From d5034cad264bb1ca76293e7a5d3f14ae7f6ea7c1 Mon Sep 17 00:00:00 2001
From: kai00 <wangchangkai@huawei.com>
Date: Tue, 22 Sep 2020 19:34:38 +0800
Subject: [PATCH] fp16 weight dequant

---
 .../arm/fp16/convolution_depthwise_fp16.cc    | 24 +++++++++++++++++++
 .../kernel/arm/fp16/convolution_fp16.cc       | 24 +++++++++++++++++++
 .../arm/fp16/deconvolution_depthwise_fp16.cc  | 24 +++++++++++++++++++
 .../kernel/arm/fp16/deconvolution_fp16.cc     | 24 +++++++++++++++++++
 .../kernel/arm/fp16/fullconnection_fp16.cc    | 23 ++++++++++++++++++
 .../runtime/kernel/arm/fp16/matmul_fp16.cc    | 22 +++++++++++++++++
 6 files changed, 141 insertions(+)

diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc
index a91c045115..8602b14df1 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc
@@ -142,6 +142,18 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::Tensor *>
                                                const mindspore::lite::PrimitiveC *primitive) {
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
+
+  auto *weight_tensor = inputs.at(kWeightIndex);
+  auto *restore_data = weight_tensor->MutableData();
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    auto *dequant_weight = kernel::LiteKernelUtil::DequantWeight(weight_tensor);
+    if (dequant_weight == nullptr) {
+      MS_LOG(ERROR) << "dequant data is nullptr.";
+      return nullptr;
+    }
+    weight_tensor->SetData(dequant_weight);
+  }
+
   auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
   kernel::LiteKernel *kernel;
   if (conv_param->input_channel_ < 32) {
@@ -152,6 +164,10 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::Tensor *>
   }
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel is nullptr.";
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -159,8 +175,16 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::Tensor *>
     delete kernel;
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    weight_tensor->FreeData();
+    weight_tensor->SetData(restore_data);
+  }
   return kernel;
 }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
index 85c59e3d15..bb78a4f124 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
@@ -218,6 +218,18 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
                                              const mindspore::lite::PrimitiveC *primitive) {
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Conv2D);
+
+  auto *weight_tensor = inputs.at(kWeightIndex);
+  auto *restore_data = weight_tensor->MutableData();
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    auto *dequant_weight = kernel::LiteKernelUtil::DequantWeight(weight_tensor);
+    if (dequant_weight == nullptr) {
+      MS_LOG(ERROR) << "dequant data is nullptr.";
+      return nullptr;
+    }
+    weight_tensor->SetData(dequant_weight);
+  }
+
   auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
   int kernel_h = conv_param->kernel_h_;
   int kernel_w = conv_param->kernel_w_;
@@ -249,6 +261,10 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
   }
   if (kernel == nullptr) {
     MS_LOG(DEBUG) << "Create conv fp16 kernel failed.";
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -256,8 +272,16 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
     delete kernel;
     MS_LOG(INFO) << "Init fp16 kernel failed, name: " << opParameter->name_
                  << ", type: " << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    weight_tensor->FreeData();
+    weight_tensor->SetData(restore_data);
+  }
   return kernel;
 }
 REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Conv2D, CpuConvFp16KernelCreator)
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc
index f087932c66..09f6627d95 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc
@@ -201,9 +201,25 @@ kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::Tensor
                                                  const mindspore::lite::PrimitiveC *primitive) {
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_DeDepthwiseConv2D);
+
+  auto *weight_tensor = inputs.at(kWeightIndex);
+  auto *restore_data = weight_tensor->MutableData();
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    auto *dequant_weight = kernel::LiteKernelUtil::DequantWeight(weight_tensor);
+    if (dequant_weight == nullptr) {
+      MS_LOG(ERROR) << "dequant data is nullptr.";
+      return nullptr;
+    }
+    weight_tensor->SetData(dequant_weight);
+  }
+
   auto kernel = new (std::nothrow) DeconvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel is nullptr.";
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -211,8 +227,16 @@ kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::Tensor
     delete kernel;
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    weight_tensor->FreeData();
