remove anakin from code, test=develop (#22420)
石晓伟
5 years ago
committed by
GitHub
parent
0a678ca045
commit
e1b0d7cbb1
@ -1,45 +0,0 @@
|
|||||||
set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
|
|
||||||
find_path(ANAKIN_INCLUDE_DIR anakin_config.h
|
|
||||||
PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
|
|
||||||
$ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
|
|
||||||
NO_DEFAULT_PATH
|
|
||||||
)
|
|
||||||
|
|
||||||
find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
|
|
||||||
PATHS ${ANAKIN_ROOT}
|
|
||||||
$ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
|
|
||||||
NO_DEFAULT_PATH
|
|
||||||
DOC "Path to ANAKIN library.")
|
|
||||||
|
|
||||||
if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
|
|
||||||
set(ANAKIN_FOUND ON)
|
|
||||||
else()
|
|
||||||
set(ANAKIN_FOUND OFF)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
if(ANAKIN_FOUND)
|
|
||||||
message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
|
|
||||||
include_directories(${ANAKIN_ROOT})
|
|
||||||
include_directories(${ANAKIN_ROOT}/include)
|
|
||||||
include_directories(${ANAKIN_ROOT}/saber)
|
|
||||||
link_directories(${ANAKIN_ROOT})
|
|
||||||
add_definitions(-DPADDLE_WITH_ANAKIN)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
if(ANAKIN_FOUND)
|
|
||||||
if (ANAKIN_MLU AND NOT WITH_GPU AND NOT ANAKIN_X86)
|
|
||||||
message(STATUS "Compile with anakin mlu place.")
|
|
||||||
add_definitions(-DANAKIN_MLU_PLACE)
|
|
||||||
elseif(ANAKIN_BM AND NOT WITH_GPU AND NOT ANAKIN_X86)
|
|
||||||
message(STATUS "Compile with anakin bm place.")
|
|
||||||
add_definitions(-DANAKIN_BM_PLACE)
|
|
||||||
elseif(ANAKIN_X86)
|
|
||||||
message(STATUS "Compile with anakin x86 place.")
|
|
||||||
add_definitions(-DANAKIN_X86_PLACE)
|
|
||||||
endif()
|
|
||||||
endif()
|
|
||||||
|
|
||||||
if(ANAKIN_FOUND AND WITH_GPU AND WITH_DSO)
|
|
||||||
message(STATUS "Compile with anakin subgraph.")
|
|
||||||
set(ANAKIN_SUBGRAPH ON)
|
|
||||||
endif()
|
|
||||||
@ -1,83 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 <memory>
|
|
||||||
#include <string>
|
|
||||||
|
|
||||||
#include "paddle/fluid/framework/ir/fillconstant_elementwisemul_fuse.h"
|
|
||||||
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace framework {
|
|
||||||
namespace ir {
|
|
||||||
|
|
||||||
#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern);
|
|
||||||
#define GET_NODES \
|
|
||||||
GET_IR_NODE(fill_constant); \
|
|
||||||
GET_IR_NODE(fill_constant_out); \
|
|
||||||
GET_IR_NODE(elementwise_mul); \
|
|
||||||
GET_IR_NODE(elementwise_mul_out);
|
|
||||||
|
|
||||||
void FillconstantElementwisemulFuse::ApplyImpl(ir::Graph* graph) const {
|
|
||||||
const std::string pattern_name = "fillconstant_elementwisemul_fuse";
|
|
||||||
FusePassBase::Init(pattern_name, graph);
|
|
||||||
|
|
||||||
GraphPatternDetector gpd;
|
|
||||||
auto* x = gpd.mutable_pattern()
|
|
||||||
->NewNode("x")
|
|
||||||
->assert_is_op_input("elementwise_mul", "X")
|
|
||||||
->AsInput();
|
|
||||||
|
|
||||||
patterns::FillConstantElementWiseMulFuse pattern(gpd.mutable_pattern(),
|
|
||||||
pattern_name);
|
|
||||||
pattern(x);
|
|
||||||
|
|
||||||
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
|
|
||||||
Graph* g) {
|
|
||||||
GET_NODES;
|
|
||||||
|
|
||||||
PADDLE_ENFORCE(subgraph.count(x));
|
|
||||||
auto* elementwise_in = subgraph.at(x);
|
|
||||||
float constant_value =
|
|
||||||
boost::get<float>(fill_constant->Op()->GetAttr("value"));
|
|
||||||
|
|
||||||
framework::OpDesc new_op_desc;
|
|
||||||
new_op_desc.SetType("scale");
|
|
||||||
new_op_desc.SetInput("X", {elementwise_in->Name()});
|
|
||||||
new_op_desc.SetAttr("scale", constant_value);
|
|
||||||
new_op_desc.SetAttr("bias", static_cast<float>(0.0));
|
|
||||||
new_op_desc.SetAttr("bias_after_scale", true);
|
|
||||||
new_op_desc.SetOutput("Out", {elementwise_mul_out->Name()});
|
|
||||||
new_op_desc.Flush();
|
|
||||||
|
|
||||||
// Create a new node for the fused op.
