!5265 [MS][LITE] arm cpu fp16 op: optimize conv depthwise

Merge pull request !5265 from yangruoqi713/lite_fp16

mindspore/lite/nnacl/assembly/opt/ConvDwFp16Row.S

@@ -0,0 +1,117 @@
+#ifdef __aarch64__
+
+.text
+.align 5
+.global ConvDwFp16Row
+#ifndef __APPLE__
+.type ConvDwFp16Row, %function
+#endif
+
+// void ConvDwFp16Row(float16_t* output_ptr, const float16_t* input_ptr,const float16_t* filter_ptr,
+//                    size_t num_pixels, size_t input_channel, size_t input_step)
+// x0: output_ptr, x1: input_ptr, x2: filter_ptr, x3: num_pixels,
+// x4: input_channel, x5: input_step
+//
+ConvDwFp16Row:
+    // registers v8 ~ v15 must be preserved by a callee across subroutine calls, according to
+    // https://github.com/ARM-software/abi-aa/blob/master/aapcs64/aapcs64.rst#simd-and-floating-point-registers
+    // x19 ~ x29 should be also preserved
+    // whereas our coding style do not permit such amount of parameters
+cmp x3, #0
+beq End
+
+mov x9, x0
+mov x12, #2 // sizeof(float16_t)
+mul x5, x5, x12
+
+LoopOutPixel:
+mov x6, x1
+mov x7, x2
+mov x8, x4
+
+LoopInputDepth32In:
+    cmp x8, #32
+    blt Loop8
+    sub x8, x8, #32
+
+    ld1 {v0.8h, v1.8h}, [x6], #32
+    ld1 {v2.8h, v3.8h}, [x7], #32
+    ld1 {v16.8h, v17.8h}, [x0], #32
+
+    cmp x8, #32
+    blt LoopInputDepth32Out
+    LoopInputDepth32:
+    fmla v16.8h, v0.8h, v2.8h
+    fmla v17.8h, v1.8h, v3.8h
+
+    st1 {v16.8h, v17.8h}, [x9], #32
+
+    ld1 {v4.8h, v5.8h}, [x6], #32
+    ld1 {v6.8h, v7.8h}, [x7], #32
+    ld1 {v18.8h, v19.8h}, [x0], #32
+
+    fmla v18.8h, v4.8h, v6.8h
+    fmla v19.8h, v5.8h, v7.8h
+
+    st1 {v18.8h, v19.8h}, [x9], #32
+
+    ld1 {v0.8h, v1.8h}, [x6], #32
+    ld1 {v2.8h, v3.8h}, [x7], #32
+    ld1 {v16.8h, v17.8h}, [x0], #32
+
+    sub x8, x8, #32
+    cmp x8, #32
+    bge LoopInputDepth32
+
+    LoopInputDepth32Out:
+    fmla v16.8h, v0.8h, v2.8h
+    fmla v17.8h, v1.8h, v3.8h
+    st1 {v16.8h, v17.8h}, [x9], #32
+
+    ld1 {v4.8h, v5.8h}, [x6], #32
+    ld1 {v6.8h, v7.8h}, [x7], #32
+    ld1 {v18.8h, v19.8h}, [x0], #32
+
+    fmla v18.8h, v4.8h, v6.8h
+    fmla v19.8h, v5.8h, v7.8h
+
+    st1 {v18.8h, v19.8h}, [x9], #32
+
+    Loop8:
+    cmp x8, #8
+    blt L0
+
+    LoopInputDepth8:
+    ld1 {v0.8h}, [x6], #16
+    ld1 {v2.8h}, [x7], #16
+    ld1 {v16.8h}, [x0], #16
+    fmla v16.8h, v0.8h, v2.8h
+    st1 {v16.8h}, [x9], #16
+    sub x8, x8, #8
+    cmp x8, #8
+    bge LoopInputDepth8
+
+    L0:
+    cmp x8, #0
+    beq Loop8LineEnd
+
+    LoopInputDepth0:
+    ldr h0, [x6], #2
+    ldr h1, [x7], #2
+    ldr h2, [x0], #2
+    fmul h0, h0, h1
+    fadd h2, h2, h0
+    str h2, [x9], #2
+    subs x8, x8, #1
+    bne LoopInputDepth0
+
+    Loop8LineEnd:
+
+subs x3, x3, #1
+add x1, x1, x5
+bne LoopOutPixel
+
+End:
+ret
+
+#endif

