forked from mindspore-Ecosystem/mindspore
!2942 change reduce node's reduce axis attr whem it's using in special foramt
Merge pull request !2942 from lianliguang/convert-reduce-axis-when-selected-hw-special-format-for-reduce-kernel
This commit is contained in:
commit
ee199007ed
|
@ -32,6 +32,8 @@
|
|||
|
||||
namespace mindspore {
|
||||
namespace kernel {
|
||||
constexpr char kAxis[] = "axis";
|
||||
constexpr char kTypeInt32[] = "Int32";
|
||||
const std::unordered_map<std::string, TypeId> type_id_maps = {
|
||||
{"float", TypeId::kNumberTypeFloat32}, {"float16", TypeId::kNumberTypeFloat16},
|
||||
{"float32", TypeId::kNumberTypeFloat32}, {"float64", TypeId::kNumberTypeFloat64},
|
||||
|
@ -989,5 +991,39 @@ void MultiThreadCompute(const MultiThreadComputeFunc &func, MultiThreadComputePa
|
|||
threads[i].join();
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<int> GetReduceAttrAxis(const CNodePtr &cnode) {
|
||||
if (AnfAlgo::GetInputTensorNum(cnode) != AnfAlgo::GetOutputTensorNum(cnode) &&
|
||||
AnfAlgo::GetInputTensorNum(cnode) != 1) {
|
||||
MS_LOG(EXCEPTION) << "the kind of reduce node [" << cnode->DebugString()
|
||||
<< "] is not single input or single output ";
|
||||
}
|
||||
std::vector<int> axis;
|
||||
auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(cnode, 0);
|
||||
auto primitive = AnfAlgo::GetCNodePrimitive(cnode);
|
||||
MS_EXCEPTION_IF_NULL(primitive);
|
||||
auto axis_attr = primitive->GetAttr(kAxis);
|
||||
if (axis_attr == nullptr) {
|
||||
MS_LOG(ERROR) << "This node does't have axie attr.";
|
||||
return std::vector<int>();
|
||||
}
|
||||
auto type = axis_attr->type();
|
||||
MS_EXCEPTION_IF_NULL(type);
|
||||
std::vector<int> axis_list;
|
||||
if (type->ToString() == kTypeInt32) {
|
||||
axis_list.emplace_back(GetValue<int>(axis_attr));
|
||||
} else {
|
||||
axis_list = GetValue<std::vector<int>>(axis_attr);
|
||||
}
|
||||
for (const auto &elem : axis_list) {
|
||||
if (elem < 0) {
|
||||
axis.emplace_back(input_shape.size() + elem);
|
||||
} else {
|
||||
axis.emplace_back(elem);
|
||||
}
|
||||
}
|
||||
AnfAlgo::SetNodeAttr(kAttrAxis, MakeValue(axis), cnode);
|
||||
return axis;
|
||||
}
|
||||
} // namespace kernel
|
||||
} // namespace mindspore
|
||||
|
|
|
@ -138,6 +138,7 @@ void ReduceMultiSparseGradient(const std::vector<std::shared_ptr<SparseGradient>
|
|||
size_t outer_dim);
|
||||
void TwoLevelReduceSparseGradient(const SparseGradient &origin_sparse_grad, SparseGradient *tmp_grad,
|
||||
SparseGradient *unique_grad, size_t first_dim, size_t outer_dim);
|
||||
std::vector<int> GetReduceAttrAxis(const CNodePtr &cnode);
|
||||
} // namespace kernel
|
||||
} // namespace mindspore
|
||||
|
||||
|
|
|
@ -20,11 +20,10 @@
|
|||
#include "utils/utils.h"
|
||||
#include "session/anf_runtime_algorithm.h"
|
||||
#include "kernel/tbe/tbe_kernel_select/common_utils.h"
|
||||
#include "kernel/common_utils.h"
|
||||
|
||||
namespace mindspore {
|
||||
namespace kernel {
|
||||
constexpr char kAxis[] = "axis";
|
||||
constexpr char kTypeInt32[] = "Int32";
|
||||
constexpr size_t kInputIndex_0 = 0;
|
||||
constexpr size_t kOutputIndex_0 = 0;
|
||||
constexpr size_t kChannelN = 0;
|
||||
|
@ -50,7 +49,7 @@ bool TbeKernelReduceSelecter::GetShapeInfo(SupportFormat *support_format) {
|
|||
// get keep dim attr
|
||||
GetReduceAttrKeepDim();
|
||||
// get axis attr
|
||||
GetReduceAttrAxis();
|
||||
axis_ = GetReduceAttrAxis(cnode_ptr_);
|
||||
AssignSupportFormat(kOpFormat_DEFAULT, support_format);
|
||||
return true;
|
||||
}
|
||||
|
@ -121,31 +120,6 @@ bool TbeKernelReduceSelecter::IsFracZAndC1HWNCoC0Common(const std::string &forma
|
|||
return true;
|
||||
}
|
||||
|