|
|
||||||
auto* scale_op = graph->CreateOpNode(&new_op_desc);
|
|
||||||
|
|
||||||
IR_NODE_LINK_TO(elementwise_in, scale_op); // Input
|
|
||||||
IR_NODE_LINK_TO(scale_op, elementwise_mul_out); // Output
|
|
||||||
|
|
||||||
// Delete the unneeded nodes.
|
|
||||||
GraphSafeRemoveNodes(graph,
|
|
||||||
{fill_constant, fill_constant_out, elementwise_mul});
|
|
||||||
};
|
|
||||||
|
|
||||||
gpd(graph, handler);
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace ir
|
|
||||||
} // namespace framework
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_PASS(fillconstant_elementwisemul_fuse,
|
|
||||||
paddle::framework::ir::FillconstantElementwisemulFuse);
|
|
||||||
@ -1,233 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 <string>
|
|
||||||
#include <vector>
|
|
||||||
|
|
||||||
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
|
|
||||||
#include "paddle/fluid/framework/ir/node.h"
|
|
||||||
#include "paddle/fluid/framework/ir/simplify_anakin_priorbox_detection_out_pass.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace framework {
|
|
||||||
namespace ir {
|
|
||||||
|
|
||||||
void RunSimplifyAnakinDetection(ir::Graph *graph, int times, bool is_density,
|
|
||||||
bool is_reshape) {
|
|
||||||
const std::string pattern_name =
|
|
||||||
"simplify_anakin_detection_pattern_pass" + std::to_string(times);
|
|
||||||
std::string priorbox_type = is_density ? "density_prior_box" : "prior_box";
|
|
||||||
|
|
||||||
GraphPatternDetector gpd;
|
|
||||||
std::vector<PDNode *> input_nodes;
|
|
||||||
for (int i = 0; i < times; i++) {
|
|
||||||
input_nodes.push_back(gpd.mutable_pattern()
|
|
||||||
->NewNode("x" + std::to_string(i))
|
|
||||||
->assert_is_op_input(priorbox_type, "Input")
|
|
||||||
->AsInput());
|
|
||||||
}
|
|
||||||
input_nodes.push_back(gpd.mutable_pattern()
|
|
||||||
->NewNode("x" + std::to_string(times))
|
|
||||||
->assert_is_op_input("box_coder", "TargetBox")
|
|
||||||
->AsInput());
|
|
||||||
|
|
||||||
input_nodes.push_back(gpd.mutable_pattern()
|
|
||||||
->NewNode("x" + std::to_string(times + 1))
|
|
||||||
->assert_is_op_input("transpose2")
|
|
||||||
->AsInput());
|
|
||||||
|
|
||||||
patterns::AnakinDetectionPattern pattern(gpd.mutable_pattern(), pattern_name);
|
|
||||||
pattern(input_nodes, times, priorbox_type, is_reshape);
|
|
||||||
|
|
||||||
auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
|
|
||||||
Graph *g) {
|
|
||||||
const int kNumFields = 7;
|
|
||||||
const int kPriorBoxLocOffset = 1;
|
|
||||||
const int kReshape1Offset = 2;
|
|
||||||
const int kReshape1OutOffset = 3;
|
|
||||||
const int kPriorBoxVarOffset = 4;
|
|
||||||
const int kReshape2Offset = 5;
|
|
||||||
const int kReshape2OutOffset = 6;
|
|
||||||
std::vector<Node *> nodes;
|
|
||||||
|
|
||||||
for (int i = 0; i < times; i++) {
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("prior_box" + std::to_string(i))));
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("box_out" + std::to_string(i))));
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape1" + std::to_string(i))));
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape1_out" + std::to_string(i))));
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape2" + std::to_string(i))));