mindspore/lite/nnacl/fp16/common_func.h

@@ -1,49 +0,0 @@
-/**
- * Copyright 2020 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_FP16_COMMON_FUNC_H_
-#define MINDSPORE_LITE_NNACL_FP16_COMMON_FUNC_H_
-
-#include <stdint.h>
-#include <stdio.h>
-#include <string.h>
-#include "nnacl/op_base.h"
-#include "nnacl/conv_parameter.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifdef ENABLE_ARM64
-void ConvDwFp16Border(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias,
-                      size_t height, size_t width, size_t in_kh_step, size_t in_kw_step, size_t kernel_w, size_t relu,
-                      size_t relu6);
-void ConvDwFp16Center(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias,
-                      size_t height, size_t width, size_t kernel_h, size_t kernel_w, size_t out_h_step,
-                      size_t block_channel, size_t in_sh_step, size_t in_sw_step, size_t in_kh_step, size_t in_kw_step,
-                      size_t relu, size_t relu6);
-void DeconvDwFp16Border(float16_t *dst, const float16_t *src, const float16_t *weight, size_t height, size_t width,
-                        size_t in_kh_step, size_t in_kw_step, size_t kernel_w);
-void DeconvDwFp16Center(float16_t *dst, const float16_t *src, const float16_t *weight, size_t height, size_t width,
-                        size_t kernel_h, size_t kernel_w, size_t out_h_step, size_t block_channel, size_t in_sh_step,
-                        size_t in_sw_step, size_t in_kh_step, size_t in_kw_step);
-#endif
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /* MINDSPORE_LITE_NNACL_FP32_COMMON_FUNC_H_ */

mindspore/lite/nnacl/fp16/conv_depthwise_fp16.c

@@ -15,8 +15,62 @@
  */
 
 #include "nnacl/fp16/conv_depthwise_fp16.h"
-#include <arm_neon.h>
-#include "nnacl/fp16/common_func.h"
+#include <string.h>
+#include "nnacl/fp16/activation_fp16.h"
+
+void ConvDwFp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
+                const float16_t *bias_data, const ConvParameter *conv_param, int task_id) {
+  int h_step = UP_DIV(conv_param->output_h_, conv_param->thread_num_);
+  int h_start = h_step * task_id;
+  int h_end = MSMIN(h_start + h_step, conv_param->output_h_);
+  bool relu = conv_param->act_type_ == ActType_Relu;
+  bool relu6 = conv_param->act_type_ == ActType_Relu6;
+  for (int b = 0; b < conv_param->output_batch_; b++) {
+    const float16_t *src = input_data + b * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
+    float16_t *dst = output_data + b * conv_param->output_h_ * conv_param->output_w_ * conv_param->output_channel_;
+    for (int oh = h_start; oh < h_end; oh++) {
+      float16_t *dst_data = dst + oh * conv_param->output_w_ * conv_param->output_channel_;
+
+      int ih_origin = oh * conv_param->stride_h_ - conv_param->pad_u_;
+      int start_kh = MSMAX(0, UP_DIV(-ih_origin, conv_param->dilation_h_));
+      int end_kh = MSMIN(conv_param->kernel_h_, UP_DIV(conv_param->input_h_ - ih_origin, conv_param->dilation_h_));
+
+      for (int ow = 0; ow < conv_param->output_w_; ow++) {
+        memcpy(dst_data + ow * conv_param->output_channel_, bias_data, conv_param->output_channel_ * sizeof(float16_t));
+      }
+      for (int kh = start_kh; kh < end_kh; kh++) {
+        int ih = ih_origin + conv_param->dilation_w_ * kh;
+
+        const float16_t *src_kh = src + ih * conv_param->input_w_ * conv_param->input_channel_;
+        const float16_t *weight_kh = weight_data + kh * conv_param->kernel_w_ * conv_param->output_channel_;
+
+        int in_sw_step = conv_param->stride_w_ * conv_param->input_channel_;
+        for (int kw = 0; kw < conv_param->kernel_w_; kw++) {
+          int out_w_start = MSMAX(
+            0, (conv_param->pad_l_ - conv_param->dilation_w_ * kw + conv_param->stride_w_ - 1) / conv_param->stride_w_);
+          int out_w_end = MSMIN(conv_param->output_w_, (conv_param->input_w_ + conv_param->pad_l_ -
+                                                        conv_param->dilation_w_ * kw + conv_param->stride_w_ - 1) /
+                                                         conv_param->stride_w_);
+
+          float16_t *dst_w = dst_data + out_w_start * conv_param->output_channel_;
+          int iw_origin = (out_w_start * conv_param->stride_w_) - conv_param->pad_l_ + conv_param->dilation_w_ * kw;
+
+          const float16_t *src_kw = src_kh + iw_origin * conv_param->input_channel_;
+          int num_pixels = out_w_end - out_w_start;
+
+          ConvDwFp16Row(dst_w, src_kw, weight_kh, num_pixels, conv_param->output_channel_, in_sw_step);
+          weight_kh += conv_param->output_channel_;
+        }
+      }
+      if (relu) {
+        ReluFp16(dst_data, dst_data, conv_param->output_w_ * conv_param->output_channel_);
+      }
+      if (relu6) {
+        Relu6Fp16(dst_data, dst_data, conv_param->output_w_ * conv_param->output_channel_);
+      }
+    }
+  }
+}
 