+    weight_tensor->SetData(restore_data);
+  }
   return kernel;
 }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc
index 2f09e6dead..dd40cb0633 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc
@@ -208,9 +208,25 @@ kernel::LiteKernel *CpuDeConvFp16KernelCreator(const std::vector<lite::Tensor *>
                                                const mindspore::lite::PrimitiveC *primitive) {
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_DeConv2D);
+
+  auto *weight_tensor = inputs.at(kWeightIndex);
+  auto *restore_data = weight_tensor->MutableData();
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    auto *dequant_weight = kernel::LiteKernelUtil::DequantWeight(weight_tensor);
+    if (dequant_weight == nullptr) {
+      MS_LOG(ERROR) << "dequant data is nullptr.";
+      return nullptr;
+    }
+    weight_tensor->SetData(dequant_weight);
+  }
+
   auto kernel = new (std::nothrow) DeConvolutionFp16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel is nullptr.";
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -218,8 +234,16 @@ kernel::LiteKernel *CpuDeConvFp16KernelCreator(const std::vector<lite::Tensor *>
     delete kernel;
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
+  if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
+    weight_tensor->FreeData();
+    weight_tensor->SetData(restore_data);
+  }
   return kernel;
 }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc
index a8b3f5ab7b..b1c4bf44ec 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc
@@ -170,9 +170,24 @@ kernel::LiteKernel *CpuFullConnectionFp16KernelCreator(const std::vector<lite::T
                                                        OpParameter *opParameter, const lite::InnerContext *ctx,
                                                        const kernel::KernelKey &desc,
                                                        const mindspore::lite::PrimitiveC *primitive) {
+  auto *weight_tensor = inputs.at(kWeightIndex);
+  // data of second tensor of fc may be nullptr
+  auto *restore_data = weight_tensor->data_c();
+  if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
+    auto *dequant_weight = kernel::LiteKernelUtil::DequantWeight(weight_tensor);
+    if (dequant_weight == nullptr) {
+      MS_LOG(ERROR) << "dequant data is nullptr.";
+      return nullptr;
+    }
+    weight_tensor->SetData(dequant_weight);
+  }
   auto *kernel = new (std::nothrow) FullconnectionFP16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel is nullptr.";
+    if (!weight_tensor->GetQuantParams().empty()) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -180,8 +195,16 @@ kernel::LiteKernel *CpuFullConnectionFp16KernelCreator(const std::vector<lite::T
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
     delete kernel;
+    if (!weight_tensor->GetQuantParams().empty()) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
+  if (!weight_tensor->GetQuantParams().empty()) {
+    weight_tensor->FreeData();
+    weight_tensor->SetData(restore_data);
+  }
   return kernel;
 }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc
index bc9fcbf698..b03bf3c31c 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc
@@ -253,9 +253,23 @@ kernel::LiteKernel *CpuMatmulFp16KernelCreator(const std::vector<lite::Tensor *>
                                                const std::vector<lite::Tensor *> &outputs, OpParameter *opParameter,
                                                const lite::InnerContext *ctx, const kernel::KernelKey &desc,
                                                const mindspore::lite::PrimitiveC *primitive) {
+  auto *weight_tensor = inputs.at(kWeightIndex);
+  auto *restore_data = weight_tensor->data_c();
+  if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
+    auto *dequant_weight = kernel::LiteKernelUtil::DequantWeight(weight_tensor);
+    if (dequant_weight == nullptr) {
+      MS_LOG(ERROR) << "dequant data is nullptr.";
+      return nullptr;
+    }
+    weight_tensor->SetData(dequant_weight);
+  }
   auto *kernel = new (std::nothrow) MatmulFP16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel is nullptr.";
+    if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -263,8 +277,16 @@ kernel::LiteKernel *CpuMatmulFp16KernelCreator(const std::vector<lite::Tensor *>
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
     delete kernel;
+    if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
+      weight_tensor->FreeData();
+      weight_tensor->SetData(restore_data);
+    }
     return nullptr;
   }
+  if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
+    weight_tensor->FreeData();
+    weight_tensor->SetData(restore_data);
+  }
   return kernel;
 }