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape2_out" + std::to_string(i))));
|
|
||||||
|
|
||||||
PADDLE_ENFORCE(
|
|
||||||
subgraph.at(pattern.GetPDNode("box_var_out" + std::to_string(i))));
|
|
||||||
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("prior_box" + std::to_string(i))));
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("box_out" + std::to_string(i))));
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape1" + std::to_string(i))));
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape1_out" + std::to_string(i))));
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("box_var_out" + std::to_string(i))));
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape2" + std::to_string(i))));
|
|
||||||
nodes.push_back(
|
|
||||||
subgraph.at(pattern.GetPDNode("reshape2_out" + std::to_string(i))));
|
|
||||||
}
|
|
||||||
|
|
||||||
Node *concat_op1 = subgraph.at(pattern.GetPDNode("concat1"));
|
|
||||||
Node *concat_out1 = subgraph.at(pattern.GetPDNode("concat1_out"));
|
|
||||||
|
|
||||||
Node *concat_op2 = subgraph.at(pattern.GetPDNode("concat2"));
|
|
||||||
Node *concat_out2 = subgraph.at(pattern.GetPDNode("concat2_out"));
|
|
||||||
|
|
||||||
Node *box_coder_third_input = subgraph.at(input_nodes[times]);
|
|
||||||
Node *box_coder_op = subgraph.at(pattern.GetPDNode("box_coder"));
|
|
||||||
Node *box_coder_out = subgraph.at(pattern.GetPDNode("box_coder_out"));
|
|
||||||
|
|
||||||
Node *multiclass_nms_second_input = subgraph.at(input_nodes[times + 1]);
|
|
||||||
Node *transpose_before_nms =
|
|
||||||
subgraph.at(pattern.GetPDNode("transpose_before_nms"));
|
|
||||||
Node *transpose_before_nms_out =
|
|
||||||
subgraph.at(pattern.GetPDNode("transpose_before_nms_out"));
|
|
||||||
|
|
||||||
Node *multiclass_nms = subgraph.at(pattern.GetPDNode("multiclass_nms"));
|
|
||||||
Node *multiclass_nms_out =
|
|
||||||
subgraph.at(pattern.GetPDNode("multiclass_nms_out"));
|
|
||||||
|
|
||||||
std::string code_type =
|
|
||||||
boost::get<std::string>(box_coder_op->Op()->GetAttr("code_type"));
|
|
||||||
bool box_normalized =
|
|
||||||
boost::get<bool>(box_coder_op->Op()->GetAttr("box_normalized"));
|
|
||||||
|
|
||||||
int background_label =
|
|
||||||
boost::get<int>(multiclass_nms->Op()->GetAttr("background_label"));
|
|
||||||
float score_threshold =
|
|
||||||
boost::get<float>(multiclass_nms->Op()->GetAttr("score_threshold"));
|
|
||||||
int nms_top_k = boost::get<int>(multiclass_nms->Op()->GetAttr("nms_top_k"));
|
|
||||||
float nms_threshold =
|
|
||||||
boost::get<float>(multiclass_nms->Op()->GetAttr("nms_threshold"));
|
|
||||||
float nms_eta = boost::get<float>(multiclass_nms->Op()->GetAttr("nms_eta"));
|
|
||||||
int keep_top_k =
|
|
||||||
boost::get<int>(multiclass_nms->Op()->GetAttr("keep_top_k"));
|
|
||||||
|
|
||||||
std::vector<std::string> concat1_input_names;
|
|
||||||
for (int i = 0; i < times; i++) {
|
|
||||||
concat1_input_names.push_back(
|
|
||||||
nodes[i * kNumFields + kPriorBoxLocOffset]->Name());
|
|
||||||
}
|
|
||||||
|
|
||||||
framework::OpDesc concat1_desc;
|
|
||||||
concat1_desc.SetType("concat");
|
|
||||||
concat1_desc.SetInput("X", concat1_input_names);
|
|
||||||
concat1_desc.SetAttr("axis", 2);
