split conv_fp32.c

mindspore/lite/nnacl/fp32/conv_common_fp32.c

@@ -0,0 +1,63 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "nnacl/fp32/conv_common_fp32.h"
+#include <string.h>
+#include "nnacl/fp32/common_func_fp32.h"
+#include "nnacl/fp32/matmul_fp32.h"
+
+// fp32 conv common
+void ConvFp32(const float *input_data, float *packed_input, const float *packed_weight, const float *bias_data,
+              float *col_major_input, float *output_data, int task_id, const ConvParameter *conv_param) {
+  int out_channel = conv_param->output_channel_;
+  int deep = conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
+  int output_count = conv_param->output_h_ * conv_param->output_w_;
+#ifdef ENABLE_AVX
+  const int cal_num = C6NUM;
+#elif defined(ENABLE_SSE)
+  const int cal_num = C4NUM;
+#else
+  const int cal_num = C12NUM;
+#endif
+  int output_tile_count = UP_DIV(output_count, cal_num);
+
+  for (int b = 0; b < conv_param->input_batch_; b++) {
+    int in_batch_offset = b * conv_param->input_channel_ * conv_param->input_h_ * conv_param->input_w_;
+    int out_batch_offset = b * out_channel * output_count;
+    for (int thread_id = task_id; thread_id < output_tile_count; thread_id += conv_param->thread_num_) {
+      int start_index = thread_id * cal_num;
+      int real_cal_num = (output_count - start_index) < cal_num ? (output_count - start_index) : cal_num;
+      float *gemm_input = packed_input + task_id * deep * cal_num;
+      float *col_major_gemm_input = col_major_input + task_id * deep * cal_num;
+      size_t packed_input_size = deep * cal_num * sizeof(float);
+      memset(gemm_input, 0, packed_input_size);
+      memset(col_major_gemm_input, 0, packed_input_size);
+      Im2ColPackUnitFp32(input_data + in_batch_offset, conv_param, gemm_input, real_cal_num, start_index);
+
+      int out_offset = thread_id * cal_num * out_channel + out_batch_offset;
+      float *gemm_output = output_data + out_offset;
+#ifdef ENABLE_AVX
+      RowMajor2Col6Major(gemm_input, col_major_gemm_input, cal_num, deep);
+#elif defined(ENABLE_SSE)
+      RowMajor2Col4Major(gemm_input, col_major_gemm_input, cal_num, deep);
+#else
+      RowMajor2Col12Major(gemm_input, col_major_gemm_input, cal_num, deep);
+#endif
+      MatMulOpt(col_major_gemm_input, packed_weight, gemm_output, bias_data, conv_param->act_type_, deep, real_cal_num,
+                out_channel, out_channel, OutType_Nhwc);
+    }
+  }
+}

mindspore/lite/nnacl/fp32/conv_common_fp32.h

@@ -0,0 +1,40 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_NNACL_FP32_CONV_COMMON_H_
+#define MINDSPORE_LITE_NNACL_FP32_CONV_COMMON_H_
+
+#ifdef ENABLE_NEON
+#include <arm_neon.h>
+#endif
+#include "nnacl/pack.h"
+#include "nnacl/op_base.h"
+#include "nnacl/common_func.h"
+#include "nnacl/conv_parameter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// fp32 convolution common (im2col+gemm)
+void ConvFp32(const float *input_data, float *packed_input, const float *packed_weight, const float *bias_data,
+              float *col_major_input, float *output_data, int task_id, const ConvParameter *conv_param);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // MINDSPORE_LITE_NNACL_FP32_CONV_COMMON_H_

mindspore/lite/nnacl/fp32/conv_depthwise_fp32.c

@@ -15,6 +15,7 @@
  */
 
 #include "nnacl/fp32/conv_depthwise_fp32.h"
+#include "nnacl/common_func.h"
 #include "nnacl/fp32/common_func_fp32.h"
 #include "nnacl/winograd_transform.h"
 #ifdef ENABLE_ARM64

mindspore/lite/nnacl/fp32/conv_fp32.c → mindspore/lite/nnacl/fp32/conv_winograd_fp32.c