 /*conv depthwise fp16 begin*/
 void DepthwiseBorderPixelFp16(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias,

mindspore/lite/nnacl/fp16/conv_depthwise_fp16.h

@@ -23,6 +23,26 @@
 #ifdef __cplusplus
 extern "C" {
 #endif
+#ifdef ENABLE_ARM64
+void ConvDwFp16Row(float16_t *output_ptr, const float16_t *input_ptr, const float16_t *filter_ptr, size_t num_pixels,
+                   size_t input_channel, size_t input_step);
+void ConvDwFp16Border(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias,
+                      size_t height, size_t width, size_t in_kh_step, size_t in_kw_step, size_t kernel_w, size_t relu,
+                      size_t relu6);
+void ConvDwFp16Center(float16_t *dst, const float16_t *src, const float16_t *weight, const float16_t *bias,
+                      size_t height, size_t width, size_t kernel_h, size_t kernel_w, size_t out_h_step,
+                      size_t block_channel, size_t in_sh_step, size_t in_sw_step, size_t in_kh_step, size_t in_kw_step,
+                      size_t relu, size_t relu6);
+void DeconvDwFp16Border(float16_t *dst, const float16_t *src, const float16_t *weight, size_t height, size_t width,
+                        size_t in_kh_step, size_t in_kw_step, size_t kernel_w);
+void DeconvDwFp16Center(float16_t *dst, const float16_t *src, const float16_t *weight, size_t height, size_t width,
+                        size_t kernel_h, size_t kernel_w, size_t out_h_step, size_t block_channel, size_t in_sh_step,
+                        size_t in_sw_step, size_t in_kh_step, size_t in_kw_step);
+#endif
+
+void ConvDwFp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
+                const float16_t *bias_data, const ConvParameter *conv_param, int task_id);
+
 void ConvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
                   const float16_t *bias_data, const ConvParameter *conv_param, const SlidingWindowParam *sliding,
                   int task_id);

mindspore/lite/nnacl/fp16/pack_fp16.c

@@ -220,6 +220,19 @@ void PackNCHWToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int
   }
 }
 
+void PackNCHWToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel) {
+  for (int n = 0; n < batch; n++) {
+    for (int c = 0; c < channel; c++) {
+      for (int hw = 0; hw < plane; hw++) {
+        int nhwc_index = n * channel * plane + hw * channel + c;
+        int nchw_index = n * channel * plane + c * plane + hw;
+        ((float16_t *)(dst))[nhwc_index] = ((const float16_t *)(src))[nchw_index];
+      }
+    }
+  }
+  return;
+}
+
 void PackNHWCToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
   int ic4 = UP_DIV(channel, C4NUM);
   int nhwc4_batch_unit_offset = ic4 * C4NUM * plane;

mindspore/lite/nnacl/fp16/pack_fp16.h

@@ -41,6 +41,8 @@ void PackNHWCToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int
 
 void PackNCHWToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel);
 
+void PackNCHWToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel);
+
 void PackNHWCToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
 
 void PackNHWCToNHWC8Fp16(const void *src, void *dst, int batch, int plane, int channel);

