!12117 [MSLITE] MAX_SHAPE_SIZE COMM_SHAPE_SIZE

From: @ling_qiao_min
Reviewed-by: 
Signed-off-by:

mindspore/lite/micro/coder/opcoders/base/conv2d_base_coder.cc

@@ -17,7 +17,7 @@
 #include "micro/coder/opcoders/base/conv2d_base_coder.h"
 #include <string>
 #include <vector>
-#include "nnacl/winograd_utils.h"
+#include "nnacl/fp32/winograd_utils.h"
 #include "nnacl/int8/quantize.h"
 #include "micro/coder/log.h"
 

mindspore/lite/nnacl/CMakeLists.txt

@@ -5,8 +5,10 @@ include_directories(NNACL_DIR)
 
 if(PLATFORM_ARM32 OR PLATFORM_ARM64)
     if("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing \
+        -ffunction-sections -fdata-sections -ffast-math")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing \
+        -ffunction-sections -fdata-sections -ffast-math")
     endif()
 endif()
 if("${X86_64_SIMD}" STREQUAL "avx")
@@ -37,13 +39,13 @@ if(PLATFORM_ARM32)
 endif()
 
 if("${X86_64_SIMD}" STREQUAL "sse")
-    file(GLOB ASSEMBLY_SRC ${NNACL_DIR}/x86_64_sse/*.c)
+    file(GLOB ASSEMBLY_SRC ${NNACL_DIR}/intrinsics/sse/*.c)
     set_property(SOURCE ${ASSEMBLY_SRC} PROPERTY LANGUAGE C)
 endif()
 
 if("${X86_64_SIMD}" STREQUAL "avx")
-    file(GLOB ASSEMBLY_SRC ${NNACL_DIR}/x86_64_sse/*.c
-        ${NNACL_DIR}/x86_64_avx/*.c
+    file(GLOB ASSEMBLY_SRC ${NNACL_DIR}/intrinsics/sse/*.c
+            ${NNACL_DIR}/intrinsics/avx/*.c
             ${NNACL_DIR}/assembly/avx/*.S)
     set_property(SOURCE ${ASSEMBLY_SRC} PROPERTY LANGUAGE C)
 endif()

mindspore/lite/nnacl/base/minimal_filtering_generator.c

@@ -13,10 +13,10 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "nnacl/minimal_filtering_generator.h"
+#include "nnacl/base/minimal_filtering_generator.h"
 #include <string.h>
 #include <math.h>
-#include "nnacl/winograd_utils.h"
+#include "nnacl/fp32/winograd_utils.h"
 #include "nnacl/errorcode.h"
 
 void Polynomial(const float *interval, float *m, int degree) {

mindspore/lite/nnacl/conv_parameter.h

@@ -72,6 +72,7 @@ typedef struct SlidingWindowParam {
   int kernel_step_;
 } SlidingWindowParam;
 
+#define OUPUT_UNIT 2
 #define DECONV_WINOGRAD_DEFAULT_UNIT 3
 #define DECONV_WINOGRAD_DEFAULT_TILE 8
 #define DECONV_WINOGRAD_BUFFER_COUNT 8

mindspore/lite/nnacl/fp16/deconv_winograd_fp16.c

@@ -15,7 +15,7 @@
  */
 
 #include "nnacl/fp16/deconv_winograd_fp16.h"
-#include "nnacl/minimal_filtering_generator.h"
+#include "nnacl/base/minimal_filtering_generator.h"
 
 void DeConvWgInputPackFp16(float16_t *src_ptr, float16_t *dst_ptr, int channel, int stride) {
   int ic4div = channel / C4NUM;
@@ -111,16 +111,16 @@ void DeConvWgMergeFp16(const float16_t *src, float16_t *dst, size_t src_stride,
 }
 
