Nigel
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!43249 [MSLITE][CPU][r1.8] fix mul_add_fusion whether InferDone 

Merge pull request !43249 from Greatpan/transpose_r1.8
This commit is contained in:
i-robot 2022-10-08 09:29:04 +00:00 committed by Gitee
parent abd85adeb1 f0db19f1ee
commit dd45074a1f
No known key found for this signature in database
GPG Key ID: 173E9B9CA92EEF8F
25 changed files with 119 additions and 62 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/infer/matmul_infer.c
View File

@ -143,7 +143,8 @@ int MatmulInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC *
for (int i = 0; i < diff; ++i) {
ShapeInsert(in->shape_, &in->shape_size_, 0, 1);
}
SetDataTypeFormat(output, input0);
TensorC *input = input1->data_ == NULL ? input1 : input0; // transfer the input which comes from the other node.
SetDataTypeFormat(output, input);
if (parameter->quant_type_ == QuantType_QUANT_DYNAMIC || parameter->quant_type_ == QuantType_QUANT_WEIGHT) {
output->data_type_ = kNumberTypeFloat32;
}

15
mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/infer/transpose_infer.c
View File

@ -1,5 +1,5 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
* Copyright 2021-2022 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.
@ -66,21 +66,17 @@ int TransposeInferShape(const TensorC *const *inputs, size_t inputs_size, Tensor
const TensorC *perm_tensor = inputs[1];
const int32_t *perm_data = (int32_t *)perm_tensor->data_;
const int perms_num = perm_tensor->shape_[0];
if (perm_tensor->shape_size_ == 0) {
return NNACL_INFER_INVALID;
}
MS_CHECK_TRUE_RET(perm_tensor->shape_size_ != 0, NNACL_INFER_INVALID);
if (perms_num != 0 && perm_data == NULL) {
return NNACL_INFER_INVALID;
}
int perm[MAX_TRANSPOSE_DIM_SIZE] = {0};
size_t perm_size = 0;
for (int i = 0; i < perms_num; i++) {
if (perm_data[i] >= perms_num) {
return NNACL_ERR;
}
MS_CHECK_TRUE_RET(perm_data[i] < perms_num, NNACL_ERR);
ShapePush(perm, &perm_size, perm_data[i]);
}
if (perms_num == 4) {
if (perms_num == PERM_NUM_FOUR) {
const int nchw2nhwc[4] = {0, 2, 3, 1};
const int nhwc2nchw[4] = {0, 3, 1, 2};
const int trans3d[3] = {0, 2, 1};
@ -95,6 +91,9 @@ int TransposeInferShape(const TensorC *const *inputs, size_t inputs_size, Tensor
ShapeSet(perm, &perm_size, trans3d, 3);
}
}
if (perms_num == PERM_NUM_THREE && perm[0] == 0 && perm[1] == 2) {
output->format_ = input->format_ == Format_NCHW ? Format_NHWC : Format_NCHW;
}
if (parameter->quant_type_ == QuantType_QUANT_WEIGHT) {
output->data_type_ = kNumberTypeFloat32;
}

2
mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/transpose.h
View File

@ -22,6 +22,8 @@
// MAX_TRANSPOSE_SERIAL_SIZE = 64 * 3 * 512 * 512
#define MAX_TRANSPOSE_SERIAL_SIZE 50331648
#define MAX_TRANSPOSE_DIM_SIZE 20
#define PERM_NUM_THREE 3
#define PERM_NUM_FOUR 4
typedef struct TransposeParameter {
// primitive parameter

4
mindspore/lite/src/runtime/delegate/npu/npu_converter_utils.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -250,6 +250,8 @@ int ConverterToNPUActivationMode(schema::ActivationType type) {
return RELU6;
case schema::ActivationType_ELU:
return ELU;
case schema::ActivationType_GELU:
return GELU;
default:
return ACTIVATION_INVALID;
}