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc

@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.h"
+#include "src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.h"
 #include "nnacl/fp16/pack_fp16.h"
 #include "nnacl/fp16/cast_fp16.h"
 #include "schema/model_generated.h"
@@ -30,72 +31,34 @@ using mindspore::schema::PrimitiveType_DepthwiseConv2D;
 
 namespace mindspore::kernel {
 ConvolutionDepthwiseFp16CPUKernel::~ConvolutionDepthwiseFp16CPUKernel() {
-  if (sliding_ != nullptr) {
-    delete sliding_;
-    sliding_ = nullptr;
-  }
   if (packed_weight_ != nullptr) {
     delete packed_weight_;
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
-}
-
-void ConvolutionDepthwiseFp16CPUKernel::FreeTmpBuffer() {
-  if (need_align_) {
-    if (packed_input_ != nullptr) {
-      delete packed_input_;
-      packed_input_ = nullptr;
-    }
-    if (packed_output_ != nullptr) {
-      delete packed_output_;
-      packed_output_ = nullptr;
-    }
-  }
-}
-
-int ConvolutionDepthwiseFp16CPUKernel::InitBuffer() {
-  if (conv_param_->input_channel_ % C4NUM != 0) {
-    need_align_ = true;
-    int C8 = UP_DIV(conv_param_->input_channel_, C8NUM);
-    int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C8NUM * C8;
-    packed_input_ = reinterpret_cast<float16_t *>(malloc(pack_input_size * sizeof(float16_t)));
-    if (packed_input_ == nullptr) {
-      MS_LOG(ERROR) << "Malloc buffer failed.";
-      return RET_ERROR;
-    }
-
-    int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C8NUM * C8;
-    packed_output_ = reinterpret_cast<float16_t *>(malloc(pack_output_size * sizeof(float16_t)));
-    if (packed_output_ == nullptr) {
-      MS_LOG(ERROR) << "Malloc buffer failed.";
-      return RET_ERROR;
-    }
-  }
-  return RET_OK;
 }
 
 int ConvolutionDepthwiseFp16CPUKernel::InitWeightBias() {
   // init weight: o, h, w, i; o == group, i == 1
+  ConvolutionBaseFP16CPUKernel::GetExecuteFilter();
+
   auto weight_tensor = in_tensors_[kWeightIndex];
-  int OC8 = UP_DIV(weight_tensor->Batch(), C8NUM);
-  auto origin_weight = reinterpret_cast<float *>(weight_tensor->Data());
-  int pack_weight_size = C8NUM * OC8 * weight_tensor->Height() * weight_tensor->Width();
+  int channel = weight_tensor->Batch();
+  int pack_weight_size = channel * weight_tensor->Height() * weight_tensor->Width();
 
   packed_weight_ = reinterpret_cast<float16_t *>(malloc(pack_weight_size * sizeof(float16_t)));
   if (packed_weight_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
   }
-  PackNCHWFp32ToNC8HW8Fp16(origin_weight, packed_weight_, 1, weight_tensor->Height() * weight_tensor->Width(),
-                           weight_tensor->Batch());
+  PackNCHWToNHWCFp16(fp16_weight_, packed_weight_, 1, weight_tensor->Height() * weight_tensor->Width(),
+                     weight_tensor->Batch());
 
-  bias_data_ = reinterpret_cast<float16_t *>(malloc(C8NUM * OC8 * sizeof(float16_t)));
+  bias_data_ = reinterpret_cast<float16_t *>(malloc(channel * sizeof(float16_t)));
   if (bias_data_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
   }
-  memset(bias_data_, 0, C8NUM * OC8 * sizeof(float16_t));
+  memset(bias_data_, 0, channel * sizeof(float16_t));
   auto bias_fp16 = reinterpret_cast<float16_t *>(bias_data_);
   if (in_tensors_.size() == kInputSize2) {
     auto bias_tensor = in_tensors_.at(kBiasIndex);
@@ -104,18 +67,10 @@ int ConvolutionDepthwiseFp16CPUKernel::InitWeightBias() {
       bias_fp16[i] = (float16_t)ori_bias[i];
     }
   }
-
-  conv_param_->thread_num_ = MSMIN(thread_count_, OC8);
   return RET_OK;
 }
 
 int ConvolutionDepthwiseFp16CPUKernel::Init() {
-  sliding_ = new (std::nothrow) SlidingWindowParam;
-  if (sliding_ == nullptr) {
-    MS_LOG(ERROR) << "new sliding window param failed.";
-    return RET_ERROR;
-  }
-
   auto ret = InitWeightBias();
   if (ret != 0) {
     MS_LOG(ERROR) << "Convolution depthwise fp16 InitWeightBias failed.";
@@ -129,24 +84,17 @@ int ConvolutionDepthwiseFp16CPUKernel::Init() {
 }
 