 void DeConvWgCalWgFp16(float16_t *tile_in, float16_t *tile_out, float16_t *weight_buf, float16_t *tmp_buf,
-                       float16_t *at_buf, float16_t *a_mid_buf, float16_t *trans_a_buf, bool *transfered,
+                       float16_t *at_buf, float16_t *a_mid_buf, float16_t *trans_a_buf, bool *transferred,
                        float16_t *bt_buf, float16_t *b_tmp_buf, int unit_size, int w_start, int h_start,
                        ConvParameter *conv_param, DeConvParam *deconv_param) {
   int winograd_plane = unit_size * unit_size;
-  if (!transfered[unit_size]) {
+  if (!transferred[unit_size]) {
     WinogradTransLeftFp16(tile_in, at_buf, a_mid_buf, DECONV_WINOGRAD_DEFAULT_UNIT, unit_size,
                           DECONV_WINOGRAD_DEFAULT_UNIT, deconv_param->ic_div4_ * DECONV_WINOGRAD_DEFAULT_TILE);
     WinogradTransRightFp16(a_mid_buf, at_buf, trans_a_buf, unit_size, unit_size, DECONV_WINOGRAD_DEFAULT_UNIT,
                            deconv_param->ic_div4_ * DECONV_WINOGRAD_DEFAULT_TILE);
-    transfered[unit_size] = true;
+    transferred[unit_size] = true;
   }
 
   for (int index = 0; index < winograd_plane; index++) {
@@ -311,7 +311,7 @@ void DeconvWgFp16(float16_t *nhwc_input_, float16_t *tile_in, float16_t *tile_ou
   }
 
   /* compute */
-  bool transfered[DECONV_WINOGRAD_BUFFER_COUNT] = {false};
+  bool transferred[DECONV_WINOGRAD_BUFFER_COUNT] = {false};
   for (int i = 0; i < deconv_param->compute_size_; i++) {
     DeConvComputeUnit *unit = &deconv_param->compute_units_[i];
     if (unit->use_winograd_) {
@@ -328,7 +328,7 @@ void DeconvWgFp16(float16_t *nhwc_input_, float16_t *tile_in, float16_t *tile_ou
                                                                     DECONV_WINOGRAD_DEFAULT_TILE *
                                                                     deconv_param->oc_up4_;
       DeConvWgCalWgFp16(tile_in, tile_out, (float16_t *)unit->weight_, tmp_buf, unit->winograd_.AT_, mid_a, dst_a,
-                        transfered, unit->winograd_.BT_, tmp_b, unit->winograd_.kh_, unit->w_start_, unit->h_start_,
+                        transferred, unit->winograd_.BT_, tmp_b, unit->winograd_.kh_, unit->w_start_, unit->h_start_,
                         conv_param, deconv_param);
     } else {
       float16_t *tmp_buf = (float16_t *)unit->tmp_buffer_ + task_id * deconv_param->oc_div4_ * unit->w_size_ *

mindspore/lite/nnacl/fp32/conv_depthwise_fp32.c

@@ -17,7 +17,7 @@
 #include "nnacl/fp32/conv_depthwise_fp32.h"
 #include "nnacl/common_func.h"
 #include "nnacl/fp32/common_func_fp32.h"
-#include "nnacl/winograd_transform.h"
+#include "nnacl/fp32/winograd_transform.h"
 #ifdef ENABLE_ARM64
 #include <arm_neon.h>
 #endif

mindspore/lite/nnacl/fp32/conv_winograd_fp32.c

@@ -17,7 +17,7 @@
 #include "nnacl/fp32/conv_winograd_fp32.h"
 #include <string.h>
 #include "nnacl/fp32/common_func_fp32.h"
-#include "nnacl/winograd_transform.h"
+#include "nnacl/fp32/winograd_transform.h"
 #include "nnacl/fp32/matmul_fp32.h"
 
 // fp32 conv winograd

mindspore/lite/nnacl/fp32/conv_winograd_fp32.h

@@ -24,7 +24,7 @@
 #include "nnacl/op_base.h"
 #include "nnacl/common_func.h"
 #include "nnacl/conv_parameter.h"
-#include "nnacl/winograd_utils.h"
+#include "nnacl/fp32/winograd_utils.h"
 #include "nnacl/fp32/conv_depthwise_fp32.h"
 
 typedef float *TmpBufferAddress;

mindspore/lite/nnacl/fp32/deconv_fp32.h

@@ -22,7 +22,7 @@
 #include "nnacl/conv_parameter.h"
 #include "nnacl/errorcode.h"
 #include "nnacl/fp32/common_func_fp32.h"
-#include "nnacl/minimal_filtering_generator.h"
+#include "nnacl/base/minimal_filtering_generator.h"
 