5
mindspore/lite/src/runtime/delegate/npu/op/activation_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -28,7 +28,8 @@ int ActivationNPUOp::IsSupport(const schema::Primitive *primitive, const std::ve
if (act_type_ != schema::ActivationType_RELU && act_type_ != schema::ActivationType_RELU6 &&
act_type_ != schema::ActivationType_SIGMOID && act_type_ != schema::ActivationType_TANH &&
act_type_ != schema::ActivationType_HSIGMOID && act_type_ != schema::ActivationType_LEAKY_RELU &&
act_type_ != schema::ActivationType_SWISH && act_type_ != schema::ActivationType_ELU) {
act_type_ != schema::ActivationType_SWISH && act_type_ != schema::ActivationType_ELU &&
act_type_ != schema::ActivationType_GELU) {
MS_LOG(WARNING) << "Unsupported activation type for activation op " << name_ << "when running npu";
return RET_NOT_SUPPORT;
}

61
mindspore/lite/src/runtime/delegate/npu/op/arithmetic_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -16,6 +16,9 @@
#include "src/runtime/delegate/npu/op/arithmetic_npu.h"
#include "src/runtime/delegate/npu/npu_converter_utils.h"
#include "src/runtime/delegate/delegate_utils.h"
#include "src/runtime/delegate/npu/transpose_kernel.h"
namespace mindspore {
constexpr int ARITHMETIC_INPUT_NUM = 2;
constexpr int MAX_HW_SIZE = 1664;
@ -24,20 +27,24 @@ int ArithmeticNPUOp::IsSupport(const schema::Primitive *primitive, const std::ve
auto in_shape_0 = in_tensors[0].Shape();
auto in_shape_1 = in_tensors[1].Shape();
auto out_shape = out_tensors[0].Shape();
if (in_shape_0.size() != 0 && in_shape_1.size() != 0 && in_shape_0.size() != in_shape_1.size()) {
MS_LOG(WARNING) << name_ << " for the two inputs, the dimension size must be same. size 1 is: " << in_shape_0.size()
<< " size 2 is: " << in_shape_1.size();
return RET_NOT_SUPPORT;
}
// a hidden limitation in npu bottom implementation
if (type_ == schema::PrimitiveType_MulFusion && out_shape.size() == NPU_SHAPE_SIZE) {
bool is_nhwc = out_tensors[0].format() == Format::NHWC;
auto out_h = is_nhwc ? out_shape.at(NHWC_H) : out_shape.at(NCHW_H);
auto out_w = is_nhwc ? out_shape.at(NHWC_W) : out_shape.at(NCHW_W);
// two inputs have different shape with the output, which means both of them need broadcast
if (in_shape_0 != out_shape && in_shape_1 != out_shape && out_h * out_w > MAX_HW_SIZE) {
MS_LOG(WARNING) << "The size of out_height * out_width is larger than the max value (1664) that npu supports "
"during broadcasting.";
if (type_ == schema::PrimitiveType_MulFusion) {
if (out_shape.size() == NPU_SHAPE_SIZE) {
bool is_nhwc = out_tensors[0].format() == Format::NHWC;
auto out_h = is_nhwc ? out_shape.at(NHWC_H) : out_shape.at(NCHW_H);
auto out_w = is_nhwc ? out_shape.at(NHWC_W) : out_shape.at(NCHW_W);
// two inputs have different shape with the output, which means both of them need broadcast
if (in_shape_0 != out_shape && in_shape_1 != out_shape && out_h * out_w > MAX_HW_SIZE) {
MS_LOG(WARNING) << "The size of out_height * out_width is larger than the max value (1664) that npu supports "
"during broadcasting.";
return RET_NOT_SUPPORT;
}
}
} else {
if (in_shape_0.size() != 0 && in_shape_1.size() != 0 && in_shape_0.size() != in_shape_1.size()) {
MS_LOG(WARNING) << name_
<< " for the two inputs, the dimension size must be same. size 1 is: " << in_shape_0.size()
<< " size 2 is: " << in_shape_1.size();
return RET_NOT_SUPPORT;
}
}
@ -240,6 +247,32 @@ ge::Operator *ArithmeticNPUOp::GetNPUOp() {
return act_;
}
int ArithmeticNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
for (size_t i = 0; i < inputs_.size(); i++) {
auto in_tensor = inputs_.at(i);
if (!in_tensor.IsConst() || in_tensor.Shape().size() != DIMENSION_4D) {
continue;
}
auto shape = in_tensor.Shape();
auto new_shape = {in_tensor.Shape().at(NHWC_N), in_tensor.Shape().at(NHWC_C), in_tensor.Shape().at(NHWC_H),
in_tensor.Shape().at(NHWC_W)};
auto nh2nc_tensor =
mindspore::MSTensor::CreateTensor(in_tensor.Name() + "_nh2nc", in_tensor.DataType(), new_shape, nullptr, 0);
if (nh2nc_tensor == nullptr) {
MS_LOG(ERROR) << "New nchw tensor failed when inserting nchw2nhwc op.";
return RET_ERROR;
}
auto dst_data = nh2nc_tensor->MutableData();
MS_CHECK_TRUE_RET(dst_data != nullptr, RET_ERROR);
// transpose dst_data to nchw.
PackNHWCToNCHWFp32(in_tensor.MutableData(), dst_data, shape[NHWC_N], shape[NHWC_H] * shape[NHWC_W], shape[NHWC_C]);
nh2nc_tensor->SetFormat(NCHW);
inputs_.at(i) = *nh2nc_tensor;
all_tensors->push_back(nh2nc_tensor);
}
return RET_OK;
}
ArithmeticNPUOp::~ArithmeticNPUOp() {
if (op_ != nullptr) {
delete op_;

