	shl.b64 	%rd6128, %rd4642, 32;

BB0_144:
	and.b64  	%rd4645, %rd111, 16777215;
	or.b64  	%rd4646, %rd6127, %rd4645;
	or.b64  	%rd125, %rd4646, %rd6128;
	shr.u64 	%rd4647, %rd93, 40;
	shl.b64 	%rd4648, %rd94, 24;
	or.b64  	%rd127, %rd4647, %rd4648;
	cvt.u16.u64	%rs123, %rd4647;
	cvt.s16.s8 	%rs124, %rs123;
	setp.lt.s16	%p81, %rs124, 0;
	mov.u64 	%rd6129, %rd6133;
	@%p81 bra 	BB0_146;

	and.b64  	%rd4649, %rd127, 15;
	add.s64 	%rd4650, %rd12, %rd4649;
	ld.local.u8 	%rd4651, [%rd4650];
	shl.b64 	%rd6129, %rd4651, 40;

BB0_146:
	shr.u64 	%rd4654, %rd93, 48;
	shl.b64 	%rd4655, %rd94, 16;
	or.b64  	%rd133, %rd4654, %rd4655;
	cvt.u16.u64	%rs125, %rd4654;
	cvt.s16.s8 	%rs126, %rs125;
	setp.lt.s16	%p82, %rs126, 0;
	mov.u64 	%rd6130, %rd6133;
	@%p82 bra 	BB0_148;

	and.b64  	%rd4656, %rd133, 15;
	add.s64 	%rd4657, %rd12, %rd4656;
	ld.local.u8 	%rd4658, [%rd4657];
	shl.b64 	%rd6130, %rd4658, 48;

BB0_148:
	and.b64  	%rd4661, %rd125, 1099511627775;
	or.b64  	%rd4662, %rd6129, %rd4661;
	or.b64  	%rd139, %rd4662, %rd6130;
	shr.u64 	%rd4663, %rd93, 56;
	shl.b64 	%rd4664, %rd94, 8;
	or.b64  	%rd141, %rd4663, %rd4664;
	cvt.u16.u64	%rs127, %rd4663;
	cvt.s16.s8 	%rs128, %rs127;
	setp.lt.s16	%p83, %rs128, 0;
	mov.u64 	%rd6131, %rd6133;
	@%p83 bra 	BB0_150;

	and.b64  	%rd4665, %rd141, 15;
	add.s64 	%rd4666, %rd12, %rd4665;
	ld.local.u8 	%rd4667, [%rd4666];
	shl.b64 	%rd6131, %rd4667, 56;

BB0_150:
	cvt.u16.u64	%rs129, %rd94;
	cvt.s16.s8 	%rs130, %rs129;
	setp.lt.s16	%p84, %rs130, 0;
	and.b64  	%rd4674, %rd139, 72057594037927935;
	or.b64  	%rd4675, %rd6131, %rd4674;
	or.b64  	%rd153, %rd4675, %rd6133;
	shr.u64 	%rd155, %rd94, 8;
	cvt.u16.u64	%rs131, %rd155;
	cvt.s16.s8 	%rs132, %rs131;
	setp.lt.s16	%p85, %rs132, 0;
	shr.u64 	%rd161, %rd94, 16;
	cvt.u16.u64	%rs133, %rd161;
	cvt.s16.s8 	%rs134, %rs133;
	setp.lt.s16	%p86, %rs134, 0;
	or.b64  	%rd4688, %rd6133, %rd153;
	or.b64  	%rd167, %rd4688, %rd6133;
	shr.u64 	%rd169, %rd94, 24;
	cvt.u16.u64	%rs135, %rd169;
	cvt.s16.s8 	%rs136, %rs135;
	setp.lt.s16	%p87, %rs136, 0;
	shr.u64 	%rd175, %rd94, 32;
	cvt.u16.u64	%rs137, %rd175;
	cvt.s16.s8 	%rs138, %rs137;
	setp.lt.s16	%p88, %rs138, 0;
	or.b64  	%rd4701, %rd6133, %rd167;
	or.b64  	%rd181, %rd4701, %rd6133;
	shr.u64 	%rd183, %rd94, 40;
	cvt.u16.u64	%rs139, %rd183;
	cvt.s16.s8 	%rs140, %rs139;
	setp.lt.s16	%p89, %rs140, 0;
	shr.u64 	%rd189, %rd94, 48;
	cvt.u16.u64	%rs141, %rd189;
	cvt.s16.s8 	%rs142, %rs141;
	setp.lt.s16	%p90, %rs142, 0;
	or.b64  	%rd4714, %rd6133, %rd181;
	or.b64  	%rd195, %rd4714, %rd6133;
	shr.u64 	%rd197, %rd94, 56;
	cvt.u16.u64	%rs143, %rd197;
	cvt.s16.s8 	%rs144, %rs143;
	setp.lt.s16	%p91, %rs144, 0;
	ld.param.u64 	%rd6067, [verus_gpu_hash_param_4];
	or.b64  	%rd4718, %rd6133, %rd195;
	xor.b32  	%r4146, %r6730, %r736;
	mov.b64	{%r4147, %r4148}, %rd4718;
	xor.b32  	%r4149, %r4146, %r4147;
	xor.b32  	%r4150, %r4149, 65536;
	xor.b32  	%r4151, %r737, %r6731;
	xor.b32  	%r4152, %r4151, %r4148;
	st.local.u8 	[%rd1+47], %r4149;
	shr.u32 	%r4153, %r4150, 8;
	shl.b32 	%r4154, %r4152, 24;
	shr.u32 	%r4155, %r4152, 8;
	shl.b32 	%r4156, %r4150, 24;
	or.b32  	%r4157, %r4155, %r4156;
	or.b32  	%r4158, %r4153, %r4154;
	st.local.v4.u32 	[%rd1+48], {%r4158, %r4157, %r4158, %r4157};
	and.b32  	%r4159, %r4150, 511;
	cvt.u64.u32	%rd4721, %r4159;
	add.s64 	%rd4722, %rd4721, %rd3;
	shl.b64 	%rd4723, %rd4722, 4;
	add.s64 	%rd4724, %rd203, %rd4723;
	and.b32  	%r4160, %r744, 255;
	mul.wide.u32 	%rd4725, %r4160, 4;
	mov.u64 	%rd4726, verus_gpu_hash$sharedMemory1;
	add.s64 	%rd4727, %rd4726, %rd4725;
	and.b32  	%r4161, %r745, 255;
	mul.wide.u32 	%rd4728, %r4161, 4;
	add.s64 	%rd4729, %rd4726, %rd4728;
	and.b32  	%r4162, %r746, 255;
	mul.wide.u32 	%rd4730, %r4162, 4;
	add.s64 	%rd4731, %rd4726, %rd4730;
	and.b32  	%r4163, %r747, 255;
	mul.wide.u32 	%rd4732, %r4163, 4;
	add.s64 	%rd4733, %rd4726, %rd4732;
	bfe.u32 	%r4164, %r745, 8, 8;
	mul.wide.u32 	%rd4734, %r4164, 4;
	add.s64 	%rd4735, %rd4734, %rd4726;
	ld.shared.u32 	%r4165, [%rd4735+1024];
	ld.shared.u32 	%r4166, [%rd4727];
	xor.b32  	%r4167, %r4165, %r4166;
	bfe.u32 	%r4168, %r746, 8, 8;
	mul.wide.u32 	%rd4736, %r4168, 4;
	add.s64 	%rd4737, %rd4736, %rd4726;
	ld.shared.u32 	%r4169, [%rd4737+1024];
	ld.shared.u32 	%r4170, [%rd4729];
	xor.b32  	%r4171, %r4169, %r4170;
	bfe.u32 	%r4172, %r747, 8, 8;
	mul.wide.u32 	%rd4738, %r4172, 4;
	add.s64 	%rd4739, %rd4738, %rd4726;
	ld.shared.u32 	%r4173, [%rd4739+1024];
	ld.shared.u32 	%r4174, [%rd4731];
	xor.b32  	%r4175, %r4173, %r4174;
	bfe.u32 	%r4176, %r744, 8, 8;
	mul.wide.u32 	%rd4740, %r4176, 4;
	add.s64 	%rd4741, %rd4740, %rd4726;
	ld.shared.u32 	%r4177, [%rd4741+1024];
	ld.shared.u32 	%r4178, [%rd4733];
	xor.b32  	%r4179, %r4177, %r4178;
	bfe.u32 	%r4180, %r746, 16, 8;
	mul.wide.u32 	%rd4742, %r4180, 4;
	add.s64 	%rd4743, %rd4742, %rd4726;
	ld.shared.u32 	%r4181, [%rd4743+2048];
	xor.b32  	%r4182, %r4167, %r4181;
	shr.u32 	%r4183, %r746, 24;
	bfe.u32 	%r4184, %r747, 16, 8;
	mul.wide.u32 	%rd4744, %r4184, 4;
	add.s64 	%rd4745, %rd4744, %rd4726;
	ld.shared.u32 	%r4185, [%rd4745+2048];
	xor.b32  	%r4186, %r4171, %r4185;
	shr.u32 	%r4187, %r747, 24;
	bfe.u32 	%r4188, %r744, 16, 8;
	mul.wide.u32 	%rd4746, %r4188, 4;
	add.s64 	%rd4747, %rd4746, %rd4726;
	ld.shared.u32 	%r4189, [%rd4747+2048];
	xor.b32  	%r4190, %r4175, %r4189;
	shr.u32 	%r4191, %r744, 24;
	bfe.u32 	%r4192, %r745, 16, 8;
	mul.wide.u32 	%rd4748, %r4192, 4;
	add.s64 	%rd4749, %rd4748, %rd4726;
	ld.shared.u32 	%r4193, [%rd4749+2048];
	xor.b32  	%r4194, %r4179, %r4193;
	shr.u32 	%r4195, %r745, 24;
	mul.wide.u32 	%rd4750, %r4187, 4;
	add.s64 	%rd4751, %rd4750, %rd4726;
	ld.shared.u32 	%r4196, [%rd4751+3072];
	mul.wide.u32 	%rd4752, %r4191, 4;
	add.s64 	%rd4753, %rd4752, %rd4726;
	ld.shared.u32 	%r4197, [%rd4753+3072];
	mul.wide.u32 	%rd4754, %r4195, 4;
	add.s64 	%rd4755, %rd4754, %rd4726;
	ld.shared.u32 	%r4198, [%rd4755+3072];
	mul.wide.u32 	%rd4756, %r4183, 4;
	add.s64 	%rd4757, %rd4756, %rd4726;
	ld.shared.u32 	%r4199, [%rd4757+3072];
	xor.b32  	%r4200, %r4194, %r4199;
	xor.b32  	%r4201, %r4190, %r4198;
	xor.b32  	%r4202, %r4186, %r4197;
	xor.b32  	%r4203, %r4182, %r4196;
	ld.global.v4.u32 	{%r4204, %r4205, %r4206, %r4207}, [%rd4724];
	and.b32  	%r4212, %r748, 255;
	mul.wide.u32 	%rd4758, %r4212, 4;
	add.s64 	%rd4759, %rd4726, %rd4758;
	and.b32  	%r4213, %r749, 255;
	mul.wide.u32 	%rd4760, %r4213, 4;
	add.s64 	%rd4761, %rd4726, %rd4760;
	and.b32  	%r4214, %r750, 255;
	mul.wide.u32 	%rd4762, %r4214, 4;
	add.s64 	%rd4763, %rd4726, %rd4762;
	and.b32  	%r4215, %r751, 255;
	mul.wide.u32 	%rd4764, %r4215, 4;
	add.s64 	%rd4765, %rd4726, %rd4764;
	bfe.u32 	%r4216, %r749, 8, 8;
	mul.wide.u32 	%rd4766, %r4216, 4;
	add.s64 	%rd4767, %rd4766, %rd4726;
	ld.shared.u32 	%r4217, [%rd4767+1024];
	ld.shared.u32 	%r4218, [%rd4759];
	xor.b32  	%r4219, %r4217, %r4218;
	bfe.u32 	%r4220, %r750, 8, 8;
	mul.wide.u32 	%rd4768, %r4220, 4;
	add.s64 	%rd4769, %rd4768, %rd4726;
	ld.shared.u32 	%r4221, [%rd4769+1024];
	ld.shared.u32 	%r4222, [%rd4761];
	xor.b32  	%r4223, %r4221, %r4222;
	bfe.u32 	%r4224, %r751, 8, 8;
	mul.wide.u32 	%rd4770, %r4224, 4;
	add.s64 	%rd4771, %rd4770, %rd4726;
	ld.shared.u32 	%r4225, [%rd4771+1024];
	ld.shared.u32 	%r4226, [%rd4763];
	xor.b32  	%r4227, %r4225, %r4226;
	bfe.u32 	%r4228, %r748, 8, 8;
	mul.wide.u32 	%rd4772, %r4228, 4;
	add.s64 	%rd4773, %rd4772, %rd4726;
	ld.shared.u32 	%r4229, [%rd4773+1024];
	ld.shared.u32 	%r4230, [%rd4765];
	xor.b32  	%r4231, %r4229, %r4230;
	bfe.u32 	%r4232, %r750, 16, 8;
	mul.wide.u32 	%rd4774, %r4232, 4;
	add.s64 	%rd4775, %rd4774, %rd4726;
	ld.shared.u32 	%r4233, [%rd4775+2048];
	xor.b32  	%r4234, %r4219, %r4233;
	shr.u32 	%r4235, %r750, 24;
	bfe.u32 	%r4236, %r751, 16, 8;
	mul.wide.u32 	%rd4776, %r4236, 4;
	add.s64 	%rd4777, %rd4776, %rd4726;
	ld.shared.u32 	%r4237, [%rd4777+2048];
	xor.b32  	%r4238, %r4223, %r4237;
	shr.u32 	%r4239, %r751, 24;
	bfe.u32 	%r4240, %r748, 16, 8;
	mul.wide.u32 	%rd4778, %r4240, 4;
	add.s64 	%rd4779, %rd4778, %rd4726;
	ld.shared.u32 	%r4241, [%rd4779+2048];
	xor.b32  	%r4242, %r4227, %r4241;
	shr.u32 	%r4243, %r748, 24;
	bfe.u32 	%r4244, %r749, 16, 8;
	mul.wide.u32 	%rd4780, %r4244, 4;
	add.s64 	%rd4781, %rd4780, %rd4726;
	ld.shared.u32 	%r4245, [%rd4781+2048];
	xor.b32  	%r4246, %r4231, %r4245;
	shr.u32 	%r4247, %r749, 24;
	mul.wide.u32 	%rd4782, %r4239, 4;
	add.s64 	%rd4783, %rd4782, %rd4726;
	ld.shared.u32 	%r4248, [%rd4783+3072];
	mul.wide.u32 	%rd4784, %r4243, 4;
	add.s64 	%rd4785, %rd4784, %rd4726;
	ld.shared.u32 	%r4249, [%rd4785+3072];
	mul.wide.u32 	%rd4786, %r4247, 4;
	add.s64 	%rd4787, %rd4786, %rd4726;
	ld.shared.u32 	%r4250, [%rd4787+3072];
	mul.wide.u32 	%rd4788, %r4235, 4;
	add.s64 	%rd4789, %rd4788, %rd4726;
	ld.shared.u32 	%r4251, [%rd4789+3072];
	xor.b32  	%r4252, %r4246, %r4251;
	xor.b32  	%r4253, %r4242, %r4250;
	xor.b32  	%r4254, %r4238, %r4249;
	xor.b32  	%r4255, %r4234, %r4248;
	ld.global.v4.u32 	{%r4256, %r4257, %r4258, %r4259}, [%rd4724+16];
	ld.local.v4.u32 	{%r4264, %r4265, %r4266, %r4267}, [%rd1+32];
	and.b32  	%r4269, %r4264, 255;
	mul.wide.u32 	%rd4790, %r4269, 4;
	add.s64 	%rd4791, %rd4726, %rd4790;
	and.b32  	%r4271, %r4265, 255;
	mul.wide.u32 	%rd4792, %r4271, 4;
	add.s64 	%rd4793, %rd4726, %rd4792;
	and.b32  	%r4273, %r4266, 255;
	mul.wide.u32 	%rd4794, %r4273, 4;
	add.s64 	%rd4795, %rd4726, %rd4794;
	and.b32  	%r4275, %r4267, 255;
	mul.wide.u32 	%rd4796, %r4275, 4;
	add.s64 	%rd4797, %rd4726, %rd4796;
	bfe.u32 	%r4276, %r4265, 8, 8;
	mul.wide.u32 	%rd4798, %r4276, 4;
	add.s64 	%rd4799, %rd4798, %rd4726;
	ld.shared.u32 	%r4277, [%rd4799+1024];
	ld.shared.u32 	%r4278, [%rd4791];
	xor.b32  	%r4279, %r4277, %r4278;
	bfe.u32 	%r4280, %r4266, 8, 8;
	mul.wide.u32 	%rd4800, %r4280, 4;
	add.s64 	%rd4801, %rd4800, %rd4726;
	ld.shared.u32 	%r4281, [%rd4801+1024];
	ld.shared.u32 	%r4282, [%rd4793];
	xor.b32  	%r4283, %r4281, %r4282;
	bfe.u32 	%r4284, %r4267, 8, 8;
	mul.wide.u32 	%rd4802, %r4284, 4;
	add.s64 	%rd4803, %rd4802, %rd4726;
	ld.shared.u32 	%r4285, [%rd4803+1024];
	ld.shared.u32 	%r4286, [%rd4795];
	xor.b32  	%r4287, %r4285, %r4286;
	bfe.u32 	%r4288, %r4264, 8, 8;
	mul.wide.u32 	%rd4804, %r4288, 4;
	add.s64 	%rd4805, %rd4804, %rd4726;
	ld.shared.u32 	%r4289, [%rd4805+1024];
	ld.shared.u32 	%r4290, [%rd4797];
	xor.b32  	%r4291, %r4289, %r4290;
	bfe.u32 	%r4292, %r4266, 16, 8;
	mul.wide.u32 	%rd4806, %r4292, 4;
	add.s64 	%rd4807, %rd4806, %rd4726;
	ld.shared.u32 	%r4293, [%rd4807+2048];
	xor.b32  	%r4294, %r4279, %r4293;
	shr.u32 	%r4295, %r4266, 24;
	bfe.u32 	%r4296, %r4267, 16, 8;
	mul.wide.u32 	%rd4808, %r4296, 4;
	add.s64 	%rd4809, %rd4808, %rd4726;
	ld.shared.u32 	%r4297, [%rd4809+2048];
	xor.b32  	%r4298, %r4283, %r4297;
	shr.u32 	%r4299, %r4267, 24;
	bfe.u32 	%r4300, %r4264, 16, 8;
	mul.wide.u32 	%rd4810, %r4300, 4;
	add.s64 	%rd4811, %rd4810, %rd4726;
	ld.shared.u32 	%r4301, [%rd4811+2048];
	xor.b32  	%r4302, %r4287, %r4301;
	shr.u32 	%r4303, %r4264, 24;
	bfe.u32 	%r4304, %r4265, 16, 8;
	mul.wide.u32 	%rd4812, %r4304, 4;
	add.s64 	%rd4813, %rd4812, %rd4726;
	ld.shared.u32 	%r4305, [%rd4813+2048];
	xor.b32  	%r4306, %r4291, %r4305;
	shr.u32 	%r4307, %r4265, 24;
	mul.wide.u32 	%rd4814, %r4299, 4;
	add.s64 	%rd4815, %rd4814, %rd4726;
	ld.shared.u32 	%r4308, [%rd4815+3072];
	mul.wide.u32 	%rd4816, %r4303, 4;
	add.s64 	%rd4817, %rd4816, %rd4726;
	ld.shared.u32 	%r4309, [%rd4817+3072];
	mul.wide.u32 	%rd4818, %r4307, 4;
	add.s64 	%rd4819, %rd4818, %rd4726;
	ld.shared.u32 	%r4310, [%rd4819+3072];
	mul.wide.u32 	%rd4820, %r4295, 4;
	add.s64 	%rd4821, %rd4820, %rd4726;
	ld.shared.u32 	%r4311, [%rd4821+3072];
	xor.b32  	%r4312, %r4306, %r4311;
	xor.b32  	%r4313, %r4302, %r4310;
	xor.b32  	%r4314, %r4298, %r4309;
	xor.b32  	%r4315, %r4294, %r4308;
	ld.global.v4.u32 	{%r4316, %r4317, %r4318, %r4319}, [%rd4724+32];
	bfe.u32 	%r4324, %r4150, 8, 8;
	mul.wide.u32 	%rd4822, %r4324, 4;
	add.s64 	%rd4823, %rd4726, %rd4822;
	bfe.u32 	%r4325, %r4152, 8, 8;
	mul.wide.u32 	%rd4824, %r4325, 4;
	add.s64 	%rd4825, %rd4726, %rd4824;
	bfe.u32 	%r4326, %r4152, 16, 8;
	mul.wide.u32 	%rd4826, %r4326, 4;
	add.s64 	%rd4827, %rd4826, %rd4726;
	ld.shared.u32 	%r4327, [%rd4827+1024];
	ld.shared.u32 	%r4328, [%rd4823];
	xor.b32  	%r4329, %r4327, %r4328;
	bfe.u32 	%r4330, %r4150, 16, 8;
	mul.wide.u32 	%rd4828, %r4330, 4;
	add.s64 	%rd4829, %rd4828, %rd4726;
	ld.shared.u32 	%r4331, [%rd4829+1024];
	ld.shared.u32 	%r4332, [%rd4825];
	xor.b32  	%r4333, %r4331, %r4332;
	shr.u32 	%r4334, %r4150, 24;
	shr.u32 	%r4335, %r4152, 24;
	mul.wide.u32 	%rd4830, %r4334, 4;
	add.s64 	%rd4831, %rd4830, %rd4726;
	ld.shared.u32 	%r4336, [%rd4831+2048];
	xor.b32  	%r4337, %r4329, %r4336;
	mul.wide.u32 	%rd4832, %r4335, 4;
	add.s64 	%rd4833, %rd4832, %rd4726;
	ld.shared.u32 	%r4338, [%rd4833+2048];
	xor.b32  	%r4339, %r4333, %r4338;
	and.b32  	%r4340, %r4150, 255;
	and.b32  	%r4341, %r4152, 255;
	mul.wide.u32 	%rd4834, %r4340, 4;
	add.s64 	%rd4835, %rd4834, %rd4726;
	ld.shared.u32 	%r4342, [%rd4835+3072];
	mul.wide.u32 	%rd4836, %r4341, 4;
	add.s64 	%rd4837, %rd4836, %rd4726;
	ld.shared.u32 	%r4343, [%rd4837+3072];
	xor.b32  	%r4344, %r4339, %r4343;
	xor.b32  	%r4345, %r4337, %r4342;
	ld.global.v4.u32 	{%r4346, %r4347, %r4348, %r4349}, [%rd4724+48];
	xor.b32  	%r4354, %r4203, %r4204;
	and.b32  	%r4355, %r4354, 255;
	mul.wide.u32 	%rd4838, %r4355, 4;
	add.s64 	%rd4839, %rd4726, %rd4838;
	xor.b32  	%r4356, %r4202, %r4205;
	and.b32  	%r4357, %r4356, 255;
	mul.wide.u32 	%rd4840, %r4357, 4;
	add.s64 	%rd4841, %rd4726, %rd4840;
	xor.b32  	%r4358, %r4201, %r4206;
	and.b32  	%r4359, %r4358, 255;
	mul.wide.u32 	%rd4842, %r4359, 4;
	add.s64 	%rd4843, %rd4726, %rd4842;
	xor.b32  	%r4360, %r4200, %r4207;
	and.b32  	%r4361, %r4360, 255;
	mul.wide.u32 	%rd4844, %r4361, 4;
	add.s64 	%rd4845, %rd4726, %rd4844;
	bfe.u32 	%r4362, %r4356, 8, 8;
	mul.wide.u32 	%rd4846, %r4362, 4;
	add.s64 	%rd4847, %rd4846, %rd4726;
	ld.shared.u32 	%r4363, [%rd4847+1024];
	ld.shared.u32 	%r4364, [%rd4839];
	xor.b32  	%r4365, %r4363, %r4364;
	bfe.u32 	%r4366, %r4358, 8, 8;
	mul.wide.u32 	%rd4848, %r4366, 4;
	add.s64 	%rd4849, %rd4848, %rd4726;
	ld.shared.u32 	%r4367, [%rd4849+1024];
	ld.shared.u32 	%r4368, [%rd4841];
	xor.b32  	%r4369, %r4367, %r4368;
	bfe.u32 	%r4370, %r4360, 8, 8;
	mul.wide.u32 	%rd4850, %r4370, 4;
	add.s64 	%rd4851, %rd4850, %rd4726;
	ld.shared.u32 	%r4371, [%rd4851+1024];
	ld.shared.u32 	%r4372, [%rd4843];
	xor.b32  	%r4373, %r4371, %r4372;
	bfe.u32 	%r4374, %r4354, 8, 8;
	mul.wide.u32 	%rd4852, %r4374, 4;
	add.s64 	%rd4853, %rd4852, %rd4726;
	ld.shared.u32 	%r4375, [%rd4853+1024];
	ld.shared.u32 	%r4376, [%rd4845];
	xor.b32  	%r4377, %r4375, %r4376;