 int ConvolutionDepthwiseFp16CPUKernel::ReSize() {
-  FreeTmpBuffer();
   auto ret = ConvolutionBaseCPUKernel::Init();
   if (ret != RET_OK) {
     return ret;
   }
-  InitSlidingParamConvDw(sliding_, conv_param_, C8NUM);
-
-  ret = InitBuffer();
-  if (ret != 0) {
-    MS_LOG(ERROR) << "Convolution depthwise fp16 InitBuffer failed.";
-    return RET_ERROR;
-  }
+  conv_param_->thread_num_ = MSMIN(thread_count_, conv_param_->output_h_);
   return RET_OK;
 }
 
 int ConvolutionDepthwiseFp16CPUKernel::Execute(int task_id) {
-  ConvDwC8Fp16(packed_output_, packed_input_, packed_weight_, reinterpret_cast<float16_t *>(bias_data_), conv_param_,
-               sliding_, task_id);
+  ConvDwFp16(execute_output_, execute_input_, packed_weight_, reinterpret_cast<float16_t *>(bias_data_), conv_param_,
+             task_id);
   return RET_OK;
 }
 
@@ -176,25 +124,13 @@ int ConvolutionDepthwiseFp16CPUKernel::Run() {
     MS_LOG(ERROR) << "Get Execute tensor failed.";
     return ret;
   }
-  if (need_align_) {
-    PackNHWCToNHWC8Fp16(execute_input_, packed_input_, conv_param_->input_batch_,
-                        conv_param_->input_h_ * conv_param_->input_w_, conv_param_->input_channel_);
-  } else {
-    packed_input_ = execute_input_;
-  }
-  if (!need_align_) {
-    packed_output_ = execute_output_;
-  }
 
   ret = LiteBackendParallelLaunch(ConvDwFp16Run, this, conv_param_->thread_num_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "ConvDwFp16Run error: error_code[" << ret << "]";
     return RET_ERROR;
   }
-  if (need_align_) {
-    PackNHWC8ToNHWCFp16(packed_output_, execute_output_, conv_param_->output_batch_,
-                        conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
-  }
+
   ConvolutionBaseFP16CPUKernel::IfCastOutput();
   ConvolutionBaseFP16CPUKernel::FreeTmpBuffer();
   return RET_OK;
@@ -207,7 +143,14 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::tensor::T
                                                const mindspore::lite::PrimitiveC *primitive) {
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
-  auto kernel = new (std::nothrow) ConvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  kernel::LiteKernel *kernel;
+  if (conv_param->input_channel_ < 32) {
+    kernel =
+      new (std::nothrow) kernel::ConvolutionDepthwiseSWFp16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  } else {
+    kernel = new (std::nothrow) kernel::ConvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  }
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel is nullptr.";
     return nullptr;

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.h

@@ -25,14 +25,12 @@
 #ifdef __cplusplus
 extern "C" {
 #endif
-void ConvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
-                  const float16_t *bias_data, const ConvParameter *conv_param, const SlidingWindowParam *sliding,
-                  int task_id);
+void ConvDwFp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
+                const float16_t *bias_data, const ConvParameter *conv_param, int task_id);
 #ifdef __cplusplus
 }
 #endif
 
-
 namespace mindspore::kernel {
 class ConvolutionDepthwiseFp16CPUKernel : public ConvolutionBaseFP16CPUKernel {
  public:
@@ -46,17 +44,11 @@ class ConvolutionDepthwiseFp16CPUKernel : public ConvolutionBaseFP16CPUKernel {
   int ReSize() override;
   int Run() override;
 
-  int InitBuffer();
   int InitWeightBias();
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
-  SlidingWindowParam *sliding_ = nullptr;
   float16_t *packed_weight_ = nullptr;
-  float16_t *packed_input_ = nullptr;
-  float16_t *packed_output_ = nullptr;
-  bool need_align_ = false;
 };
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.cc