|
|
||||||
concat1_desc.SetOutput("Out", {concat_out1->Name()});
|
|
||||||
|
|
||||||
auto *new_add_concat_op = graph->CreateOpNode(&concat1_desc);
|
|
||||||
|
|
||||||
for (int i = 0; i < times; i++) {
|
|
||||||
nodes[i * kNumFields + kPriorBoxLocOffset]->outputs.push_back(
|
|
||||||
new_add_concat_op);
|
|
||||||
new_add_concat_op->inputs.push_back(
|
|
||||||
nodes[i * kNumFields + kPriorBoxLocOffset]);
|
|
||||||
}
|
|
||||||
|
|
||||||
framework::OpDesc new_op_desc;
|
|
||||||
new_op_desc.SetType("detection_out");
|
|
||||||
new_op_desc.SetInput("PriorBox", {concat_out1->Name()});
|
|
||||||
new_op_desc.SetInput("TargetBox", {box_coder_third_input->Name()});
|
|
||||||
new_op_desc.SetInput("Scores", {multiclass_nms_second_input->Name()});
|
|
||||||
new_op_desc.SetAttr("code_type", code_type);
|
|
||||||
new_op_desc.SetAttr("box_normalized", box_normalized);
|
|
||||||
new_op_desc.SetAttr("background_label", background_label);
|
|
||||||
new_op_desc.SetAttr("score_threshold", score_threshold);
|
|
||||||
new_op_desc.SetAttr("nms_top_k", nms_top_k);
|
|
||||||
new_op_desc.SetAttr("nms_threshold", nms_threshold);
|
|
||||||
new_op_desc.SetAttr("nms_eta", nms_eta);
|
|
||||||
new_op_desc.SetAttr("keep_top_k", keep_top_k);
|
|
||||||
new_op_desc.SetOutput("Out", {multiclass_nms_out->Name()});
|
|
||||||
new_op_desc.Flush();
|
|
||||||
|
|
||||||
// Create a new node for the fused op.
|
|
||||||
auto *detection_out_op = graph->CreateOpNode(&new_op_desc);
|
|
||||||
|
|
||||||
std::unordered_set<const Node *> delete_nodes;
|
|
||||||
|
|
||||||
for (int i = 0; i < times; i++) {
|
|
||||||
nodes[i * kNumFields + kPriorBoxLocOffset]->outputs.push_back(concat_op1);
|
|
||||||
delete_nodes.insert(nodes[i * kNumFields + kReshape1Offset]);
|
|
||||||
delete_nodes.insert(nodes[i * kNumFields + kReshape1OutOffset]);
|
|
||||||
delete_nodes.insert(nodes[i * kNumFields + kPriorBoxVarOffset]);
|
|
||||||
delete_nodes.insert(nodes[i * kNumFields + kReshape2Offset]);
|
|
||||||
delete_nodes.insert(nodes[i * kNumFields + kReshape2OutOffset]);
|
|
||||||
}
|
|
||||||
|
|
||||||
delete_nodes.insert(concat_op1);
|
|
||||||
delete_nodes.insert(concat_op2);
|
|
||||||
delete_nodes.insert(concat_out2);
|
|
||||||
delete_nodes.insert(box_coder_op);
|
|
||||||
delete_nodes.insert(box_coder_out);
|
|
||||||
delete_nodes.insert(transpose_before_nms);
|
|
||||||
delete_nodes.insert(transpose_before_nms_out);
|
|
||||||
delete_nodes.insert(multiclass_nms);
|
|
||||||
|
|
||||||
new_add_concat_op->outputs.push_back(concat_out1);
|
|
||||||
concat_out1->inputs.push_back(new_add_concat_op);
|
|
||||||
|
|
||||||
detection_out_op->inputs.push_back(concat_out1);
|
|
||||||
detection_out_op->inputs.push_back(box_coder_third_input);
|
|
||||||
detection_out_op->inputs.push_back(multiclass_nms_second_input);
|
|
||||||
detection_out_op->outputs.push_back(multiclass_nms_out);
|
|
||||||
|
|
||||||
concat_out1->outputs.push_back(detection_out_op);
|
|
||||||
box_coder_third_input->outputs.push_back(detection_out_op);
|
|
||||||
multiclass_nms_second_input->outputs.push_back(detection_out_op);
|
|
||||||
multiclass_nms_out->inputs.push_back(detection_out_op);
|
|
||||||
|
|
||||||
// Delete the unneeded nodes.