 #ifdef __cplusplus
 extern "C" {

mindspore/lite/nnacl/fp32/deconv_winograd_fp32.h

@@ -22,7 +22,7 @@
 #include "nnacl/conv_parameter.h"
 #include "nnacl/errorcode.h"
 #include "nnacl/fp32/common_func_fp32.h"
-#include "nnacl/minimal_filtering_generator.h"
+#include "nnacl/base/minimal_filtering_generator.h"
 
 #ifdef __cplusplus
 extern "C" {

mindspore/lite/nnacl/fp32/winograd_transform.c

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "nnacl/winograd_transform.h"
+#include "nnacl/fp32/winograd_transform.h"
 #include "nnacl/op_base.h"
 
 // fp32 conv winograd

mindspore/lite/nnacl/fp32/winograd_transform.h

@@ -22,10 +22,7 @@
 #endif
 #include <string.h>
 #include "nnacl/pack.h"
-#include "nnacl/winograd_utils.h"
-#include "mindspore/lite/nnacl/int8/fixed_point.h"
-
-#define OUPUT_UNIT 2
+#include "nnacl/fp32/winograd_utils.h"
 
 #ifdef __cplusplus
 extern "C" {

mindspore/lite/nnacl/int8/conv3x3_int8.h

@@ -24,11 +24,10 @@
 #include "nnacl/op_base.h"
 #include "nnacl/common_func.h"
 #include "nnacl/conv_parameter.h"
-#include "nnacl/winograd_utils.h"
+#include "nnacl/int8/fixed_point.h"
 #include "nnacl/int8/quantize.h"
 #include "nnacl/matmul_parameter.h"
 #include "nnacl/int8/matmul_int8.h"
-#include "nnacl/winograd_transform.h"
 #include "nnacl/int8/common_func_int8.h"
 
 #ifdef __cplusplus

mindspore/lite/nnacl/int8/conv_int8.h

@@ -24,7 +24,6 @@
 #include "nnacl/op_base.h"
 #include "nnacl/common_func.h"
 #include "nnacl/conv_parameter.h"
-#include "nnacl/winograd_utils.h"
 #include "nnacl/int8/quantize.h"
 #include "nnacl/matmul_parameter.h"
 #include "nnacl/int8/matmul_int8.h"

mindspore/lite/nnacl/intrinsics/avx/common_utils.c

@@ -13,7 +13,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "nnacl/x86_64_avx/common_utils.h"
+#include "nnacl/intrinsics/avx/common_utils.h"
 #ifdef WIN32
 #ifdef ENABLE_AVX
 #include <stdint.h>

mindspore/lite/nnacl/intrinsics/sse/DepthwiseFp32_Sse.c

@@ -17,6 +17,7 @@
 #ifdef ENABLE_SSE
 #include <x86intrin.h>
 #include "nnacl/fp32/conv_depthwise_fp32.h"
+#include "nnacl/intrinsics/sse/sse_common.h"
 
 void ConvDwFp32Border(float *dst, const float *src, const float *weight, const float *bias, size_t height, size_t width,
                       size_t in_kh_step, size_t in_kw_step, size_t kernel_w_step, size_t relu, size_t relu6) {
@@ -123,18 +124,16 @@ void ConvDwFp32Center(float *dst, const float *src, const float *weight, const f
     int c2 = DOWN_DIV(width, C2NUM) * C2NUM;
     int c1 = 0;
     // c4 loop
-    for (; c1 < c4; c1 += C4NUM) {
-      const float *src_kh = src_w;
-      const float *weight_kh = weight;
+    for (; c1 < c4; c1 += C4NUM, dst_w += C4NUM * block_channel, src_w += C4NUM * in_sw_step) {
+      const float *src_kh = src_w, *weight_kh = weight;
       __m128 dst_w_ma1 = _mm_setzero_ps();
       __m128 dst_w_ma2 = _mm_setzero_ps();
       __m128 dst_w_ma3 = _mm_setzero_ps();
       __m128 dst_w_ma4 = _mm_setzero_ps();
 
