!19336 [MSLITE][DEVELOP] fix bug of npu+fp16, convert const tensor data from fp16 -> fp32

Merge pull request !19336 from yangruoqi713/npu

mindspore/lite/src/delegate/npu/npu_converter_utils.cc

@@ -15,8 +15,45 @@
  */
 
 #include "src/delegate/npu/npu_converter_utils.h"
+#include <arm_neon.h>
 #include "src/common/log_adapter.h"
 namespace mindspore {
+#define C8NUM 8
+#ifdef ENABLE_ARM64
+void Float32ToFloat16(const float *__restrict input, float16_t *__restrict output, int number) {
+  int count = (number & ~(C8NUM - 1));
+  int i = 0;
+  for (; i < count; i += C8NUM) {
+    float32x4_t in1 = vld1q_f32(input + i);
+    float16x4_t out1 = vcvt_f16_f32(in1);
+    float32x4_t in2 = vld1q_f32(input + i + 4);
+    float16x4_t out2 = vcvt_f16_f32(in2);
+    float16x8_t out = vcombine_f16(out1, out2);
+    vst1q_f16(output + i, out);
+  }
+  for (; i < number; ++i) {
+    output[i] = static_cast<float16_t>(input[i]);
+  }
+}
+
+void Float16ToFloat32(const float16_t *__restrict input, float *__restrict output, int number) {
+  int count = number & ~(C8NUM - 1);
+  int i = 0;
+  for (; i < count; i += C8NUM) {
+    float16x8_t in = vld1q_f16(input + i);
+    float16x4_t in1 = vget_low_f16(in);
+    float16x4_t in2 = vget_high_f16(in);
+    float32x4_t out1 = vcvt_f32_f16(in1);
+    vst1q_f32(output + i, out1);
+    float32x4_t out2 = vcvt_f32_f16(in2);
+    vst1q_f32(output + i + 4, out2);
+  }
+  for (; i < number; ++i) {
+    output[i] = static_cast<float>(input[i]);
+  }
+}
+#endif
+
 ge::Shape ConverterToNPUShape(const std::vector<int> &src_shape) {
   vector<int64_t> shapes;
   shapes.reserve(src_shape.size());
@@ -50,10 +87,8 @@ ge::DataType ConverterToNPUDataType(TypeId type_id) {
   switch (type_id) {
     case kNumberTypeFloat:
     case kNumberTypeFloat32:
-      data_type = ge::DT_FLOAT;
-      break;
     case kNumberTypeFloat16:
-      data_type = ge::DT_FLOAT16;
+      data_type = ge::DT_FLOAT;
       break;
     case kNumberTypeInt8:
       data_type = ge::DT_INT8;
@@ -93,7 +128,7 @@ std::shared_ptr<ge::Tensor> ConverterToNPUTensor(tensor::MSTensor *src) {
   std::shared_ptr<ge::Tensor> ge_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
   if (ge_tensor == nullptr) {
     MS_LOG(ERROR) << "new ge_tensor failed.";
-    return ge_tensor;
+    return nullptr;
   }
   ge::TensorDesc tensor_desc(ConverterToNPUShape(src->shape()), ge::FORMAT_NCHW,
                              ConverterToNPUDataType(src->data_type()));
@@ -101,8 +136,21 @@ std::shared_ptr<ge::Tensor> ConverterToNPUTensor(tensor::MSTensor *src) {
   ge_tensor->SetTensorDesc(tensor_desc);
 
   if (src->data() != nullptr) {
+    if (src->data_type() == kNumberTypeFloat16) {
+#ifdef ENABLE_ARM64
+      auto fp32_data = malloc(src->ElementsNum() * sizeof(float));
+      Float16ToFloat32(reinterpret_cast<float16_t *>(src->data()), reinterpret_cast<float *>(fp32_data),
+                       src->ElementsNum());
+      ge_tensor->SetData(reinterpret_cast<const uint8_t *>(fp32_data), src->ElementsNum() * sizeof(float));
+      free(fp32_data);
+#else
+      MS_LOG(ERROR) << "This platform does not support fp16.";
+      return nullptr;
+#endif
+    } else {
       ge_tensor->SetData(reinterpret_cast<const uint8_t *>(src->data()), src->Size());
     }
+  }
   return ge_tensor;
 }
 