4
mindspore/lite/src/runtime/delegate/npu/op/arithmetic_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -47,6 +47,8 @@ class ArithmeticNPUOp : public NPUOp {
ge::Operator *GetNPUOp() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
private:
int SetActivation();
schema::ActivationType act_type_ = schema::ActivationType_NO_ACTIVATION;

4
mindspore/lite/src/runtime/delegate/npu/op/concat_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -57,7 +57,7 @@ int ConcatNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors
ge::Operator *ConcatNPUOp::GetNPUOp() { return this->concat_; }
int ConcatNPUOp::HandleAxis() {
int ConcatNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
axis_ = TransFormAxis(axis_);
if (axis_ == NCHW_INVALID) {
MS_LOG(ERROR) << "Transform axis for concat op failed.";

4
mindspore/lite/src/runtime/delegate/npu/op/concat_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -45,7 +45,7 @@ class ConcatNPUOp : public NPUOp {
ge::Operator *GetNPUOp() override;
int HandleAxis() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
private:
hiai::op::ConcatD *concat_ = nullptr;

4
mindspore/lite/src/runtime/delegate/npu/op/npu_op.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -77,7 +77,7 @@ class NPUOp {
virtual ge::Operator *GetNPUOp() { return nullptr; }
virtual int HandleAxis() { return RET_OK; }
virtual int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) { return RET_OK; }
void set_inputs(const std::vector<mindspore::MSTensor> &in_tensors) { this->inputs_ = in_tensors; }

4
mindspore/lite/src/runtime/delegate/npu/op/pad_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -116,7 +116,7 @@ int PadNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
ge::Operator *PadNPUOp::GetNPUOp() { return this->pad_; }
int PadNPUOp::HandleAxis() {
int PadNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
if (paddings_vec_.size() != PAD_SIZE) {
return RET_ERROR;
}

4
mindspore/lite/src/runtime/delegate/npu/op/pad_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -42,7 +42,7 @@ class PadNPUOp : public NPUOp {
ge::Operator *GetNPUOp() override;
int HandleAxis() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
private:
hiai::op::PadV2 *pad_ = nullptr;

4
mindspore/lite/src/runtime/delegate/npu/op/reduce_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -88,7 +88,7 @@ int ReduceNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors
ge::Operator *ReduceNPUOp::GetNPUOp() { return this->reduce_; }
int ReduceNPUOp::HandleAxis() {
int ReduceNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
auto reduce_axes = inputs_.at(1);
int num_axes = reduce_axes.Shape().at(0);
MS_CHECK_TRUE_RET(reduce_axes.MutableData() != nullptr, RET_ERROR);