	bfe.u32 	%r4378, %r4358, 16, 8;
	mul.wide.u32 	%rd4854, %r4378, 4;
	add.s64 	%rd4855, %rd4854, %rd4726;
	ld.shared.u32 	%r4379, [%rd4855+2048];
	xor.b32  	%r4380, %r4365, %r4379;
	shr.u32 	%r4381, %r4358, 24;
	bfe.u32 	%r4382, %r4360, 16, 8;
	mul.wide.u32 	%rd4856, %r4382, 4;
	add.s64 	%rd4857, %rd4856, %rd4726;
	ld.shared.u32 	%r4383, [%rd4857+2048];
	xor.b32  	%r4384, %r4369, %r4383;
	shr.u32 	%r4385, %r4360, 24;
	bfe.u32 	%r4386, %r4354, 16, 8;
	mul.wide.u32 	%rd4858, %r4386, 4;
	add.s64 	%rd4859, %rd4858, %rd4726;
	ld.shared.u32 	%r4387, [%rd4859+2048];
	xor.b32  	%r4388, %r4373, %r4387;
	shr.u32 	%r4389, %r4354, 24;
	bfe.u32 	%r4390, %r4356, 16, 8;
	mul.wide.u32 	%rd4860, %r4390, 4;
	add.s64 	%rd4861, %rd4860, %rd4726;
	ld.shared.u32 	%r4391, [%rd4861+2048];
	xor.b32  	%r4392, %r4377, %r4391;
	shr.u32 	%r4393, %r4356, 24;
	mul.wide.u32 	%rd4862, %r4385, 4;
	add.s64 	%rd4863, %rd4862, %rd4726;
	ld.shared.u32 	%r4394, [%rd4863+3072];
	mul.wide.u32 	%rd4864, %r4389, 4;
	add.s64 	%rd4865, %rd4864, %rd4726;
	ld.shared.u32 	%r4395, [%rd4865+3072];
	mul.wide.u32 	%rd4866, %r4393, 4;
	add.s64 	%rd4867, %rd4866, %rd4726;
	ld.shared.u32 	%r4396, [%rd4867+3072];
	mul.wide.u32 	%rd4868, %r4381, 4;
	add.s64 	%rd4869, %rd4868, %rd4726;
	ld.shared.u32 	%r4397, [%rd4869+3072];
	xor.b32  	%r4398, %r4388, %r4396;
	xor.b32  	%r4399, %r4384, %r4395;
	xor.b32  	%r4400, %r4380, %r4394;
	xor.b32  	%r4401, %r4392, %r4397;
	ld.global.v4.u32 	{%r4402, %r4403, %r4404, %r4405}, [%rd4724+64];
	xor.b32  	%r4410, %r4255, %r4256;
	and.b32  	%r4411, %r4410, 255;
	mul.wide.u32 	%rd4870, %r4411, 4;
	add.s64 	%rd4871, %rd4726, %rd4870;
	xor.b32  	%r4412, %r4254, %r4257;
	and.b32  	%r4413, %r4412, 255;
	mul.wide.u32 	%rd4872, %r4413, 4;
	add.s64 	%rd4873, %rd4726, %rd4872;
	xor.b32  	%r4414, %r4253, %r4258;
	and.b32  	%r4415, %r4414, 255;
	mul.wide.u32 	%rd4874, %r4415, 4;
	add.s64 	%rd4875, %rd4726, %rd4874;
	xor.b32  	%r4416, %r4252, %r4259;
	and.b32  	%r4417, %r4416, 255;
	mul.wide.u32 	%rd4876, %r4417, 4;
	add.s64 	%rd4877, %rd4726, %rd4876;
	bfe.u32 	%r4418, %r4412, 8, 8;
	mul.wide.u32 	%rd4878, %r4418, 4;
	add.s64 	%rd4879, %rd4878, %rd4726;
	ld.shared.u32 	%r4419, [%rd4879+1024];
	ld.shared.u32 	%r4420, [%rd4871];
	xor.b32  	%r4421, %r4419, %r4420;
	bfe.u32 	%r4422, %r4414, 8, 8;
	mul.wide.u32 	%rd4880, %r4422, 4;
	add.s64 	%rd4881, %rd4880, %rd4726;
	ld.shared.u32 	%r4423, [%rd4881+1024];
	ld.shared.u32 	%r4424, [%rd4873];
	xor.b32  	%r4425, %r4423, %r4424;
	bfe.u32 	%r4426, %r4416, 8, 8;
	mul.wide.u32 	%rd4882, %r4426, 4;
	add.s64 	%rd4883, %rd4882, %rd4726;
	ld.shared.u32 	%r4427, [%rd4883+1024];
	ld.shared.u32 	%r4428, [%rd4875];
	xor.b32  	%r4429, %r4427, %r4428;
	bfe.u32 	%r4430, %r4410, 8, 8;
	mul.wide.u32 	%rd4884, %r4430, 4;
	add.s64 	%rd4885, %rd4884, %rd4726;
	ld.shared.u32 	%r4431, [%rd4885+1024];
	ld.shared.u32 	%r4432, [%rd4877];
	xor.b32  	%r4433, %r4431, %r4432;
	bfe.u32 	%r4434, %r4414, 16, 8;
	mul.wide.u32 	%rd4886, %r4434, 4;
	add.s64 	%rd4887, %rd4886, %rd4726;
	ld.shared.u32 	%r4435, [%rd4887+2048];
	xor.b32  	%r4436, %r4421, %r4435;
	shr.u32 	%r4437, %r4414, 24;
	bfe.u32 	%r4438, %r4416, 16, 8;
	mul.wide.u32 	%rd4888, %r4438, 4;
	add.s64 	%rd4889, %rd4888, %rd4726;
	ld.shared.u32 	%r4439, [%rd4889+2048];
	xor.b32  	%r4440, %r4425, %r4439;
	shr.u32 	%r4441, %r4416, 24;
	bfe.u32 	%r4442, %r4410, 16, 8;
	mul.wide.u32 	%rd4890, %r4442, 4;
	add.s64 	%rd4891, %rd4890, %rd4726;
	ld.shared.u32 	%r4443, [%rd4891+2048];
	xor.b32  	%r4444, %r4429, %r4443;
	shr.u32 	%r4445, %r4410, 24;
	bfe.u32 	%r4446, %r4412, 16, 8;
	mul.wide.u32 	%rd4892, %r4446, 4;
	add.s64 	%rd4893, %rd4892, %rd4726;
	ld.shared.u32 	%r4447, [%rd4893+2048];
	xor.b32  	%r4448, %r4433, %r4447;
	shr.u32 	%r4449, %r4412, 24;
	mul.wide.u32 	%rd4894, %r4441, 4;
	add.s64 	%rd4895, %rd4894, %rd4726;
	ld.shared.u32 	%r4450, [%rd4895+3072];
	mul.wide.u32 	%rd4896, %r4445, 4;
	add.s64 	%rd4897, %rd4896, %rd4726;
	ld.shared.u32 	%r4451, [%rd4897+3072];
	mul.wide.u32 	%rd4898, %r4449, 4;
	add.s64 	%rd4899, %rd4898, %rd4726;
	ld.shared.u32 	%r4452, [%rd4899+3072];
	mul.wide.u32 	%rd4900, %r4437, 4;
	add.s64 	%rd4901, %rd4900, %rd4726;
	ld.shared.u32 	%r4453, [%rd4901+3072];
	xor.b32  	%r4454, %r4444, %r4452;
	xor.b32  	%r4455, %r4440, %r4451;
	xor.b32  	%r4456, %r4436, %r4450;
	xor.b32  	%r4457, %r4448, %r4453;
	ld.global.v4.u32 	{%r4458, %r4459, %r4460, %r4461}, [%rd4724+80];
	xor.b32  	%r4466, %r4315, %r4316;
	and.b32  	%r4467, %r4466, 255;
	mul.wide.u32 	%rd4902, %r4467, 4;
	add.s64 	%rd4903, %rd4726, %rd4902;
	xor.b32  	%r4468, %r4314, %r4317;
	and.b32  	%r4469, %r4468, 255;
	mul.wide.u32 	%rd4904, %r4469, 4;
	add.s64 	%rd4905, %rd4726, %rd4904;
	xor.b32  	%r4470, %r4313, %r4318;
	and.b32  	%r4471, %r4470, 255;
	mul.wide.u32 	%rd4906, %r4471, 4;
	add.s64 	%rd4907, %rd4726, %rd4906;
	xor.b32  	%r4472, %r4312, %r4319;
	and.b32  	%r4473, %r4472, 255;
	mul.wide.u32 	%rd4908, %r4473, 4;
	add.s64 	%rd4909, %rd4726, %rd4908;
	bfe.u32 	%r4474, %r4468, 8, 8;
	mul.wide.u32 	%rd4910, %r4474, 4;
	add.s64 	%rd4911, %rd4910, %rd4726;
	ld.shared.u32 	%r4475, [%rd4911+1024];
	ld.shared.u32 	%r4476, [%rd4903];
	xor.b32  	%r4477, %r4475, %r4476;
	bfe.u32 	%r4478, %r4470, 8, 8;
	mul.wide.u32 	%rd4912, %r4478, 4;
	add.s64 	%rd4913, %rd4912, %rd4726;
	ld.shared.u32 	%r4479, [%rd4913+1024];
	ld.shared.u32 	%r4480, [%rd4905];
	xor.b32  	%r4481, %r4479, %r4480;
	bfe.u32 	%r4482, %r4472, 8, 8;
	mul.wide.u32 	%rd4914, %r4482, 4;
	add.s64 	%rd4915, %rd4914, %rd4726;
	ld.shared.u32 	%r4483, [%rd4915+1024];
	ld.shared.u32 	%r4484, [%rd4907];
	xor.b32  	%r4485, %r4483, %r4484;
	bfe.u32 	%r4486, %r4466, 8, 8;
	mul.wide.u32 	%rd4916, %r4486, 4;
	add.s64 	%rd4917, %rd4916, %rd4726;
	ld.shared.u32 	%r4487, [%rd4917+1024];
	ld.shared.u32 	%r4488, [%rd4909];
	xor.b32  	%r4489, %r4487, %r4488;
	bfe.u32 	%r4490, %r4470, 16, 8;
	mul.wide.u32 	%rd4918, %r4490, 4;
	add.s64 	%rd4919, %rd4918, %rd4726;
	ld.shared.u32 	%r4491, [%rd4919+2048];
	xor.b32  	%r4492, %r4477, %r4491;
	shr.u32 	%r4493, %r4470, 24;
	bfe.u32 	%r4494, %r4472, 16, 8;
	mul.wide.u32 	%rd4920, %r4494, 4;
	add.s64 	%rd4921, %rd4920, %rd4726;
	ld.shared.u32 	%r4495, [%rd4921+2048];
	xor.b32  	%r4496, %r4481, %r4495;
	shr.u32 	%r4497, %r4472, 24;
	bfe.u32 	%r4498, %r4466, 16, 8;
	mul.wide.u32 	%rd4922, %r4498, 4;
	add.s64 	%rd4923, %rd4922, %rd4726;
	ld.shared.u32 	%r4499, [%rd4923+2048];
	xor.b32  	%r4500, %r4485, %r4499;
	shr.u32 	%r4501, %r4466, 24;
	bfe.u32 	%r4502, %r4468, 16, 8;
	mul.wide.u32 	%rd4924, %r4502, 4;
	add.s64 	%rd4925, %rd4924, %rd4726;
	ld.shared.u32 	%r4503, [%rd4925+2048];
	xor.b32  	%r4504, %r4489, %r4503;
	shr.u32 	%r4505, %r4468, 24;
	mul.wide.u32 	%rd4926, %r4497, 4;
	add.s64 	%rd4927, %rd4926, %rd4726;
	ld.shared.u32 	%r4506, [%rd4927+3072];
	mul.wide.u32 	%rd4928, %r4501, 4;
	add.s64 	%rd4929, %rd4928, %rd4726;
	ld.shared.u32 	%r4507, [%rd4929+3072];
	mul.wide.u32 	%rd4930, %r4505, 4;
	add.s64 	%rd4931, %rd4930, %rd4726;
	ld.shared.u32 	%r4508, [%rd4931+3072];
	mul.wide.u32 	%rd4932, %r4493, 4;
	add.s64 	%rd4933, %rd4932, %rd4726;
	ld.shared.u32 	%r4509, [%rd4933+3072];
	xor.b32  	%r4510, %r4500, %r4508;
	xor.b32  	%r4511, %r4496, %r4507;
	xor.b32  	%r4512, %r4492, %r4506;
	xor.b32  	%r4513, %r4504, %r4509;
	ld.global.v4.u32 	{%r4514, %r4515, %r4516, %r4517}, [%rd4724+96];
	xor.b32  	%r4522, %r4345, %r4346;
	and.b32  	%r4523, %r4522, 255;
	mul.wide.u32 	%rd4934, %r4523, 4;
	add.s64 	%rd4935, %rd4726, %rd4934;
	xor.b32  	%r4524, %r4344, %r4347;
	and.b32  	%r4525, %r4524, 255;
	mul.wide.u32 	%rd4936, %r4525, 4;
	add.s64 	%rd4937, %rd4726, %rd4936;
	xor.b32  	%r4526, %r4345, %r4348;
	and.b32  	%r4527, %r4526, 255;
	mul.wide.u32 	%rd4938, %r4527, 4;
	add.s64 	%rd4939, %rd4726, %rd4938;
	xor.b32  	%r4528, %r4344, %r4349;
	and.b32  	%r4529, %r4528, 255;
	mul.wide.u32 	%rd4940, %r4529, 4;
	add.s64 	%rd4941, %rd4726, %rd4940;
	bfe.u32 	%r4530, %r4524, 8, 8;
	mul.wide.u32 	%rd4942, %r4530, 4;
	add.s64 	%rd4943, %rd4942, %rd4726;
	ld.shared.u32 	%r4531, [%rd4943+1024];
	ld.shared.u32 	%r4532, [%rd4935];
	xor.b32  	%r4533, %r4531, %r4532;
	bfe.u32 	%r4534, %r4526, 8, 8;
	mul.wide.u32 	%rd4944, %r4534, 4;
	add.s64 	%rd4945, %rd4944, %rd4726;
	ld.shared.u32 	%r4535, [%rd4945+1024];
	ld.shared.u32 	%r4536, [%rd4937];
	xor.b32  	%r4537, %r4535, %r4536;
	bfe.u32 	%r4538, %r4528, 8, 8;
	mul.wide.u32 	%rd4946, %r4538, 4;
	add.s64 	%rd4947, %rd4946, %rd4726;
	ld.shared.u32 	%r4539, [%rd4947+1024];
	ld.shared.u32 	%r4540, [%rd4939];
	xor.b32  	%r4541, %r4539, %r4540;
	bfe.u32 	%r4542, %r4522, 8, 8;
	mul.wide.u32 	%rd4948, %r4542, 4;
	add.s64 	%rd4949, %rd4948, %rd4726;
	ld.shared.u32 	%r4543, [%rd4949+1024];
	ld.shared.u32 	%r4544, [%rd4941];
	xor.b32  	%r4545, %r4543, %r4544;
	bfe.u32 	%r4546, %r4526, 16, 8;
	mul.wide.u32 	%rd4950, %r4546, 4;
	add.s64 	%rd4951, %rd4950, %rd4726;
	ld.shared.u32 	%r4547, [%rd4951+2048];
	xor.b32  	%r4548, %r4533, %r4547;
	shr.u32 	%r4549, %r4526, 24;
	bfe.u32 	%r4550, %r4528, 16, 8;
	mul.wide.u32 	%rd4952, %r4550, 4;
	add.s64 	%rd4953, %rd4952, %rd4726;
	ld.shared.u32 	%r4551, [%rd4953+2048];
	xor.b32  	%r4552, %r4537, %r4551;
	shr.u32 	%r4553, %r4528, 24;
	bfe.u32 	%r4554, %r4522, 16, 8;
	mul.wide.u32 	%rd4954, %r4554, 4;
	add.s64 	%rd4955, %rd4954, %rd4726;
	ld.shared.u32 	%r4555, [%rd4955+2048];
	xor.b32  	%r4556, %r4541, %r4555;
	shr.u32 	%r4557, %r4522, 24;
	bfe.u32 	%r4558, %r4524, 16, 8;
	mul.wide.u32 	%rd4956, %r4558, 4;
	add.s64 	%rd4957, %rd4956, %rd4726;
	ld.shared.u32 	%r4559, [%rd4957+2048];
	xor.b32  	%r4560, %r4545, %r4559;
	shr.u32 	%r4561, %r4524, 24;
	mul.wide.u32 	%rd4958, %r4553, 4;
	add.s64 	%rd4959, %rd4958, %rd4726;
	ld.shared.u32 	%r4562, [%rd4959+3072];
	mul.wide.u32 	%rd4960, %r4557, 4;
	add.s64 	%rd4961, %rd4960, %rd4726;
	ld.shared.u32 	%r4563, [%rd4961+3072];
	mul.wide.u32 	%rd4962, %r4561, 4;
	add.s64 	%rd4963, %rd4962, %rd4726;
	ld.shared.u32 	%r4564, [%rd4963+3072];
	mul.wide.u32 	%rd4964, %r4549, 4;
	add.s64 	%rd4965, %rd4964, %rd4726;
	ld.shared.u32 	%r4565, [%rd4965+3072];
	xor.b32  	%r4566, %r4556, %r4564;
	xor.b32  	%r4567, %r4552, %r4563;
	xor.b32  	%r4568, %r4548, %r4562;
	xor.b32  	%r4569, %r4560, %r4565;
	ld.global.v4.u32 	{%r4570, %r4571, %r4572, %r4573}, [%rd4724+112];
	xor.b32  	%r4578, %r4401, %r4405;
	and.b32  	%r4579, %r4578, 255;
	mul.wide.u32 	%rd4966, %r4579, 4;
	add.s64 	%rd4967, %rd4726, %rd4966;
	xor.b32  	%r4580, %r4513, %r4517;
	and.b32  	%r4581, %r4580, 255;
	mul.wide.u32 	%rd4968, %r4581, 4;
	add.s64 	%rd4969, %rd4726, %rd4968;
	xor.b32  	%r4582, %r4457, %r4461;
	and.b32  	%r4583, %r4582, 255;
	mul.wide.u32 	%rd4970, %r4583, 4;
	add.s64 	%rd4971, %rd4726, %rd4970;
	xor.b32  	%r4584, %r4569, %r4573;
	and.b32  	%r4585, %r4584, 255;
	mul.wide.u32 	%rd4972, %r4585, 4;
	add.s64 	%rd4973, %rd4726, %rd4972;
	bfe.u32 	%r4586, %r4580, 8, 8;
	mul.wide.u32 	%rd4974, %r4586, 4;
	add.s64 	%rd4975, %rd4974, %rd4726;
	ld.shared.u32 	%r4587, [%rd4975+1024];
	ld.shared.u32 	%r4588, [%rd4967];
	xor.b32  	%r4589, %r4587, %r4588;
	bfe.u32 	%r4590, %r4582, 8, 8;
	mul.wide.u32 	%rd4976, %r4590, 4;
	add.s64 	%rd4977, %rd4976, %rd4726;
	ld.shared.u32 	%r4591, [%rd4977+1024];
	ld.shared.u32 	%r4592, [%rd4969];
	xor.b32  	%r4593, %r4591, %r4592;
	bfe.u32 	%r4594, %r4584, 8, 8;
	mul.wide.u32 	%rd4978, %r4594, 4;
	add.s64 	%rd4979, %rd4978, %rd4726;
	ld.shared.u32 	%r4595, [%rd4979+1024];
	ld.shared.u32 	%r4596, [%rd4971];
	xor.b32  	%r4597, %r4595, %r4596;
	bfe.u32 	%r4598, %r4578, 8, 8;
	mul.wide.u32 	%rd4980, %r4598, 4;
	add.s64 	%rd4981, %rd4980, %rd4726;
	ld.shared.u32 	%r4599, [%rd4981+1024];
	ld.shared.u32 	%r4600, [%rd4973];
	xor.b32  	%r4601, %r4599, %r4600;
	bfe.u32 	%r4602, %r4582, 16, 8;
	mul.wide.u32 	%rd4982, %r4602, 4;
	add.s64 	%rd4983, %rd4982, %rd4726;
	ld.shared.u32 	%r4603, [%rd4983+2048];
	xor.b32  	%r4604, %r4589, %r4603;
	shr.u32 	%r4605, %r4582, 24;
	bfe.u32 	%r4606, %r4584, 16, 8;
	mul.wide.u32 	%rd4984, %r4606, 4;
	add.s64 	%rd4985, %rd4984, %rd4726;
	ld.shared.u32 	%r4607, [%rd4985+2048];
	xor.b32  	%r4608, %r4593, %r4607;
	shr.u32 	%r4609, %r4584, 24;
	bfe.u32 	%r4610, %r4578, 16, 8;
	mul.wide.u32 	%rd4986, %r4610, 4;
	add.s64 	%rd4987, %rd4986, %rd4726;
	ld.shared.u32 	%r4611, [%rd4987+2048];
	xor.b32  	%r4612, %r4597, %r4611;
	shr.u32 	%r4613, %r4578, 24;
	bfe.u32 	%r4614, %r4580, 16, 8;
	mul.wide.u32 	%rd4988, %r4614, 4;
	add.s64 	%rd4989, %rd4988, %rd4726;
	ld.shared.u32 	%r4615, [%rd4989+2048];
	xor.b32  	%r4616, %r4601, %r4615;
	shr.u32 	%r4617, %r4580, 24;
	mul.wide.u32 	%rd4990, %r4609, 4;
	add.s64 	%rd4991, %rd4990, %rd4726;
	ld.shared.u32 	%r4618, [%rd4991+3072];
	mul.wide.u32 	%rd4992, %r4613, 4;
	add.s64 	%rd4993, %rd4992, %rd4726;
	ld.shared.u32 	%r4619, [%rd4993+3072];
	mul.wide.u32 	%rd4994, %r4617, 4;
	add.s64 	%rd4995, %rd4994, %rd4726;
	ld.shared.u32 	%r4620, [%rd4995+3072];
	mul.wide.u32 	%rd4996, %r4605, 4;
	add.s64 	%rd4997, %rd4996, %rd4726;
	ld.shared.u32 	%r4621, [%rd4997+3072];
	xor.b32  	%r4622, %r4616, %r4621;
	xor.b32  	%r4623, %r4612, %r4620;
	xor.b32  	%r4624, %r4608, %r4619;
	xor.b32  	%r4625, %r4604, %r4618;
	ld.global.v4.u32 	{%r4626, %r4627, %r4628, %r4629}, [%rd4724+128];
	xor.b32  	%r4634, %r4512, %r4514;
	and.b32  	%r4635, %r4634, 255;
	mul.wide.u32 	%rd4998, %r4635, 4;
	add.s64 	%rd4999, %rd4726, %rd4998;
	xor.b32  	%r4636, %r4400, %r4402;
	and.b32  	%r4637, %r4636, 255;
	mul.wide.u32 	%rd5000, %r4637, 4;
	add.s64 	%rd5001, %rd4726, %rd5000;
	xor.b32  	%r4638, %r4568, %r4570;
	and.b32  	%r4639, %r4638, 255;
	mul.wide.u32 	%rd5002, %r4639, 4;
	add.s64 	%rd5003, %rd4726, %rd5002;
	xor.b32  	%r4640, %r4456, %r4458;
	and.b32  	%r4641, %r4640, 255;
	mul.wide.u32 	%rd5004, %r4641, 4;
	add.s64 	%rd5005, %rd4726, %rd5004;
	bfe.u32 	%r4642, %r4636, 8, 8;
	mul.wide.u32 	%rd5006, %r4642, 4;
	add.s64 	%rd5007, %rd5006, %rd4726;
	ld.shared.u32 	%r4643, [%rd5007+1024];
	ld.shared.u32 	%r4644, [%rd4999];
	xor.b32  	%r4645, %r4643, %r4644;
	bfe.u32 	%r4646, %r4638, 8, 8;
	mul.wide.u32 	%rd5008, %r4646, 4;
	add.s64 	%rd5009, %rd5008, %rd4726;
	ld.shared.u32 	%r4647, [%rd5009+1024];
	ld.shared.u32 	%r4648, [%rd5001];
	xor.b32  	%r4649, %r4647, %r4648;
	bfe.u32 	%r4650, %r4640, 8, 8;
	mul.wide.u32 	%rd5010, %r4650, 4;
	add.s64 	%rd5011, %rd5010, %rd4726;
	ld.shared.u32 	%r4651, [%rd5011+1024];
	ld.shared.u32 	%r4652, [%rd5003];
	xor.b32  	%r4653, %r4651, %r4652;
	bfe.u32 	%r4654, %r4634, 8, 8;
	mul.wide.u32 	%rd5012, %r4654, 4;
	add.s64 	%rd5013, %rd5012, %rd4726;
	ld.shared.u32 	%r4655, [%rd5013+1024];
	ld.shared.u32 	%r4656, [%rd5005];
	xor.b32  	%r4657, %r4655, %r4656;
	bfe.u32 	%r4658, %r4638, 16, 8;
	mul.wide.u32 	%rd5014, %r4658, 4;