||||
void TbeKernelReduceSelecter::GetReduceAttrAxis() {
|
||||
auto primitive = AnfAlgo::GetCNodePrimitive(cnode_ptr_);
|
||||
MS_EXCEPTION_IF_NULL(primitive);
|
||||
auto axis = primitive->GetAttr(kAxis);
|
||||
if (axis == nullptr) {
|
||||
MS_LOG(INFO) << "This node does't have axie attr.";
|
||||
return;
|
||||
}
|
||||
auto type = axis->type();
|
||||
MS_EXCEPTION_IF_NULL(type);
|
||||
std::vector<int> axis_list;
|
||||
if (type->ToString() == kTypeInt32) {
|
||||
axis_list.emplace_back(GetValue<int>(axis));
|
||||
} else {
|
||||
axis_list = GetValue<std::vector<int>>(axis);
|
||||
}
|
||||
for (const auto &elem : axis_list) {
|
||||
if (elem < 0) {
|
||||
axis_.emplace_back(input_shape_.size() + elem);
|
||||
} else {
|
||||
axis_.emplace_back(IntToSize(elem));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void TbeKernelReduceSelecter::GetReduceAttrKeepDim() {
|
||||
if (!AnfAlgo::HasNodeAttr(kAttrKeepDims, cnode_ptr_)) {
|
||||
MS_LOG(INFO) << "This node does't have keep_attr.";
|
||||
|
|
|
@ -36,7 +36,6 @@ class TbeKernelReduceSelecter {
|
|||
|
||||
private:
|
||||
bool IsFracZAndC1HWNCoC0Common(const std::string &format, SupportFormat *support_format) const;
|
||||
void GetReduceAttrAxis();
|
||||
void GetReduceAttrKeepDim();
|
||||
void AssignSupportFormat(const std::string &support_format_str, SupportFormat *support_format) const;
|
||||
bool Is4DShape(const std::vector<size_t> &shape) const;
|
||||
|
@ -44,7 +43,7 @@ class TbeKernelReduceSelecter {
|
|||
CNodePtr cnode_ptr_;
|
||||
std::vector<size_t> input_shape_{};
|
||||
std::vector<size_t> output_shape_{};
|
||||
std::vector<size_t> axis_{};
|
||||
std::vector<int> axis_{};
|
||||
bool keep_dims_ = false;
|
||||
};
|
||||
} // namespace kernel
|
||||
|
|
|
@ -57,6 +57,7 @@
|
|||
#include "pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.h"
|
||||
#include "pre_activate/ascend/format_type/insert_trans_op.h"
|
||||
#include "pre_activate/ascend/format_type/rectify_do_mask_kernel_info.h"
|
||||
#include "pre_activate/ascend/format_type/chang_axis_of_reduce_kernel.h"
|
||||
#include "pre_activate/pass/getitem_tuple.h"
|
||||
#include "pre_activate/pass/optimize_dependence.h"
|
||||
#include "pre_activate/pass/erase_visit_attr.h"
|
||||
|
@ -158,6 +159,7 @@ void RunOpAscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_g
|
|||
MS_EXCEPTION_IF_NULL(kernel_graph);
|
||||
auto optimizer = std::make_shared<GraphOptimizer>();
|
||||
auto data_layout_pm = std::make_shared<PassManager>("pynative_transop_pm");
|
||||
data_layout_pm->AddPass(std::make_shared<ChangeAxisOfReduceKernel>());
|
||||
data_layout_pm->AddPass(std::make_shared<RectifyDoMaskKernelInfo>());
|
||||
data_layout_pm->AddPass(std::make_shared<RunOpInsertTransData>());
|
||||
data_layout_pm->AddPass(std::make_shared<GetitemTuple>());
|
||||
|
|
|
@ -0,0 +1,103 @@
|
|||
/**
|
||||
* Copyright 2019 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 "pre_activate/ascend/format_type/chang_axis_of_reduce_kernel.h"
|
||||
|
||||
#include <string>
|
||||
#include <memory>
|
||||
#include <vector>
|
||||
#include <map>
|
||||
|
||||
#include "utils/utils.h"
|
||||
#include "session/anf_runtime_algorithm.h"
|
||||
#include "common/utils.h"
|
||||
#include "kernel/common_utils.h"
|
||||
|
||||
namespace mindspore {
|
||||
namespace opt {
|
||||
namespace {
|
||||
using ConvertFunction = std::function<void(const CNodePtr &)>;
|
||||
|
||||
void ConvertReduceAttrFraczAnd6HD(const CNodePtr &cnode);
|
||||
const size_t kAxis_H = 2;
|
||||
const size_t kAxis_W = 3;
|
||||
const size_t kAxis_6HD_H = 1;
|
||||
const size_t kAxis_6HD_W = 2;
|
||||