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2021 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -14,55 +14,12 @@
  * limitations under the License.
  */
 
-#include "nnacl/fp32/conv_fp32.h"
+#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/matmul_fp32.h"
 
-// fp32 conv common
-void ConvFp32(const float *input_data, float *packed_input, const float *packed_weight, const float *bias_data,
-              float *col_major_input, float *output_data, int task_id, const ConvParameter *conv_param) {
-  int out_channel = conv_param->output_channel_;
-  int deep = conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
-  int output_count = conv_param->output_h_ * conv_param->output_w_;
-#ifdef ENABLE_AVX
-  const int cal_num = C6NUM;
-#elif defined(ENABLE_SSE)
-  const int cal_num = C4NUM;
-#else
-  const int cal_num = C12NUM;
-#endif
-  int output_tile_count = UP_DIV(output_count, cal_num);
-
-  for (int b = 0; b < conv_param->input_batch_; b++) {
-    int in_batch_offset = b * conv_param->input_channel_ * conv_param->input_h_ * conv_param->input_w_;
-    int out_batch_offset = b * out_channel * output_count;
-    for (int thread_id = task_id; thread_id < output_tile_count; thread_id += conv_param->thread_num_) {
-      int start_index = thread_id * cal_num;
-      int real_cal_num = (output_count - start_index) < cal_num ? (output_count - start_index) : cal_num;
-      float *gemm_input = packed_input + task_id * deep * cal_num;
-      float *col_major_gemm_input = col_major_input + task_id * deep * cal_num;
-      size_t packed_input_size = deep * cal_num * sizeof(float);
-      memset(gemm_input, 0, packed_input_size);
-      memset(col_major_gemm_input, 0, packed_input_size);
-      Im2ColPackUnitFp32(input_data + in_batch_offset, conv_param, gemm_input, real_cal_num, start_index);
-
-      int out_offset = thread_id * cal_num * out_channel + out_batch_offset;
-      float *gemm_output = output_data + out_offset;
-#ifdef ENABLE_AVX
-      RowMajor2Col6Major(gemm_input, col_major_gemm_input, cal_num, deep);
-#elif defined(ENABLE_SSE)
-      RowMajor2Col4Major(gemm_input, col_major_gemm_input, cal_num, deep);
-#else
-      RowMajor2Col12Major(gemm_input, col_major_gemm_input, cal_num, deep);
-#endif
-      MatMulOpt(col_major_gemm_input, packed_weight, gemm_output, bias_data, conv_param->act_type_, deep, real_cal_num,
-                out_channel, out_channel, OutType_Nhwc);
-    }
-  }
-}
-
 // fp32 conv winograd
 void ConvWinogardFp32(const float *input_data, const float *trans_weight, const float *bias_data, float *output_data,
                       TmpBufferAddress *buffer_list, int task_id, const ConvParameter *conv_param,

mindspore/lite/nnacl/fp32/conv_fp32.h → mindspore/lite/nnacl/fp32/conv_winograd_fp32.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2020 Huawei Technologies Co., Ltd
+ * Copyright 2021 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_NNACL_FP32_CONV_H_
-#define MINDSPORE_LITE_NNACL_FP32_CONV_H_
+#ifndef MINDSPORE_LITE_NNACL_FP32_CONV_WINOGRAD_H_
+#define MINDSPORE_LITE_NNACL_FP32_CONV_WINOGRAD_H_
 
 #ifdef ENABLE_NEON
 #include <arm_neon.h>
@@ -33,10 +33,6 @@ typedef float *TmpBufferAddress;
 extern "C" {
 #endif
 