	add.s64 	%rd5015, %rd5014, %rd4726;
	ld.shared.u32 	%r4659, [%rd5015+2048];
	xor.b32  	%r4660, %r4645, %r4659;
	shr.u32 	%r4661, %r4638, 24;
	bfe.u32 	%r4662, %r4640, 16, 8;
	mul.wide.u32 	%rd5016, %r4662, 4;
	add.s64 	%rd5017, %rd5016, %rd4726;
	ld.shared.u32 	%r4663, [%rd5017+2048];
	xor.b32  	%r4664, %r4649, %r4663;
	shr.u32 	%r4665, %r4640, 24;
	bfe.u32 	%r4666, %r4634, 16, 8;
	mul.wide.u32 	%rd5018, %r4666, 4;
	add.s64 	%rd5019, %rd5018, %rd4726;
	ld.shared.u32 	%r4667, [%rd5019+2048];
	xor.b32  	%r4668, %r4653, %r4667;
	shr.u32 	%r4669, %r4634, 24;
	bfe.u32 	%r4670, %r4636, 16, 8;
	mul.wide.u32 	%rd5020, %r4670, 4;
	add.s64 	%rd5021, %rd5020, %rd4726;
	ld.shared.u32 	%r4671, [%rd5021+2048];
	xor.b32  	%r4672, %r4657, %r4671;
	shr.u32 	%r4673, %r4636, 24;
	mul.wide.u32 	%rd5022, %r4665, 4;
	add.s64 	%rd5023, %rd5022, %rd4726;
	ld.shared.u32 	%r4674, [%rd5023+3072];
	mul.wide.u32 	%rd5024, %r4669, 4;
	add.s64 	%rd5025, %rd5024, %rd4726;
	ld.shared.u32 	%r4675, [%rd5025+3072];
	mul.wide.u32 	%rd5026, %r4673, 4;
	add.s64 	%rd5027, %rd5026, %rd4726;
	ld.shared.u32 	%r4676, [%rd5027+3072];
	mul.wide.u32 	%rd5028, %r4661, 4;
	add.s64 	%rd5029, %rd5028, %rd4726;
	ld.shared.u32 	%r4677, [%rd5029+3072];
	xor.b32  	%r4678, %r4672, %r4677;
	xor.b32  	%r4679, %r4668, %r4676;
	xor.b32  	%r4680, %r4664, %r4675;
	xor.b32  	%r4681, %r4660, %r4674;
	ld.global.v4.u32 	{%r4682, %r4683, %r4684, %r4685}, [%rd4724+144];
	xor.b32  	%r4690, %r4511, %r4515;
	and.b32  	%r4691, %r4690, 255;
	mul.wide.u32 	%rd5030, %r4691, 4;
	add.s64 	%rd5031, %rd4726, %rd5030;
	xor.b32  	%r4692, %r4399, %r4403;
	and.b32  	%r4693, %r4692, 255;
	mul.wide.u32 	%rd5032, %r4693, 4;
	add.s64 	%rd5033, %rd4726, %rd5032;
	xor.b32  	%r4694, %r4567, %r4571;
	and.b32  	%r4695, %r4694, 255;
	mul.wide.u32 	%rd5034, %r4695, 4;
	add.s64 	%rd5035, %rd4726, %rd5034;
	xor.b32  	%r4696, %r4455, %r4459;
	and.b32  	%r4697, %r4696, 255;
	mul.wide.u32 	%rd5036, %r4697, 4;
	add.s64 	%rd5037, %rd4726, %rd5036;
	bfe.u32 	%r4698, %r4692, 8, 8;
	mul.wide.u32 	%rd5038, %r4698, 4;
	add.s64 	%rd5039, %rd5038, %rd4726;
	ld.shared.u32 	%r4699, [%rd5039+1024];
	ld.shared.u32 	%r4700, [%rd5031];
	xor.b32  	%r4701, %r4699, %r4700;
	bfe.u32 	%r4702, %r4694, 8, 8;
	mul.wide.u32 	%rd5040, %r4702, 4;
	add.s64 	%rd5041, %rd5040, %rd4726;
	ld.shared.u32 	%r4703, [%rd5041+1024];
	ld.shared.u32 	%r4704, [%rd5033];
	xor.b32  	%r4705, %r4703, %r4704;
	bfe.u32 	%r4706, %r4696, 8, 8;
	mul.wide.u32 	%rd5042, %r4706, 4;
	add.s64 	%rd5043, %rd5042, %rd4726;
	ld.shared.u32 	%r4707, [%rd5043+1024];
	ld.shared.u32 	%r4708, [%rd5035];
	xor.b32  	%r4709, %r4707, %r4708;
	bfe.u32 	%r4710, %r4690, 8, 8;
	mul.wide.u32 	%rd5044, %r4710, 4;
	add.s64 	%rd5045, %rd5044, %rd4726;
	ld.shared.u32 	%r4711, [%rd5045+1024];
	ld.shared.u32 	%r4712, [%rd5037];
	xor.b32  	%r4713, %r4711, %r4712;
	bfe.u32 	%r4714, %r4694, 16, 8;
	mul.wide.u32 	%rd5046, %r4714, 4;
	add.s64 	%rd5047, %rd5046, %rd4726;
	ld.shared.u32 	%r4715, [%rd5047+2048];
	xor.b32  	%r4716, %r4701, %r4715;
	shr.u32 	%r4717, %r4694, 24;
	bfe.u32 	%r4718, %r4696, 16, 8;
	mul.wide.u32 	%rd5048, %r4718, 4;
	add.s64 	%rd5049, %rd5048, %rd4726;
	ld.shared.u32 	%r4719, [%rd5049+2048];
	xor.b32  	%r4720, %r4705, %r4719;
	shr.u32 	%r4721, %r4696, 24;
	bfe.u32 	%r4722, %r4690, 16, 8;
	mul.wide.u32 	%rd5050, %r4722, 4;
	add.s64 	%rd5051, %rd5050, %rd4726;
	ld.shared.u32 	%r4723, [%rd5051+2048];
	xor.b32  	%r4724, %r4709, %r4723;
	shr.u32 	%r4725, %r4690, 24;
	bfe.u32 	%r4726, %r4692, 16, 8;
	mul.wide.u32 	%rd5052, %r4726, 4;
	add.s64 	%rd5053, %rd5052, %rd4726;
	ld.shared.u32 	%r4727, [%rd5053+2048];
	xor.b32  	%r4728, %r4713, %r4727;
	shr.u32 	%r4729, %r4692, 24;
	mul.wide.u32 	%rd5054, %r4721, 4;
	add.s64 	%rd5055, %rd5054, %rd4726;
	ld.shared.u32 	%r4730, [%rd5055+3072];
	mul.wide.u32 	%rd5056, %r4725, 4;
	add.s64 	%rd5057, %rd5056, %rd4726;
	ld.shared.u32 	%r4731, [%rd5057+3072];
	mul.wide.u32 	%rd5058, %r4729, 4;
	add.s64 	%rd5059, %rd5058, %rd4726;
	ld.shared.u32 	%r4732, [%rd5059+3072];
	mul.wide.u32 	%rd5060, %r4717, 4;
	add.s64 	%rd5061, %rd5060, %rd4726;
	ld.shared.u32 	%r4733, [%rd5061+3072];
	xor.b32  	%r4734, %r4728, %r4733;
	xor.b32  	%r4735, %r4724, %r4732;
	xor.b32  	%r4736, %r4720, %r4731;
	xor.b32  	%r4737, %r4716, %r4730;
	ld.global.v4.u32 	{%r4738, %r4739, %r4740, %r4741}, [%rd4724+160];
	xor.b32  	%r4746, %r4398, %r4404;
	and.b32  	%r4747, %r4746, 255;
	mul.wide.u32 	%rd5062, %r4747, 4;
	add.s64 	%rd5063, %rd4726, %rd5062;
	xor.b32  	%r4748, %r4510, %r4516;
	and.b32  	%r4749, %r4748, 255;
	mul.wide.u32 	%rd5064, %r4749, 4;
	add.s64 	%rd5065, %rd4726, %rd5064;
	xor.b32  	%r4750, %r4454, %r4460;
	and.b32  	%r4751, %r4750, 255;
	mul.wide.u32 	%rd5066, %r4751, 4;
	add.s64 	%rd5067, %rd4726, %rd5066;
	xor.b32  	%r4752, %r4566, %r4572;
	and.b32  	%r4753, %r4752, 255;
	mul.wide.u32 	%rd5068, %r4753, 4;
	add.s64 	%rd5069, %rd4726, %rd5068;
	bfe.u32 	%r4754, %r4748, 8, 8;
	mul.wide.u32 	%rd5070, %r4754, 4;
	add.s64 	%rd5071, %rd5070, %rd4726;
	ld.shared.u32 	%r4755, [%rd5071+1024];
	ld.shared.u32 	%r4756, [%rd5063];
	xor.b32  	%r4757, %r4755, %r4756;
	bfe.u32 	%r4758, %r4750, 8, 8;
	mul.wide.u32 	%rd5072, %r4758, 4;
	add.s64 	%rd5073, %rd5072, %rd4726;
	ld.shared.u32 	%r4759, [%rd5073+1024];
	ld.shared.u32 	%r4760, [%rd5065];
	xor.b32  	%r4761, %r4759, %r4760;
	bfe.u32 	%r4762, %r4752, 8, 8;
	mul.wide.u32 	%rd5074, %r4762, 4;
	add.s64 	%rd5075, %rd5074, %rd4726;
	ld.shared.u32 	%r4763, [%rd5075+1024];
	ld.shared.u32 	%r4764, [%rd5067];
	xor.b32  	%r4765, %r4763, %r4764;
	bfe.u32 	%r4766, %r4746, 8, 8;
	mul.wide.u32 	%rd5076, %r4766, 4;
	add.s64 	%rd5077, %rd5076, %rd4726;
	ld.shared.u32 	%r4767, [%rd5077+1024];
	ld.shared.u32 	%r4768, [%rd5069];
	xor.b32  	%r4769, %r4767, %r4768;
	bfe.u32 	%r4770, %r4750, 16, 8;
	mul.wide.u32 	%rd5078, %r4770, 4;
	add.s64 	%rd5079, %rd5078, %rd4726;
	ld.shared.u32 	%r4771, [%rd5079+2048];
	xor.b32  	%r4772, %r4757, %r4771;
	shr.u32 	%r4773, %r4750, 24;
	bfe.u32 	%r4774, %r4752, 16, 8;
	mul.wide.u32 	%rd5080, %r4774, 4;
	add.s64 	%rd5081, %rd5080, %rd4726;
	ld.shared.u32 	%r4775, [%rd5081+2048];
	xor.b32  	%r4776, %r4761, %r4775;
	shr.u32 	%r4777, %r4752, 24;
	bfe.u32 	%r4778, %r4746, 16, 8;
	mul.wide.u32 	%rd5082, %r4778, 4;
	add.s64 	%rd5083, %rd5082, %rd4726;
	ld.shared.u32 	%r4779, [%rd5083+2048];
	xor.b32  	%r4780, %r4765, %r4779;
	shr.u32 	%r4781, %r4746, 24;
	bfe.u32 	%r4782, %r4748, 16, 8;
	mul.wide.u32 	%rd5084, %r4782, 4;
	add.s64 	%rd5085, %rd5084, %rd4726;
	ld.shared.u32 	%r4783, [%rd5085+2048];
	xor.b32  	%r4784, %r4769, %r4783;
	shr.u32 	%r4785, %r4748, 24;
	mul.wide.u32 	%rd5086, %r4777, 4;
	add.s64 	%rd5087, %rd5086, %rd4726;
	ld.shared.u32 	%r4786, [%rd5087+3072];
	mul.wide.u32 	%rd5088, %r4781, 4;
	add.s64 	%rd5089, %rd5088, %rd4726;
	ld.shared.u32 	%r4787, [%rd5089+3072];
	mul.wide.u32 	%rd5090, %r4785, 4;
	add.s64 	%rd5091, %rd5090, %rd4726;
	ld.shared.u32 	%r4788, [%rd5091+3072];
	mul.wide.u32 	%rd5092, %r4773, 4;
	add.s64 	%rd5093, %rd5092, %rd4726;
	ld.shared.u32 	%r4789, [%rd5093+3072];
	xor.b32  	%r4790, %r4784, %r4789;
	xor.b32  	%r4791, %r4780, %r4788;
	xor.b32  	%r4792, %r4776, %r4787;
	xor.b32  	%r4793, %r4772, %r4786;
	ld.global.v4.u32 	{%r4794, %r4795, %r4796, %r4797}, [%rd4724+176];
	xor.b32  	%r4802, %r4625, %r4626;
	and.b32  	%r4803, %r4802, 255;
	mul.wide.u32 	%rd5094, %r4803, 4;
	add.s64 	%rd5095, %rd4726, %rd5094;
	xor.b32  	%r4804, %r4624, %r4627;
	and.b32  	%r4805, %r4804, 255;
	mul.wide.u32 	%rd5096, %r4805, 4;
	add.s64 	%rd5097, %rd4726, %rd5096;
	xor.b32  	%r4806, %r4623, %r4628;
	and.b32  	%r4807, %r4806, 255;
	mul.wide.u32 	%rd5098, %r4807, 4;
	add.s64 	%rd5099, %rd4726, %rd5098;
	xor.b32  	%r4808, %r4622, %r4629;
	and.b32  	%r4809, %r4808, 255;
	mul.wide.u32 	%rd5100, %r4809, 4;
	add.s64 	%rd5101, %rd4726, %rd5100;
	bfe.u32 	%r4810, %r4804, 8, 8;
	mul.wide.u32 	%rd5102, %r4810, 4;
	add.s64 	%rd5103, %rd5102, %rd4726;
	ld.shared.u32 	%r4811, [%rd5103+1024];
	ld.shared.u32 	%r4812, [%rd5095];
	xor.b32  	%r4813, %r4811, %r4812;
	bfe.u32 	%r4814, %r4806, 8, 8;
	mul.wide.u32 	%rd5104, %r4814, 4;
	add.s64 	%rd5105, %rd5104, %rd4726;
	ld.shared.u32 	%r4815, [%rd5105+1024];
	ld.shared.u32 	%r4816, [%rd5097];
	xor.b32  	%r4817, %r4815, %r4816;
	bfe.u32 	%r4818, %r4808, 8, 8;
	mul.wide.u32 	%rd5106, %r4818, 4;
	add.s64 	%rd5107, %rd5106, %rd4726;
	ld.shared.u32 	%r4819, [%rd5107+1024];
	ld.shared.u32 	%r4820, [%rd5099];
	xor.b32  	%r4821, %r4819, %r4820;
	bfe.u32 	%r4822, %r4802, 8, 8;
	mul.wide.u32 	%rd5108, %r4822, 4;
	add.s64 	%rd5109, %rd5108, %rd4726;
	ld.shared.u32 	%r4823, [%rd5109+1024];
	ld.shared.u32 	%r4824, [%rd5101];
	xor.b32  	%r4825, %r4823, %r4824;
	bfe.u32 	%r4826, %r4806, 16, 8;
	mul.wide.u32 	%rd5110, %r4826, 4;
	add.s64 	%rd5111, %rd5110, %rd4726;
	ld.shared.u32 	%r4827, [%rd5111+2048];
	xor.b32  	%r4828, %r4813, %r4827;
	shr.u32 	%r4829, %r4806, 24;
	bfe.u32 	%r4830, %r4808, 16, 8;
	mul.wide.u32 	%rd5112, %r4830, 4;
	add.s64 	%rd5113, %rd5112, %rd4726;
	ld.shared.u32 	%r4831, [%rd5113+2048];
	xor.b32  	%r4832, %r4817, %r4831;
	shr.u32 	%r4833, %r4808, 24;
	bfe.u32 	%r4834, %r4802, 16, 8;
	mul.wide.u32 	%rd5114, %r4834, 4;
	add.s64 	%rd5115, %rd5114, %rd4726;
	ld.shared.u32 	%r4835, [%rd5115+2048];
	xor.b32  	%r4836, %r4821, %r4835;
	shr.u32 	%r4837, %r4802, 24;
	bfe.u32 	%r4838, %r4804, 16, 8;
	mul.wide.u32 	%rd5116, %r4838, 4;
	add.s64 	%rd5117, %rd5116, %rd4726;
	ld.shared.u32 	%r4839, [%rd5117+2048];
	xor.b32  	%r4840, %r4825, %r4839;
	shr.u32 	%r4841, %r4804, 24;
	mul.wide.u32 	%rd5118, %r4833, 4;
	add.s64 	%rd5119, %rd5118, %rd4726;
	ld.shared.u32 	%r4842, [%rd5119+3072];
	mul.wide.u32 	%rd5120, %r4837, 4;
	add.s64 	%rd5121, %rd5120, %rd4726;
	ld.shared.u32 	%r4843, [%rd5121+3072];
	mul.wide.u32 	%rd5122, %r4841, 4;
	add.s64 	%rd5123, %rd5122, %rd4726;
	ld.shared.u32 	%r4844, [%rd5123+3072];
	mul.wide.u32 	%rd5124, %r4829, 4;
	add.s64 	%rd5125, %rd5124, %rd4726;
	ld.shared.u32 	%r4845, [%rd5125+3072];
	xor.b32  	%r4846, %r4836, %r4844;
	xor.b32  	%r4847, %r4832, %r4843;
	xor.b32  	%r4848, %r4828, %r4842;
	xor.b32  	%r4849, %r4840, %r4845;
	ld.global.v4.u32 	{%r4850, %r4851, %r4852, %r4853}, [%rd4724+192];
	xor.b32  	%r4858, %r4681, %r4682;
	and.b32  	%r4859, %r4858, 255;
	mul.wide.u32 	%rd5126, %r4859, 4;
	add.s64 	%rd5127, %rd4726, %rd5126;
	xor.b32  	%r4860, %r4680, %r4683;
	and.b32  	%r4861, %r4860, 255;
	mul.wide.u32 	%rd5128, %r4861, 4;
	add.s64 	%rd5129, %rd4726, %rd5128;
	xor.b32  	%r4862, %r4679, %r4684;
	and.b32  	%r4863, %r4862, 255;
	mul.wide.u32 	%rd5130, %r4863, 4;
	add.s64 	%rd5131, %rd4726, %rd5130;
	xor.b32  	%r4864, %r4678, %r4685;
	and.b32  	%r4865, %r4864, 255;
	mul.wide.u32 	%rd5132, %r4865, 4;
	add.s64 	%rd5133, %rd4726, %rd5132;
	bfe.u32 	%r4866, %r4860, 8, 8;
	mul.wide.u32 	%rd5134, %r4866, 4;
	add.s64 	%rd5135, %rd5134, %rd4726;
	ld.shared.u32 	%r4867, [%rd5135+1024];
	ld.shared.u32 	%r4868, [%rd5127];
	xor.b32  	%r4869, %r4867, %r4868;
	bfe.u32 	%r4870, %r4862, 8, 8;
	mul.wide.u32 	%rd5136, %r4870, 4;
	add.s64 	%rd5137, %rd5136, %rd4726;
	ld.shared.u32 	%r4871, [%rd5137+1024];
	ld.shared.u32 	%r4872, [%rd5129];
	xor.b32  	%r4873, %r4871, %r4872;
	bfe.u32 	%r4874, %r4864, 8, 8;
	mul.wide.u32 	%rd5138, %r4874, 4;
	add.s64 	%rd5139, %rd5138, %rd4726;
	ld.shared.u32 	%r4875, [%rd5139+1024];
	ld.shared.u32 	%r4876, [%rd5131];
	xor.b32  	%r4877, %r4875, %r4876;
	bfe.u32 	%r4878, %r4858, 8, 8;
	mul.wide.u32 	%rd5140, %r4878, 4;
	add.s64 	%rd5141, %rd5140, %rd4726;
	ld.shared.u32 	%r4879, [%rd5141+1024];
	ld.shared.u32 	%r4880, [%rd5133];
	xor.b32  	%r4881, %r4879, %r4880;
	bfe.u32 	%r4882, %r4862, 16, 8;
	mul.wide.u32 	%rd5142, %r4882, 4;
	add.s64 	%rd5143, %rd5142, %rd4726;
	ld.shared.u32 	%r4883, [%rd5143+2048];
	xor.b32  	%r4884, %r4869, %r4883;
	shr.u32 	%r4885, %r4862, 24;
	bfe.u32 	%r4886, %r4864, 16, 8;
	mul.wide.u32 	%rd5144, %r4886, 4;
	add.s64 	%rd5145, %rd5144, %rd4726;
	ld.shared.u32 	%r4887, [%rd5145+2048];
	xor.b32  	%r4888, %r4873, %r4887;
	shr.u32 	%r4889, %r4864, 24;
	bfe.u32 	%r4890, %r4858, 16, 8;
	mul.wide.u32 	%rd5146, %r4890, 4;
	add.s64 	%rd5147, %rd5146, %rd4726;
	ld.shared.u32 	%r4891, [%rd5147+2048];
	xor.b32  	%r4892, %r4877, %r4891;
	shr.u32 	%r4893, %r4858, 24;
	bfe.u32 	%r4894, %r4860, 16, 8;
	mul.wide.u32 	%rd5148, %r4894, 4;
	add.s64 	%rd5149, %rd5148, %rd4726;
	ld.shared.u32 	%r4895, [%rd5149+2048];
	xor.b32  	%r4896, %r4881, %r4895;
	shr.u32 	%r4897, %r4860, 24;
	mul.wide.u32 	%rd5150, %r4889, 4;
	add.s64 	%rd5151, %rd5150, %rd4726;
	ld.shared.u32 	%r4898, [%rd5151+3072];
	mul.wide.u32 	%rd5152, %r4893, 4;
	add.s64 	%rd5153, %rd5152, %rd4726;
	ld.shared.u32 	%r4899, [%rd5153+3072];
	mul.wide.u32 	%rd5154, %r4897, 4;
	add.s64 	%rd5155, %rd5154, %rd4726;
	ld.shared.u32 	%r4900, [%rd5155+3072];
	mul.wide.u32 	%rd5156, %r4885, 4;
	add.s64 	%rd5157, %rd5156, %rd4726;
	ld.shared.u32 	%r4901, [%rd5157+3072];
	xor.b32  	%r4902, %r4892, %r4900;
	xor.b32  	%r4903, %r4888, %r4899;
	xor.b32  	%r4904, %r4884, %r4898;
	xor.b32  	%r4905, %r4896, %r4901;
	ld.global.v4.u32 	{%r4906, %r4907, %r4908, %r4909}, [%rd4724+208];
	xor.b32  	%r4914, %r4737, %r4738;
	and.b32  	%r4915, %r4914, 255;
	mul.wide.u32 	%rd5158, %r4915, 4;
	add.s64 	%rd5159, %rd4726, %rd5158;
	xor.b32  	%r4916, %r4736, %r4739;
	and.b32  	%r4917, %r4916, 255;
	mul.wide.u32 	%rd5160, %r4917, 4;
	add.s64 	%rd5161, %rd4726, %rd5160;
	xor.b32  	%r4918, %r4735, %r4740;
	and.b32  	%r4919, %r4918, 255;
	mul.wide.u32 	%rd5162, %r4919, 4;
	add.s64 	%rd5163, %rd4726, %rd5162;
	xor.b32  	%r4920, %r4734, %r4741;
	and.b32  	%r4921, %r4920, 255;
	mul.wide.u32 	%rd5164, %r4921, 4;
	add.s64 	%rd5165, %rd4726, %rd5164;
	bfe.u32 	%r4922, %r4916, 8, 8;
	mul.wide.u32 	%rd5166, %r4922, 4;
	add.s64 	%rd5167, %rd5166, %rd4726;
	ld.shared.u32 	%r4923, [%rd5167+1024];
	ld.shared.u32 	%r4924, [%rd5159];
	xor.b32  	%r4925, %r4923, %r4924;
	bfe.u32 	%r4926, %r4918, 8, 8;
	mul.wide.u32 	%rd5168, %r4926, 4;
	add.s64 	%rd5169, %rd5168, %rd4726;
	ld.shared.u32 	%r4927, [%rd5169+1024];
	ld.shared.u32 	%r4928, [%rd5161];
	xor.b32  	%r4929, %r4927, %r4928;
	bfe.u32 	%r4930, %r4920, 8, 8;
	mul.wide.u32 	%rd5170, %r4930, 4;
	add.s64 	%rd5171, %rd5170, %rd4726;
	ld.shared.u32 	%r4931, [%rd5171+1024];
	ld.shared.u32 	%r4932, [%rd5163];
	xor.b32  	%r4933, %r4931, %r4932;
	bfe.u32 	%r4934, %r4914, 8, 8;
	mul.wide.u32 	%rd5172, %r4934, 4;
	add.s64 	%rd5173, %rd5172, %rd4726;
	ld.shared.u32 	%r4935, [%rd5173+1024];
	ld.shared.u32 	%r4936, [%rd5165];
	xor.b32  	%r4937, %r4935, %r4936;
	bfe.u32 	%r4938, %r4918, 16, 8;
	mul.wide.u32 	%rd5174, %r4938, 4;
	add.s64 	%rd5175, %rd5174, %rd4726;
