Updata internal ovxlib to rel1.2.6

Internal ovxlib commit hash: c5d3e69356579fc7b595a7c0939fc7e4e0aaab5a

Type: Code Improvement
Signed-off-by: Feiyue Chen <Feiyue.Chen@verisilicon.com>

VERSION

@@ -1 +1 @@
-1.1.50
+1.2.6

src/tim/vx/internal/include/interface/ops.def

@@ -197,3 +197,5 @@ DEF_OP(RESIZE_3D)
 DEF_OP(REDUCEL2)
 DEF_OP(CROP_AND_RESIZE)
 DEF_OP(TAN)
+DEF_OP(RMSNORM)
+DEF_OP(SHAPE)

src/tim/vx/internal/include/kernel/vsi_nn_kernel.h

@@ -26,6 +26,8 @@
 #define _VSI_NN_KERNEL_H
 
 #include <stdint.h>
+#include <stdarg.h>
+
 #include "vsi_nn_log.h"
 #include "vsi_nn_ops.h"
 #include "vsi_nn_graph.h"
@@ -81,6 +83,7 @@ typedef enum
     U4,
     FP8_E4M3,
     FP8_E5M2,
+    INVALID_DTYPE,
 } VSI_PUBLIC_TYPE vsi_nn_kernel_dtype_e;
 
 typedef enum
@@ -532,9 +535,8 @@ static VSI_INLINE_API vsi_nn_kernel_dtype_e vsi_nn_kernel_map_dtype
         return FP8_E5M2;
     default:
         VSILOGE("error data type %d", dtype);
-        break;
+        return INVALID_DTYPE;
     }
-    return I8;
 } /* vsi_nn_kernel_map_dtype() */
 
 static VSI_INLINE_API  vsi_nn_type_e vsi_nn_dtype_map_kernel

src/tim/vx/internal/include/ops/vsi_nn_op_resize_internal.h

@@ -43,6 +43,7 @@ typedef struct _vsi_nn_resize_internal_param
     vsi_bool    half_pixel_centers;
     float       factor;
     vsi_enum    layout;
+    vsi_enum    type;
 } vsi_nn_resize_internal_param;
 
 #ifdef __cplusplus

src/tim/vx/internal/include/ops/vsi_nn_op_rmsnorm.h

@@ -0,0 +1,54 @@
+/****************************************************************************
+*
+*    Copyright (c) 2020 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+
+#ifndef _VSI_NN_OP_RMSNORM_H
+#define _VSI_NN_OP_RMSNORM_H
+
+#include "vsi_nn_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+typedef struct _rmsnorm_local_data_t {
+    int32_t placeholder;
+} rmsnorm_local_data_t;
+
+typedef struct _vsi_nn_rmsnorm_param
+{
+    struct _rmsnorm_local_data_t* local;
+    float eps;
+    int32_t axis;
+} vsi_nn_rmsnorm_param;
+
+_compiler_assert(offsetof(vsi_nn_rmsnorm_param, local) == 0, \
+    vsi_nn_rmsnorm_h );
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
+

src/tim/vx/internal/include/ops/vsi_nn_op_shape.h

@@ -0,0 +1,47 @@
+/****************************************************************************
+*
+*    Copyright (c) 2020 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+
+#ifndef _VSI_NN_OP_SHAPE_H
+#define _VSI_NN_OP_SHAPE_H
+
+#include "vsi_nn_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct _vsi_nn_shape_param
+{
+    struct _shape_local_data_t* local;
+    // Add parameters here
+} vsi_nn_shape_param;
+_compiler_assert(offsetof(vsi_nn_shape_param, local) == 0, \
+    vsi_nn_shape_h );
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
+

src/tim/vx/internal/include/utils/vsi_nn_dtype_util_prv.h

@@ -33,6 +33,14 @@
 extern "C" {
 #endif
 
+/*
+ * A helper union for fp32 bit casting.
+ */
+typedef union {
+    float val;
+    uint32_t data;
+} fp32_bit_cast_t;
+
 static VSI_INLINE_API vsi_bool type_is_integer
     (
     const vsi_nn_type_e type
@@ -203,9 +211,11 @@ static VSI_INLINE_API vsi_bool fp32_is_inf
         float val
     )
 {
-    uint32_t u_value = *(uint32_t*)&val;
+    fp32_bit_cast_t fp32_bit_cast;
+    fp32_bit_cast.val = val;
+    uint32_t fp32_data = fp32_bit_cast.data;
 