@@ -0,0 +1,203 @@
+/**
+ * Copyright 2020 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 "src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.h"
+#include "nnacl/fp16/pack_fp16.h"
+#include "nnacl/fp16/cast_fp16.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "src/runtime/runtime_api.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_DepthwiseConv2D;
+
+namespace mindspore::kernel {
+ConvolutionDepthwiseSWFp16CPUKernel::~ConvolutionDepthwiseSWFp16CPUKernel() {
+  if (sliding_ != nullptr) {
+    delete sliding_;
+    sliding_ = nullptr;
+  }
+  if (packed_weight_ != nullptr) {
+    delete packed_weight_;
+    packed_weight_ = nullptr;
+  }
+  FreeTmpBuffer();
+}
+
+void ConvolutionDepthwiseSWFp16CPUKernel::FreeTmpBuffer() {
+  if (need_align_) {
+    if (packed_input_ != nullptr) {
+      delete packed_input_;
+      packed_input_ = nullptr;
+    }
+    if (packed_output_ != nullptr) {
+      delete packed_output_;
+      packed_output_ = nullptr;
+    }
+  }
+}
+
+int ConvolutionDepthwiseSWFp16CPUKernel::InitBuffer() {
+  if (conv_param_->input_channel_ % C4NUM != 0) {
+    need_align_ = true;
+    int C8 = UP_DIV(conv_param_->input_channel_, C8NUM);
+    int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C8NUM * C8;
+    packed_input_ = reinterpret_cast<float16_t *>(malloc(pack_input_size * sizeof(float16_t)));
+    if (packed_input_ == nullptr) {
+      MS_LOG(ERROR) << "Malloc buffer failed.";
+      return RET_ERROR;
+    }
+
+    int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C8NUM * C8;
+    packed_output_ = reinterpret_cast<float16_t *>(malloc(pack_output_size * sizeof(float16_t)));
+    if (packed_output_ == nullptr) {
+      MS_LOG(ERROR) << "Malloc buffer failed.";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+int ConvolutionDepthwiseSWFp16CPUKernel::InitWeightBias() {
+  // init weight: o, h, w, i; o == group, i == 1
+  auto weight_tensor = in_tensors_[kWeightIndex];
+  int OC8 = UP_DIV(weight_tensor->Batch(), C8NUM);
+  auto origin_weight = reinterpret_cast<float *>(weight_tensor->Data());
+  int pack_weight_size = C8NUM * OC8 * weight_tensor->Height() * weight_tensor->Width();
+
+  packed_weight_ = reinterpret_cast<float16_t *>(malloc(pack_weight_size * sizeof(float16_t)));
+  if (packed_weight_ == nullptr) {
+    MS_LOG(ERROR) << "Malloc buffer failed.";
+    return RET_ERROR;
+  }
+  PackNCHWFp32ToNC8HW8Fp16(origin_weight, packed_weight_, 1, weight_tensor->Height() * weight_tensor->Width(),
+                           weight_tensor->Batch());
+
+  bias_data_ = reinterpret_cast<float16_t *>(malloc(C8NUM * OC8 * sizeof(float16_t)));
+  if (bias_data_ == nullptr) {
+    MS_LOG(ERROR) << "Malloc buffer failed.";
+    return RET_ERROR;
+  }
+  memset(bias_data_, 0, C8NUM * OC8 * sizeof(float16_t));
+  auto bias_fp16 = reinterpret_cast<float16_t *>(bias_data_);
+  if (in_tensors_.size() == kInputSize2) {
+    auto bias_tensor = in_tensors_.at(kBiasIndex);
+    auto ori_bias = reinterpret_cast<float *>(bias_tensor->Data());
+    for (int i = 0; i < bias_tensor->ElementsNum(); i++) {
+      bias_fp16[i] = (float16_t)ori_bias[i];
+    }
+  }
+
+  conv_param_->thread_num_ = MSMIN(thread_count_, OC8);
+  return RET_OK;
+}
+
+int ConvolutionDepthwiseSWFp16CPUKernel::Init() {
+  sliding_ = new (std::nothrow) SlidingWindowParam;
+  if (sliding_ == nullptr) {
+    MS_LOG(ERROR) << "new sliding window param failed.";
+    return RET_ERROR;
+  }
+
+  auto ret = InitWeightBias();
+  if (ret != 0) {
+    MS_LOG(ERROR) << "Convolution depthwise fp16 InitWeightBias failed.";
+    return RET_ERROR;
+  }
+
+  if (!InferShapeDone()) {
+    return RET_OK;
+  }
+  return ReSize();
+}
+
+int ConvolutionDepthwiseSWFp16CPUKernel::ReSize() {
+  FreeTmpBuffer();
+  auto ret = ConvolutionBaseCPUKernel::Init();
+  if (ret != RET_OK) {
+    return ret;
+  }
+  InitSlidingParamConvDw(sliding_, conv_param_, C8NUM);
+
+  ret = InitBuffer();
+  if (ret != 0) {
+    MS_LOG(ERROR) << "Convolution depthwise fp16 InitBuffer failed.";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ConvolutionDepthwiseSWFp16CPUKernel::Execute(int task_id) {
+  ConvDwC8Fp16(packed_output_, packed_input_, packed_weight_, reinterpret_cast<float16_t *>(bias_data_), conv_param_,
+               sliding_, task_id);
+  return RET_OK;
+}
+
+static int ConvDwSWFp16Run(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
+  auto conv_dw_fp16 = reinterpret_cast<ConvolutionDepthwiseSWFp16CPUKernel *>(cdata);
+  auto ret = conv_dw_fp16->Execute(task_id);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "ConvolutionDepthwiseSWFp16Run error task_id[" << task_id << "] error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ConvolutionDepthwiseSWFp16CPUKernel::Run() {
+  auto ret = Prepare();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare failed.";
+    return RET_ERROR;
+  }
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
+
+  ret = ConvolutionBaseFP16CPUKernel::GetExecuteTensor();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Get Execute tensor failed.";
+    return ret;
+  }
+  if (need_align_) {
+    PackNHWCToNHWC8Fp16(execute_input_, packed_input_, conv_param_->input_batch_,
+                        conv_param_->input_h_ * conv_param_->input_w_, conv_param_->input_channel_);
+  } else {
+    packed_input_ = execute_input_;
+  }
+  if (!need_align_) {
+    packed_output_ = execute_output_;
+  }
+
+  ret = LiteBackendParallelLaunch(ConvDwSWFp16Run, this, conv_param_->thread_num_);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "ConvDwSWFp16Run error: error_code[" << ret << "]";
+    return RET_ERROR;
+  }
+  if (need_align_) {
+    PackNHWC8ToNHWCFp16(packed_output_, execute_output_, conv_param_->output_batch_,
+                        conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+  }
+  ConvolutionBaseFP16CPUKernel::IfCastOutput();
+  ConvolutionBaseFP16CPUKernel::FreeTmpBuffer();
+  return RET_OK;
+}
+
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.h

