From 797221b1444b3510d273e6a5ce9626ae122ac7bc Mon Sep 17 00:00:00 2001
From: kai00 <wangchangkai@huawei.com>
Date: Thu, 24 Sep 2020 21:24:23 +0800
Subject: [PATCH] float16 quant fix

---
 mindspore/lite/src/lite_kernel.cc                             | 1 +
 .../src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc | 4 ++++
 .../lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc      | 4 ++++
 .../runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc   | 4 ++++
 .../lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc    | 4 ++++
 .../lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc   | 4 ++++
 mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc     | 4 ++++
 7 files changed, 25 insertions(+)

diff --git a/mindspore/lite/src/lite_kernel.cc b/mindspore/lite/src/lite_kernel.cc
index c20ce8fe9d4..43a7a07f698 100644
--- a/mindspore/lite/src/lite_kernel.cc
+++ b/mindspore/lite/src/lite_kernel.cc
@@ -214,6 +214,7 @@ float *LiteKernelUtil::DequantWeight(lite::Tensor *input_tensor) {
       dequant_datas[j] = static_cast<float>((quant_datas[j] - zero_point) * scale);
     }
   }
+
   return dequant_datas;
 }
 }  // namespace mindspore::kernel
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 8602b14df1c..2c2b060a7a5 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
@@ -151,6 +151,7 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::Tensor *>
       MS_LOG(ERROR) << "dequant data is nullptr.";
       return nullptr;
     }
+    weight_tensor->set_data_type(kNumberTypeFloat32);
     weight_tensor->SetData(dequant_weight);
   }
 
@@ -166,6 +167,7 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::Tensor *>
     MS_LOG(ERROR) << "kernel is nullptr.";
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
@@ -177,12 +179,14 @@ kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::Tensor *>
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
   }
   if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
     weight_tensor->FreeData();
+    weight_tensor->set_data_type(kNumberTypeInt8);
     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 e7337cefe81..a7ae4dd4c70 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
@@ -196,6 +196,7 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
       MS_LOG(ERROR) << "dequant data is nullptr.";
       return nullptr;
     }
+    weight_tensor->set_data_type(kNumberTypeFloat32);
     weight_tensor->SetData(dequant_weight);
   }
 
@@ -232,6 +233,7 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
     MS_LOG(DEBUG) << "Create conv fp16 kernel failed.";
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
@@ -243,12 +245,14 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::Tensor *> &
                  << ", 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->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
   }
   if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
     weight_tensor->FreeData();
+    weight_tensor->set_data_type(kNumberTypeInt8);
     weight_tensor->SetData(restore_data);
   }
   return kernel;
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 09f6627d95f..6c1b6b58446 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
@@ -210,6 +210,7 @@ kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::Tensor
       MS_LOG(ERROR) << "dequant data is nullptr.";
       return nullptr;
     }
+    weight_tensor->set_data_type(kNumberTypeFloat32);
     weight_tensor->SetData(dequant_weight);
   }
 
@@ -218,6 +219,7 @@ kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::Tensor
     MS_LOG(ERROR) << "kernel is nullptr.";
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
@@ -229,12 +231,14 @@ kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::Tensor
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
   }
   if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
     weight_tensor->FreeData();
+    weight_tensor->set_data_type(kNumberTypeInt8);
     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 dd40cb06338..cb1e52bfc6d 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc
@@ -217,6 +217,7 @@ kernel::LiteKernel *CpuDeConvFp16KernelCreator(const std::vector<lite::Tensor *>
       MS_LOG(ERROR) << "dequant data is nullptr.";
       return nullptr;
     }
+    weight_tensor->set_data_type(kNumberTypeFloat32);
     weight_tensor->SetData(dequant_weight);
   }
 
@@ -225,6 +226,7 @@ kernel::LiteKernel *CpuDeConvFp16KernelCreator(const std::vector<lite::Tensor *>
     MS_LOG(ERROR) << "kernel is nullptr.";
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
@@ -236,12 +238,14 @@ kernel::LiteKernel *CpuDeConvFp16KernelCreator(const std::vector<lite::Tensor *>
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
     if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
   }
   if (weight_tensor->data_type() == kNumberTypeInt8 || primitive->GetQuantType() == schema::QuantType_WeightQuant) {
     weight_tensor->FreeData();
+    weight_tensor->set_data_type(kNumberTypeInt8);
     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 bfded12ce2d..78316811520 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/fullconnection_fp16.cc
@@ -198,6 +198,7 @@ kernel::LiteKernel *CpuFullConnectionFp16KernelCreator(const std::vector<lite::T
       MS_LOG(ERROR) << "dequant data is nullptr.";
       return nullptr;
     }
+    weight_tensor->set_data_type(kNumberTypeFloat32);
     weight_tensor->SetData(dequant_weight);
   }
   auto *kernel = new (std::nothrow) FullconnectionFP16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
@@ -205,6 +206,7 @@ kernel::LiteKernel *CpuFullConnectionFp16KernelCreator(const std::vector<lite::T
     MS_LOG(ERROR) << "kernel is nullptr.";
     if (!weight_tensor->GetQuantParams().empty()) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
@@ -216,12 +218,14 @@ kernel::LiteKernel *CpuFullConnectionFp16KernelCreator(const std::vector<lite::T
     delete kernel;
     if (!weight_tensor->GetQuantParams().empty()) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
   }
   if (!weight_tensor->GetQuantParams().empty()) {
     weight_tensor->FreeData();
+    weight_tensor->set_data_type(kNumberTypeInt8);
     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 b03bf3c31ca..5ba2e0bae16 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc
@@ -261,6 +261,7 @@ kernel::LiteKernel *CpuMatmulFp16KernelCreator(const std::vector<lite::Tensor *>
       MS_LOG(ERROR) << "dequant data is nullptr.";
       return nullptr;
     }
+    weight_tensor->set_data_type(kNumberTypeFloat32);
     weight_tensor->SetData(dequant_weight);
   }
   auto *kernel = new (std::nothrow) MatmulFP16CPUKernel(opParameter, inputs, outputs, ctx, primitive);
@@ -268,6 +269,7 @@ kernel::LiteKernel *CpuMatmulFp16KernelCreator(const std::vector<lite::Tensor *>
     MS_LOG(ERROR) << "kernel is nullptr.";
     if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
@@ -279,12 +281,14 @@ kernel::LiteKernel *CpuMatmulFp16KernelCreator(const std::vector<lite::Tensor *>
     delete kernel;
     if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
       weight_tensor->FreeData();
+      weight_tensor->set_data_type(kNumberTypeInt8);
       weight_tensor->SetData(restore_data);
     }
     return nullptr;
   }
   if (!weight_tensor->GetQuantParams().empty() && restore_data != nullptr) {
     weight_tensor->FreeData();
+    weight_tensor->set_data_type(kNumberTypeInt8);
     weight_tensor->SetData(restore_data);
   }
   return kernel;