-    if ((u_value & (uint32_t)VSI_NN_INT32_MAX) == (uint32_t)VSI_NN_FLOAT32_INF)
+    if ((fp32_data & (uint32_t)VSI_NN_INT32_MAX) == (uint32_t)VSI_NN_FLOAT32_INF)
     {
         return TRUE;
     }
@@ -232,7 +242,9 @@ static VSI_INLINE_API int32_t fp32_to_affine
 
     if (fp32_is_inf(in) != 0)
     {
-        uint32_t sign = (*(uint32_t*)&in) >> 31;
+        fp32_bit_cast_t fp32_bit_cast;
+        fp32_bit_cast.val = in;
+        uint32_t sign = fp32_bit_cast.data >> 31;
         data = sign == 1 ? (int32_t)min_range : (int32_t)max_range;
     }
 
@@ -277,7 +289,9 @@ static VSI_INLINE_API int32_t fp32_to_dfp
 
     if (fp32_is_inf(in) != 0)
     {
-        uint32_t sign = (*(uint32_t*)&in) >> 31;
+        fp32_bit_cast_t fp32_bit_cast;
+        fp32_bit_cast.val = in;
+        uint32_t sign = fp32_bit_cast.data >> 31;
         data = sign == 1 ? (int32_t)min_range : (int32_t) max_range;
     }
 
@@ -373,8 +387,9 @@ static VSI_INLINE_API float bfp16_to_fp32
     int16_t in
     )
 {
-    int32_t t1, t2, t3;
+    uint32_t t1, t2, t3;
     float out;
+    fp32_bit_cast_t fp32_bit_cast;
 
     t1 = in & 0x00FF;                       // Mantissa
     t2 = in & 0xFF00;                       // Sign bit + Exponent
@@ -384,20 +399,23 @@ static VSI_INLINE_API float bfp16_to_fp32
     t2 <<= 16;                              // Shift (sign + Exponent) bit into position
     t1 |= t2;                               // Re-insert (sign + Exponent) bit
 
-    *((uint32_t*)&out) = t1;
+    fp32_bit_cast.data = t1;
+    out = fp32_bit_cast.val;
 
-    return t3 == 0 ? 0 : out;
+    return t3 == 0 ? 0.0f : out;
 } /* bfp16_to_fp32() */
 
 static VSI_INLINE_API uint16_t fp32_to_fp16
     (
     float in
     )
 {
-    uint32_t fp32 = *((uint32_t *) &in);
-    uint32_t t1 = (fp32 & 0x80000000u) >> 16;  /* sign bit. */
-    uint32_t t2 = (fp32 & 0x7F800000u) >> 13;  /* Exponent bits */
-    uint32_t t3 = (fp32 & 0x007FE000u) >> 13;  /* Mantissa bits, no rounding */
+    fp32_bit_cast_t fp32_bit_cast;
+    fp32_bit_cast.val = in;
+    uint32_t fp32_data = fp32_bit_cast.data;
+    uint32_t t1 = (fp32_data & 0x80000000u) >> 16;  /* sign bit. */
+    uint32_t t2 = (fp32_data & 0x7F800000u) >> 13;  /* Exponent bits */
+    uint32_t t3 = (fp32_data & 0x007FE000u) >> 13;  /* Mantissa bits, no rounding */
     uint32_t fp16 = 0u;
     if( t2 >= 0x023c00u )
     {
@@ -420,8 +438,10 @@ static VSI_INLINE_API uint16_t fp32_to_bfp16
     float in
     )
 {
-    uint32_t fp32 = *((unsigned int *) &in);
-    uint32_t t1 = fp32 >> 16;
+    fp32_bit_cast_t fp32_bit_cast;
+    fp32_bit_cast.val = in;
+    uint32_t fp32_data = fp32_bit_cast.data;
+    uint32_t t1 = fp32_data >> 16;
 