mindspore/lite/src/delegate/npu/npu_delegate.cc

@@ -217,6 +217,17 @@ NPUOp *NPUDelegate::GetOP(kernel::Kernel *kernel, const schema::Primitive *primi
       return nullptr;
     }
   }
+
+  for (auto tensor : in_tensors) {
+    if (tensor->data_type() == kNumberTypeFloat16 && tensor->data() == nullptr) {
+      tensor->set_data_type(kNumberTypeFloat32);
+    }
+  }
+  for (auto tensor : out_tensors) {
+    if (tensor->data_type() == kNumberTypeFloat16) {
+      tensor->set_data_type(kNumberTypeFloat32);
+    }
+  }
   return npu_op;
 }
 

mindspore/lite/src/delegate/npu/op/batchnorm_npu.cc

@@ -45,7 +45,7 @@ int BatchnormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tenso
     MS_LOG(ERROR) << "New scale const failed.";
     return RET_ERROR;
   }
-  auto scale_tensor = mindspore::ConverterToNPUTensor(in_tensors[1]);
+  auto scale_tensor = ConverterToNPUTensor(in_tensors[1]);
   scale->set_attr_value(scale_tensor);
   batchnorm_->set_input_scale(*scale);
 
@@ -54,7 +54,7 @@ int BatchnormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tenso
     MS_LOG(ERROR) << "New offset const failed.";
     return RET_ERROR;
   }
-  auto offset_tensor = mindspore::ConverterToNPUTensor(in_tensors[2]);
+  auto offset_tensor = ConverterToNPUTensor(in_tensors[2]);
   offset->set_attr_value(offset_tensor);
   batchnorm_->set_input_offset(*offset);
 
@@ -63,7 +63,7 @@ int BatchnormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tenso
     MS_LOG(ERROR) << "New mean const failed.";
     return RET_ERROR;
   }
-  auto mean_tensor = mindspore::ConverterToNPUTensor(in_tensors[3]);
+  auto mean_tensor = ConverterToNPUTensor(in_tensors[3]);
   mean->set_attr_value(mean_tensor);
   batchnorm_->set_input_mean(*mean);
 
@@ -72,7 +72,7 @@ int BatchnormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tenso
     MS_LOG(ERROR) << "New variance const failed.";
     return RET_ERROR;
   }
-  auto variance_tensor = mindspore::ConverterToNPUTensor(in_tensors[4]);
+  auto variance_tensor = ConverterToNPUTensor(in_tensors[4]);
   variance->set_attr_value(variance_tensor);
   batchnorm_->set_input_variance(*variance);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/convolution_base_npu.cc

@@ -17,6 +17,8 @@
 #include "src/delegate/npu/op/convolution_base_npu.h"
 #include "src/delegate/npu/npu_converter_utils.h"
 #include "src/delegate/npu/transpose_kernel.h"
+#include "nnacl/fp16/cast_fp16.h"
+
 namespace mindspore {
 ConvolutionBaseNPUOp::~ConvolutionBaseNPUOp() {
   if (act_ != nullptr) {
@@ -40,13 +42,20 @@ int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<tensor::MSTensor *>
     return RET_ERROR;
   }
   auto w_shape = inputs[1]->shape();
-  auto nhwc_data = inputs[1]->data();
+  auto origin_data = inputs[1]->data();
+  auto fp32_data = origin_data;
+  if (inputs[1]->data_type() == kNumberTypeFloat16) {
+    fp32_data = reinterpret_cast<float *>(malloc(inputs[1]->ElementsNum() * sizeof(float)));
+    // fp16->fp32
+    Float16ToFloat32(reinterpret_cast<float16_t *>(origin_data), reinterpret_cast<float *>(fp32_data),
+                     inputs[1]->ElementsNum());
+  }
   auto nchw_data = reinterpret_cast<float *>(malloc(inputs[1]->ElementsNum() * sizeof(float)));
   if (nchw_data == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
   }
-  PackNHWCToNCHWFp32(nhwc_data, nchw_data, w_shape[0], w_shape[1] * w_shape[2], w_shape[3]);
+  PackNHWCToNCHWFp32(fp32_data, nchw_data, w_shape[0], w_shape[1] * w_shape[2], w_shape[3]);
 