4
mindspore/lite/src/runtime/delegate/npu/op/reduce_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -42,7 +42,7 @@ class ReduceNPUOp : public NPUOp {
ge::Operator *GetNPUOp() override;
int HandleAxis() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
private:
schema::ReduceMode reduce_mode_ = schema::ReduceMode_ReduceMean;

4
mindspore/lite/src/runtime/delegate/npu/op/softmax_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -47,7 +47,7 @@ int SoftmaxNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensor
ge::Operator *SoftmaxNPUOp::GetNPUOp() { return this->softmax_; }
int SoftmaxNPUOp::HandleAxis() {
int SoftmaxNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
axis_ = TransFormAxis(axis_);
if (axis_ == NCHW_INVALID) {
MS_LOG(ERROR) << "Transform axis for Softmax op failed.";

4
mindspore/lite/src/runtime/delegate/npu/op/softmax_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -42,7 +42,7 @@ class SoftmaxNPUOp : public NPUOp {
const std::vector<mindspore::MSTensor> &out_tensors,
const std::vector<ge::Operator *> &npu_inputs) override;
int HandleAxis() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
ge::Operator *GetNPUOp() override;

4
mindspore/lite/src/runtime/delegate/npu/op/split_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -86,7 +86,7 @@ int SplitNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
ge::Operator *SplitNPUOp::GetNPUOp() { return this->split_; }
int SplitNPUOp::HandleAxis() {
int SplitNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
axis_ = TransFormAxis(axis_);
if (axis_ == NCHW_INVALID) {
MS_LOG(ERROR) << "Transform axis for split op failed.";

4
mindspore/lite/src/runtime/delegate/npu/op/split_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -42,7 +42,7 @@ class SplitNPUOp : public NPUOp {
const std::vector<mindspore::MSTensor> &out_tensors,
const std::vector<ge::Operator *> &npu_inputs) override;
int HandleAxis() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
ge::Operator *GetNPUOp() override;

4
mindspore/lite/src/runtime/delegate/npu/op/strided_slice_npu.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -110,7 +110,7 @@ ge::Operator *StridedSliceNPUOp::GetNPUOp() {
}
}
int StridedSliceNPUOp::HandleAxis() {
int StridedSliceNPUOp::HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) {
if (inputs_.size() < MIN_INPUT_SIZE) {
MS_LOG(ERROR) << "StridedSlice in tensors size < " << MIN_INPUT_SIZE;
return RET_ERROR;

4
mindspore/lite/src/runtime/delegate/npu/op/strided_slice_npu.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020-2021 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -49,7 +49,7 @@ class StridedSliceNPUOp : public NPUOp {
ge::Operator *GetNPUOp() override;
int HandleAxis() override;
int HandleAxisAndConstantInputs(std::vector<mindspore::MSTensor *> *all_tensors) override;
private:
hiai::op::StridedSlice *strided_slice_ = nullptr;

7
mindspore/lite/src/runtime/delegate/npu/pass/npu_fusion_pass.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
* Copyright 2021-2022 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.
@ -248,9 +248,10 @@ int NPUFusionPass::CommonFusion(NPUOp *cur_op) {
MS_LOG(ERROR) << "UpdateOp failed.";
return RET_ERROR;
}
ret = cur_op->HandleAxis();
auto all_tensors = subgraph_->GetInsertTensors();
ret = cur_op->HandleAxisAndConstantInputs(all_tensors);
if (ret != RET_OK) {
MS_LOG(ERROR) << "HandleAxis failed.";
MS_LOG(ERROR) << "HandleAxisAndConstantInputs failed.";
return ret;
}
return RET_OK;