     return (uint16_t) t1;
 } /* fp32_to_bfp16() */
@@ -435,10 +455,12 @@ static VSI_INLINE_API uint16_t fp32_to_bfp16_rtne
     Convert a float point to bfloat16, with round-nearest-to-even as rounding method.
     */
 
-    uint32_t fp32 = *((unsigned int *) &in);
+    fp32_bit_cast_t fp32_bit_cast;
+    fp32_bit_cast.val = in;
+    uint32_t fp32_data = fp32_bit_cast.data;
     uint16_t out;
 
-    uint32_t lsb = (fp32 >> 16) & 1;    /* Least significant bit of resulting bfloat. */
+    uint32_t lsb = (fp32_data >> 16) & 1;    /* Least significant bit of resulting bfloat. */
     uint32_t rounding_bias = 0x7fff + lsb;
 
     if ( VSI_NN_FLOAT32_NAN == in )
@@ -447,8 +469,8 @@ static VSI_INLINE_API uint16_t fp32_to_bfp16_rtne
     }
     else
     {
-        fp32 += rounding_bias;
-        out = (uint16_t) (fp32 >> 16);
+        fp32_data += rounding_bias;
+        out = (uint16_t) (fp32_data >> 16);
     }
 
     return out;
@@ -466,7 +488,9 @@ static VSI_INLINE_API uint16_t fp32_to_bfp16_rtne
 
 static VSI_INLINE_API uint8_t fp32_to_fp8_e4m3(float in, const float scale) {
     float fp8_f32 = in / scale;
-    int32_t in_val = *((int32_t*)&fp8_f32);
+    fp32_bit_cast_t fp32_bit_cast;
+    fp32_bit_cast.val = fp8_f32;
+    uint32_t in_val = fp32_bit_cast.data;
 
     uint32_t in_sign = (in_val >> (FLOAT_EXPONENT_SIZE + FLOAT_MANTISSA_SIZE)) & 0x1; /* bit 31 is sign */
     uint32_t in_exp = (in_val >> FLOAT_MANTISSA_SIZE) & 0xFF; /* bit[30: 24] is exp */
@@ -512,7 +536,9 @@ static VSI_INLINE_API uint8_t fp32_to_fp8_e4m3(float in, const float scale) {
 
 static VSI_INLINE_API uint8_t fp32_to_fp8_e5m2(float in, const float scale) {
     float fp8_f32 = in / scale;
-    int32_t in_val = *((int32_t*)&fp8_f32);
+    fp32_bit_cast_t fp32_bit_cast;
+    fp32_bit_cast.val = fp8_f32;
+    uint32_t in_val = fp32_bit_cast.data;
     uint32_t in_sign = (in_val >> (FLOAT_EXPONENT_SIZE + FLOAT_MANTISSA_SIZE)) & 0x1; /* bit 31 is sign */
     uint32_t in_exp = (in_val >> FLOAT_MANTISSA_SIZE) & 0xFF; /* bit[30: 24] is exp */
     uint32_t in_man = (in_val & 0x7FFFFF);   /* low 23 bits is man */
@@ -561,6 +587,7 @@ static VSI_INLINE_API float fp8_e4m3_to_fp32(uint8_t in, const float scale) {
     uint32_t exponentOut = 0;
     uint32_t mantissaOut = 0;
     uint32_t out_u = 0;
+    fp32_bit_cast_t fp32_bit_cast;
 