const std::map<std::string, ConvertFunction> kReduceConvertMap = {{kOpFormat_FRAC_Z, ConvertReduceAttrFraczAnd6HD},
|
||||
{kOpFormat_C1HWNCoC0, ConvertReduceAttrFraczAnd6HD}};
|
||||
void SafeCheckFunction(const CNodePtr &cnode, const std::vector<int> &reduce_axis) {
|
||||
if (reduce_axis.empty()) {
|
||||
MS_LOG(EXCEPTION) << "The node " << cnode->DebugString() << "'s reduce axis got a empty vector";
|
||||
}
|
||||
if (AnfAlgo::GetInputTensorNum(cnode) != AnfAlgo::GetOutputTensorNum(cnode) &&
|
||||
AnfAlgo::GetInputTensorNum(cnode) != 1) {
|
||||
MS_LOG(EXCEPTION) << "the kind of reduce node [" << cnode->DebugString()
|
||||
<< "] is not single input or single output ";
|
||||
}
|
||||
for (auto elem : reduce_axis) {
|
||||
if (elem > 4) {
|
||||
MS_LOG(INFO) << "reduce axis is larger than 4 dims reduce axis : [" << elem << "]";
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void ConvertReduceAttrFraczAnd6HD(const CNodePtr &cnode) {
|
||||
auto axis = kernel::GetReduceAttrAxis(cnode);
|
||||
std::vector<int> convert_axis;
|
||||
SafeCheckFunction(cnode, axis);
|
||||
auto format = AnfAlgo::GetInputFormat(cnode, 0);
|
||||
if (format != kOpFormat_FRAC_Z || format != kOpFormat_C1HWNCoC0) {
|
||||
MS_LOG(EXCEPTION) << "The node [" << cnode->DebugString() << "] format " << format << " is not 5hd";
|
||||
}
|
||||
for (auto elem : axis) {
|
||||
switch (elem) {
|
||||
case kAxis_H:
|
||||
convert_axis.emplace_back(kAxis_6HD_H);
|
||||
break;
|
||||
case kAxis_W:
|
||||
convert_axis.emplace_back(kAxis_6HD_W);
|
||||
break;
|
||||
default:
|
||||
MS_LOG(INFO) << "reduce axis is axis : [" << elem << "]"
|
||||
<< " but the format is not supported this reduce axis";
|
||||
}
|
||||
}
|
||||
AnfAlgo::SetNodeAttr(kAttrAxis, MakeValue(convert_axis), cnode);
|
||||
}
|
||||
} // namespace
|
||||
|
||||
const BaseRef ChangeAxisOfReduceKernel::DefinePattern() const {
|
||||
VarPtr X = std::make_shared<Var>();
|
||||
VarPtr Xs = std::make_shared<SeqVar>();
|
||||
return VectorRef({X, Xs});
|
||||
}
|
||||
|
||||
const AnfNodePtr ChangeAxisOfReduceKernel::Process(const FuncGraphPtr &, const AnfNodePtr &node,
|
||||
const EquivPtr &) const {
|
||||
if (node == nullptr || !node->isa<CNode>() || !AnfAlgo::IsRealKernel(node)) {
|
||||
return nullptr;
|
||||
}
|
||||
if (AnfAlgo::GetOpPattern(node) != kernel::kReducePattern) {
|
||||
return nullptr;
|
||||
}
|
||||
auto convert_map = kReduceConvertMap.find(AnfAlgo::GetInputFormat(node, 0));
|
||||
if (convert_map == kReduceConvertMap.end()) {
|
||||
return nullptr;
|
||||
}
|
||||
convert_map->second(node->cast<CNodePtr>());
|
||||
return nullptr;
|
||||
}
|
||||
} // namespace opt
|
||||
} // namespace mindspore
|
|
@ -0,0 +1,33 @@
|
|||
/**
|
||||
* Copyright 2019 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_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_CHANGE_AXIS_OF_REDUCE_KENRNEL_H_
|
||||
#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_CHANGE_AXIS_OF_REDUCE_KENRNEL_H_
|
||||
|
||||
#include "pre_activate/common/optimizer.h"
|
||||
|
||||
namespace mindspore {
|
||||
namespace opt {
|
||||
class ChangeAxisOfReduceKernel : public PatternProcessPass {
|
||||
public:
|
||||
explicit ChangeAxisOfReduceKernel(bool multigraph = true)
|
||||
: PatternProcessPass("change_axis_of_reduce_kernel", multigraph) {}
|
||||
~ChangeAxisOfReduceKernel() override = default;
|
||||
const BaseRef DefinePattern() const override;
|
||||
const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
|
||||
};
|
||||
} // namespace opt
|
||||
} // namespace mindspore
|
||||
#endif // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_CHANGE_AXIS_OF_REDUCE_KENRNEL_H_
|
Loading…
Reference in New Issue