	ld.shared.u32 	%r4939, [%rd5175+2048];
	xor.b32  	%r4940, %r4925, %r4939;
	shr.u32 	%r4941, %r4918, 24;
	bfe.u32 	%r4942, %r4920, 16, 8;
	mul.wide.u32 	%rd5176, %r4942, 4;
	add.s64 	%rd5177, %rd5176, %rd4726;
	ld.shared.u32 	%r4943, [%rd5177+2048];
	xor.b32  	%r4944, %r4929, %r4943;
	shr.u32 	%r4945, %r4920, 24;
	bfe.u32 	%r4946, %r4914, 16, 8;
	mul.wide.u32 	%rd5178, %r4946, 4;
	add.s64 	%rd5179, %rd5178, %rd4726;
	ld.shared.u32 	%r4947, [%rd5179+2048];
	xor.b32  	%r4948, %r4933, %r4947;
	shr.u32 	%r4949, %r4914, 24;
	bfe.u32 	%r4950, %r4916, 16, 8;
	mul.wide.u32 	%rd5180, %r4950, 4;
	add.s64 	%rd5181, %rd5180, %rd4726;
	ld.shared.u32 	%r4951, [%rd5181+2048];
	xor.b32  	%r4952, %r4937, %r4951;
	shr.u32 	%r4953, %r4916, 24;
	mul.wide.u32 	%rd5182, %r4945, 4;
	add.s64 	%rd5183, %rd5182, %rd4726;
	ld.shared.u32 	%r4954, [%rd5183+3072];
	mul.wide.u32 	%rd5184, %r4949, 4;
	add.s64 	%rd5185, %rd5184, %rd4726;
	ld.shared.u32 	%r4955, [%rd5185+3072];
	mul.wide.u32 	%rd5186, %r4953, 4;
	add.s64 	%rd5187, %rd5186, %rd4726;
	ld.shared.u32 	%r4956, [%rd5187+3072];
	mul.wide.u32 	%rd5188, %r4941, 4;
	add.s64 	%rd5189, %rd5188, %rd4726;
	ld.shared.u32 	%r4957, [%rd5189+3072];
	xor.b32  	%r4958, %r4948, %r4956;
	xor.b32  	%r4959, %r4944, %r4955;
	xor.b32  	%r4960, %r4940, %r4954;
	xor.b32  	%r4961, %r4952, %r4957;
	ld.global.v4.u32 	{%r4962, %r4963, %r4964, %r4965}, [%rd4724+224];
	xor.b32  	%r4970, %r4793, %r4794;
	and.b32  	%r4971, %r4970, 255;
	mul.wide.u32 	%rd5190, %r4971, 4;
	add.s64 	%rd5191, %rd4726, %rd5190;
	xor.b32  	%r4972, %r4792, %r4795;
	and.b32  	%r4973, %r4972, 255;
	mul.wide.u32 	%rd5192, %r4973, 4;
	add.s64 	%rd5193, %rd4726, %rd5192;
	xor.b32  	%r4974, %r4791, %r4796;
	and.b32  	%r4975, %r4974, 255;
	mul.wide.u32 	%rd5194, %r4975, 4;
	add.s64 	%rd5195, %rd4726, %rd5194;
	xor.b32  	%r4976, %r4790, %r4797;
	and.b32  	%r4977, %r4976, 255;
	mul.wide.u32 	%rd5196, %r4977, 4;
	add.s64 	%rd5197, %rd4726, %rd5196;
	bfe.u32 	%r4978, %r4972, 8, 8;
	mul.wide.u32 	%rd5198, %r4978, 4;
	add.s64 	%rd5199, %rd5198, %rd4726;
	ld.shared.u32 	%r4979, [%rd5199+1024];
	ld.shared.u32 	%r4980, [%rd5191];
	xor.b32  	%r4981, %r4979, %r4980;
	bfe.u32 	%r4982, %r4974, 8, 8;
	mul.wide.u32 	%rd5200, %r4982, 4;
	add.s64 	%rd5201, %rd5200, %rd4726;
	ld.shared.u32 	%r4983, [%rd5201+1024];
	ld.shared.u32 	%r4984, [%rd5193];
	xor.b32  	%r4985, %r4983, %r4984;
	bfe.u32 	%r4986, %r4976, 8, 8;
	mul.wide.u32 	%rd5202, %r4986, 4;
	add.s64 	%rd5203, %rd5202, %rd4726;
	ld.shared.u32 	%r4987, [%rd5203+1024];
	ld.shared.u32 	%r4988, [%rd5195];
	xor.b32  	%r4989, %r4987, %r4988;
	bfe.u32 	%r4990, %r4970, 8, 8;
	mul.wide.u32 	%rd5204, %r4990, 4;
	add.s64 	%rd5205, %rd5204, %rd4726;
	ld.shared.u32 	%r4991, [%rd5205+1024];
	ld.shared.u32 	%r4992, [%rd5197];
	xor.b32  	%r4993, %r4991, %r4992;
	bfe.u32 	%r4994, %r4974, 16, 8;
	mul.wide.u32 	%rd5206, %r4994, 4;
	add.s64 	%rd5207, %rd5206, %rd4726;
	ld.shared.u32 	%r4995, [%rd5207+2048];
	xor.b32  	%r4996, %r4981, %r4995;
	shr.u32 	%r4997, %r4974, 24;
	bfe.u32 	%r4998, %r4976, 16, 8;
	mul.wide.u32 	%rd5208, %r4998, 4;
	add.s64 	%rd5209, %rd5208, %rd4726;
	ld.shared.u32 	%r4999, [%rd5209+2048];
	xor.b32  	%r5000, %r4985, %r4999;
	shr.u32 	%r5001, %r4976, 24;
	bfe.u32 	%r5002, %r4970, 16, 8;
	mul.wide.u32 	%rd5210, %r5002, 4;
	add.s64 	%rd5211, %rd5210, %rd4726;
	ld.shared.u32 	%r5003, [%rd5211+2048];
	xor.b32  	%r5004, %r4989, %r5003;
	shr.u32 	%r5005, %r4970, 24;
	bfe.u32 	%r5006, %r4972, 16, 8;
	mul.wide.u32 	%rd5212, %r5006, 4;
	add.s64 	%rd5213, %rd5212, %rd4726;
	ld.shared.u32 	%r5007, [%rd5213+2048];
	xor.b32  	%r5008, %r4993, %r5007;
	shr.u32 	%r5009, %r4972, 24;
	mul.wide.u32 	%rd5214, %r5001, 4;
	add.s64 	%rd5215, %rd5214, %rd4726;
	ld.shared.u32 	%r5010, [%rd5215+3072];
	mul.wide.u32 	%rd5216, %r5005, 4;
	add.s64 	%rd5217, %rd5216, %rd4726;
	ld.shared.u32 	%r5011, [%rd5217+3072];
	mul.wide.u32 	%rd5218, %r5009, 4;
	add.s64 	%rd5219, %rd5218, %rd4726;
	ld.shared.u32 	%r5012, [%rd5219+3072];
	mul.wide.u32 	%rd5220, %r4997, 4;
	add.s64 	%rd5221, %rd5220, %rd4726;
	ld.shared.u32 	%r5013, [%rd5221+3072];
	xor.b32  	%r5014, %r5004, %r5012;
	xor.b32  	%r5015, %r5000, %r5011;
	xor.b32  	%r5016, %r4996, %r5010;
	xor.b32  	%r5017, %r5008, %r5013;
	ld.global.v4.u32 	{%r5018, %r5019, %r5020, %r5021}, [%rd4724+240];
	xor.b32  	%r5026, %r4849, %r4853;
	and.b32  	%r5027, %r5026, 255;
	mul.wide.u32 	%rd5222, %r5027, 4;
	add.s64 	%rd5223, %rd4726, %rd5222;
	xor.b32  	%r5028, %r4961, %r4965;
	and.b32  	%r5029, %r5028, 255;
	mul.wide.u32 	%rd5224, %r5029, 4;
	add.s64 	%rd5225, %rd4726, %rd5224;
	xor.b32  	%r5030, %r4905, %r4909;
	and.b32  	%r5031, %r5030, 255;
	mul.wide.u32 	%rd5226, %r5031, 4;
	add.s64 	%rd5227, %rd4726, %rd5226;
	xor.b32  	%r5032, %r5017, %r5021;
	and.b32  	%r5033, %r5032, 255;
	mul.wide.u32 	%rd5228, %r5033, 4;
	add.s64 	%rd5229, %rd4726, %rd5228;
	bfe.u32 	%r5034, %r5028, 8, 8;
	mul.wide.u32 	%rd5230, %r5034, 4;
	add.s64 	%rd5231, %rd5230, %rd4726;
	ld.shared.u32 	%r5035, [%rd5231+1024];
	ld.shared.u32 	%r5036, [%rd5223];
	xor.b32  	%r5037, %r5035, %r5036;
	bfe.u32 	%r5038, %r5030, 8, 8;
	mul.wide.u32 	%rd5232, %r5038, 4;
	add.s64 	%rd5233, %rd5232, %rd4726;
	ld.shared.u32 	%r5039, [%rd5233+1024];
	ld.shared.u32 	%r5040, [%rd5225];
	xor.b32  	%r5041, %r5039, %r5040;
	bfe.u32 	%r5042, %r5032, 8, 8;
	mul.wide.u32 	%rd5234, %r5042, 4;
	add.s64 	%rd5235, %rd5234, %rd4726;
	ld.shared.u32 	%r5043, [%rd5235+1024];
	ld.shared.u32 	%r5044, [%rd5227];
	xor.b32  	%r5045, %r5043, %r5044;
	bfe.u32 	%r5046, %r5026, 8, 8;
	mul.wide.u32 	%rd5236, %r5046, 4;
	add.s64 	%rd5237, %rd5236, %rd4726;
	ld.shared.u32 	%r5047, [%rd5237+1024];
	ld.shared.u32 	%r5048, [%rd5229];
	xor.b32  	%r5049, %r5047, %r5048;
	bfe.u32 	%r5050, %r5030, 16, 8;
	mul.wide.u32 	%rd5238, %r5050, 4;
	add.s64 	%rd5239, %rd5238, %rd4726;
	ld.shared.u32 	%r5051, [%rd5239+2048];
	xor.b32  	%r5052, %r5037, %r5051;
	shr.u32 	%r5053, %r5030, 24;
	bfe.u32 	%r5054, %r5032, 16, 8;
	mul.wide.u32 	%rd5240, %r5054, 4;
	add.s64 	%rd5241, %rd5240, %rd4726;
	ld.shared.u32 	%r5055, [%rd5241+2048];
	xor.b32  	%r5056, %r5041, %r5055;
	shr.u32 	%r5057, %r5032, 24;
	bfe.u32 	%r5058, %r5026, 16, 8;
	mul.wide.u32 	%rd5242, %r5058, 4;
	add.s64 	%rd5243, %rd5242, %rd4726;
	ld.shared.u32 	%r5059, [%rd5243+2048];
	xor.b32  	%r5060, %r5045, %r5059;
	shr.u32 	%r5061, %r5026, 24;
	bfe.u32 	%r5062, %r5028, 16, 8;
	mul.wide.u32 	%rd5244, %r5062, 4;
	add.s64 	%rd5245, %rd5244, %rd4726;
	ld.shared.u32 	%r5063, [%rd5245+2048];
	xor.b32  	%r5064, %r5049, %r5063;
	shr.u32 	%r5065, %r5028, 24;
	mul.wide.u32 	%rd5246, %r5057, 4;
	add.s64 	%rd5247, %rd5246, %rd4726;
	ld.shared.u32 	%r5066, [%rd5247+3072];
	mul.wide.u32 	%rd5248, %r5061, 4;
	add.s64 	%rd5249, %rd5248, %rd4726;
	ld.shared.u32 	%r5067, [%rd5249+3072];
	mul.wide.u32 	%rd5250, %r5065, 4;
	add.s64 	%rd5251, %rd5250, %rd4726;
	ld.shared.u32 	%r5068, [%rd5251+3072];
	mul.wide.u32 	%rd5252, %r5053, 4;
	add.s64 	%rd5253, %rd5252, %rd4726;
	ld.shared.u32 	%r5069, [%rd5253+3072];
	xor.b32  	%r5070, %r5064, %r5069;
	xor.b32  	%r5071, %r5060, %r5068;
	xor.b32  	%r5072, %r5056, %r5067;
	xor.b32  	%r5073, %r5052, %r5066;
	ld.global.v4.u32 	{%r5074, %r5075, %r5076, %r5077}, [%rd4724+256];
	xor.b32  	%r5082, %r4960, %r4962;
	and.b32  	%r5083, %r5082, 255;
	mul.wide.u32 	%rd5254, %r5083, 4;
	add.s64 	%rd5255, %rd4726, %rd5254;
	xor.b32  	%r5084, %r4848, %r4850;
	and.b32  	%r5085, %r5084, 255;
	mul.wide.u32 	%rd5256, %r5085, 4;
	add.s64 	%rd5257, %rd4726, %rd5256;
	xor.b32  	%r5086, %r5016, %r5018;
	and.b32  	%r5087, %r5086, 255;
	mul.wide.u32 	%rd5258, %r5087, 4;
	add.s64 	%rd5259, %rd4726, %rd5258;
	xor.b32  	%r5088, %r4904, %r4906;
	and.b32  	%r5089, %r5088, 255;
	mul.wide.u32 	%rd5260, %r5089, 4;
	add.s64 	%rd5261, %rd4726, %rd5260;
	bfe.u32 	%r5090, %r5084, 8, 8;
	mul.wide.u32 	%rd5262, %r5090, 4;
	add.s64 	%rd5263, %rd5262, %rd4726;
	ld.shared.u32 	%r5091, [%rd5263+1024];
	ld.shared.u32 	%r5092, [%rd5255];
	xor.b32  	%r5093, %r5091, %r5092;
	bfe.u32 	%r5094, %r5086, 8, 8;
	mul.wide.u32 	%rd5264, %r5094, 4;
	add.s64 	%rd5265, %rd5264, %rd4726;
	ld.shared.u32 	%r5095, [%rd5265+1024];
	ld.shared.u32 	%r5096, [%rd5257];
	xor.b32  	%r5097, %r5095, %r5096;
	bfe.u32 	%r5098, %r5088, 8, 8;
	mul.wide.u32 	%rd5266, %r5098, 4;
	add.s64 	%rd5267, %rd5266, %rd4726;
	ld.shared.u32 	%r5099, [%rd5267+1024];
	ld.shared.u32 	%r5100, [%rd5259];
	xor.b32  	%r5101, %r5099, %r5100;
	bfe.u32 	%r5102, %r5082, 8, 8;
	mul.wide.u32 	%rd5268, %r5102, 4;
	add.s64 	%rd5269, %rd5268, %rd4726;
	ld.shared.u32 	%r5103, [%rd5269+1024];
	ld.shared.u32 	%r5104, [%rd5261];
	xor.b32  	%r5105, %r5103, %r5104;
	bfe.u32 	%r5106, %r5086, 16, 8;
	mul.wide.u32 	%rd5270, %r5106, 4;
	add.s64 	%rd5271, %rd5270, %rd4726;
	ld.shared.u32 	%r5107, [%rd5271+2048];
	xor.b32  	%r5108, %r5093, %r5107;
	shr.u32 	%r5109, %r5086, 24;
	bfe.u32 	%r5110, %r5088, 16, 8;
	mul.wide.u32 	%rd5272, %r5110, 4;
	add.s64 	%rd5273, %rd5272, %rd4726;
	ld.shared.u32 	%r5111, [%rd5273+2048];
	xor.b32  	%r5112, %r5097, %r5111;
	shr.u32 	%r5113, %r5088, 24;
	bfe.u32 	%r5114, %r5082, 16, 8;
	mul.wide.u32 	%rd5274, %r5114, 4;
	add.s64 	%rd5275, %rd5274, %rd4726;
	ld.shared.u32 	%r5115, [%rd5275+2048];
	xor.b32  	%r5116, %r5101, %r5115;
	shr.u32 	%r5117, %r5082, 24;
	bfe.u32 	%r5118, %r5084, 16, 8;
	mul.wide.u32 	%rd5276, %r5118, 4;
	add.s64 	%rd5277, %rd5276, %rd4726;
	ld.shared.u32 	%r5119, [%rd5277+2048];
	xor.b32  	%r5120, %r5105, %r5119;
	shr.u32 	%r5121, %r5084, 24;
	mul.wide.u32 	%rd5278, %r5113, 4;
	add.s64 	%rd5279, %rd5278, %rd4726;
	ld.shared.u32 	%r5122, [%rd5279+3072];
	mul.wide.u32 	%rd5280, %r5117, 4;
	add.s64 	%rd5281, %rd5280, %rd4726;
	ld.shared.u32 	%r5123, [%rd5281+3072];
	mul.wide.u32 	%rd5282, %r5121, 4;
	add.s64 	%rd5283, %rd5282, %rd4726;
	ld.shared.u32 	%r5124, [%rd5283+3072];
	mul.wide.u32 	%rd5284, %r5109, 4;
	add.s64 	%rd5285, %rd5284, %rd4726;
	ld.shared.u32 	%r5125, [%rd5285+3072];
	xor.b32  	%r5126, %r5120, %r5125;
	xor.b32  	%r5127, %r5116, %r5124;
	xor.b32  	%r5128, %r5112, %r5123;
	xor.b32  	%r5129, %r5108, %r5122;
	ld.global.v4.u32 	{%r5130, %r5131, %r5132, %r5133}, [%rd4724+272];
	xor.b32  	%r5138, %r4959, %r4963;
	and.b32  	%r5139, %r5138, 255;
	mul.wide.u32 	%rd5286, %r5139, 4;
	add.s64 	%rd5287, %rd4726, %rd5286;
	xor.b32  	%r5140, %r4847, %r4851;
	and.b32  	%r5141, %r5140, 255;
	mul.wide.u32 	%rd5288, %r5141, 4;
	add.s64 	%rd5289, %rd4726, %rd5288;
	xor.b32  	%r5142, %r5015, %r5019;
	and.b32  	%r5143, %r5142, 255;
	mul.wide.u32 	%rd5290, %r5143, 4;
	add.s64 	%rd5291, %rd4726, %rd5290;
	xor.b32  	%r5144, %r4903, %r4907;
	and.b32  	%r5145, %r5144, 255;
	mul.wide.u32 	%rd5292, %r5145, 4;
	add.s64 	%rd5293, %rd4726, %rd5292;
	bfe.u32 	%r5146, %r5140, 8, 8;
	mul.wide.u32 	%rd5294, %r5146, 4;
	add.s64 	%rd5295, %rd5294, %rd4726;
	ld.shared.u32 	%r5147, [%rd5295+1024];
	ld.shared.u32 	%r5148, [%rd5287];
	xor.b32  	%r5149, %r5147, %r5148;
	bfe.u32 	%r5150, %r5142, 8, 8;
	mul.wide.u32 	%rd5296, %r5150, 4;
	add.s64 	%rd5297, %rd5296, %rd4726;
	ld.shared.u32 	%r5151, [%rd5297+1024];
	ld.shared.u32 	%r5152, [%rd5289];
	xor.b32  	%r5153, %r5151, %r5152;
	bfe.u32 	%r5154, %r5144, 8, 8;
	mul.wide.u32 	%rd5298, %r5154, 4;
	add.s64 	%rd5299, %rd5298, %rd4726;
	ld.shared.u32 	%r5155, [%rd5299+1024];
	ld.shared.u32 	%r5156, [%rd5291];
	xor.b32  	%r5157, %r5155, %r5156;
	bfe.u32 	%r5158, %r5138, 8, 8;
	mul.wide.u32 	%rd5300, %r5158, 4;
	add.s64 	%rd5301, %rd5300, %rd4726;
	ld.shared.u32 	%r5159, [%rd5301+1024];
	ld.shared.u32 	%r5160, [%rd5293];
	xor.b32  	%r5161, %r5159, %r5160;
	bfe.u32 	%r5162, %r5142, 16, 8;
	mul.wide.u32 	%rd5302, %r5162, 4;
	add.s64 	%rd5303, %rd5302, %rd4726;
	ld.shared.u32 	%r5163, [%rd5303+2048];
	xor.b32  	%r5164, %r5149, %r5163;
	shr.u32 	%r5165, %r5142, 24;
	bfe.u32 	%r5166, %r5144, 16, 8;
	mul.wide.u32 	%rd5304, %r5166, 4;
	add.s64 	%rd5305, %rd5304, %rd4726;
	ld.shared.u32 	%r5167, [%rd5305+2048];
	xor.b32  	%r5168, %r5153, %r5167;
	shr.u32 	%r5169, %r5144, 24;
	bfe.u32 	%r5170, %r5138, 16, 8;
	mul.wide.u32 	%rd5306, %r5170, 4;
	add.s64 	%rd5307, %rd5306, %rd4726;
	ld.shared.u32 	%r5171, [%rd5307+2048];
	xor.b32  	%r5172, %r5157, %r5171;
	shr.u32 	%r5173, %r5138, 24;
	bfe.u32 	%r5174, %r5140, 16, 8;
	mul.wide.u32 	%rd5308, %r5174, 4;
	add.s64 	%rd5309, %rd5308, %rd4726;
	ld.shared.u32 	%r5175, [%rd5309+2048];
	xor.b32  	%r5176, %r5161, %r5175;
	shr.u32 	%r5177, %r5140, 24;
	mul.wide.u32 	%rd5310, %r5169, 4;
	add.s64 	%rd5311, %rd5310, %rd4726;
	ld.shared.u32 	%r5178, [%rd5311+3072];
	mul.wide.u32 	%rd5312, %r5173, 4;
	add.s64 	%rd5313, %rd5312, %rd4726;
	ld.shared.u32 	%r5179, [%rd5313+3072];
	mul.wide.u32 	%rd5314, %r5177, 4;
	add.s64 	%rd5315, %rd5314, %rd4726;
	ld.shared.u32 	%r5180, [%rd5315+3072];
	mul.wide.u32 	%rd5316, %r5165, 4;
	add.s64 	%rd5317, %rd5316, %rd4726;
	ld.shared.u32 	%r5181, [%rd5317+3072];
	xor.b32  	%r5182, %r5176, %r5181;
	xor.b32  	%r5183, %r5172, %r5180;
	xor.b32  	%r5184, %r5168, %r5179;
	xor.b32  	%r5185, %r5164, %r5178;
	ld.global.v4.u32 	{%r5186, %r5187, %r5188, %r5189}, [%rd4724+288];
	xor.b32  	%r5194, %r4846, %r4852;
	and.b32  	%r5195, %r5194, 255;
	mul.wide.u32 	%rd5318, %r5195, 4;
	add.s64 	%rd5319, %rd4726, %rd5318;
	xor.b32  	%r5196, %r4958, %r4964;
	and.b32  	%r5197, %r5196, 255;
	mul.wide.u32 	%rd5320, %r5197, 4;
	add.s64 	%rd5321, %rd4726, %rd5320;
	xor.b32  	%r5198, %r4902, %r4908;
	and.b32  	%r5199, %r5198, 255;
	mul.wide.u32 	%rd5322, %r5199, 4;
	add.s64 	%rd5323, %rd4726, %rd5322;
	xor.b32  	%r5200, %r5014, %r5020;
	and.b32  	%r5201, %r5200, 255;
	mul.wide.u32 	%rd5324, %r5201, 4;
	add.s64 	%rd5325, %rd4726, %rd5324;
	bfe.u32 	%r5202, %r5196, 8, 8;
	mul.wide.u32 	%rd5326, %r5202, 4;
	add.s64 	%rd5327, %rd5326, %rd4726;
	ld.shared.u32 	%r5203, [%rd5327+1024];
	ld.shared.u32 	%r5204, [%rd5319];
	xor.b32  	%r5205, %r5203, %r5204;
	bfe.u32 	%r5206, %r5198, 8, 8;
	mul.wide.u32 	%rd5328, %r5206, 4;
	add.s64 	%rd5329, %rd5328, %rd4726;
	ld.shared.u32 	%r5207, [%rd5329+1024];
	ld.shared.u32 	%r5208, [%rd5321];
	xor.b32  	%r5209, %r5207, %r5208;
	bfe.u32 	%r5210, %r5200, 8, 8;
	mul.wide.u32 	%rd5330, %r5210, 4;
	add.s64 	%rd5331, %rd5330, %rd4726;
	ld.shared.u32 	%r5211, [%rd5331+1024];
	ld.shared.u32 	%r5212, [%rd5323];
	xor.b32  	%r5213, %r5211, %r5212;
	bfe.u32 	%r5214, %r5194, 8, 8;
	mul.wide.u32 	%rd5332, %r5214, 4;
	add.s64 	%rd5333, %rd5332, %rd4726;
	ld.shared.u32 	%r5215, [%rd5333+1024];
	ld.shared.u32 	%r5216, [%rd5325];
	xor.b32  	%r5217, %r5215, %r5216;
	bfe.u32 	%r5218, %r5198, 16, 8;
	mul.wide.u32 	%rd5334, %r5218, 4;
	add.s64 	%rd5335, %rd5334, %rd4726;
	ld.shared.u32 	%r5219, [%rd5335+2048];
	xor.b32  	%r5220, %r5205, %r5219;