|
|
||||||
GraphSafeRemoveNodes(graph, delete_nodes);
|
|
||||||
};
|
|
||||||
|
|
||||||
gpd(graph, handler);
|
|
||||||
}
|
|
||||||
|
|
||||||
void SimplifyAnakinDetectionPatternPass::ApplyImpl(ir::Graph *graph) const {
|
|
||||||
const int pattern_nums = 6;
|
|
||||||
const std::string pattern_name = "simplify_anakin_detection_pattern_pass";
|
|
||||||
FusePassBase::Init(pattern_name, graph);
|
|
||||||
std::vector<bool> options = {true, false};
|
|
||||||
for (const auto &is_density : options) {
|
|
||||||
for (const auto &is_reshape : options) {
|
|
||||||
for (int i = 1; i <= pattern_nums; i++) {
|
|
||||||
RunSimplifyAnakinDetection(graph, i, is_density, is_reshape);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace ir
|
|
||||||
} // namespace framework
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
typedef paddle::framework::ir::SimplifyAnakinDetectionPatternPass
|
|
||||||
priorbox_pattern;
|
|
||||||
REGISTER_PASS(simplify_anakin_priorbox_detection_out_pass, priorbox_pattern);
|
|
||||||
@ -1,39 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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.
|
|
||||||
|
|
||||||
#pragma once
|
|
||||||
#include <memory>
|
|
||||||
#include <unordered_set>
|
|
||||||
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
|
|
||||||
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace framework {
|
|
||||||
namespace ir {
|
|
||||||
|
|
||||||
// There may be many transpose-flatten structures in a model, and the output of
|
|
||||||
// these structures will be used as inputs to the concat Op. This pattern will
|
|
||||||
// be detected by our pass. The times here represents the repeat times of this
|
|
||||||
// structure.
|
|
||||||
class SimplifyAnakinDetectionPatternPass : public FusePassBase {
|
|
||||||
public:
|
|
||||||
virtual ~SimplifyAnakinDetectionPatternPass() {}
|
|
||||||
|
|
||||||
protected:
|
|
||||||
void ApplyImpl(ir::Graph* graph) const override;
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace ir
|
|
||||||
} // namespace framework
|
|
||||||
} // namespace paddle
|
|
||||||
@ -1,5 +0,0 @@
|
|||||||
cc_library(anakin_engine SRCS engine.cc DEPS framework_proto boost)
|
|
||||||
cc_library(anakin_op_teller SRCS op_teller.cc DEPS framework_proto device_context boost)
|
|
||||||
target_link_libraries(anakin_engine anakin anakin_saber_common)
|
|
||||||
cc_test(test_anakin_engine SRCS test_anakin_engine.cc DEPS anakin_engine)
|
|
||||||
add_subdirectory(convert)
|
|
||||||
@ -1,23 +0,0 @@
|
|||||||
cc_library(anakin_op_converter SRCS fc.cc conv2d.cc conv2d_fusion.cc
|
|
||||||
elementwise.cc activation.cc pool2d.cc concat.cc split.cc relu.cc softmax.cc
|
|
||||||
batch_norm.cc reshape.cc flatten.cc transpose.cc density_prior_box.cc
|
|
||||||
detection_out.cc scale.cc dropout.cc im2sequence.cc sum.cc affine_channel.cc
|
|
||||||
roi_align.cc shuffle_channel.cc helper.cc DEPS anakin_engine framework_proto
|
|
||||||
scope op_registry gtest gflags)
|
|
||||||
|
|
||||||
cc_test(test_anakin_fc SRCS test_fc_op.cc DEPS anakin_op_converter mul_op)
|
|
||||||
cc_test(test_anakin_conv2d SRCS test_conv2d_op.cc DEPS anakin_op_converter conv_op im2col vol2col depthwise_conv)
|
|
||||||
cc_test(test_anakin_activation SRCS test_activation_op.cc DEPS activation_op anakin_op_converter)
|
|
||||||
cc_test(test_anakin_pool2d SRCS test_pool2d_op.cc DEPS anakin_op_converter pool_op pooling)
|
|
||||||
cc_test(test_anakin_concat SRCS test_concat_op.cc DEPS anakin_op_converter concat_op concat_and_split)
|
|
||||||
cc_test(test_anakin_split SRCS test_split_op.cc DEPS anakin_op_converter split_op concat_and_split)
|
|
||||||
cc_test(test_anakin_elementwise SRCS test_elementwise_op.cc DEPS anakin_op_converter elementwise_add_op elementwise_mul_op)
|
|
||||||
cc_test(test_anakin_relu SRCS test_relu_op.cc DEPS activation_op anakin_op_converter)
|
|
||||||
cc_test(test_anakin_softmax SRCS test_softmax_op.cc DEPS anakin_op_converter softmax_op softmax)
|
|
||||||
cc_test(test_anakin_reshape SRCS test_reshape_op.cc DEPS anakin_op_converter reshape_op)
|
|
||||||
cc_test(test_anakin_flatten SRCS test_flatten_op.cc DEPS anakin_op_converter flatten_op reshape_op)
|
|
||||||
cc_test(test_anakin_transpose SRCS test_transpose_op.cc DEPS anakin_op_converter transpose_op)
|
|
||||||
cc_test(test_anakin_batch_norm SRCS test_batch_norm_op.cc DEPS anakin_op_converter batch_norm_op)
|
|
||||||
cc_test(test_anakin_dropout SRCS test_dropout_op.cc DEPS anakin_op_converter dropout_op)
|
|
||||||
cc_test(test_anakin_sum SRCS test_sum_op.cc DEPS anakin_op_converter sum_op selected_rows_functor)
|
|
||||||
cc_test(test_anakin_affine_channel SRCS test_affine_channel_op.cc DEPS anakin_op_converter affine_channel_op)
|
|
||||||