@@ -0,0 +1,62 @@
+/**
+ * Copyright 2020 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_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_DEPTHWISE_SW_FP16_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_DEPTHWISE_SW_FP16_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "src/runtime/kernel/arm/fp16/convolution_base_fp16.h"
+#include "nnacl/fp16/conv_depthwise_fp16.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+void ConvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
+                  const float16_t *bias_data, const ConvParameter *conv_param, const SlidingWindowParam *sliding,
+                  int task_id);
+#ifdef __cplusplus
+}
+#endif
+
+namespace mindspore::kernel {
+class ConvolutionDepthwiseSWFp16CPUKernel : public ConvolutionBaseFP16CPUKernel {
+ public:
+  ConvolutionDepthwiseSWFp16CPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                      const std::vector<lite::tensor::Tensor *> &outputs, const Context *ctx,
+                                      const mindspore::lite::PrimitiveC *primitive)
+      : ConvolutionBaseFP16CPUKernel(parameter, inputs, outputs, ctx, primitive) {}
+  ~ConvolutionDepthwiseSWFp16CPUKernel() override;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+  int InitBuffer();
+  int InitWeightBias();
+  int Execute(int task_id);
+
+ private:
+  void FreeTmpBuffer();
+  SlidingWindowParam *sliding_ = nullptr;
+  float16_t *packed_weight_ = nullptr;
+  float16_t *packed_input_ = nullptr;
+  float16_t *packed_output_ = nullptr;
+  bool need_align_ = false;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP16_CONVOLUTION_DEPTHWISE_SW_FP16_H_

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_winograd_fp16.cc

@@ -17,7 +17,6 @@
 #include "src/runtime/kernel/arm/fp16/convolution_winograd_fp16.h"
 #include "src/runtime/kernel/arm/fp16/matrix_fp16.h"
 #include "nnacl/fp16/conv_fp16.h"
-#include "nnacl/fp16/common_func.h"
 #include "nnacl/fp16/cast_fp16.h"
 #include "nnacl/fp16/pack_fp16.h"
 #include "nnacl/fp16/winograd_transform_fp16.h"