   std::shared_ptr<ge::Tensor> weight_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
   if (weight_tensor == nullptr) {
@@ -56,7 +65,7 @@ int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<tensor::MSTensor *>
   ge::TensorDesc tensor_desc(ConverterToNPUShape({w_shape[0], w_shape[3], w_shape[1], w_shape[2]}), ge::FORMAT_NCHW,
                              ConverterToNPUDataType(inputs[1]->data_type()));
   weight_tensor->SetTensorDesc(tensor_desc);
-  weight_tensor->SetData(reinterpret_cast<const uint8_t *>(nchw_data), inputs[1]->Size());
+  weight_tensor->SetData(reinterpret_cast<const uint8_t *>(nchw_data), inputs[1]->ElementsNum() * sizeof(float));
 
   weight_->set_attr_value(weight_tensor);
   free(nchw_data);
@@ -70,15 +79,11 @@ int ConvolutionBaseNPUOp::InitBiasConst(const std::vector<tensor::MSTensor *> &i
       MS_LOG(ERROR) << "New bias const failed.";
       return RET_ERROR;
     }
-    std::shared_ptr<ge::Tensor> bias_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
+    std::shared_ptr<ge::Tensor> bias_tensor = ConverterToNPUTensor(inputs[2]);
     if (bias_tensor == nullptr) {
-      MS_LOG(ERROR) << "new bias_tensor failed.";
+      MS_LOG(ERROR) << "Get bias_tensor failed.";
       return RET_ERROR;
     }
-    ge::TensorDesc tensor_desc(ConverterToNPUShape({inputs[2]->shape()[0]}), ge::FORMAT_NCHW,
-                               ConverterToNPUDataType(inputs[2]->data_type()));
-    bias_tensor->SetTensorDesc(tensor_desc);
-    bias_tensor->SetData(reinterpret_cast<const uint8_t *>(inputs[2]->data()), inputs[2]->Size());
     bias_->set_attr_value(bias_tensor);
   }
   return RET_OK;

mindspore/lite/src/delegate/npu/op/convolution_npu.cc

@@ -139,6 +139,11 @@ NPUOp *GetNPUConvOp(const schema::Primitive *primitive, const std::vector<tensor
     return nullptr;
   }
 
+  if (in_tensors[0]->data_type() != kNumberTypeFloat32 && in_tensors[0]->data_type() != kNumberTypeFloat16) {
+    MS_LOG(ERROR) << "Npu does not support datatype " << in_tensors[0]->data_type();
+    return nullptr;
+  }
+
   NPUOp *op = nullptr;
   auto conv_prim = primitive->value_as_Conv2DFusion();
   auto group = static_cast<int>(conv_prim->group());

mindspore/lite/src/delegate/npu/op/fullconnection_npu.cc

@@ -65,7 +65,7 @@ int FullconnectionNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_
     MS_LOG(ERROR) << "New weight const failed.";
     return RET_ERROR;
   }
-  auto weight_tensor = mindspore::ConverterToNPUTensor(in_tensors[1]);
+  auto weight_tensor = ConverterToNPUTensor(in_tensors[1]);
   weight_->set_attr_value(weight_tensor);
   fc_->set_input_x2(*weight_).set_attr_transpose_x2(true);
 

mindspore/lite/src/delegate/npu/op/instance_norm_npu.cc

@@ -41,15 +41,13 @@ int InstanceNormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_te
   instance_norm_->set_input_x(*npu_inputs[0]);
 