-// fp32 convolution common (im2col+gemm)
-void ConvFp32(const float *input_data, float *packed_input, const float *packed_weight, const float *bias_data,
-              float *col_major_input, float *output_data, int task_id, const ConvParameter *conv_param);
-
 // fp32 convolution winograd
 void ConvWinogardFp32(const float *input_data, const float *trans_weight, const float *bias_data, float *output_data,
                       TmpBufferAddress *buffer_list, int task_id, const ConvParameter *conv_param,
@@ -45,4 +41,4 @@ void ConvWinogardFp32(const float *input_data, const float *trans_weight, const
 }
 #endif
 
-#endif  // MINDSPORE_LITE_NNACL_FP32_CONV_H_
+#endif  // MINDSPORE_LITE_NNACL_FP32_CONV_WINOGRAD_H_

mindspore/lite/nnacl/fp32/deconv_fp32.h

@@ -22,7 +22,6 @@
 #include "nnacl/conv_parameter.h"
 #include "nnacl/errorcode.h"
 #include "nnacl/fp32/common_func_fp32.h"
-#include "nnacl/fp32/conv_fp32.h"
 #include "nnacl/minimal_filtering_generator.h"
 
 #ifdef __cplusplus

mindspore/lite/nnacl/winograd_transform.h

@@ -22,7 +22,6 @@
 #endif
 #include <string.h>
 #include "nnacl/pack.h"
-#include "nnacl/fp32/conv_fp32.h"
 #include "nnacl/winograd_utils.h"
 #include "mindspore/lite/nnacl/int8/fixed_point.h"
 

mindspore/lite/src/runtime/kernel/arm/fp32/adder_fp32.h

@@ -21,7 +21,6 @@
 #include "src/lite_kernel.h"
 #include "nnacl/op_base.h"
 #include "src/runtime/kernel/arm/fp32/convolution_fp32.h"
-#include "nnacl/fp32/conv_fp32.h"
 
 namespace mindspore::kernel {
 class AdderCPUKernel : public ConvolutionCPUKernel {

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_fp32.cc

@@ -19,7 +19,7 @@
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "src/runtime/runtime_api.h"
-#include "nnacl/fp32/conv_fp32.h"
+#include "nnacl/fp32/conv_common_fp32.h"
 #include "nnacl/fp32/matmul_fp32.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_fp32.h

@@ -21,7 +21,6 @@
 #include "src/lite_kernel.h"
 #include "nnacl/op_base.h"
 #include "src/runtime/kernel/arm/base/convolution_base.h"
-#include "nnacl/fp32/conv_fp32.h"
 
 namespace mindspore::kernel {
 class ConvolutionCPUKernel : public ConvolutionBaseCPUKernel {

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd_fp32.cc

@@ -15,7 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/fp32/convolution_winograd_fp32.h"
-#include "nnacl/fp32/conv_fp32.h"
+#include "nnacl/fp32/conv_winograd_fp32.h"
 #include "nnacl/pack.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
@@ -85,7 +85,7 @@ int ConvolutionWinogradCPUKernel::InitWeightBias() {
   }
   ret = WinogradFilterTransform(origin_weight_, matrix_g, matrix_gt, oc_block);
   if (ret != RET_OK) {
-    MS_LOG(ERROR) << "winograd filter transfrom failed.";
+    MS_LOG(ERROR) << "winograd filter transform failed.";
     return ret;
   }
 

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd_fp32.h

@@ -21,6 +21,7 @@
 #include "src/lite_kernel.h"
 #include "nnacl/winograd_transform.h"
 #include "nnacl/minimal_filtering_generator.h"
+#include "nnacl/fp32/conv_winograd_fp32.h"
 #include "src/runtime/kernel/arm/base/convolution_base.h"
 
 namespace mindspore::kernel {

mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.h

@@ -22,7 +22,7 @@
 #include "src/lite_kernel.h"
 #include "nnacl/op_base.h"
 #include "src/runtime/kernel/arm/base/convolution_base.h"
-#include "nnacl/fp32/conv_fp32.h"
+#include "nnacl/fp32/conv_common_fp32.h"
 
 namespace mindspore::kernel {
 class GroupConvolutionCPUKernel : public ConvolutionBaseCPUKernel {