-      for (int kh = 0; kh < kernel_h; kh++) {
-        const float *src_kw = src_kh;
-        const float *weight_kw = weight_kh;
-        for (int kw = 0; kw < kernel_w; kw++) {
+      for (int kh = 0; kh < kernel_h; kh++, src_kh += in_kh_step, weight_kh += kernel_w * C4NUM) {
+        const float *src_kw = src_kh, *weight_kw = weight_kh;
+        for (int kw = 0; kw < kernel_w; kw++, src_kw += in_kw_step, weight_kw += C4NUM) {
           __m128 src_kw_ma1 = _mm_loadu_ps(src_kw);
           __m128 weight_kw_ma1 = _mm_loadu_ps(weight_kw);
           __m128 tmp_ma1 = _mm_mul_ps(src_kw_ma1, weight_kw_ma1);
@@ -154,13 +153,9 @@ void ConvDwFp32Center(float *dst, const float *src, const float *weight, const f
           __m128 weight_kw_ma4 = _mm_loadu_ps(weight_kw);
           __m128 tmp_ma4 = _mm_mul_ps(src_kw_ma4, weight_kw_ma4);
           dst_w_ma4 = _mm_add_ps(dst_w_ma4, tmp_ma4);
-
-          src_kw += in_kw_step;
-          weight_kw += C4NUM;
         }  // kernel_w loop
-        src_kh += in_kh_step;
-        weight_kh += kernel_w * C4NUM;
       }    // kernel_h loop
+
       // add bias relu
       __m128 bias_ma = _mm_loadu_ps(bias);
       dst_w_ma1 = _mm_add_ps(dst_w_ma1, bias_ma);
@@ -168,39 +163,23 @@ void ConvDwFp32Center(float *dst, const float *src, const float *weight, const f
       dst_w_ma3 = _mm_add_ps(dst_w_ma3, bias_ma);
       dst_w_ma4 = _mm_add_ps(dst_w_ma4, bias_ma);
 
-      __m128 zero_ma = _mm_setzero_ps();
-      if (relu || relu6) {
-        dst_w_ma1 = _mm_max_ps(zero_ma, dst_w_ma1);
-        dst_w_ma2 = _mm_max_ps(zero_ma, dst_w_ma2);
-        dst_w_ma3 = _mm_max_ps(zero_ma, dst_w_ma3);
-        dst_w_ma4 = _mm_max_ps(zero_ma, dst_w_ma4);
-        if (relu6) {
-          __m128 const_ma = _mm_set_ps(6.0f, 6.0f, 6.0f, 6.0f);
-          dst_w_ma1 = _mm_min_ps(const_ma, dst_w_ma1);
-          dst_w_ma2 = _mm_min_ps(const_ma, dst_w_ma2);
-          dst_w_ma3 = _mm_min_ps(const_ma, dst_w_ma3);
-          dst_w_ma4 = _mm_min_ps(const_ma, dst_w_ma4);
-        }
-      }
+      ActBlock4(&dst_w_ma1, &dst_w_ma2, &dst_w_ma3, &dst_w_ma4, relu, relu6);
+
       _mm_storeu_ps(dst_w, dst_w_ma1);
       _mm_storeu_ps(dst_w + block_channel, dst_w_ma2);
       _mm_storeu_ps(dst_w + 2 * block_channel, dst_w_ma3);
       _mm_storeu_ps(dst_w + 3 * block_channel, dst_w_ma4);
-
-      dst_w += C4NUM * block_channel;
-      src_w += C4NUM * in_sw_step;
     }  // dst_width loop
+
     // c2 loop
-    for (; c1 < c2; c1 += C2NUM) {
-      const float *src_kh = src_w;
-      const float *weight_kh = weight;
+    for (; c1 < c2; c1 += C2NUM, dst_w += C2NUM * block_channel, src_w += C2NUM * in_sw_step) {
+      const float *src_kh = src_w, *weight_kh = weight;
       __m128 dst_w_ma1 = _mm_setzero_ps();
       __m128 dst_w_ma2 = _mm_setzero_ps();
 