   auto gamma_shape = in_tensors[1]->shape();
-  std::shared_ptr<ge::Tensor> gamma_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
+  auto gamma_tensor = ConverterToNPUTensor(in_tensors[1]);
   if (gamma_tensor == nullptr) {
-    MS_LOG(ERROR) << "new gamma_tensor failed.";
+    MS_LOG(ERROR) << "Get gamma_tensor failed.";
     return RET_ERROR;
   }
-  ge::TensorDesc gamma_tensor_desc(ConverterToNPUShape({1, gamma_shape[0], 1, 1}), ge::FORMAT_NCHW,
+  gamma_tensor->SetTensorDesc(ge::TensorDesc(ConverterToNPUShape({1, gamma_shape[0], 1, 1})));
-                                   ConverterToNPUDataType(in_tensors[1]->data_type()));
+
-  gamma_tensor->SetTensorDesc(gamma_tensor_desc);
-  gamma_tensor->SetData(reinterpret_cast<const uint8_t *>(in_tensors[1]->data()), in_tensors[1]->Size());
   gamma_ = new (std::nothrow) hiai::op::Const(name_ + "_gamma");
   if (gamma_ == nullptr) {
     MS_LOG(ERROR) << "New gamma_ const failed.";
@@ -59,15 +57,13 @@ int InstanceNormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_te
   instance_norm_->set_input_gamma(*gamma_);
 
   auto beta_shape = in_tensors[2]->shape();
-  std::shared_ptr<ge::Tensor> beta_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
+  auto beta_tensor = ConverterToNPUTensor(in_tensors[2]);
   if (beta_tensor == nullptr) {
-    MS_LOG(ERROR) << "new beta_tensor failed.";
+    MS_LOG(ERROR) << "Get beta_tensor failed.";
     return RET_ERROR;
   }
-  ge::TensorDesc beta_tensor_desc(ConverterToNPUShape({1, beta_shape[0], 1, 1}), ge::FORMAT_NCHW,
+  beta_tensor->SetTensorDesc(ge::TensorDesc(ConverterToNPUShape({1, beta_shape[0], 1, 1})));
-                                  ConverterToNPUDataType(in_tensors[2]->data_type()));
+
-  beta_tensor->SetTensorDesc(beta_tensor_desc);
-  beta_tensor->SetData(reinterpret_cast<const uint8_t *>(in_tensors[2]->data()), in_tensors[2]->Size());
   beta_ = new (std::nothrow) hiai::op::Const(name_ + "_beta");
   if (beta_ == nullptr) {
     MS_LOG(ERROR) << "New beta_ const failed.";

mindspore/lite/src/delegate/npu/op/matmul_npu.cc

@@ -61,18 +61,18 @@ int MatMulNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
     }
     add_op_->set_input_x1(*matmul_);
     auto bias_shape = in_tensors[2]->shape();
-    auto bias_tensor = std::make_shared<ge::Tensor>();
+    auto bias_tensor = ConverterToNPUTensor(in_tensors[2]);
     if (bias_tensor == nullptr) {
-      MS_LOG(ERROR) << "new bias_tensor failed.";
+      MS_LOG(ERROR) << "Get bias_tensor failed.";
       return RET_ERROR;
     }
-    ge::TensorDesc bias_tensor_desc(ConverterToNPUShape({1, bias_shape[0], 1, 1}), ge::FORMAT_NCHW,
+
-                                    ConverterToNPUDataType(in_tensors[2]->data_type()));
+    ge::TensorDesc bias_tensor_desc(ConverterToNPUShape({1, bias_shape[0], 1, 1}));
     if (out_tensors[0]->shape().size() == 2) {
       bias_tensor_desc.SetShape(ConverterToNPUShape({1, bias_shape[0]}));
     }
     bias_tensor->SetTensorDesc(bias_tensor_desc);
-    bias_tensor->SetData(reinterpret_cast<const uint8_t *>(in_tensors[2]->data()), in_tensors[2]->Size());
+
     bias_ = new (std::nothrow) hiai::op::Const(name_ + "_bias");
     if (bias_ == nullptr) {
       MS_LOG(ERROR) << "new bias const failed.";