	shr.u32 	%r5221, %r5198, 24;
	bfe.u32 	%r5222, %r5200, 16, 8;
	mul.wide.u32 	%rd5336, %r5222, 4;
	add.s64 	%rd5337, %rd5336, %rd4726;
	ld.shared.u32 	%r5223, [%rd5337+2048];
	xor.b32  	%r5224, %r5209, %r5223;
	shr.u32 	%r5225, %r5200, 24;
	bfe.u32 	%r5226, %r5194, 16, 8;
	mul.wide.u32 	%rd5338, %r5226, 4;
	add.s64 	%rd5339, %rd5338, %rd4726;
	ld.shared.u32 	%r5227, [%rd5339+2048];
	xor.b32  	%r5228, %r5213, %r5227;
	shr.u32 	%r5229, %r5194, 24;
	bfe.u32 	%r5230, %r5196, 16, 8;
	mul.wide.u32 	%rd5340, %r5230, 4;
	add.s64 	%rd5341, %rd5340, %rd4726;
	ld.shared.u32 	%r5231, [%rd5341+2048];
	xor.b32  	%r5232, %r5217, %r5231;
	shr.u32 	%r5233, %r5196, 24;
	mul.wide.u32 	%rd5342, %r5225, 4;
	add.s64 	%rd5343, %rd5342, %rd4726;
	ld.shared.u32 	%r5234, [%rd5343+3072];
	mul.wide.u32 	%rd5344, %r5229, 4;
	add.s64 	%rd5345, %rd5344, %rd4726;
	ld.shared.u32 	%r5235, [%rd5345+3072];
	mul.wide.u32 	%rd5346, %r5233, 4;
	add.s64 	%rd5347, %rd5346, %rd4726;
	ld.shared.u32 	%r5236, [%rd5347+3072];
	mul.wide.u32 	%rd5348, %r5221, 4;
	add.s64 	%rd5349, %rd5348, %rd4726;
	ld.shared.u32 	%r5237, [%rd5349+3072];
	xor.b32  	%r5238, %r5232, %r5237;
	xor.b32  	%r5239, %r5228, %r5236;
	xor.b32  	%r5240, %r5224, %r5235;
	xor.b32  	%r5241, %r5220, %r5234;
	ld.global.v4.u32 	{%r5242, %r5243, %r5244, %r5245}, [%rd4724+304];
	xor.b32  	%r5250, %r5073, %r5074;
	and.b32  	%r5251, %r5250, 255;
	mul.wide.u32 	%rd5350, %r5251, 4;
	add.s64 	%rd5351, %rd4726, %rd5350;
	xor.b32  	%r5252, %r5072, %r5075;
	and.b32  	%r5253, %r5252, 255;
	mul.wide.u32 	%rd5352, %r5253, 4;
	add.s64 	%rd5353, %rd4726, %rd5352;
	xor.b32  	%r5254, %r5071, %r5076;
	and.b32  	%r5255, %r5254, 255;
	mul.wide.u32 	%rd5354, %r5255, 4;
	add.s64 	%rd5355, %rd4726, %rd5354;
	xor.b32  	%r5256, %r5070, %r5077;
	and.b32  	%r5257, %r5256, 255;
	mul.wide.u32 	%rd5356, %r5257, 4;
	add.s64 	%rd5357, %rd4726, %rd5356;
	bfe.u32 	%r5258, %r5252, 8, 8;
	mul.wide.u32 	%rd5358, %r5258, 4;
	add.s64 	%rd5359, %rd5358, %rd4726;
	ld.shared.u32 	%r5259, [%rd5359+1024];
	ld.shared.u32 	%r5260, [%rd5351];
	xor.b32  	%r5261, %r5259, %r5260;
	bfe.u32 	%r5262, %r5254, 8, 8;
	mul.wide.u32 	%rd5360, %r5262, 4;
	add.s64 	%rd5361, %rd5360, %rd4726;
	ld.shared.u32 	%r5263, [%rd5361+1024];
	ld.shared.u32 	%r5264, [%rd5353];
	xor.b32  	%r5265, %r5263, %r5264;
	bfe.u32 	%r5266, %r5256, 8, 8;
	mul.wide.u32 	%rd5362, %r5266, 4;
	add.s64 	%rd5363, %rd5362, %rd4726;
	ld.shared.u32 	%r5267, [%rd5363+1024];
	ld.shared.u32 	%r5268, [%rd5355];
	xor.b32  	%r5269, %r5267, %r5268;
	bfe.u32 	%r5270, %r5250, 8, 8;
	mul.wide.u32 	%rd5364, %r5270, 4;
	add.s64 	%rd5365, %rd5364, %rd4726;
	ld.shared.u32 	%r5271, [%rd5365+1024];
	ld.shared.u32 	%r5272, [%rd5357];
	xor.b32  	%r5273, %r5271, %r5272;
	bfe.u32 	%r5274, %r5254, 16, 8;
	mul.wide.u32 	%rd5366, %r5274, 4;
	add.s64 	%rd5367, %rd5366, %rd4726;
	ld.shared.u32 	%r5275, [%rd5367+2048];
	xor.b32  	%r5276, %r5261, %r5275;
	shr.u32 	%r5277, %r5254, 24;
	bfe.u32 	%r5278, %r5256, 16, 8;
	mul.wide.u32 	%rd5368, %r5278, 4;
	add.s64 	%rd5369, %rd5368, %rd4726;
	ld.shared.u32 	%r5279, [%rd5369+2048];
	xor.b32  	%r5280, %r5265, %r5279;
	shr.u32 	%r5281, %r5256, 24;
	bfe.u32 	%r5282, %r5250, 16, 8;
	mul.wide.u32 	%rd5370, %r5282, 4;
	add.s64 	%rd5371, %rd5370, %rd4726;
	ld.shared.u32 	%r5283, [%rd5371+2048];
	xor.b32  	%r5284, %r5269, %r5283;
	shr.u32 	%r5285, %r5250, 24;
	bfe.u32 	%r5286, %r5252, 16, 8;
	mul.wide.u32 	%rd5372, %r5286, 4;
	add.s64 	%rd5373, %rd5372, %rd4726;
	ld.shared.u32 	%r5287, [%rd5373+2048];
	xor.b32  	%r5288, %r5273, %r5287;
	shr.u32 	%r5289, %r5252, 24;
	mul.wide.u32 	%rd5374, %r5281, 4;
	add.s64 	%rd5375, %rd5374, %rd4726;
	ld.shared.u32 	%r5290, [%rd5375+3072];
	mul.wide.u32 	%rd5376, %r5285, 4;
	add.s64 	%rd5377, %rd5376, %rd4726;
	ld.shared.u32 	%r5291, [%rd5377+3072];
	mul.wide.u32 	%rd5378, %r5289, 4;
	add.s64 	%rd5379, %rd5378, %rd4726;
	ld.shared.u32 	%r5292, [%rd5379+3072];
	mul.wide.u32 	%rd5380, %r5277, 4;
	add.s64 	%rd5381, %rd5380, %rd4726;
	ld.shared.u32 	%r5293, [%rd5381+3072];
	xor.b32  	%r5294, %r5284, %r5292;
	xor.b32  	%r5295, %r5280, %r5291;
	xor.b32  	%r5296, %r5276, %r5290;
	xor.b32  	%r5297, %r5288, %r5293;
	ld.global.v4.u32 	{%r5298, %r5299, %r5300, %r5301}, [%rd4724+320];
	xor.b32  	%r5306, %r5129, %r5130;
	and.b32  	%r5307, %r5306, 255;
	mul.wide.u32 	%rd5382, %r5307, 4;
	add.s64 	%rd5383, %rd4726, %rd5382;
	xor.b32  	%r5308, %r5128, %r5131;
	and.b32  	%r5309, %r5308, 255;
	mul.wide.u32 	%rd5384, %r5309, 4;
	add.s64 	%rd5385, %rd4726, %rd5384;
	xor.b32  	%r5310, %r5127, %r5132;
	and.b32  	%r5311, %r5310, 255;
	mul.wide.u32 	%rd5386, %r5311, 4;
	add.s64 	%rd5387, %rd4726, %rd5386;
	xor.b32  	%r5312, %r5126, %r5133;
	and.b32  	%r5313, %r5312, 255;
	mul.wide.u32 	%rd5388, %r5313, 4;
	add.s64 	%rd5389, %rd4726, %rd5388;
	bfe.u32 	%r5314, %r5308, 8, 8;
	mul.wide.u32 	%rd5390, %r5314, 4;
	add.s64 	%rd5391, %rd5390, %rd4726;
	ld.shared.u32 	%r5315, [%rd5391+1024];
	ld.shared.u32 	%r5316, [%rd5383];
	xor.b32  	%r5317, %r5315, %r5316;
	bfe.u32 	%r5318, %r5310, 8, 8;
	mul.wide.u32 	%rd5392, %r5318, 4;
	add.s64 	%rd5393, %rd5392, %rd4726;
	ld.shared.u32 	%r5319, [%rd5393+1024];
	ld.shared.u32 	%r5320, [%rd5385];
	xor.b32  	%r5321, %r5319, %r5320;
	bfe.u32 	%r5322, %r5312, 8, 8;
	mul.wide.u32 	%rd5394, %r5322, 4;
	add.s64 	%rd5395, %rd5394, %rd4726;
	ld.shared.u32 	%r5323, [%rd5395+1024];
	ld.shared.u32 	%r5324, [%rd5387];
	xor.b32  	%r5325, %r5323, %r5324;
	bfe.u32 	%r5326, %r5306, 8, 8;
	mul.wide.u32 	%rd5396, %r5326, 4;
	add.s64 	%rd5397, %rd5396, %rd4726;
	ld.shared.u32 	%r5327, [%rd5397+1024];
	ld.shared.u32 	%r5328, [%rd5389];
	xor.b32  	%r5329, %r5327, %r5328;
	bfe.u32 	%r5330, %r5310, 16, 8;
	mul.wide.u32 	%rd5398, %r5330, 4;
	add.s64 	%rd5399, %rd5398, %rd4726;
	ld.shared.u32 	%r5331, [%rd5399+2048];
	xor.b32  	%r5332, %r5317, %r5331;
	shr.u32 	%r5333, %r5310, 24;
	bfe.u32 	%r5334, %r5312, 16, 8;
	mul.wide.u32 	%rd5400, %r5334, 4;
	add.s64 	%rd5401, %rd5400, %rd4726;
	ld.shared.u32 	%r5335, [%rd5401+2048];
	xor.b32  	%r5336, %r5321, %r5335;
	shr.u32 	%r5337, %r5312, 24;
	bfe.u32 	%r5338, %r5306, 16, 8;
	mul.wide.u32 	%rd5402, %r5338, 4;
	add.s64 	%rd5403, %rd5402, %rd4726;
	ld.shared.u32 	%r5339, [%rd5403+2048];
	xor.b32  	%r5340, %r5325, %r5339;
	shr.u32 	%r5341, %r5306, 24;
	bfe.u32 	%r5342, %r5308, 16, 8;
	mul.wide.u32 	%rd5404, %r5342, 4;
	add.s64 	%rd5405, %rd5404, %rd4726;
	ld.shared.u32 	%r5343, [%rd5405+2048];
	xor.b32  	%r5344, %r5329, %r5343;
	shr.u32 	%r5345, %r5308, 24;
	mul.wide.u32 	%rd5406, %r5337, 4;
	add.s64 	%rd5407, %rd5406, %rd4726;
	ld.shared.u32 	%r5346, [%rd5407+3072];
	mul.wide.u32 	%rd5408, %r5341, 4;
	add.s64 	%rd5409, %rd5408, %rd4726;
	ld.shared.u32 	%r5347, [%rd5409+3072];
	mul.wide.u32 	%rd5410, %r5345, 4;
	add.s64 	%rd5411, %rd5410, %rd4726;
	ld.shared.u32 	%r5348, [%rd5411+3072];
	mul.wide.u32 	%rd5412, %r5333, 4;
	add.s64 	%rd5413, %rd5412, %rd4726;
	ld.shared.u32 	%r5349, [%rd5413+3072];
	xor.b32  	%r5350, %r5340, %r5348;
	xor.b32  	%r5351, %r5336, %r5347;
	xor.b32  	%r5352, %r5332, %r5346;
	xor.b32  	%r5353, %r5344, %r5349;
	ld.global.v4.u32 	{%r5354, %r5355, %r5356, %r5357}, [%rd4724+336];
	xor.b32  	%r5362, %r5185, %r5186;
	and.b32  	%r5363, %r5362, 255;
	mul.wide.u32 	%rd5414, %r5363, 4;
	add.s64 	%rd5415, %rd4726, %rd5414;
	xor.b32  	%r5364, %r5184, %r5187;
	and.b32  	%r5365, %r5364, 255;
	mul.wide.u32 	%rd5416, %r5365, 4;
	add.s64 	%rd5417, %rd4726, %rd5416;
	xor.b32  	%r5366, %r5183, %r5188;
	and.b32  	%r5367, %r5366, 255;
	mul.wide.u32 	%rd5418, %r5367, 4;
	add.s64 	%rd5419, %rd4726, %rd5418;
	xor.b32  	%r5368, %r5182, %r5189;
	and.b32  	%r5369, %r5368, 255;
	mul.wide.u32 	%rd5420, %r5369, 4;
	add.s64 	%rd5421, %rd4726, %rd5420;
	bfe.u32 	%r5370, %r5364, 8, 8;
	mul.wide.u32 	%rd5422, %r5370, 4;
	add.s64 	%rd5423, %rd5422, %rd4726;
	ld.shared.u32 	%r5371, [%rd5423+1024];
	ld.shared.u32 	%r5372, [%rd5415];
	xor.b32  	%r5373, %r5371, %r5372;
	bfe.u32 	%r5374, %r5366, 8, 8;
	mul.wide.u32 	%rd5424, %r5374, 4;
	add.s64 	%rd5425, %rd5424, %rd4726;
	ld.shared.u32 	%r5375, [%rd5425+1024];
	ld.shared.u32 	%r5376, [%rd5417];
	xor.b32  	%r5377, %r5375, %r5376;
	bfe.u32 	%r5378, %r5368, 8, 8;
	mul.wide.u32 	%rd5426, %r5378, 4;
	add.s64 	%rd5427, %rd5426, %rd4726;
	ld.shared.u32 	%r5379, [%rd5427+1024];
	ld.shared.u32 	%r5380, [%rd5419];
	xor.b32  	%r5381, %r5379, %r5380;
	bfe.u32 	%r5382, %r5362, 8, 8;
	mul.wide.u32 	%rd5428, %r5382, 4;
	add.s64 	%rd5429, %rd5428, %rd4726;
	ld.shared.u32 	%r5383, [%rd5429+1024];
	ld.shared.u32 	%r5384, [%rd5421];
	xor.b32  	%r5385, %r5383, %r5384;
	bfe.u32 	%r5386, %r5366, 16, 8;
	mul.wide.u32 	%rd5430, %r5386, 4;
	add.s64 	%rd5431, %rd5430, %rd4726;
	ld.shared.u32 	%r5387, [%rd5431+2048];
	xor.b32  	%r5388, %r5373, %r5387;
	shr.u32 	%r5389, %r5366, 24;
	bfe.u32 	%r5390, %r5368, 16, 8;
	mul.wide.u32 	%rd5432, %r5390, 4;
	add.s64 	%rd5433, %rd5432, %rd4726;
	ld.shared.u32 	%r5391, [%rd5433+2048];
	xor.b32  	%r5392, %r5377, %r5391;
	shr.u32 	%r5393, %r5368, 24;
	bfe.u32 	%r5394, %r5362, 16, 8;
	mul.wide.u32 	%rd5434, %r5394, 4;
	add.s64 	%rd5435, %rd5434, %rd4726;
	ld.shared.u32 	%r5395, [%rd5435+2048];
	xor.b32  	%r5396, %r5381, %r5395;
	shr.u32 	%r5397, %r5362, 24;
	bfe.u32 	%r5398, %r5364, 16, 8;
	mul.wide.u32 	%rd5436, %r5398, 4;
	add.s64 	%rd5437, %rd5436, %rd4726;
	ld.shared.u32 	%r5399, [%rd5437+2048];
	xor.b32  	%r5400, %r5385, %r5399;
	shr.u32 	%r5401, %r5364, 24;
	mul.wide.u32 	%rd5438, %r5393, 4;
	add.s64 	%rd5439, %rd5438, %rd4726;
	ld.shared.u32 	%r5402, [%rd5439+3072];
	mul.wide.u32 	%rd5440, %r5397, 4;
	add.s64 	%rd5441, %rd5440, %rd4726;
	ld.shared.u32 	%r5403, [%rd5441+3072];
	mul.wide.u32 	%rd5442, %r5401, 4;
	add.s64 	%rd5443, %rd5442, %rd4726;
	ld.shared.u32 	%r5404, [%rd5443+3072];
	mul.wide.u32 	%rd5444, %r5389, 4;
	add.s64 	%rd5445, %rd5444, %rd4726;
	ld.shared.u32 	%r5405, [%rd5445+3072];
	xor.b32  	%r5406, %r5396, %r5404;
	xor.b32  	%r5407, %r5392, %r5403;
	xor.b32  	%r5408, %r5388, %r5402;
	xor.b32  	%r5409, %r5400, %r5405;
	ld.global.v4.u32 	{%r5410, %r5411, %r5412, %r5413}, [%rd4724+352];
	xor.b32  	%r5418, %r5241, %r5242;
	and.b32  	%r5419, %r5418, 255;
	mul.wide.u32 	%rd5446, %r5419, 4;
	add.s64 	%rd5447, %rd4726, %rd5446;
	xor.b32  	%r5420, %r5240, %r5243;
	and.b32  	%r5421, %r5420, 255;
	mul.wide.u32 	%rd5448, %r5421, 4;
	add.s64 	%rd5449, %rd4726, %rd5448;
	xor.b32  	%r5422, %r5239, %r5244;
	and.b32  	%r5423, %r5422, 255;
	mul.wide.u32 	%rd5450, %r5423, 4;
	add.s64 	%rd5451, %rd4726, %rd5450;
	xor.b32  	%r5424, %r5238, %r5245;
	and.b32  	%r5425, %r5424, 255;
	mul.wide.u32 	%rd5452, %r5425, 4;
	add.s64 	%rd5453, %rd4726, %rd5452;
	bfe.u32 	%r5426, %r5420, 8, 8;
	mul.wide.u32 	%rd5454, %r5426, 4;
	add.s64 	%rd5455, %rd5454, %rd4726;
	ld.shared.u32 	%r5427, [%rd5455+1024];
	ld.shared.u32 	%r5428, [%rd5447];
	xor.b32  	%r5429, %r5427, %r5428;
	bfe.u32 	%r5430, %r5422, 8, 8;
	mul.wide.u32 	%rd5456, %r5430, 4;
	add.s64 	%rd5457, %rd5456, %rd4726;
	ld.shared.u32 	%r5431, [%rd5457+1024];
	ld.shared.u32 	%r5432, [%rd5449];
	xor.b32  	%r5433, %r5431, %r5432;
	bfe.u32 	%r5434, %r5424, 8, 8;
	mul.wide.u32 	%rd5458, %r5434, 4;
	add.s64 	%rd5459, %rd5458, %rd4726;
	ld.shared.u32 	%r5435, [%rd5459+1024];
	ld.shared.u32 	%r5436, [%rd5451];
	xor.b32  	%r5437, %r5435, %r5436;
	bfe.u32 	%r5438, %r5418, 8, 8;
	mul.wide.u32 	%rd5460, %r5438, 4;
	add.s64 	%rd5461, %rd5460, %rd4726;
	ld.shared.u32 	%r5439, [%rd5461+1024];
	ld.shared.u32 	%r5440, [%rd5453];
	xor.b32  	%r5441, %r5439, %r5440;
	bfe.u32 	%r5442, %r5422, 16, 8;
	mul.wide.u32 	%rd5462, %r5442, 4;
	add.s64 	%rd5463, %rd5462, %rd4726;
	ld.shared.u32 	%r5443, [%rd5463+2048];
	xor.b32  	%r5444, %r5429, %r5443;
	shr.u32 	%r5445, %r5422, 24;
	bfe.u32 	%r5446, %r5424, 16, 8;
	mul.wide.u32 	%rd5464, %r5446, 4;
	add.s64 	%rd5465, %rd5464, %rd4726;
	ld.shared.u32 	%r5447, [%rd5465+2048];
	xor.b32  	%r5448, %r5433, %r5447;
	shr.u32 	%r5449, %r5424, 24;
	bfe.u32 	%r5450, %r5418, 16, 8;
	mul.wide.u32 	%rd5466, %r5450, 4;
	add.s64 	%rd5467, %rd5466, %rd4726;
	ld.shared.u32 	%r5451, [%rd5467+2048];
	xor.b32  	%r5452, %r5437, %r5451;
	shr.u32 	%r5453, %r5418, 24;
	bfe.u32 	%r5454, %r5420, 16, 8;
	mul.wide.u32 	%rd5468, %r5454, 4;
	add.s64 	%rd5469, %rd5468, %rd4726;
	ld.shared.u32 	%r5455, [%rd5469+2048];
	xor.b32  	%r5456, %r5441, %r5455;
	shr.u32 	%r5457, %r5420, 24;
	mul.wide.u32 	%rd5470, %r5449, 4;
	add.s64 	%rd5471, %rd5470, %rd4726;
	ld.shared.u32 	%r5458, [%rd5471+3072];
	mul.wide.u32 	%rd5472, %r5453, 4;
	add.s64 	%rd5473, %rd5472, %rd4726;
	ld.shared.u32 	%r5459, [%rd5473+3072];
	mul.wide.u32 	%rd5474, %r5457, 4;
	add.s64 	%rd5475, %rd5474, %rd4726;
	ld.shared.u32 	%r5460, [%rd5475+3072];
	mul.wide.u32 	%rd5476, %r5445, 4;
	add.s64 	%rd5477, %rd5476, %rd4726;
	ld.shared.u32 	%r5461, [%rd5477+3072];
	xor.b32  	%r5462, %r5452, %r5460;
	xor.b32  	%r5463, %r5448, %r5459;
	xor.b32  	%r5464, %r5444, %r5458;
	xor.b32  	%r5465, %r5456, %r5461;
	ld.global.v4.u32 	{%r5466, %r5467, %r5468, %r5469}, [%rd4724+368];
	xor.b32  	%r5474, %r5297, %r5301;
	and.b32  	%r5475, %r5474, 255;
	mul.wide.u32 	%rd5478, %r5475, 4;
	add.s64 	%rd5479, %rd4726, %rd5478;
	xor.b32  	%r5476, %r5409, %r5413;
	and.b32  	%r5477, %r5476, 255;
	mul.wide.u32 	%rd5480, %r5477, 4;
	add.s64 	%rd5481, %rd4726, %rd5480;
	xor.b32  	%r5478, %r5353, %r5357;
	and.b32  	%r5479, %r5478, 255;
	mul.wide.u32 	%rd5482, %r5479, 4;
	add.s64 	%rd5483, %rd4726, %rd5482;
	xor.b32  	%r5480, %r5465, %r5469;
	and.b32  	%r5481, %r5480, 255;
	mul.wide.u32 	%rd5484, %r5481, 4;
	add.s64 	%rd5485, %rd4726, %rd5484;
	bfe.u32 	%r5482, %r5476, 8, 8;
	mul.wide.u32 	%rd5486, %r5482, 4;
	add.s64 	%rd5487, %rd5486, %rd4726;
	ld.shared.u32 	%r5483, [%rd5487+1024];
	ld.shared.u32 	%r5484, [%rd5479];
	xor.b32  	%r5485, %r5483, %r5484;
	bfe.u32 	%r5486, %r5478, 8, 8;
	mul.wide.u32 	%rd5488, %r5486, 4;
	add.s64 	%rd5489, %rd5488, %rd4726;
	ld.shared.u32 	%r5487, [%rd5489+1024];
	ld.shared.u32 	%r5488, [%rd5481];
	xor.b32  	%r5489, %r5487, %r5488;
	bfe.u32 	%r5490, %r5480, 8, 8;
	mul.wide.u32 	%rd5490, %r5490, 4;
	add.s64 	%rd5491, %rd5490, %rd4726;
	ld.shared.u32 	%r5491, [%rd5491+1024];
	ld.shared.u32 	%r5492, [%rd5483];
	xor.b32  	%r5493, %r5491, %r5492;
	bfe.u32 	%r5494, %r5474, 8, 8;
	mul.wide.u32 	%rd5492, %r5494, 4;
	add.s64 	%rd5493, %rd5492, %rd4726;
	ld.shared.u32 	%r5495, [%rd5493+1024];
	ld.shared.u32 	%r5496, [%rd5485];
	xor.b32  	%r5497, %r5495, %r5496;
	bfe.u32 	%r5498, %r5478, 16, 8;
	mul.wide.u32 	%rd5494, %r5498, 4;
	add.s64 	%rd5495, %rd5494, %rd4726;
	ld.shared.u32 	%r5499, [%rd5495+2048];
	xor.b32  	%r5500, %r5485, %r5499;
	shr.u32 	%r5501, %r5478, 24;
	bfe.u32 	%r5502, %r5480, 16, 8;