@ -1,64 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 "paddle/fluid/inference/anakin/convert/activation.h"
|
|
||||||
#include <algorithm>
|
|
||||||
#include <map>
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
ActivationOpConverter<TargetT, PrecisionT>::ActivationOpConverter(
|
|
||||||
const std::string &op_type)
|
|
||||||
: op_type_(op_type) {
|
|
||||||
auto it = anakin_op_types_.find(op_type_);
|
|
||||||
PADDLE_ENFORCE(it != anakin_op_types_.end(),
|
|
||||||
"activation op type is not support");
|
|
||||||
anakin_op_type_ = it->second;
|
|
||||||
}
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
void ActivationOpConverter<TargetT, PrecisionT>::operator()(
|
|
||||||
const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope, bool test_mode) {
|
|
||||||
framework::OpDesc op_desc(op, nullptr);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
|
|
||||||
|
|
||||||
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
|
|
||||||
auto input_name = op_desc.Input("X").front();
|
|
||||||
auto output_name = op_desc.Output("Out").front();
|
|
||||||
this->engine_->AddOp(op_name, "Activation", {input_name}, {output_name});
|
|
||||||
this->engine_->AddOpAttr(op_name, "type", anakin_op_type_);
|
|
||||||
|
|
||||||
if (op_type_ == "swish") {
|
|
||||||
float beta = boost::get<float>(op_desc.GetAttr("beta"));
|
|
||||||
this->engine_->AddOpAttr(op_name, "clip_relu_num", beta);
|
|
||||||
}
|
|
||||||
if (op_type_ == "relu6") {
|
|
||||||
float threshold = boost::get<float>(op_desc.GetAttr("threshold"));
|
|
||||||
this->engine_->AddOpAttr(op_name, "clip_relu_num", threshold);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(sigmoid, SigmoidOpConverter);
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(tanh, TanhOpConverter);
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(swish, SwishOpConverter);
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(relu6, Relu6OpConverter);
|
|
||||||
@ -1,72 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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.
|
|
||||||
|
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include <map>
|
|
||||||
#include <string>
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class ActivationOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
explicit ActivationOpConverter(const std::string &op_type);
|
|
||||||
|
|
||||||
virtual void operator()(const framework::proto::OpDesc &op,
|
|
||||||
const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope,
|
|
||||||
bool test_mode) override;
|
|
||||||
virtual ~ActivationOpConverter() {}
|
|
||||||
|
|
||||||
private:
|
|
||||||
std::string op_type_;
|
|
||||||
std::string anakin_op_type_;
|
|
||||||
std::map<std::string, std::string> anakin_op_types_{{"tanh", "TanH"},
|
|
||||||
{"sigmoid", "Sigmoid"},
|
|
||||||
{"relu6", "ClippedRelu"},
|
|
||||||
{"swish", "Swish"}};
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class TanhOpConverter : public ActivationOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
TanhOpConverter() : ActivationOpConverter<TargetT, PrecisionT>("tanh") {}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class SigmoidOpConverter : public ActivationOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
SigmoidOpConverter()
|
|
||||||
: ActivationOpConverter<TargetT, PrecisionT>("sigmoid") {}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class Relu6OpConverter : public ActivationOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
Relu6OpConverter() : ActivationOpConverter<TargetT, PrecisionT>("relu6") {}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class SwishOpConverter : public ActivationOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
SwishOpConverter() : ActivationOpConverter<TargetT, PrecisionT>("swish") {}
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
@ -1,55 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 "paddle/fluid/inference/anakin/convert/affine_channel.h"
|
|
||||||
#include <algorithm>
|
|
||||||
#include <string>
|
|
||||||
#include <vector>
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/helper.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
void AffineChannelOpConverter<TargetT, PrecisionT>::operator()(
|
|
||||||
const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope, bool test_mode) {
|
|
||||||
framework::OpDesc op_desc(op, nullptr);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
|
|
||||||
|
|
||||||
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
|
|
||||||
auto input_name = op_desc.Input("X").front();
|
|
||||||
auto output_name = op_desc.Output("Out").front();
|
|
||||||
this->engine_->AddOp(op_name, "AffineChannel", {input_name}, {output_name});
|
|
||||||
|
|
||||||
// Copy the Scale to CPUPlace and get the pointer.