mindspore/lite/src/delegate/npu/op/npu_op.h

@@ -126,7 +126,8 @@ NPUOp *GetNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MS
   std::set<schema::PrimitiveType> int32_lists = {schema::PrimitiveType_Cast, schema::PrimitiveType_StridedSlice};
   auto support_int32 = in_tensors[0]->data_type() == kNumberTypeInt32 &&
                        find(int32_lists.begin(), int32_lists.end(), primitive->value_type()) != int32_lists.end();
-  if (in_tensors[0]->data_type() != kNumberTypeFloat32 && !support_int32) {
+  if (in_tensors[0]->data_type() != kNumberTypeFloat32 && in_tensors[0]->data_type() != kNumberTypeFloat16 &&
+      !support_int32) {
     MS_LOG(ERROR) << "Npu does not support datatype " << in_tensors[0]->data_type() << " for op type "
                   << primitive->value_type();
     return nullptr;

mindspore/lite/src/delegate/npu/op/scale_npu.cc

@@ -67,16 +67,14 @@ int ScaleNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
                              const std::vector<ge::Operator *> &npu_inputs) {
   op_->set_input_x(*npu_inputs.at(0));
   MS_ASSERT(in_tensors.size() > 1);
-  auto scale_shape = in_tensors.at(1)->shape();
+  auto scale_shape = in_tensors[1]->shape();
-  std::shared_ptr<ge::Tensor> scale_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
+  auto scale_tensor = ConverterToNPUTensor(in_tensors[1]);
   if (scale_tensor == nullptr) {
-    MS_LOG(ERROR) << "new scale_tensor failed.";
+    MS_LOG(ERROR) << "Get scale_tensor failed.";
     return RET_ERROR;
   }
-  ge::TensorDesc scale_tensor_desc(ConverterToNPUShape({1, scale_shape[0], 1, 1}), ge::FORMAT_NCHW,
+  scale_tensor->SetTensorDesc(ge::TensorDesc(ConverterToNPUShape({1, scale_shape[0], 1, 1})));
-                                   ConverterToNPUDataType(in_tensors[1]->data_type()));
+
-  scale_tensor->SetTensorDesc(scale_tensor_desc);
-  scale_tensor->SetData(reinterpret_cast<const uint8_t *>(in_tensors[1]->data()), in_tensors[1]->Size());
   scale_ = new (std::nothrow) hiai::op::Const(name_ + "_scale");
   if (scale_ == nullptr) {
     MS_LOG(ERROR) << "New scale_ const failed.";
@@ -87,15 +85,13 @@ int ScaleNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
 
   if (in_tensors.size() > 2 && in_tensors[2] != nullptr) {
     auto bias_shape = in_tensors[2]->shape();
-    std::shared_ptr<ge::Tensor> bias_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
+    auto bias_tensor = ConverterToNPUTensor(in_tensors[2]);
     if (bias_tensor == nullptr) {
-      MS_LOG(ERROR) << "new bias_tensor failed.";
+      MS_LOG(ERROR) << "Get bias_tensor failed.";
       return RET_ERROR;
     }
-    ge::TensorDesc bias_tensor_desc(ConverterToNPUShape({1, bias_shape[0], 1, 1}), ge::FORMAT_NCHW,
+    scale_tensor->SetTensorDesc(ge::TensorDesc(ConverterToNPUShape({1, bias_shape[0], 1, 1})));
-                                    ConverterToNPUDataType(in_tensors[2]->data_type()));
+
-    bias_tensor->SetTensorDesc(bias_tensor_desc);
-    bias_tensor->SetData(reinterpret_cast<const uint8_t *>(in_tensors[2]->data()), in_tensors[2]->Size());
     bias_ = new (std::nothrow) hiai::op::Const(name_ + "_beta");
     if (bias_ == nullptr) {
       MS_LOG(ERROR) << "New beta_ const failed.";

mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_winograd_fp16.cc

@@ -28,6 +28,9 @@ DeConvWinogradFp16CPUKernel::~DeConvWinogradFp16CPUKernel() {
 }
 
 void DeConvWinogradFp16CPUKernel::FreeResizeBuf() {
+  if (deconv_param_ == nullptr) {
+    return;
+  }
   for (int i = 0; i < deconv_param_->compute_size_; i++) {
     DeConvComputeUnit &unit = deconv_param_->compute_units_[i];
     if (unit.tmp_buffer_ != nullptr) {