	mul.wide.u32 	%rd5496, %r5502, 4;
	add.s64 	%rd5497, %rd5496, %rd4726;
	ld.shared.u32 	%r5503, [%rd5497+2048];
	xor.b32  	%r5504, %r5489, %r5503;
	shr.u32 	%r5505, %r5480, 24;
	bfe.u32 	%r5506, %r5474, 16, 8;
	mul.wide.u32 	%rd5498, %r5506, 4;
	add.s64 	%rd5499, %rd5498, %rd4726;
	ld.shared.u32 	%r5507, [%rd5499+2048];
	xor.b32  	%r5508, %r5493, %r5507;
	shr.u32 	%r5509, %r5474, 24;
	bfe.u32 	%r5510, %r5476, 16, 8;
	mul.wide.u32 	%rd5500, %r5510, 4;
	add.s64 	%rd5501, %rd5500, %rd4726;
	ld.shared.u32 	%r5511, [%rd5501+2048];
	xor.b32  	%r5512, %r5497, %r5511;
	shr.u32 	%r5513, %r5476, 24;
	mul.wide.u32 	%rd5502, %r5505, 4;
	add.s64 	%rd5503, %rd5502, %rd4726;
	ld.shared.u32 	%r5514, [%rd5503+3072];
	mul.wide.u32 	%rd5504, %r5509, 4;
	add.s64 	%rd5505, %rd5504, %rd4726;
	ld.shared.u32 	%r5515, [%rd5505+3072];
	mul.wide.u32 	%rd5506, %r5513, 4;
	add.s64 	%rd5507, %rd5506, %rd4726;
	ld.shared.u32 	%r5516, [%rd5507+3072];
	mul.wide.u32 	%rd5508, %r5501, 4;
	add.s64 	%rd5509, %rd5508, %rd4726;
	ld.shared.u32 	%r5517, [%rd5509+3072];
	xor.b32  	%r5518, %r5512, %r5517;
	xor.b32  	%r5519, %r5508, %r5516;
	xor.b32  	%r5520, %r5504, %r5515;
	xor.b32  	%r5521, %r5500, %r5514;
	ld.global.v4.u32 	{%r5522, %r5523, %r5524, %r5525}, [%rd4724+384];
	xor.b32  	%r5530, %r5408, %r5410;
	and.b32  	%r5531, %r5530, 255;
	mul.wide.u32 	%rd5510, %r5531, 4;
	add.s64 	%rd5511, %rd4726, %rd5510;
	xor.b32  	%r5532, %r5296, %r5298;
	and.b32  	%r5533, %r5532, 255;
	mul.wide.u32 	%rd5512, %r5533, 4;
	add.s64 	%rd5513, %rd4726, %rd5512;
	xor.b32  	%r5534, %r5464, %r5466;
	and.b32  	%r5535, %r5534, 255;
	mul.wide.u32 	%rd5514, %r5535, 4;
	add.s64 	%rd5515, %rd4726, %rd5514;
	xor.b32  	%r5536, %r5352, %r5354;
	and.b32  	%r5537, %r5536, 255;
	mul.wide.u32 	%rd5516, %r5537, 4;
	add.s64 	%rd5517, %rd4726, %rd5516;
	bfe.u32 	%r5538, %r5532, 8, 8;
	mul.wide.u32 	%rd5518, %r5538, 4;
	add.s64 	%rd5519, %rd5518, %rd4726;
	ld.shared.u32 	%r5539, [%rd5519+1024];
	ld.shared.u32 	%r5540, [%rd5511];
	xor.b32  	%r5541, %r5539, %r5540;
	bfe.u32 	%r5542, %r5534, 8, 8;
	mul.wide.u32 	%rd5520, %r5542, 4;
	add.s64 	%rd5521, %rd5520, %rd4726;
	ld.shared.u32 	%r5543, [%rd5521+1024];
	ld.shared.u32 	%r5544, [%rd5513];
	xor.b32  	%r5545, %r5543, %r5544;
	bfe.u32 	%r5546, %r5536, 8, 8;
	mul.wide.u32 	%rd5522, %r5546, 4;
	add.s64 	%rd5523, %rd5522, %rd4726;
	ld.shared.u32 	%r5547, [%rd5523+1024];
	ld.shared.u32 	%r5548, [%rd5515];
	xor.b32  	%r5549, %r5547, %r5548;
	bfe.u32 	%r5550, %r5530, 8, 8;
	mul.wide.u32 	%rd5524, %r5550, 4;
	add.s64 	%rd5525, %rd5524, %rd4726;
	ld.shared.u32 	%r5551, [%rd5525+1024];
	ld.shared.u32 	%r5552, [%rd5517];
	xor.b32  	%r5553, %r5551, %r5552;
	bfe.u32 	%r5554, %r5534, 16, 8;
	mul.wide.u32 	%rd5526, %r5554, 4;
	add.s64 	%rd5527, %rd5526, %rd4726;
	ld.shared.u32 	%r5555, [%rd5527+2048];
	xor.b32  	%r5556, %r5541, %r5555;
	shr.u32 	%r5557, %r5534, 24;
	bfe.u32 	%r5558, %r5536, 16, 8;
	mul.wide.u32 	%rd5528, %r5558, 4;
	add.s64 	%rd5529, %rd5528, %rd4726;
	ld.shared.u32 	%r5559, [%rd5529+2048];
	xor.b32  	%r5560, %r5545, %r5559;
	shr.u32 	%r5561, %r5536, 24;
	bfe.u32 	%r5562, %r5530, 16, 8;
	mul.wide.u32 	%rd5530, %r5562, 4;
	add.s64 	%rd5531, %rd5530, %rd4726;
	ld.shared.u32 	%r5563, [%rd5531+2048];
	xor.b32  	%r5564, %r5549, %r5563;
	shr.u32 	%r5565, %r5530, 24;
	bfe.u32 	%r5566, %r5532, 16, 8;
	mul.wide.u32 	%rd5532, %r5566, 4;
	add.s64 	%rd5533, %rd5532, %rd4726;
	ld.shared.u32 	%r5567, [%rd5533+2048];
	xor.b32  	%r5568, %r5553, %r5567;
	shr.u32 	%r5569, %r5532, 24;
	mul.wide.u32 	%rd5534, %r5561, 4;
	add.s64 	%rd5535, %rd5534, %rd4726;
	ld.shared.u32 	%r5570, [%rd5535+3072];
	mul.wide.u32 	%rd5536, %r5565, 4;
	add.s64 	%rd5537, %rd5536, %rd4726;
	ld.shared.u32 	%r5571, [%rd5537+3072];
	mul.wide.u32 	%rd5538, %r5569, 4;
	add.s64 	%rd5539, %rd5538, %rd4726;
	ld.shared.u32 	%r5572, [%rd5539+3072];
	mul.wide.u32 	%rd5540, %r5557, 4;
	add.s64 	%rd5541, %rd5540, %rd4726;
	ld.shared.u32 	%r5573, [%rd5541+3072];
	xor.b32  	%r5574, %r5568, %r5573;
	xor.b32  	%r5575, %r5564, %r5572;
	xor.b32  	%r5576, %r5560, %r5571;
	xor.b32  	%r5577, %r5556, %r5570;
	ld.global.v4.u32 	{%r5578, %r5579, %r5580, %r5581}, [%rd4724+400];
	xor.b32  	%r5586, %r5407, %r5411;
	and.b32  	%r5587, %r5586, 255;
	mul.wide.u32 	%rd5542, %r5587, 4;
	add.s64 	%rd5543, %rd4726, %rd5542;
	xor.b32  	%r5588, %r5295, %r5299;
	and.b32  	%r5589, %r5588, 255;
	mul.wide.u32 	%rd5544, %r5589, 4;
	add.s64 	%rd5545, %rd4726, %rd5544;
	xor.b32  	%r5590, %r5463, %r5467;
	and.b32  	%r5591, %r5590, 255;
	mul.wide.u32 	%rd5546, %r5591, 4;
	add.s64 	%rd5547, %rd4726, %rd5546;
	xor.b32  	%r5592, %r5351, %r5355;
	and.b32  	%r5593, %r5592, 255;
	mul.wide.u32 	%rd5548, %r5593, 4;
	add.s64 	%rd5549, %rd4726, %rd5548;
	bfe.u32 	%r5594, %r5588, 8, 8;
	mul.wide.u32 	%rd5550, %r5594, 4;
	add.s64 	%rd5551, %rd5550, %rd4726;
	ld.shared.u32 	%r5595, [%rd5551+1024];
	ld.shared.u32 	%r5596, [%rd5543];
	xor.b32  	%r5597, %r5595, %r5596;
	bfe.u32 	%r5598, %r5590, 8, 8;
	mul.wide.u32 	%rd5552, %r5598, 4;
	add.s64 	%rd5553, %rd5552, %rd4726;
	ld.shared.u32 	%r5599, [%rd5553+1024];
	ld.shared.u32 	%r5600, [%rd5545];
	xor.b32  	%r5601, %r5599, %r5600;
	bfe.u32 	%r5602, %r5592, 8, 8;
	mul.wide.u32 	%rd5554, %r5602, 4;
	add.s64 	%rd5555, %rd5554, %rd4726;
	ld.shared.u32 	%r5603, [%rd5555+1024];
	ld.shared.u32 	%r5604, [%rd5547];
	xor.b32  	%r5605, %r5603, %r5604;
	bfe.u32 	%r5606, %r5586, 8, 8;
	mul.wide.u32 	%rd5556, %r5606, 4;
	add.s64 	%rd5557, %rd5556, %rd4726;
	ld.shared.u32 	%r5607, [%rd5557+1024];
	ld.shared.u32 	%r5608, [%rd5549];
	xor.b32  	%r5609, %r5607, %r5608;
	bfe.u32 	%r5610, %r5590, 16, 8;
	mul.wide.u32 	%rd5558, %r5610, 4;
	add.s64 	%rd5559, %rd5558, %rd4726;
	ld.shared.u32 	%r5611, [%rd5559+2048];
	xor.b32  	%r5612, %r5597, %r5611;
	shr.u32 	%r5613, %r5590, 24;
	bfe.u32 	%r5614, %r5592, 16, 8;
	mul.wide.u32 	%rd5560, %r5614, 4;
	add.s64 	%rd5561, %rd5560, %rd4726;
	ld.shared.u32 	%r5615, [%rd5561+2048];
	xor.b32  	%r5616, %r5601, %r5615;
	shr.u32 	%r5617, %r5592, 24;
	bfe.u32 	%r5618, %r5586, 16, 8;
	mul.wide.u32 	%rd5562, %r5618, 4;
	add.s64 	%rd5563, %rd5562, %rd4726;
	ld.shared.u32 	%r5619, [%rd5563+2048];
	xor.b32  	%r5620, %r5605, %r5619;
	shr.u32 	%r5621, %r5586, 24;
	bfe.u32 	%r5622, %r5588, 16, 8;
	mul.wide.u32 	%rd5564, %r5622, 4;
	add.s64 	%rd5565, %rd5564, %rd4726;
	ld.shared.u32 	%r5623, [%rd5565+2048];
	xor.b32  	%r5624, %r5609, %r5623;
	shr.u32 	%r5625, %r5588, 24;
	mul.wide.u32 	%rd5566, %r5617, 4;
	add.s64 	%rd5567, %rd5566, %rd4726;
	ld.shared.u32 	%r5626, [%rd5567+3072];
	mul.wide.u32 	%rd5568, %r5621, 4;
	add.s64 	%rd5569, %rd5568, %rd4726;
	ld.shared.u32 	%r5627, [%rd5569+3072];
	mul.wide.u32 	%rd5570, %r5625, 4;
	add.s64 	%rd5571, %rd5570, %rd4726;
	ld.shared.u32 	%r5628, [%rd5571+3072];
	mul.wide.u32 	%rd5572, %r5613, 4;
	add.s64 	%rd5573, %rd5572, %rd4726;
	ld.shared.u32 	%r5629, [%rd5573+3072];
	xor.b32  	%r5630, %r5624, %r5629;
	xor.b32  	%r5631, %r5620, %r5628;
	xor.b32  	%r5632, %r5616, %r5627;
	xor.b32  	%r5633, %r5612, %r5626;
	ld.global.v4.u32 	{%r5634, %r5635, %r5636, %r5637}, [%rd4724+416];
	xor.b32  	%r5642, %r5294, %r5300;
	and.b32  	%r5643, %r5642, 255;
	mul.wide.u32 	%rd5574, %r5643, 4;
	add.s64 	%rd5575, %rd4726, %rd5574;
	xor.b32  	%r5644, %r5406, %r5412;
	and.b32  	%r5645, %r5644, 255;
	mul.wide.u32 	%rd5576, %r5645, 4;
	add.s64 	%rd5577, %rd4726, %rd5576;
	xor.b32  	%r5646, %r5350, %r5356;
	and.b32  	%r5647, %r5646, 255;
	mul.wide.u32 	%rd5578, %r5647, 4;
	add.s64 	%rd5579, %rd4726, %rd5578;
	xor.b32  	%r5648, %r5462, %r5468;
	and.b32  	%r5649, %r5648, 255;
	mul.wide.u32 	%rd5580, %r5649, 4;
	add.s64 	%rd5581, %rd4726, %rd5580;
	bfe.u32 	%r5650, %r5644, 8, 8;
	mul.wide.u32 	%rd5582, %r5650, 4;
	add.s64 	%rd5583, %rd5582, %rd4726;
	ld.shared.u32 	%r5651, [%rd5583+1024];
	ld.shared.u32 	%r5652, [%rd5575];
	xor.b32  	%r5653, %r5651, %r5652;
	bfe.u32 	%r5654, %r5646, 8, 8;
	mul.wide.u32 	%rd5584, %r5654, 4;
	add.s64 	%rd5585, %rd5584, %rd4726;
	ld.shared.u32 	%r5655, [%rd5585+1024];
	ld.shared.u32 	%r5656, [%rd5577];
	xor.b32  	%r5657, %r5655, %r5656;
	bfe.u32 	%r5658, %r5648, 8, 8;
	mul.wide.u32 	%rd5586, %r5658, 4;
	add.s64 	%rd5587, %rd5586, %rd4726;
	ld.shared.u32 	%r5659, [%rd5587+1024];
	ld.shared.u32 	%r5660, [%rd5579];
	xor.b32  	%r5661, %r5659, %r5660;
	bfe.u32 	%r5662, %r5642, 8, 8;
	mul.wide.u32 	%rd5588, %r5662, 4;
	add.s64 	%rd5589, %rd5588, %rd4726;
	ld.shared.u32 	%r5663, [%rd5589+1024];
	ld.shared.u32 	%r5664, [%rd5581];
	xor.b32  	%r5665, %r5663, %r5664;
	bfe.u32 	%r5666, %r5646, 16, 8;
	mul.wide.u32 	%rd5590, %r5666, 4;
	add.s64 	%rd5591, %rd5590, %rd4726;
	ld.shared.u32 	%r5667, [%rd5591+2048];
	xor.b32  	%r5668, %r5653, %r5667;
	shr.u32 	%r5669, %r5646, 24;
	bfe.u32 	%r5670, %r5648, 16, 8;
	mul.wide.u32 	%rd5592, %r5670, 4;
	add.s64 	%rd5593, %rd5592, %rd4726;
	ld.shared.u32 	%r5671, [%rd5593+2048];
	xor.b32  	%r5672, %r5657, %r5671;
	shr.u32 	%r5673, %r5648, 24;
	bfe.u32 	%r5674, %r5642, 16, 8;
	mul.wide.u32 	%rd5594, %r5674, 4;
	add.s64 	%rd5595, %rd5594, %rd4726;
	ld.shared.u32 	%r5675, [%rd5595+2048];
	xor.b32  	%r5676, %r5661, %r5675;
	shr.u32 	%r5677, %r5642, 24;
	bfe.u32 	%r5678, %r5644, 16, 8;
	mul.wide.u32 	%rd5596, %r5678, 4;
	add.s64 	%rd5597, %rd5596, %rd4726;
	ld.shared.u32 	%r5679, [%rd5597+2048];
	xor.b32  	%r5680, %r5665, %r5679;
	shr.u32 	%r5681, %r5644, 24;
	mul.wide.u32 	%rd5598, %r5673, 4;
	add.s64 	%rd5599, %rd5598, %rd4726;
	ld.shared.u32 	%r5682, [%rd5599+3072];
	mul.wide.u32 	%rd5600, %r5677, 4;
	add.s64 	%rd5601, %rd5600, %rd4726;
	ld.shared.u32 	%r5683, [%rd5601+3072];
	mul.wide.u32 	%rd5602, %r5681, 4;
	add.s64 	%rd5603, %rd5602, %rd4726;
	ld.shared.u32 	%r5684, [%rd5603+3072];
	mul.wide.u32 	%rd5604, %r5669, 4;
	add.s64 	%rd5605, %rd5604, %rd4726;
	ld.shared.u32 	%r5685, [%rd5605+3072];
	xor.b32  	%r5686, %r5680, %r5685;
	xor.b32  	%r5687, %r5676, %r5684;
	xor.b32  	%r5688, %r5672, %r5683;
	xor.b32  	%r5689, %r5668, %r5682;
	ld.global.v4.u32 	{%r5690, %r5691, %r5692, %r5693}, [%rd4724+432];
	xor.b32  	%r5698, %r5521, %r5522;
	xor.b32  	%r5699, %r5520, %r5523;
	and.b32  	%r5700, %r5699, 255;
	mul.wide.u32 	%rd5606, %r5700, 4;
	add.s64 	%rd5607, %rd4726, %rd5606;
	xor.b32  	%r5701, %r5519, %r5524;
	xor.b32  	%r5702, %r5518, %r5525;
	bfe.u32 	%r5703, %r5701, 8, 8;
	mul.wide.u32 	%rd5608, %r5703, 4;
	add.s64 	%rd5609, %rd5608, %rd4726;
	ld.shared.u32 	%r5704, [%rd5609+1024];
	ld.shared.u32 	%r5705, [%rd5607];
	xor.b32  	%r5706, %r5704, %r5705;
	bfe.u32 	%r5707, %r5702, 16, 8;
	mul.wide.u32 	%rd5610, %r5707, 4;
	add.s64 	%rd5611, %rd5610, %rd4726;
	ld.shared.u32 	%r5708, [%rd5611+2048];
	xor.b32  	%r5709, %r5706, %r5708;
	shr.u32 	%r5710, %r5698, 24;
	mul.wide.u32 	%rd5612, %r5710, 4;
	add.s64 	%rd5613, %rd5612, %rd4726;
	ld.shared.u32 	%r5711, [%rd5613+3072];
	xor.b32  	%r5712, %r5709, %r5711;
	xor.b32  	%r5713, %r5577, %r5578;
	xor.b32  	%r5714, %r5576, %r5579;
	and.b32  	%r5715, %r5714, 255;
	mul.wide.u32 	%rd5614, %r5715, 4;
	add.s64 	%rd5615, %rd4726, %rd5614;
	xor.b32  	%r5716, %r5575, %r5580;
	xor.b32  	%r5717, %r5574, %r5581;
	bfe.u32 	%r5718, %r5716, 8, 8;
	mul.wide.u32 	%rd5616, %r5718, 4;
	add.s64 	%rd5617, %rd5616, %rd4726;
	ld.shared.u32 	%r5719, [%rd5617+1024];
	ld.shared.u32 	%r5720, [%rd5615];
	xor.b32  	%r5721, %r5719, %r5720;
	bfe.u32 	%r5722, %r5717, 16, 8;
	mul.wide.u32 	%rd5618, %r5722, 4;
	add.s64 	%rd5619, %rd5618, %rd4726;
	ld.shared.u32 	%r5723, [%rd5619+2048];
	xor.b32  	%r5724, %r5721, %r5723;
	shr.u32 	%r5725, %r5713, 24;
	mul.wide.u32 	%rd5620, %r5725, 4;
	add.s64 	%rd5621, %rd5620, %rd4726;
	ld.shared.u32 	%r5726, [%rd5621+3072];
	xor.b32  	%r5727, %r5724, %r5726;
	xor.b32  	%r5728, %r5633, %r5634;
	xor.b32  	%r5729, %r5632, %r5635;
	and.b32  	%r5730, %r5729, 255;
	mul.wide.u32 	%rd5622, %r5730, 4;
	add.s64 	%rd5623, %rd4726, %rd5622;
	xor.b32  	%r5731, %r5631, %r5636;
	xor.b32  	%r5732, %r5630, %r5637;
	bfe.u32 	%r5733, %r5731, 8, 8;
	mul.wide.u32 	%rd5624, %r5733, 4;
	add.s64 	%rd5625, %rd5624, %rd4726;
	ld.shared.u32 	%r5734, [%rd5625+1024];
	ld.shared.u32 	%r5735, [%rd5623];
	xor.b32  	%r5736, %r5734, %r5735;
	bfe.u32 	%r5737, %r5732, 16, 8;
	mul.wide.u32 	%rd5626, %r5737, 4;
	add.s64 	%rd5627, %rd5626, %rd4726;
	ld.shared.u32 	%r5738, [%rd5627+2048];
	xor.b32  	%r5739, %r5736, %r5738;
	shr.u32 	%r5740, %r5728, 24;
	mul.wide.u32 	%rd5628, %r5740, 4;
	add.s64 	%rd5629, %rd5628, %rd4726;
	ld.shared.u32 	%r5741, [%rd5629+3072];
	xor.b32  	%r5742, %r5739, %r5741;
	xor.b32  	%r5743, %r5689, %r5690;
	xor.b32  	%r5744, %r5688, %r5691;
	and.b32  	%r5745, %r5744, 255;
	mul.wide.u32 	%rd5630, %r5745, 4;
	add.s64 	%rd5631, %rd4726, %rd5630;
	xor.b32  	%r5746, %r5687, %r5692;
	xor.b32  	%r5747, %r5686, %r5693;
	bfe.u32 	%r5748, %r5746, 8, 8;
	mul.wide.u32 	%rd5632, %r5748, 4;
	add.s64 	%rd5633, %rd5632, %rd4726;
	ld.shared.u32 	%r5749, [%rd5633+1024];
	ld.shared.u32 	%r5750, [%rd5631];
	xor.b32  	%r5751, %r5749, %r5750;
	bfe.u32 	%r5752, %r5747, 16, 8;
	mul.wide.u32 	%rd5634, %r5752, 4;
	add.s64 	%rd5635, %rd5634, %rd4726;
	ld.shared.u32 	%r5753, [%rd5635+2048];
	xor.b32  	%r5754, %r5751, %r5753;
	shr.u32 	%r5755, %r5743, 24;
	mul.wide.u32 	%rd5636, %r5755, 4;
	add.s64 	%rd5637, %rd5636, %rd4726;
	ld.shared.u32 	%r5756, [%rd5637+3072];
	xor.b32  	%r5757, %r5754, %r5756;
	ld.global.v4.u32 	{%r5758, %r5759, %r5760, %r5761}, [%rd4724+480];
	xor.b32  	%r5763, %r5742, %r5759;
	and.b32  	%r5764, %r5763, 255;
	mul.wide.u32 	%rd5638, %r5764, 4;
	add.s64 	%rd5639, %rd4726, %rd5638;
	ld.global.v4.u32 	{%r5765, %r5766, %r5767, %r5768}, [%rd4724+448];
	xor.b32  	%r5770, %r5712, %r5766;
	and.b32  	%r5771, %r5770, 255;
	mul.wide.u32 	%rd5640, %r5771, 4;
	add.s64 	%rd5641, %rd4726, %rd5640;
	ld.global.v4.u32 	{%r5772, %r5773, %r5774, %r5775}, [%rd4724+496];
	xor.b32  	%r5777, %r5757, %r5773;
	and.b32  	%r5778, %r5777, 255;
	mul.wide.u32 	%rd5642, %r5778, 4;
	add.s64 	%rd5643, %rd4726, %rd5642;
	ld.global.v4.u32 	{%r5779, %r5780, %r5781, %r5782}, [%rd4724+464];
	xor.b32  	%r5784, %r5727, %r5780;
	and.b32  	%r5785, %r5784, 255;
	mul.wide.u32 	%rd5644, %r5785, 4;
	add.s64 	%rd5645, %rd4726, %rd5644;
	bfe.u32 	%r5786, %r5770, 8, 8;
	mul.wide.u32 	%rd5646, %r5786, 4;
	add.s64 	%rd5647, %rd5646, %rd4726;
	ld.shared.u32 	%r5787, [%rd5647+1024];
	ld.shared.u32 	%r5788, [%rd5639];
	xor.b32  	%r5789, %r5787, %r5788;
	bfe.u32 	%r5790, %r5777, 8, 8;
	mul.wide.u32 	%rd5648, %r5790, 4;
	add.s64 	%rd5649, %rd5648, %rd4726;
	ld.shared.u32 	%r5791, [%rd5649+1024];
	ld.shared.u32 	%r5792, [%rd5641];
	xor.b32  	%r5793, %r5791, %r5792;
	bfe.u32 	%r5794, %r5784, 8, 8;
	mul.wide.u32 	%rd5650, %r5794, 4;
	add.s64 	%rd5651, %rd5650, %rd4726;