|
|
||||||
auto *scale_v = scope.FindVar(op_desc.Input("Scale").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(scale_v);
|
|
||||||
auto weight1 = pblock_from_var<TargetT, PrecisionT>(*scale_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
|
|
||||||
|
|
||||||
// Copy the Bias to CPUPlace and get the pointer.
|
|
||||||
auto *bias_v = scope.FindVar(op_desc.Input("Bias").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(bias_v);
|
|
||||||
auto weight2 = pblock_from_var<TargetT, PrecisionT>(*bias_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(affine_channel, AffineChannelOpConverter);
|
|
||||||
@ -1,40 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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.
|
|
||||||
|
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include <memory>
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class AffineChannelOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
AffineChannelOpConverter() = default;
|
|
||||||
|
|
||||||
virtual void operator()(const framework::proto::OpDesc &op,
|
|
||||||
const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope,
|
|
||||||
bool test_mode) override;
|
|
||||||
virtual ~AffineChannelOpConverter() {}
|
|
||||||
|
|
||||||
private:
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
@ -1,85 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 "paddle/fluid/inference/anakin/convert/batch_norm.h"
|
|
||||||
#include <math.h>
|
|
||||||
#include <algorithm>
|
|
||||||
#include <map>
|
|
||||||
#include <string>
|
|
||||||
#include <vector>
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/helper.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
void BatchNormOpConverter<TargetT, PrecisionT>::operator()(
|
|
||||||
const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope, bool test_mode) {
|
|
||||||
framework::OpDesc op_desc(op, nullptr);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Output("Y").size(), 1);
|
|
||||||
std::map<std::string, std::string> inputs;
|
|
||||||
for (auto k : {"X", "Scale", "Bias", "Mean", "Variance"}) {
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input(k).size(), 1UL);
|
|
||||||
}
|
|
||||||
|
|
||||||
auto input = op_desc.Input("X").front();
|
|
||||||
auto output = op_desc.Output("Y").front();
|
|
||||||
auto op_name = op_desc.Type() + ":" + op_desc.Output("Y").front();
|
|
||||||
auto epsilon = boost::get<float>(op_desc.GetAttr("epsilon"));
|
|
||||||
|
|
||||||
auto bn_op_name = op_name + ":bn";
|
|
||||||
auto bn_output = bn_op_name + "_output";
|
|
||||||
this->engine_->AddOp(bn_op_name, "BatchNorm", {input}, {bn_output});
|
|
||||||
this->engine_->AddOpAttr(bn_op_name, "epsilon", epsilon);
|
|
||||||
this->engine_->AddOpAttr(bn_op_name, "momentum", static_cast<float>(1.0));
|
|
||||||
|
|
||||||
auto scale_op_name = op_name + ":scale";
|
|
||||||
this->engine_->AddOp(scale_op_name, "Scale", {bn_output}, {output});
|
|
||||||
this->engine_->AddOpAttr(scale_op_name, "axis", 1);
|
|
||||||
this->engine_->AddOpAttr(scale_op_name, "num_axes", 1);
|
|
||||||
this->engine_->AddOpAttr(scale_op_name, "bias_term", true);
|
|
||||||
|
|
||||||
auto *mean_v = scope.FindVar(op_desc.Input("Mean").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(mean_v);
|
|
||||||
auto weight1 = pblock_from_var<TargetT, PrecisionT>(*mean_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(bn_op_name, "weight_1", *weight1);
|
|
||||||
|
|
||||||
auto *variance_v = scope.FindVar(op_desc.Input("Variance").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(variance_v);
|
|
||||||
auto weight2 =
|
|
||||||
pblock_from_var<TargetT, PrecisionT>(*variance_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(bn_op_name, "weight_2", *weight2);
|
|
||||||
|
|
||||||
auto *weight3 = pblock_from_vector<TargetT, PrecisionT>(
|
|
||||||
std::vector<float>({1}), this->engine_);
|
|
||||||
this->engine_->AddOpAttr(bn_op_name, "weight_3", *weight3);
|
|
||||||
|
|
||||||
auto *scale_v = scope.FindVar(op_desc.Input("Scale").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(scale_v);
|
|
||||||
auto scale = pblock_from_var<TargetT, PrecisionT>(*scale_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(scale_op_name, "weight_1", *scale);
|
|
||||||
|
|
||||||
auto *bias_v = scope.FindVar(op_desc.Input("Bias").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(bias_v);
|
|
||||||
auto bias = pblock_from_var<TargetT, PrecisionT>(*bias_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(scale_op_name, "weight_2", *bias);
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(batch_norm, BatchNormOpConverter);
|
|
||||||
@ -1,37 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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.