-      for (int kh = 0; kh < kernel_h; kh++) {
-        const float *src_kw = src_kh;
-        const float *weight_kw = weight_kh;
-        for (int kw = 0; kw < kernel_w; kw++) {
+      for (int kh = 0; kh < kernel_h; kh++, src_kh += in_kh_step, weight_kh += kernel_w * C4NUM) {
+        const float *src_kw = src_kh, *weight_kw = weight_kh;
+        for (int kw = 0; kw < kernel_w; kw++, src_kw += in_kw_step, weight_kw += C4NUM) {
           __m128 src_kw_ma1 = _mm_loadu_ps(src_kw);
           __m128 weight_kw_ma1 = _mm_loadu_ps(weight_kw);
           __m128 tmp_ma1 = _mm_mul_ps(src_kw_ma1, weight_kw_ma1);
@@ -210,68 +189,38 @@ void ConvDwFp32Center(float *dst, const float *src, const float *weight, const f
           __m128 weight_kw_ma2 = _mm_loadu_ps(weight_kw);
           __m128 tmp_ma2 = _mm_mul_ps(src_kw_ma2, weight_kw_ma2);
           dst_w_ma2 = _mm_add_ps(dst_w_ma2, tmp_ma2);
-
-          src_kw += in_kw_step;
-          weight_kw += C4NUM;
         }  // kernel_w loop
-        src_kh += in_kh_step;
-        weight_kh += kernel_w * C4NUM;
       }    // kernel_h loop
       // add bias relu
       __m128 bias_ma = _mm_loadu_ps(bias);
       dst_w_ma1 = _mm_add_ps(dst_w_ma1, bias_ma);
       dst_w_ma2 = _mm_add_ps(dst_w_ma2, bias_ma);
-      __m128 zero_ma = _mm_setzero_ps();
-      if (relu || relu6) {
-        dst_w_ma1 = _mm_max_ps(zero_ma, dst_w_ma1);
-        dst_w_ma2 = _mm_max_ps(zero_ma, dst_w_ma2);
-        if (relu6) {
-          __m128 const_ma = _mm_set_ps(6.0f, 6.0f, 6.0f, 6.0f);
-          dst_w_ma1 = _mm_min_ps(const_ma, dst_w_ma1);
-          dst_w_ma2 = _mm_min_ps(const_ma, dst_w_ma2);
-        }
-      }
+
+      ActBlock2(&dst_w_ma1, &dst_w_ma2, relu, relu6);
+
       _mm_storeu_ps(dst_w, dst_w_ma1);
       _mm_storeu_ps(dst_w + block_channel, dst_w_ma2);
-
-      dst_w += C2NUM * block_channel;
-      src_w += C2NUM * in_sw_step;
     }
+
     // remaining
-    for (; c1 < width; c1++) {
-      const float *src_kh = src_w;
-      const float *weight_kh = weight;
+    for (; c1 < width; c1++, dst_w += block_channel, src_w += in_sw_step) {
+      const float *src_kh = src_w, *weight_kh = weight;
       __m128 dst_w_ma1 = _mm_setzero_ps();
-      for (int kh = 0; kh < kernel_h; kh++) {
-        const float *src_kw = src_kh;
-        const float *weight_kw = weight_kh;
-        for (int kw = 0; kw < kernel_w; kw++) {
+      for (int kh = 0; kh < kernel_h; kh++, src_kh += in_kh_step, weight_kh += kernel_w * C4NUM) {
+        const float *src_kw = src_kh, *weight_kw = weight_kh;
+        for (int kw = 0; kw < kernel_w; kw++, src_kw += in_kw_step, weight_kw += C4NUM) {
           __m128 src_kw_ma1 = _mm_loadu_ps(src_kw);
           __m128 weight_kw_ma1 = _mm_loadu_ps(weight_kw);
           __m128 tmp_ma1 = _mm_mul_ps(src_kw_ma1, weight_kw_ma1);
           dst_w_ma1 = _mm_add_ps(dst_w_ma1, tmp_ma1);
-
-          src_kw += in_kw_step;
-          weight_kw += C4NUM;
         }  // kernel_w loop
-        src_kh += in_kh_step;
-        weight_kh += kernel_w * C4NUM;
       }    // kernel_h loop
+
       // add bias relu
       __m128 bias_ma = _mm_loadu_ps(bias);
       dst_w_ma1 = _mm_add_ps(dst_w_ma1, bias_ma);
-      __m128 zero_ma = _mm_setzero_ps();
-      if (relu || relu6) {
-        dst_w_ma1 = _mm_max_ps(zero_ma, dst_w_ma1);
-        if (relu6) {
-          __m128 const_ma = _mm_set_ps(6.0f, 6.0f, 6.0f, 6.0f);
-          dst_w_ma1 = _mm_min_ps(const_ma, dst_w_ma1);
-        }
-      }
+      ActBlock1(&dst_w_ma1, relu, relu6);
       _mm_storeu_ps(dst_w, dst_w_ma1);
-
-      dst_w += block_channel;
-      src_w += in_sw_step;
     }
     dst_h += out_h_step;
     src_h += in_sh_step;