	ld.shared.u32 	%r5795, [%rd5651+1024];
	ld.shared.u32 	%r5796, [%rd5643];
	xor.b32  	%r5797, %r5795, %r5796;
	bfe.u32 	%r5798, %r5763, 8, 8;
	mul.wide.u32 	%rd5652, %r5798, 4;
	add.s64 	%rd5653, %rd5652, %rd4726;
	ld.shared.u32 	%r5799, [%rd5653+1024];
	ld.shared.u32 	%r5800, [%rd5645];
	xor.b32  	%r5801, %r5799, %r5800;
	bfe.u32 	%r5802, %r5777, 16, 8;
	mul.wide.u32 	%rd5654, %r5802, 4;
	add.s64 	%rd5655, %rd5654, %rd4726;
	ld.shared.u32 	%r5803, [%rd5655+2048];
	xor.b32  	%r5804, %r5789, %r5803;
	shr.u32 	%r5805, %r5777, 24;
	bfe.u32 	%r5806, %r5784, 16, 8;
	mul.wide.u32 	%rd5656, %r5806, 4;
	add.s64 	%rd5657, %rd5656, %rd4726;
	ld.shared.u32 	%r5807, [%rd5657+2048];
	xor.b32  	%r5808, %r5793, %r5807;
	shr.u32 	%r5809, %r5784, 24;
	bfe.u32 	%r5810, %r5763, 16, 8;
	mul.wide.u32 	%rd5658, %r5810, 4;
	add.s64 	%rd5659, %rd5658, %rd4726;
	ld.shared.u32 	%r5811, [%rd5659+2048];
	xor.b32  	%r5812, %r5797, %r5811;
	shr.u32 	%r5813, %r5763, 24;
	bfe.u32 	%r5814, %r5770, 16, 8;
	mul.wide.u32 	%rd5660, %r5814, 4;
	add.s64 	%rd5661, %rd5660, %rd4726;
	ld.shared.u32 	%r5815, [%rd5661+2048];
	xor.b32  	%r5816, %r5801, %r5815;
	shr.u32 	%r5817, %r5770, 24;
	mul.wide.u32 	%rd5662, %r5809, 4;
	add.s64 	%rd5663, %rd5662, %rd4726;
	ld.shared.u32 	%r5818, [%rd5663+3072];
	mul.wide.u32 	%rd5664, %r5813, 4;
	add.s64 	%rd5665, %rd5664, %rd4726;
	ld.shared.u32 	%r5819, [%rd5665+3072];
	mul.wide.u32 	%rd5666, %r5817, 4;
	add.s64 	%rd5667, %rd5666, %rd4726;
	ld.shared.u32 	%r5820, [%rd5667+3072];
	mul.wide.u32 	%rd5668, %r5805, 4;
	add.s64 	%rd5669, %rd5668, %rd4726;
	ld.shared.u32 	%r5821, [%rd5669+3072];
	xor.b32  	%r5822, %r5816, %r5821;
	xor.b32  	%r5823, %r5812, %r5820;
	xor.b32  	%r5824, %r5808, %r5819;
	xor.b32  	%r5825, %r5804, %r5818;
	ld.global.v4.u32 	{%r5826, %r5827, %r5828, %r5829}, [%rd4724+544];
	xor.b32  	%r5834, %r5825, %r5826;
	xor.b32  	%r5835, %r5824, %r5827;
	xor.b32  	%r5836, %r5823, %r5828;
	and.b32  	%r5837, %r5836, 255;
	mul.wide.u32 	%rd5670, %r5837, 4;
	add.s64 	%rd5671, %rd4726, %rd5670;
	xor.b32  	%r5838, %r5822, %r5829;
	bfe.u32 	%r5839, %r5838, 8, 8;
	mul.wide.u32 	%rd5672, %r5839, 4;
	add.s64 	%rd5673, %rd5672, %rd4726;
	ld.shared.u32 	%r5840, [%rd5673+1024];
	ld.shared.u32 	%r5841, [%rd5671];
	xor.b32  	%r5842, %r5840, %r5841;
	bfe.u32 	%r5843, %r5834, 16, 8;
	mul.wide.u32 	%rd5674, %r5843, 4;
	add.s64 	%rd5675, %rd5674, %rd4726;
	ld.shared.u32 	%r5844, [%rd5675+2048];
	xor.b32  	%r5845, %r5842, %r5844;
	shr.u32 	%r5846, %r5835, 24;
	mul.wide.u32 	%rd5676, %r5846, 4;
	add.s64 	%rd5677, %rd5676, %rd4726;
	ld.shared.u32 	%r5847, [%rd5677+3072];
	xor.b32  	%r5848, %r5845, %r5847;
	ld.global.v4.u32 	{%r5849, %r5850, %r5851, %r5852}, [%rd4724+608];
	xor.b32  	%r5854, %r5848, %r5851;
	xor.b32  	%r5855, %r5854, %r4157;
	ld.global.u32 	%r5856, [%rd6067+28];
	setp.ge.u32	%p92, %r5855, %r5856;
	@%p92 bra 	BB0_152;

	mov.b32	%r6443, %envreg3;
	mov.u32 	%r6442, %ntid.x;
	mov.u32 	%r6441, %ctaid.x;
	mov.u32 	%r6440, %tid.x;
	mad.lo.s32 	%r6439, %r6441, %r6442, %r6443;
	ld.param.u64 	%rd6070, [verus_gpu_hash_param_0];
	ld.const.u32 	%r6438, [%rd6070];
	add.s32 	%r6437, %r6439, %r6440;
	add.s32 	%r6436, %r6437, %r6438;
	ld.param.u64 	%rd6069, [verus_gpu_hash_param_5];
	st.global.u32 	[%rd6069], %r6436;

BB0_152:
	ld.param.u64 	%rd6068, [verus_gpu_hash_param_3];
	mov.u32 	%r6435, %tid.x;
	shl.b32 	%r5857, %r6435, 5;
	mul.wide.u32 	%rd5678, %r5857, 2;
	mov.u64 	%rd5679, verus_gpu_hash$sharedrand;
	add.s64 	%rd5680, %rd5679, %rd5678;
	ld.shared.u16 	%rs145, [%rd5680];
	cvt.u64.u16	%rd5681, %rs145;
	cvt.u32.u16	%r5858, %rs145;
	mul.wide.u32 	%rd5682, %r5858, 16;
	add.s64 	%rd5683, %rd6068, %rd5682;
	ld.const.v4.u32 	{%r5859, %r5860, %r5861, %r5862}, [%rd5683];
	add.s64 	%rd5684, %rd5681, %rd3;
	shl.b64 	%rd5685, %rd5684, 4;
	add.s64 	%rd5686, %rd203, %rd5685;
	st.global.v4.u32 	[%rd5686], {%r5859, %r5860, %r5861, %r5862};
	mov.u64 	%rd5687, verus_gpu_hash$sharedrandex;
	add.s64 	%rd5688, %rd5687, %rd5678;
	ld.shared.u16 	%rs146, [%rd5688];
	cvt.u64.u16	%rd5689, %rs146;
	cvt.u32.u16	%r5867, %rs146;
	mul.wide.u32 	%rd5690, %r5867, 16;
	add.s64 	%rd5691, %rd6068, %rd5690;
	ld.const.v4.u32 	{%r5868, %r5869, %r5870, %r5871}, [%rd5691];
	add.s64 	%rd5692, %rd5689, %rd3;
	shl.b64 	%rd5693, %rd5692, 4;
	add.s64 	%rd5694, %rd203, %rd5693;
	st.global.v4.u32 	[%rd5694], {%r5868, %r5869, %r5870, %r5871};
	ld.shared.u16 	%rs147, [%rd5680+2];
	cvt.u64.u16	%rd5695, %rs147;
	cvt.u32.u16	%r5876, %rs147;
	mul.wide.u32 	%rd5696, %r5876, 16;
	add.s64 	%rd5697, %rd6068, %rd5696;
	ld.const.v4.u32 	{%r5877, %r5878, %r5879, %r5880}, [%rd5697];
	add.s64 	%rd5698, %rd5695, %rd3;
	shl.b64 	%rd5699, %rd5698, 4;
	add.s64 	%rd5700, %rd203, %rd5699;
	st.global.v4.u32 	[%rd5700], {%r5877, %r5878, %r5879, %r5880};
	ld.shared.u16 	%rs148, [%rd5688+2];
	cvt.u64.u16	%rd5701, %rs148;
	cvt.u32.u16	%r5885, %rs148;
	mul.wide.u32 	%rd5702, %r5885, 16;
	add.s64 	%rd5703, %rd6068, %rd5702;
	ld.const.v4.u32 	{%r5886, %r5887, %r5888, %r5889}, [%rd5703];
	add.s64 	%rd5704, %rd5701, %rd3;
	shl.b64 	%rd5705, %rd5704, 4;
	add.s64 	%rd5706, %rd203, %rd5705;
	st.global.v4.u32 	[%rd5706], {%r5886, %r5887, %r5888, %r5889};
	ld.shared.u16 	%rs149, [%rd5680+4];
	cvt.u64.u16	%rd5707, %rs149;
	cvt.u32.u16	%r5894, %rs149;
	mul.wide.u32 	%rd5708, %r5894, 16;
	add.s64 	%rd5709, %rd6068, %rd5708;
	ld.const.v4.u32 	{%r5895, %r5896, %r5897, %r5898}, [%rd5709];
	add.s64 	%rd5710, %rd5707, %rd3;
	shl.b64 	%rd5711, %rd5710, 4;
	add.s64 	%rd5712, %rd203, %rd5711;
	st.global.v4.u32 	[%rd5712], {%r5895, %r5896, %r5897, %r5898};
	ld.shared.u16 	%rs150, [%rd5688+4];
	cvt.u64.u16	%rd5713, %rs150;
	cvt.u32.u16	%r5903, %rs150;
	mul.wide.u32 	%rd5714, %r5903, 16;
	add.s64 	%rd5715, %rd6068, %rd5714;
	ld.const.v4.u32 	{%r5904, %r5905, %r5906, %r5907}, [%rd5715];
	add.s64 	%rd5716, %rd5713, %rd3;
	shl.b64 	%rd5717, %rd5716, 4;
	add.s64 	%rd5718, %rd203, %rd5717;
	st.global.v4.u32 	[%rd5718], {%r5904, %r5905, %r5906, %r5907};
	ld.shared.u16 	%rs151, [%rd5680+6];
	cvt.u64.u16	%rd5719, %rs151;
	cvt.u32.u16	%r5912, %rs151;
	mul.wide.u32 	%rd5720, %r5912, 16;
	add.s64 	%rd5721, %rd6068, %rd5720;
	ld.const.v4.u32 	{%r5913, %r5914, %r5915, %r5916}, [%rd5721];
	add.s64 	%rd5722, %rd5719, %rd3;
	shl.b64 	%rd5723, %rd5722, 4;
	add.s64 	%rd5724, %rd203, %rd5723;
	st.global.v4.u32 	[%rd5724], {%r5913, %r5914, %r5915, %r5916};
	ld.shared.u16 	%rs152, [%rd5688+6];
	cvt.u64.u16	%rd5725, %rs152;
	cvt.u32.u16	%r5921, %rs152;
	mul.wide.u32 	%rd5726, %r5921, 16;
	add.s64 	%rd5727, %rd6068, %rd5726;
	ld.const.v4.u32 	{%r5922, %r5923, %r5924, %r5925}, [%rd5727];
	add.s64 	%rd5728, %rd5725, %rd3;
	shl.b64 	%rd5729, %rd5728, 4;
	add.s64 	%rd5730, %rd203, %rd5729;
	st.global.v4.u32 	[%rd5730], {%r5922, %r5923, %r5924, %r5925};
	ld.shared.u16 	%rs153, [%rd5680+8];
	cvt.u64.u16	%rd5731, %rs153;
	cvt.u32.u16	%r5930, %rs153;
	mul.wide.u32 	%rd5732, %r5930, 16;
	add.s64 	%rd5733, %rd6068, %rd5732;
	ld.const.v4.u32 	{%r5931, %r5932, %r5933, %r5934}, [%rd5733];
	add.s64 	%rd5734, %rd5731, %rd3;
	shl.b64 	%rd5735, %rd5734, 4;
	add.s64 	%rd5736, %rd203, %rd5735;
	st.global.v4.u32 	[%rd5736], {%r5931, %r5932, %r5933, %r5934};
	ld.shared.u16 	%rs154, [%rd5688+8];
	cvt.u64.u16	%rd5737, %rs154;
	cvt.u32.u16	%r5939, %rs154;
	mul.wide.u32 	%rd5738, %r5939, 16;
	add.s64 	%rd5739, %rd6068, %rd5738;
	ld.const.v4.u32 	{%r5940, %r5941, %r5942, %r5943}, [%rd5739];
	add.s64 	%rd5740, %rd5737, %rd3;
	shl.b64 	%rd5741, %rd5740, 4;
	add.s64 	%rd5742, %rd203, %rd5741;
	st.global.v4.u32 	[%rd5742], {%r5940, %r5941, %r5942, %r5943};
	ld.shared.u16 	%rs155, [%rd5680+10];
	cvt.u64.u16	%rd5743, %rs155;
	cvt.u32.u16	%r5948, %rs155;
	mul.wide.u32 	%rd5744, %r5948, 16;
	add.s64 	%rd5745, %rd6068, %rd5744;
	ld.const.v4.u32 	{%r5949, %r5950, %r5951, %r5952}, [%rd5745];
	add.s64 	%rd5746, %rd5743, %rd3;
	shl.b64 	%rd5747, %rd5746, 4;
	add.s64 	%rd5748, %rd203, %rd5747;
	st.global.v4.u32 	[%rd5748], {%r5949, %r5950, %r5951, %r5952};
	ld.shared.u16 	%rs156, [%rd5688+10];
	cvt.u64.u16	%rd5749, %rs156;
	cvt.u32.u16	%r5957, %rs156;
	mul.wide.u32 	%rd5750, %r5957, 16;
	add.s64 	%rd5751, %rd6068, %rd5750;
	ld.const.v4.u32 	{%r5958, %r5959, %r5960, %r5961}, [%rd5751];
	add.s64 	%rd5752, %rd5749, %rd3;
	shl.b64 	%rd5753, %rd5752, 4;
	add.s64 	%rd5754, %rd203, %rd5753;
	st.global.v4.u32 	[%rd5754], {%r5958, %r5959, %r5960, %r5961};
	ld.shared.u16 	%rs157, [%rd5680+12];
	cvt.u64.u16	%rd5755, %rs157;
	cvt.u32.u16	%r5966, %rs157;
	mul.wide.u32 	%rd5756, %r5966, 16;
	add.s64 	%rd5757, %rd6068, %rd5756;
	ld.const.v4.u32 	{%r5967, %r5968, %r5969, %r5970}, [%rd5757];
	add.s64 	%rd5758, %rd5755, %rd3;
	shl.b64 	%rd5759, %rd5758, 4;
	add.s64 	%rd5760, %rd203, %rd5759;
	st.global.v4.u32 	[%rd5760], {%r5967, %r5968, %r5969, %r5970};
	ld.shared.u16 	%rs158, [%rd5688+12];
	cvt.u64.u16	%rd5761, %rs158;
	cvt.u32.u16	%r5975, %rs158;
	mul.wide.u32 	%rd5762, %r5975, 16;
	add.s64 	%rd5763, %rd6068, %rd5762;
	ld.const.v4.u32 	{%r5976, %r5977, %r5978, %r5979}, [%rd5763];
	add.s64 	%rd5764, %rd5761, %rd3;
	shl.b64 	%rd5765, %rd5764, 4;
	add.s64 	%rd5766, %rd203, %rd5765;
	st.global.v4.u32 	[%rd5766], {%r5976, %r5977, %r5978, %r5979};
	ld.shared.u16 	%rs159, [%rd5680+14];
	cvt.u64.u16	%rd5767, %rs159;
	cvt.u32.u16	%r5984, %rs159;
	mul.wide.u32 	%rd5768, %r5984, 16;
	add.s64 	%rd5769, %rd6068, %rd5768;
	ld.const.v4.u32 	{%r5985, %r5986, %r5987, %r5988}, [%rd5769];
	add.s64 	%rd5770, %rd5767, %rd3;
	shl.b64 	%rd5771, %rd5770, 4;
	add.s64 	%rd5772, %rd203, %rd5771;
	st.global.v4.u32 	[%rd5772], {%r5985, %r5986, %r5987, %r5988};
	ld.shared.u16 	%rs160, [%rd5688+14];
	cvt.u64.u16	%rd5773, %rs160;
	cvt.u32.u16	%r5993, %rs160;
	mul.wide.u32 	%rd5774, %r5993, 16;
	add.s64 	%rd5775, %rd6068, %rd5774;
	ld.const.v4.u32 	{%r5994, %r5995, %r5996, %r5997}, [%rd5775];
	add.s64 	%rd5776, %rd5773, %rd3;
	shl.b64 	%rd5777, %rd5776, 4;
	add.s64 	%rd5778, %rd203, %rd5777;
	st.global.v4.u32 	[%rd5778], {%r5994, %r5995, %r5996, %r5997};
	ld.shared.u16 	%rs161, [%rd5680+16];
	cvt.u64.u16	%rd5779, %rs161;
	cvt.u32.u16	%r6002, %rs161;
	mul.wide.u32 	%rd5780, %r6002, 16;
	add.s64 	%rd5781, %rd6068, %rd5780;
	ld.const.v4.u32 	{%r6003, %r6004, %r6005, %r6006}, [%rd5781];
	add.s64 	%rd5782, %rd5779, %rd3;
	shl.b64 	%rd5783, %rd5782, 4;
	add.s64 	%rd5784, %rd203, %rd5783;
	st.global.v4.u32 	[%rd5784], {%r6003, %r6004, %r6005, %r6006};
	ld.shared.u16 	%rs162, [%rd5688+16];
	cvt.u64.u16	%rd5785, %rs162;
	cvt.u32.u16	%r6011, %rs162;
	mul.wide.u32 	%rd5786, %r6011, 16;
	add.s64 	%rd5787, %rd6068, %rd5786;
	ld.const.v4.u32 	{%r6012, %r6013, %r6014, %r6015}, [%rd5787];
	add.s64 	%rd5788, %rd5785, %rd3;
	shl.b64 	%rd5789, %rd5788, 4;
	add.s64 	%rd5790, %rd203, %rd5789;
	st.global.v4.u32 	[%rd5790], {%r6012, %r6013, %r6014, %r6015};
	ld.shared.u16 	%rs163, [%rd5680+18];
	cvt.u64.u16	%rd5791, %rs163;
	cvt.u32.u16	%r6020, %rs163;
	mul.wide.u32 	%rd5792, %r6020, 16;
	add.s64 	%rd5793, %rd6068, %rd5792;
	ld.const.v4.u32 	{%r6021, %r6022, %r6023, %r6024}, [%rd5793];
	add.s64 	%rd5794, %rd5791, %rd3;
	shl.b64 	%rd5795, %rd5794, 4;
	add.s64 	%rd5796, %rd203, %rd5795;
	st.global.v4.u32 	[%rd5796], {%r6021, %r6022, %r6023, %r6024};
	ld.shared.u16 	%rs164, [%rd5688+18];
	cvt.u64.u16	%rd5797, %rs164;
	cvt.u32.u16	%r6029, %rs164;
	mul.wide.u32 	%rd5798, %r6029, 16;
	add.s64 	%rd5799, %rd6068, %rd5798;
	ld.const.v4.u32 	{%r6030, %r6031, %r6032, %r6033}, [%rd5799];
	add.s64 	%rd5800, %rd5797, %rd3;
	shl.b64 	%rd5801, %rd5800, 4;
	add.s64 	%rd5802, %rd203, %rd5801;
	st.global.v4.u32 	[%rd5802], {%r6030, %r6031, %r6032, %r6033};
	ld.shared.u16 	%rs165, [%rd5680+20];
	cvt.u64.u16	%rd5803, %rs165;
	cvt.u32.u16	%r6038, %rs165;
	mul.wide.u32 	%rd5804, %r6038, 16;
	add.s64 	%rd5805, %rd6068, %rd5804;
	ld.const.v4.u32 	{%r6039, %r6040, %r6041, %r6042}, [%rd5805];
	add.s64 	%rd5806, %rd5803, %rd3;
	shl.b64 	%rd5807, %rd5806, 4;
	add.s64 	%rd5808, %rd203, %rd5807;
	st.global.v4.u32 	[%rd5808], {%r6039, %r6040, %r6041, %r6042};
	ld.shared.u16 	%rs166, [%rd5688+20];
	cvt.u64.u16	%rd5809, %rs166;
	cvt.u32.u16	%r6047, %rs166;
	mul.wide.u32 	%rd5810, %r6047, 16;
	add.s64 	%rd5811, %rd6068, %rd5810;
	ld.const.v4.u32 	{%r6048, %r6049, %r6050, %r6051}, [%rd5811];
	add.s64 	%rd5812, %rd5809, %rd3;
	shl.b64 	%rd5813, %rd5812, 4;
	add.s64 	%rd5814, %rd203, %rd5813;
	st.global.v4.u32 	[%rd5814], {%r6048, %r6049, %r6050, %r6051};
	ld.shared.u16 	%rs167, [%rd5680+22];
	cvt.u64.u16	%rd5815, %rs167;
	cvt.u32.u16	%r6056, %rs167;
	mul.wide.u32 	%rd5816, %r6056, 16;
	add.s64 	%rd5817, %rd6068, %rd5816;
	ld.const.v4.u32 	{%r6057, %r6058, %r6059, %r6060}, [%rd5817];
	add.s64 	%rd5818, %rd5815, %rd3;
	shl.b64 	%rd5819, %rd5818, 4;
	add.s64 	%rd5820, %rd203, %rd5819;
	st.global.v4.u32 	[%rd5820], {%r6057, %r6058, %r6059, %r6060};
	ld.shared.u16 	%rs168, [%rd5688+22];
	cvt.u64.u16	%rd5821, %rs168;
	cvt.u32.u16	%r6065, %rs168;
	mul.wide.u32 	%rd5822, %r6065, 16;
	add.s64 	%rd5823, %rd6068, %rd5822;
	ld.const.v4.u32 	{%r6066, %r6067, %r6068, %r6069}, [%rd5823];
	add.s64 	%rd5824, %rd5821, %rd3;
	shl.b64 	%rd5825, %rd5824, 4;
	add.s64 	%rd5826, %rd203, %rd5825;
	st.global.v4.u32 	[%rd5826], {%r6066, %r6067, %r6068, %r6069};
	ld.shared.u16 	%rs169, [%rd5680+24];
	cvt.u64.u16	%rd5827, %rs169;
	cvt.u32.u16	%r6074, %rs169;
	mul.wide.u32 	%rd5828, %r6074, 16;
	add.s64 	%rd5829, %rd6068, %rd5828;
	ld.const.v4.u32 	{%r6075, %r6076, %r6077, %r6078}, [%rd5829];
	add.s64 	%rd5830, %rd5827, %rd3;
	shl.b64 	%rd5831, %rd5830, 4;
	add.s64 	%rd5832, %rd203, %rd5831;
	st.global.v4.u32 	[%rd5832], {%r6075, %r6076, %r6077, %r6078};
	ld.shared.u16 	%rs170, [%rd5688+24];
	cvt.u64.u16	%rd5833, %rs170;
	cvt.u32.u16	%r6083, %rs170;
	mul.wide.u32 	%rd5834, %r6083, 16;
	add.s64 	%rd5835, %rd6068, %rd5834;
	ld.const.v4.u32 	{%r6084, %r6085, %r6086, %r6087}, [%rd5835];
	add.s64 	%rd5836, %rd5833, %rd3;
	shl.b64 	%rd5837, %rd5836, 4;
	add.s64 	%rd5838, %rd203, %rd5837;
	st.global.v4.u32 	[%rd5838], {%r6084, %r6085, %r6086, %r6087};
	ld.shared.u16 	%rs171, [%rd5680+26];
	cvt.u64.u16	%rd5839, %rs171;
	cvt.u32.u16	%r6092, %rs171;
	mul.wide.u32 	%rd5840, %r6092, 16;
	add.s64 	%rd5841, %rd6068, %rd5840;
	ld.const.v4.u32 	{%r6093, %r6094, %r6095, %r6096}, [%rd5841];
	add.s64 	%rd5842, %rd5839, %rd3;
	shl.b64 	%rd5843, %rd5842, 4;
	add.s64 	%rd5844, %rd203, %rd5843;
	st.global.v4.u32 	[%rd5844], {%r6093, %r6094, %r6095, %r6096};
	ld.shared.u16 	%rs172, [%rd5688+26];
	cvt.u64.u16	%rd5845, %rs172;
	cvt.u32.u16	%r6101, %rs172;
	mul.wide.u32 	%rd5846, %r6101, 16;
	add.s64 	%rd5847, %rd6068, %rd5846;
	ld.const.v4.u32 	{%r6102, %r6103, %r6104, %r6105}, [%rd5847];
	add.s64 	%rd5848, %rd5845, %rd3;
