better support. Just uses avx2

SCsub

@@ -15,7 +15,7 @@ if(env["PLATFORM"] != "win32"):
     env_rwkv.add_source_files(env.modules_sources, "*.cpp")
 
     # add -march=avx512
-    env_rwkv.Append(CCFLAGS=['-march=skylake-avx512'])
+    env_rwkv.Append(CCFLAGS=['-march=haswell'])
 
     # -fopenmp  -flto  -fopenmp -funroll-loops -D_GLIBCXX_PARALLEL
 
@@ -46,7 +46,7 @@ else:
     env_rwkv.Append(CPPPATH=[path + "/rwkv.hpp/include"])
 
     # set avx512 
-    env_rwkv.Append(CCFLAGS=['/arch:AVX512'])
+    env_rwkv.Append(CCFLAGS=['/arch:AVX2'])
 
 
     env_rwkv.add_source_files(env.modules_sources, "*.cpp")

rwkv.hpp/build.sh

@@ -22,7 +22,6 @@ fi
 # build with intel compiler
 # check if intel compiler installed
 if [ "$(icpx --version)" ]; then
-source /opt/intel/oneapi/2024.0/oneapi-vars.sh 
 icpx -m64 ./rwkv.cpp -I ./include/ -o ./build/rwkv -march=native -std=c++17 $vulk -ffast-math -O3 -pthread
 
 else

rwkv.hpp/include/hvml/intrinsics.hpp

@@ -6,12 +6,12 @@
 #include <cmath>
 #define UINT8THREADALLOC 64
 
-#ifdef __AVX512F__ // This macro is defined if AVX-512 is supported
+#if defined(__AVX512F__) && defined(HVMLUSEAVX512) // This macro is defined if AVX-512 is supported
 #include <immintrin.h>
 
 #define SIMD_WIDTH 16
-#define LOAD(x) _mm512_loadu_ps(x)
-#define STORE(x, y) _mm512_storeu_ps(x, y)
+#define LOAD(x) _mm512_load_ps(x)
+#define STORE(x, y) _mm512_store_ps(x, y)
 #define SET1(x) _mm512_set1_ps(x)
 #define MULTIPLY(x, y) _mm512_mul_ps(x, y)
 #define MULTADD(x, y, z) _mm512_fmadd_ps(x, y, z)
@@ -79,8 +79,8 @@ SIMDTYPE exp_ps_fill(SIMDTYPE x)
 #ifdef __AVX2__
 #include <immintrin.h>
 #define SIMD_WIDTH 8
-#define LOAD(x) _mm256_loadu_ps(x)
-#define STORE(x, y) _mm256_storeu_ps(x, y)
+#define LOAD(x) _mm256_load_ps(x)
+#define STORE(x, y) _mm256_store_ps((float*)x, y)
 #define SET1(x) _mm256_set1_ps(x)
 #define MULTIPLY(x, y) _mm256_mul_ps(x, y)
 #define MULTADD(x, y, z) _mm256_fmadd_ps(x, y, z)
@@ -93,6 +93,10 @@ SIMDTYPE exp_ps_fill(SIMDTYPE x)
 #define MAX(x, y) _mm256_max_ps(x, y)
 #define DIVIDE(x, y) _mm256_div_ps(x, y)
 #define SIMDTYPE __m256
+#if defined(__INTEL_LLVM_COMPILER)
+#pragma message("AVX-2 exp is supported")
+#define EXP(x) _mm256_exp_ps(x)
+#else
 #define EXP(x) exp_ps_fill(x)
 SIMDTYPE exp_ps_fill(SIMDTYPE x)
 {
@@ -103,6 +107,7 @@ SIMDTYPE exp_ps_fill(SIMDTYPE x)
     }
     return result;
 }
+#endif
 // print out the SIMD width
 #pragma message("AVX-2 is supported")
 

rwkv.hpp/include/hvml/operations/avx512/matmul8.hpp

@@ -93,7 +93,7 @@ void dopartial(MatMulJob job) {
 	auto Ar = job.Ar;
 	auto bbt = job.bbt;
 	auto ii = job.ii;
-	#ifdef __AVX512F__ 
+	#if defined(__AVX512F__) && defined(HVMLUSEAVX512)
 		const auto Ario = _mm512_load_ps(Ar + ii);
 		const auto Aoioo = _mm512_div_ps(_mm512_load_ps(Ao + ii), Ario);
 		__m512 zz = _mm512_setzero_ps();
@@ -118,9 +118,52 @@ void dopartial(MatMulJob job) {
 		_mm512_store_ps(
 				(void *)(C + bbt * OUT + ii),
 				zz * Ario);
-	#endif
+	#elif defined(__AVX2__)
+		for (ulong b = 0; b < 16; b+= 8){
+		const auto Ario1 = LOAD(Ar + ii+b);
+		const auto Aoio1 = DIVIDE(LOAD(Ao + ii + b),Ario1);
+
+		auto zz1 = SET1(0.0);
+
+		for (uint32_t i = ii+b; i < ii + b+8; i += 1) {
+			auto Aoio = Aoio1[i&7];
+
+			const auto IAIN = A + i * IN;
+
+			auto sum1 = SET1(0.0);
+			auto sum2 = SET1(0.0);
+
+			for (uint32_t k = 0; k < IN; k += 16) {
+				// avx2
+				auto w = _mm256_cvtepu8_epi32(_mm_loadu_si128((__m128i *)(IAIN + k)));  // Load the input uint8_t vector
+				// convert uint32_t to float32x8_t
+				auto u = _mm256_cvtepi32_ps(w)+Aoio;   // Convert uint32_t to float32_t
+				// Load the input float vector
+				// Perform the multiplication with inp vector
+				sum1 = MULTADD(u, LOAD(B + bbt * IN + k),sum1);
+
+				auto w1 = _mm256_cvtepu8_epi32(_mm_loadu_si128((__m128i *)(IAIN + k + 8)));  // Load the input uint8_t vector
+
+				auto u1 = _mm256_cvtepi32_ps(w1)+Aoio;   // Convert uint32_t to float32_t
+
+				sum2 = MULTADD(u1, LOAD(B + bbt * IN + k + 8),sum2);
+
+			}
+
+			sum1 = sum1+sum2;
+
+			zz1[i&7]= REDUCE(sum1);
+
+
+		}
+
+
+		STORE(
+				(void *)(C + bbt * OUT + ii + b),
+				zz1 * Ario1);
+		}
 
-	#if defined(__ARM_NEON) || defined(__ARM_NEON__)
+	#elif defined(__ARM_NEON) || defined(__ARM_NEON__)
 	 	for (ulong b = 0; b < 16; b+= 4){
 		const auto Ario1 = LOAD(Ar + ii+b);
 		const auto Aoio1 = DIVIDE(LOAD(Ao + ii + b),Ario1);

rwkv.hpp/rwkv.cpp

@@ -4,6 +4,7 @@
 #include "rwkv.hpp"
 #include "sampler/sample.hpp"
 #include "tokenizer/tokenizer.hpp"
+
 int main( int argc, char** argv ){
 
     std::cout << "Hello World" << std::endl;