4
mindspore/lite/tools/converter/parser/onnx/onnx_matmul_parser.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
* Copyright 2020-2022 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.
@ -40,7 +40,7 @@ PrimitiveCPtr OnnxMatmulParser::Parse(const onnx::GraphProto &onnx_graph, const
}
}
if (!FloatCompare(alpha, 1.0f) || (!FloatCompare(beta, 1.0f) && !(onnx_node.input().size() == 2 &&
!FloatCompare(beta)))) { // 2: input num is A and B
FloatCompare(beta)))) { // 2: input num is A and B
MS_LOG(ERROR) << "not support alpha * A * B + beta * C";
return nullptr;
}

17
mindspore/lite/tools/optimizer/fusion/mul_add_fusion.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
* Copyright 2021-2022 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.
@ -189,6 +189,15 @@ bool MulAddFusion::ScaleInputShapeValid(size_t *axis_offset) const {
return true;
}
bool MulAddFusion::MulInputAnodeIsInferred(const AnfNodePtr &mul_input_anode) const {
auto mul_input_cnode = mul_input_anode->cast<CNodePtr>();
MS_EXCEPTION_IF_NULL(mul_input_cnode);
auto prim = GetValueNode<PrimitivePtr>(mul_input_cnode->input(0));
MS_CHECK_TRUE_RET(prim != nullptr, false);
auto is_inferred = prim->GetAttr(kInferDone) != nullptr && GetValue<bool>(prim->GetAttr(kInferDone));
return is_inferred;
}
AnfNodePtr MulAddFusion::Process(const std::string &pattern_name, const mindspore::FuncGraphPtr &func_graph,
const mindspore::AnfNodePtr &node, const mindspore::EquivPtr &equiv) const {
if (func_graph == nullptr || node == nullptr) {
@ -232,6 +241,12 @@ AnfNodePtr MulAddFusion::Process(const std::string &pattern_name, const mindspor
}
MS_CHECK_TRUE_RET(mul_input_anode != nullptr, nullptr);
if (mul_input_anode->isa<CNode>()) {
if (!MulInputAnodeIsInferred(mul_input_anode)) {
MS_LOG(DEBUG) << "mul_input_anode is not inferred, don't perform the ScaleInputShapeValid method.";
return nullptr;
}
}
if (FetchShapeFromAbstract(mul_input_anode->abstract(), &mul_input_shape_) != lite::RET_OK) {
return nullptr;
}

3
mindspore/lite/tools/optimizer/fusion/mul_add_fusion.h
View File

@ -1,5 +1,5 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
* Copyright 2021-2022 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.
@ -45,6 +45,7 @@ class MulAddFusion : public MultiplePatternProcessPass {
bool CheckAddNode(const mindspore::CNodePtr &cnode) const;
bool CheckMulNode(const mindspore::FuncGraphPtr &func_graph, const mindspore::CNodePtr &cnode) const;
bool ScaleInputShapeValid(size_t *axis_offset) const;
bool MulInputAnodeIsInferred(const AnfNodePtr &mul_input_anode) const;
bool AdjustScaleBiasTensorShape(size_t *axis_offset) const;
private:

4
mindspore/lite/tools/optimizer/graph/decrease_transpose_algo.cc
View File

@ -1,5 +1,5 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
* Copyright 2021-2022 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.
@ -332,7 +332,7 @@ STATUS DecreaseTransposeAlgo::DoPreInsert(const FuncGraphPtr &func_graph, const
auto HandleFunc = [this, &shape](const FuncGraphPtr &func_graph, const CNodePtr &cnode, size_t index,
FormatTransNodeType trans_type) -> STATUS {
auto before_perm = trans_type == kNHWC2NCHW ? kNH2NC : kNC2NH;
if (shape.size() == kInputSizeFour && !cnode->input(index)->isa<CNode>()) {
if (!cnode->input(index)->isa<CNode>()) {
if (ConvertTensorToNCOrNH(func_graph, cnode, index, fmk_type_, train_flag_, trans_type) != lite::RET_OK) {
MS_LOG(ERROR) << "ConvertTensorToNCOrNH failed.";
return lite::RET_ERROR;