	shl.b64 	%rd5849, %rd5848, 4;
	add.s64 	%rd5850, %rd203, %rd5849;
	st.global.v4.u32 	[%rd5850], {%r6102, %r6103, %r6104, %r6105};
	ld.shared.u16 	%rs173, [%rd5680+28];
	cvt.u64.u16	%rd5851, %rs173;
	cvt.u32.u16	%r6110, %rs173;
	mul.wide.u32 	%rd5852, %r6110, 16;
	add.s64 	%rd5853, %rd6068, %rd5852;
	ld.const.v4.u32 	{%r6111, %r6112, %r6113, %r6114}, [%rd5853];
	add.s64 	%rd5854, %rd5851, %rd3;
	shl.b64 	%rd5855, %rd5854, 4;
	add.s64 	%rd5856, %rd203, %rd5855;
	st.global.v4.u32 	[%rd5856], {%r6111, %r6112, %r6113, %r6114};
	ld.shared.u16 	%rs174, [%rd5688+28];
	cvt.u64.u16	%rd5857, %rs174;
	cvt.u32.u16	%r6119, %rs174;
	mul.wide.u32 	%rd5858, %r6119, 16;
	add.s64 	%rd5859, %rd6068, %rd5858;
	ld.const.v4.u32 	{%r6120, %r6121, %r6122, %r6123}, [%rd5859];
	add.s64 	%rd5860, %rd5857, %rd3;
	shl.b64 	%rd5861, %rd5860, 4;
	add.s64 	%rd5862, %rd203, %rd5861;
	st.global.v4.u32 	[%rd5862], {%r6120, %r6121, %r6122, %r6123};
	ld.shared.u16 	%rs175, [%rd5680+30];
	cvt.u64.u16	%rd5863, %rs175;
	cvt.u32.u16	%r6128, %rs175;
	mul.wide.u32 	%rd5864, %r6128, 16;
	add.s64 	%rd5865, %rd6068, %rd5864;
	ld.const.v4.u32 	{%r6129, %r6130, %r6131, %r6132}, [%rd5865];
	add.s64 	%rd5866, %rd5863, %rd3;
	shl.b64 	%rd5867, %rd5866, 4;
	add.s64 	%rd5868, %rd203, %rd5867;
	st.global.v4.u32 	[%rd5868], {%r6129, %r6130, %r6131, %r6132};
	ld.shared.u16 	%rs176, [%rd5688+30];
	cvt.u64.u16	%rd5869, %rs176;
	cvt.u32.u16	%r6137, %rs176;
	mul.wide.u32 	%rd5870, %r6137, 16;
	add.s64 	%rd5871, %rd6068, %rd5870;
	ld.const.v4.u32 	{%r6138, %r6139, %r6140, %r6141}, [%rd5871];
	add.s64 	%rd5872, %rd5869, %rd3;
	shl.b64 	%rd5873, %rd5872, 4;
	add.s64 	%rd5874, %rd203, %rd5873;
	st.global.v4.u32 	[%rd5874], {%r6138, %r6139, %r6140, %r6141};
	ld.shared.u16 	%rs177, [%rd5680+32];
	cvt.u64.u16	%rd5875, %rs177;
	cvt.u32.u16	%r6146, %rs177;
	mul.wide.u32 	%rd5876, %r6146, 16;
	add.s64 	%rd5877, %rd6068, %rd5876;
	ld.const.v4.u32 	{%r6147, %r6148, %r6149, %r6150}, [%rd5877];
	add.s64 	%rd5878, %rd5875, %rd3;
	shl.b64 	%rd5879, %rd5878, 4;
	add.s64 	%rd5880, %rd203, %rd5879;
	st.global.v4.u32 	[%rd5880], {%r6147, %r6148, %r6149, %r6150};
	ld.shared.u16 	%rs178, [%rd5688+32];
	cvt.u64.u16	%rd5881, %rs178;
	cvt.u32.u16	%r6155, %rs178;
	mul.wide.u32 	%rd5882, %r6155, 16;
	add.s64 	%rd5883, %rd6068, %rd5882;
	ld.const.v4.u32 	{%r6156, %r6157, %r6158, %r6159}, [%rd5883];
	add.s64 	%rd5884, %rd5881, %rd3;
	shl.b64 	%rd5885, %rd5884, 4;
	add.s64 	%rd5886, %rd203, %rd5885;
	st.global.v4.u32 	[%rd5886], {%r6156, %r6157, %r6158, %r6159};
	ld.shared.u16 	%rs179, [%rd5680+34];
	cvt.u64.u16	%rd5887, %rs179;
	cvt.u32.u16	%r6164, %rs179;
	mul.wide.u32 	%rd5888, %r6164, 16;
	add.s64 	%rd5889, %rd6068, %rd5888;
	ld.const.v4.u32 	{%r6165, %r6166, %r6167, %r6168}, [%rd5889];
	add.s64 	%rd5890, %rd5887, %rd3;
	shl.b64 	%rd5891, %rd5890, 4;
	add.s64 	%rd5892, %rd203, %rd5891;
	st.global.v4.u32 	[%rd5892], {%r6165, %r6166, %r6167, %r6168};
	ld.shared.u16 	%rs180, [%rd5688+34];
	cvt.u64.u16	%rd5893, %rs180;
	cvt.u32.u16	%r6173, %rs180;
	mul.wide.u32 	%rd5894, %r6173, 16;
	add.s64 	%rd5895, %rd6068, %rd5894;
	ld.const.v4.u32 	{%r6174, %r6175, %r6176, %r6177}, [%rd5895];
	add.s64 	%rd5896, %rd5893, %rd3;
	shl.b64 	%rd5897, %rd5896, 4;
	add.s64 	%rd5898, %rd203, %rd5897;
	st.global.v4.u32 	[%rd5898], {%r6174, %r6175, %r6176, %r6177};
	ld.shared.u16 	%rs181, [%rd5680+36];
	cvt.u64.u16	%rd5899, %rs181;
	cvt.u32.u16	%r6182, %rs181;
	mul.wide.u32 	%rd5900, %r6182, 16;
	add.s64 	%rd5901, %rd6068, %rd5900;
	ld.const.v4.u32 	{%r6183, %r6184, %r6185, %r6186}, [%rd5901];
	add.s64 	%rd5902, %rd5899, %rd3;
	shl.b64 	%rd5903, %rd5902, 4;
	add.s64 	%rd5904, %rd203, %rd5903;
	st.global.v4.u32 	[%rd5904], {%r6183, %r6184, %r6185, %r6186};
	ld.shared.u16 	%rs182, [%rd5688+36];
	cvt.u64.u16	%rd5905, %rs182;
	cvt.u32.u16	%r6191, %rs182;
	mul.wide.u32 	%rd5906, %r6191, 16;
	add.s64 	%rd5907, %rd6068, %rd5906;
	ld.const.v4.u32 	{%r6192, %r6193, %r6194, %r6195}, [%rd5907];
	add.s64 	%rd5908, %rd5905, %rd3;
	shl.b64 	%rd5909, %rd5908, 4;
	add.s64 	%rd5910, %rd203, %rd5909;
	st.global.v4.u32 	[%rd5910], {%r6192, %r6193, %r6194, %r6195};
	ld.shared.u16 	%rs183, [%rd5680+38];
	cvt.u64.u16	%rd5911, %rs183;
	cvt.u32.u16	%r6200, %rs183;
	mul.wide.u32 	%rd5912, %r6200, 16;
	add.s64 	%rd5913, %rd6068, %rd5912;
	ld.const.v4.u32 	{%r6201, %r6202, %r6203, %r6204}, [%rd5913];
	add.s64 	%rd5914, %rd5911, %rd3;
	shl.b64 	%rd5915, %rd5914, 4;
	add.s64 	%rd5916, %rd203, %rd5915;
	st.global.v4.u32 	[%rd5916], {%r6201, %r6202, %r6203, %r6204};
	ld.shared.u16 	%rs184, [%rd5688+38];
	cvt.u64.u16	%rd5917, %rs184;
	cvt.u32.u16	%r6209, %rs184;
	mul.wide.u32 	%rd5918, %r6209, 16;
	add.s64 	%rd5919, %rd6068, %rd5918;
	ld.const.v4.u32 	{%r6210, %r6211, %r6212, %r6213}, [%rd5919];
	add.s64 	%rd5920, %rd5917, %rd3;
	shl.b64 	%rd5921, %rd5920, 4;
	add.s64 	%rd5922, %rd203, %rd5921;
	st.global.v4.u32 	[%rd5922], {%r6210, %r6211, %r6212, %r6213};
	ld.shared.u16 	%rs185, [%rd5680+40];
	cvt.u64.u16	%rd5923, %rs185;
	cvt.u32.u16	%r6218, %rs185;
	mul.wide.u32 	%rd5924, %r6218, 16;
	add.s64 	%rd5925, %rd6068, %rd5924;
	ld.const.v4.u32 	{%r6219, %r6220, %r6221, %r6222}, [%rd5925];
	add.s64 	%rd5926, %rd5923, %rd3;
	shl.b64 	%rd5927, %rd5926, 4;
	add.s64 	%rd5928, %rd203, %rd5927;
	st.global.v4.u32 	[%rd5928], {%r6219, %r6220, %r6221, %r6222};
	ld.shared.u16 	%rs186, [%rd5688+40];
	cvt.u64.u16	%rd5929, %rs186;
	cvt.u32.u16	%r6227, %rs186;
	mul.wide.u32 	%rd5930, %r6227, 16;
	add.s64 	%rd5931, %rd6068, %rd5930;
	ld.const.v4.u32 	{%r6228, %r6229, %r6230, %r6231}, [%rd5931];
	add.s64 	%rd5932, %rd5929, %rd3;
	shl.b64 	%rd5933, %rd5932, 4;
	add.s64 	%rd5934, %rd203, %rd5933;
	st.global.v4.u32 	[%rd5934], {%r6228, %r6229, %r6230, %r6231};
	ld.shared.u16 	%rs187, [%rd5680+42];
	cvt.u64.u16	%rd5935, %rs187;
	cvt.u32.u16	%r6236, %rs187;
	mul.wide.u32 	%rd5936, %r6236, 16;
	add.s64 	%rd5937, %rd6068, %rd5936;
	ld.const.v4.u32 	{%r6237, %r6238, %r6239, %r6240}, [%rd5937];
	add.s64 	%rd5938, %rd5935, %rd3;
	shl.b64 	%rd5939, %rd5938, 4;
	add.s64 	%rd5940, %rd203, %rd5939;
	st.global.v4.u32 	[%rd5940], {%r6237, %r6238, %r6239, %r6240};
	ld.shared.u16 	%rs188, [%rd5688+42];
	cvt.u64.u16	%rd5941, %rs188;
	cvt.u32.u16	%r6245, %rs188;
	mul.wide.u32 	%rd5942, %r6245, 16;
	add.s64 	%rd5943, %rd6068, %rd5942;
	ld.const.v4.u32 	{%r6246, %r6247, %r6248, %r6249}, [%rd5943];
	add.s64 	%rd5944, %rd5941, %rd3;
	shl.b64 	%rd5945, %rd5944, 4;
	add.s64 	%rd5946, %rd203, %rd5945;
	st.global.v4.u32 	[%rd5946], {%r6246, %r6247, %r6248, %r6249};
	ld.shared.u16 	%rs189, [%rd5680+44];
	cvt.u64.u16	%rd5947, %rs189;
	cvt.u32.u16	%r6254, %rs189;
	mul.wide.u32 	%rd5948, %r6254, 16;
	add.s64 	%rd5949, %rd6068, %rd5948;
	ld.const.v4.u32 	{%r6255, %r6256, %r6257, %r6258}, [%rd5949];
	add.s64 	%rd5950, %rd5947, %rd3;
	shl.b64 	%rd5951, %rd5950, 4;
	add.s64 	%rd5952, %rd203, %rd5951;
	st.global.v4.u32 	[%rd5952], {%r6255, %r6256, %r6257, %r6258};
	ld.shared.u16 	%rs190, [%rd5688+44];
	cvt.u64.u16	%rd5953, %rs190;
	cvt.u32.u16	%r6263, %rs190;
	mul.wide.u32 	%rd5954, %r6263, 16;
	add.s64 	%rd5955, %rd6068, %rd5954;
	ld.const.v4.u32 	{%r6264, %r6265, %r6266, %r6267}, [%rd5955];
	add.s64 	%rd5956, %rd5953, %rd3;
	shl.b64 	%rd5957, %rd5956, 4;
	add.s64 	%rd5958, %rd203, %rd5957;
	st.global.v4.u32 	[%rd5958], {%r6264, %r6265, %r6266, %r6267};
	ld.shared.u16 	%rs191, [%rd5680+46];
	cvt.u64.u16	%rd5959, %rs191;
	cvt.u32.u16	%r6272, %rs191;
	mul.wide.u32 	%rd5960, %r6272, 16;
	add.s64 	%rd5961, %rd6068, %rd5960;
	ld.const.v4.u32 	{%r6273, %r6274, %r6275, %r6276}, [%rd5961];
	add.s64 	%rd5962, %rd5959, %rd3;
	shl.b64 	%rd5963, %rd5962, 4;
	add.s64 	%rd5964, %rd203, %rd5963;
	st.global.v4.u32 	[%rd5964], {%r6273, %r6274, %r6275, %r6276};
	ld.shared.u16 	%rs192, [%rd5688+46];
	cvt.u64.u16	%rd5965, %rs192;
	cvt.u32.u16	%r6281, %rs192;
	mul.wide.u32 	%rd5966, %r6281, 16;
	add.s64 	%rd5967, %rd6068, %rd5966;
	ld.const.v4.u32 	{%r6282, %r6283, %r6284, %r6285}, [%rd5967];
	add.s64 	%rd5968, %rd5965, %rd3;
	shl.b64 	%rd5969, %rd5968, 4;
	add.s64 	%rd5970, %rd203, %rd5969;
	st.global.v4.u32 	[%rd5970], {%r6282, %r6283, %r6284, %r6285};
	ld.shared.u16 	%rs193, [%rd5680+48];
	cvt.u64.u16	%rd5971, %rs193;
	cvt.u32.u16	%r6290, %rs193;
	mul.wide.u32 	%rd5972, %r6290, 16;
	add.s64 	%rd5973, %rd6068, %rd5972;
	ld.const.v4.u32 	{%r6291, %r6292, %r6293, %r6294}, [%rd5973];
	add.s64 	%rd5974, %rd5971, %rd3;
	shl.b64 	%rd5975, %rd5974, 4;
	add.s64 	%rd5976, %rd203, %rd5975;
	st.global.v4.u32 	[%rd5976], {%r6291, %r6292, %r6293, %r6294};
	ld.shared.u16 	%rs194, [%rd5688+48];
	cvt.u64.u16	%rd5977, %rs194;
	cvt.u32.u16	%r6299, %rs194;
	mul.wide.u32 	%rd5978, %r6299, 16;
	add.s64 	%rd5979, %rd6068, %rd5978;
	ld.const.v4.u32 	{%r6300, %r6301, %r6302, %r6303}, [%rd5979];
	add.s64 	%rd5980, %rd5977, %rd3;
	shl.b64 	%rd5981, %rd5980, 4;
	add.s64 	%rd5982, %rd203, %rd5981;
	st.global.v4.u32 	[%rd5982], {%r6300, %r6301, %r6302, %r6303};
	ld.shared.u16 	%rs195, [%rd5680+50];
	cvt.u64.u16	%rd5983, %rs195;
	cvt.u32.u16	%r6308, %rs195;
	mul.wide.u32 	%rd5984, %r6308, 16;
	add.s64 	%rd5985, %rd6068, %rd5984;
	ld.const.v4.u32 	{%r6309, %r6310, %r6311, %r6312}, [%rd5985];
	add.s64 	%rd5986, %rd5983, %rd3;
	shl.b64 	%rd5987, %rd5986, 4;
	add.s64 	%rd5988, %rd203, %rd5987;
	st.global.v4.u32 	[%rd5988], {%r6309, %r6310, %r6311, %r6312};
	ld.shared.u16 	%rs196, [%rd5688+50];
	cvt.u64.u16	%rd5989, %rs196;
	cvt.u32.u16	%r6317, %rs196;
	mul.wide.u32 	%rd5990, %r6317, 16;
	add.s64 	%rd5991, %rd6068, %rd5990;
	ld.const.v4.u32 	{%r6318, %r6319, %r6320, %r6321}, [%rd5991];
	add.s64 	%rd5992, %rd5989, %rd3;
	shl.b64 	%rd5993, %rd5992, 4;
	add.s64 	%rd5994, %rd203, %rd5993;
	st.global.v4.u32 	[%rd5994], {%r6318, %r6319, %r6320, %r6321};
	ld.shared.u16 	%rs197, [%rd5680+52];
	cvt.u64.u16	%rd5995, %rs197;
	cvt.u32.u16	%r6326, %rs197;
	mul.wide.u32 	%rd5996, %r6326, 16;
	add.s64 	%rd5997, %rd6068, %rd5996;
	ld.const.v4.u32 	{%r6327, %r6328, %r6329, %r6330}, [%rd5997];
	add.s64 	%rd5998, %rd5995, %rd3;
	shl.b64 	%rd5999, %rd5998, 4;
	add.s64 	%rd6000, %rd203, %rd5999;
	st.global.v4.u32 	[%rd6000], {%r6327, %r6328, %r6329, %r6330};
	ld.shared.u16 	%rs198, [%rd5688+52];
	cvt.u64.u16	%rd6001, %rs198;
	cvt.u32.u16	%r6335, %rs198;
	mul.wide.u32 	%rd6002, %r6335, 16;
	add.s64 	%rd6003, %rd6068, %rd6002;
	ld.const.v4.u32 	{%r6336, %r6337, %r6338, %r6339}, [%rd6003];
	add.s64 	%rd6004, %rd6001, %rd3;
	shl.b64 	%rd6005, %rd6004, 4;
	add.s64 	%rd6006, %rd203, %rd6005;
	st.global.v4.u32 	[%rd6006], {%r6336, %r6337, %r6338, %r6339};
	ld.shared.u16 	%rs199, [%rd5680+54];
	cvt.u64.u16	%rd6007, %rs199;
	cvt.u32.u16	%r6344, %rs199;
	mul.wide.u32 	%rd6008, %r6344, 16;
	add.s64 	%rd6009, %rd6068, %rd6008;
	ld.const.v4.u32 	{%r6345, %r6346, %r6347, %r6348}, [%rd6009];
	add.s64 	%rd6010, %rd6007, %rd3;
	shl.b64 	%rd6011, %rd6010, 4;
	add.s64 	%rd6012, %rd203, %rd6011;
	st.global.v4.u32 	[%rd6012], {%r6345, %r6346, %r6347, %r6348};
	ld.shared.u16 	%rs200, [%rd5688+54];
	cvt.u64.u16	%rd6013, %rs200;
	cvt.u32.u16	%r6353, %rs200;
	mul.wide.u32 	%rd6014, %r6353, 16;
	add.s64 	%rd6015, %rd6068, %rd6014;
	ld.const.v4.u32 	{%r6354, %r6355, %r6356, %r6357}, [%rd6015];
	add.s64 	%rd6016, %rd6013, %rd3;
	shl.b64 	%rd6017, %rd6016, 4;
	add.s64 	%rd6018, %rd203, %rd6017;
	st.global.v4.u32 	[%rd6018], {%r6354, %r6355, %r6356, %r6357};
	ld.shared.u16 	%rs201, [%rd5680+56];
	cvt.u64.u16	%rd6019, %rs201;
	cvt.u32.u16	%r6362, %rs201;
	mul.wide.u32 	%rd6020, %r6362, 16;
	add.s64 	%rd6021, %rd6068, %rd6020;
	ld.const.v4.u32 	{%r6363, %r6364, %r6365, %r6366}, [%rd6021];
	add.s64 	%rd6022, %rd6019, %rd3;
	shl.b64 	%rd6023, %rd6022, 4;
	add.s64 	%rd6024, %rd203, %rd6023;
	st.global.v4.u32 	[%rd6024], {%r6363, %r6364, %r6365, %r6366};
	ld.shared.u16 	%rs202, [%rd5688+56];
	cvt.u64.u16	%rd6025, %rs202;
	cvt.u32.u16	%r6371, %rs202;
	mul.wide.u32 	%rd6026, %r6371, 16;
	add.s64 	%rd6027, %rd6068, %rd6026;
	ld.const.v4.u32 	{%r6372, %r6373, %r6374, %r6375}, [%rd6027];
	add.s64 	%rd6028, %rd6025, %rd3;
	shl.b64 	%rd6029, %rd6028, 4;
	add.s64 	%rd6030, %rd203, %rd6029;
	st.global.v4.u32 	[%rd6030], {%r6372, %r6373, %r6374, %r6375};
	ld.shared.u16 	%rs203, [%rd5680+58];
	cvt.u64.u16	%rd6031, %rs203;
	cvt.u32.u16	%r6380, %rs203;
	mul.wide.u32 	%rd6032, %r6380, 16;
	add.s64 	%rd6033, %rd6068, %rd6032;
	ld.const.v4.u32 	{%r6381, %r6382, %r6383, %r6384}, [%rd6033];
	add.s64 	%rd6034, %rd6031, %rd3;
	shl.b64 	%rd6035, %rd6034, 4;
	add.s64 	%rd6036, %rd203, %rd6035;
	st.global.v4.u32 	[%rd6036], {%r6381, %r6382, %r6383, %r6384};
	ld.shared.u16 	%rs204, [%rd5688+58];
	cvt.u64.u16	%rd6037, %rs204;
	cvt.u32.u16	%r6389, %rs204;
	mul.wide.u32 	%rd6038, %r6389, 16;
	add.s64 	%rd6039, %rd6068, %rd6038;
	ld.const.v4.u32 	{%r6390, %r6391, %r6392, %r6393}, [%rd6039];
	add.s64 	%rd6040, %rd6037, %rd3;
	shl.b64 	%rd6041, %rd6040, 4;
	add.s64 	%rd6042, %rd203, %rd6041;
	st.global.v4.u32 	[%rd6042], {%r6390, %r6391, %r6392, %r6393};
	ld.shared.u16 	%rs205, [%rd5680+60];
	cvt.u64.u16	%rd6043, %rs205;
	cvt.u32.u16	%r6398, %rs205;
	mul.wide.u32 	%rd6044, %r6398, 16;
	add.s64 	%rd6045, %rd6068, %rd6044;
	ld.const.v4.u32 	{%r6399, %r6400, %r6401, %r6402}, [%rd6045];
	add.s64 	%rd6046, %rd6043, %rd3;
	shl.b64 	%rd6047, %rd6046, 4;
	add.s64 	%rd6048, %rd203, %rd6047;
	st.global.v4.u32 	[%rd6048], {%r6399, %r6400, %r6401, %r6402};
	ld.shared.u16 	%rs206, [%rd5688+60];
	cvt.u64.u16	%rd6049, %rs206;
	cvt.u32.u16	%r6407, %rs206;
	mul.wide.u32 	%rd6050, %r6407, 16;
	add.s64 	%rd6051, %rd6068, %rd6050;
	ld.const.v4.u32 	{%r6408, %r6409, %r6410, %r6411}, [%rd6051];
	add.s64 	%rd6052, %rd6049, %rd3;
	shl.b64 	%rd6053, %rd6052, 4;
	add.s64 	%rd6054, %rd203, %rd6053;
	st.global.v4.u32 	[%rd6054], {%r6408, %r6409, %r6410, %r6411};
	ld.shared.u16 	%rs207, [%rd5680+62];
	cvt.u64.u16	%rd6055, %rs207;
	cvt.u32.u16	%r6416, %rs207;
	mul.wide.u32 	%rd6056, %r6416, 16;
	add.s64 	%rd6057, %rd6068, %rd6056;
	ld.const.v4.u32 	{%r6417, %r6418, %r6419, %r6420}, [%rd6057];
	add.s64 	%rd6058, %rd6055, %rd3;
	shl.b64 	%rd6059, %rd6058, 4;
	add.s64 	%rd6060, %rd203, %rd6059;
	st.global.v4.u32 	[%rd6060], {%r6417, %r6418, %r6419, %r6420};
	ld.shared.u16 	%rs208, [%rd5688+62];
	cvt.u64.u16	%rd6061, %rs208;
	cvt.u32.u16	%r6425, %rs208;
	mul.wide.u32 	%rd6062, %r6425, 16;
	add.s64 	%rd6063, %rd6068, %rd6062;
	ld.const.v4.u32 	{%r6426, %r6427, %r6428, %r6429}, [%rd6063];
	add.s64 	%rd6064, %rd6061, %rd3;
	shl.b64 	%rd6065, %rd6064, 4;
	add.s64 	%rd6066, %rd203, %rd6065;
	st.global.v4.u32 	[%rd6066], {%r6426, %r6427, %r6428, %r6429};
	ret;
}


.metadata_section {

.metadata 0 {
	"cl_kernel_attributes",
	"verus_gpu_hash",
	"reqd_work_group_size(64,1,1)"
}

} // end of .metadata_section