     {
         uint32_t signIn;
@@ -610,7 +637,8 @@ static VSI_INLINE_API float fp8_e4m3_to_fp32(uint8_t in, const float scale) {
     }
 final:
     out_u = signOut << 31 | exponentOut << 23 | mantissaOut;
-    val_fp32 = *((float*)&out_u);
+    fp32_bit_cast.data = out_u;
+    val_fp32 = fp32_bit_cast.val;
 
     return val_fp32 * scale;
 } /* fp8_e4m3_to_fp32() */
@@ -621,6 +649,7 @@ static VSI_INLINE_API float fp8_e5m2_to_fp32(int8_t in, const float scale) {
     uint32_t exponentOut = 0;
     uint32_t mantissaOut = 0;
     uint32_t out_u = 0;
+    fp32_bit_cast_t fp32_bit_cast;
 
     {
         uint32_t signIn;
@@ -670,7 +699,8 @@ static VSI_INLINE_API float fp8_e5m2_to_fp32(int8_t in, const float scale) {
     }
 final:
     out_u = signOut << 31 | exponentOut << 23 | mantissaOut;
-    val_fp32 = *((float*)&out_u);
+    fp32_bit_cast.data = out_u;
+    val_fp32 = fp32_bit_cast.val;
     return val_fp32 * scale;
 } /* fp8_e5m2_to_fp32() */
 

src/tim/vx/internal/include/vsi_nn_context.h

@@ -60,9 +60,7 @@ typedef struct _vsi_nn_hw_config_t
 {
     char target_name[VSI_NN_MAX_TARGET_NAME];
     vsi_nn_hw_evis_t evis;
-#if VX_HARDWARE_CAPS_PARAMS_EXT_SUPPORT
     uint32_t subGroupSize;
-#endif
     uint32_t use_40bits_va;
     uint32_t support_stream_processor;
     uint32_t sp_exec_count;

src/tim/vx/internal/include/vsi_nn_node_type.h

@@ -210,6 +210,8 @@
 #include "ops/vsi_nn_op_resize_3d.h"
 #include "ops/vsi_nn_op_reducel2.h"
 #include "ops/vsi_nn_op_crop_and_resize.h"
+#include "ops/vsi_nn_op_rmsnorm.h"
+#include "ops/vsi_nn_op_shape.h"
 /* custom node head define define */
 #include "custom/vsi_nn_custom_node_type.h"
 #include "ops/vsi_nn_op_inverse_sigmoid.h"
@@ -408,6 +410,8 @@ typedef union _vsi_nn_nn_param
     vsi_nn_resize_3d_param          resize_3d;
     vsi_nn_reducel2_param           reducel2;
     vsi_nn_crop_and_resize_param    crop_and_resize;
+    vsi_nn_rmsnorm_param            rmsnorm;
+    vsi_nn_shape_param              shape;
     void*                         client_param;
 
     /* custom node data struct define */

src/tim/vx/internal/include/vsi_nn_version.h

@@ -33,7 +33,7 @@ extern "C"{
 
 #define VSI_NN_VERSION_MAJOR 1
 #define VSI_NN_VERSION_MINOR 2
-#define VSI_NN_VERSION_PATCH 2
+#define VSI_NN_VERSION_PATCH 6
 #define VSI_NN_VERSION \
     (VSI_NN_VERSION_MAJOR * 10000 + VSI_NN_VERSION_MINOR * 100 + VSI_NN_VERSION_PATCH)
 

src/tim/vx/internal/src/custom/ops/kernel/evis/custom_tiny_yolov4_postprocess_confidence_evis.c

@@ -267,7 +267,7 @@ static vsi_nn_kernel_node_t _setup
     vsi_status status = VSI_FAILURE;
     vsi_nn_kernel_node_param_t node_params[_TINY_YOLOV4_POSTPROCESS_CONFIDENCE_PARAM_NUM];
     vsi_nn_kernel_node_t node = NULL;
-    vsi_size_t shape[2][VSI_NN_MAX_DIM_NUM] = { 0 };
+    vsi_size_t shape[2][VSI_NN_MAX_DIM_NUM] = { { 0 } };
     vsi_nn_tensor_t* reshape_tensors[2] = { NULL };
 
     VSI_UNREFERENCED(params);