|
|
||||||
|
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class BatchNormOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
BatchNormOpConverter() = default;
|
|
||||||
|
|
||||||
virtual void operator()(const framework::proto::OpDesc &op,
|
|
||||||
const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope,
|
|
||||||
bool test_mode) override;
|
|
||||||
virtual ~BatchNormOpConverter() {}
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
@ -1,41 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 "paddle/fluid/inference/anakin/convert/concat.h"
|
|
||||||
#include <algorithm>
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
void ConcatOpConverter<TargetT, PrecisionT>::operator()(
|
|
||||||
const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope, bool test_mode) {
|
|
||||||
framework::OpDesc op_desc(op, nullptr);
|
|
||||||
int axis = boost::get<int>(op_desc.GetAttr("axis"));
|
|
||||||
auto input_names = op_desc.Input("X");
|
|
||||||
|
|
||||||
auto y_name = op_desc.Output("Out").front();
|
|
||||||
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
|
|
||||||
|
|
||||||
this->engine_->AddOp(op_name, "Concat", input_names, {y_name});
|
|
||||||
this->engine_->AddOpAttr(op_name, "axis", axis);
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(concat, ConcatOpConverter);
|
|
||||||
@ -1,39 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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.
|
|
||||||
|
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class ConcatOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
ConcatOpConverter() = default;
|
|
||||||
|
|
||||||
virtual void operator()(const framework::proto::OpDesc &op,
|
|
||||||
const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope,
|
|
||||||
bool test_mode) override;
|
|
||||||
virtual ~ConcatOpConverter() {}
|
|
||||||
|
|
||||||
private:
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
@ -1,109 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 "paddle/fluid/inference/anakin/convert/conv2d.h"
|
|
||||||
#include <algorithm>
|
|
||||||
#include <memory>
|
|
||||||
#include <vector>
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/helper.h"
|
|
||||||
|
|
||||||
using anakin::PTuple;
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
void Conv2dOpConverter<TargetT, PrecisionT>::operator()(
|
|
||||||
const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope, bool test_mode) {
|
|
||||||
framework::OpDesc op_desc(op, nullptr);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1UL);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("Filter").size(), 1UL);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1UL);
|
|
||||||
|
|
||||||
auto input_name = op_desc.Input("Input").front();
|
|
||||||
auto output_name = op_desc.Output("Output").front();
|
|
||||||
auto op_name = op_desc.Type() + ":" + op_desc.Output("Output").front();
|
|
||||||
this->engine_->AddOp(op_name, "Convolution", {input_name}, {output_name});
|
|
||||||
|
|
||||||
auto *filter_v = scope.FindVar(op_desc.Input("Filter").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(filter_v);
|
|
||||||
auto weight_tensor = tensor_from_var(*filter_v, platform::CPUPlace());
|
|
||||||
auto weight_shape = framework::vectorize<int>(weight_tensor->dims());
|
|
||||||
|
|
||||||
PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
|
|
||||||
|
|
||||||
const int filter_h = weight_tensor->dims()[2];
|
|
||||||
const int filter_w = weight_tensor->dims()[3];
|
|
||||||
|
|
||||||
auto filter_num = weight_tensor->dims()[0];
|
|
||||||
this->engine_->template AddOpAttr<int>(op_name, "filter_num", filter_num);
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "kernel_size",
|
|
||||||
{filter_h, filter_w});
|
|
||||||
auto strides = boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
|
|
||||||
auto paddings = boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
|
|
||||||
auto dilations = boost::get<std::vector<int>>(op_desc.GetAttr("dilations"));
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dilation_rate",
|
|
||||||
dilations);
|
|
||||||
const int groups = boost::get<int>(op_desc.GetAttr("groups"));
|
|
||||||
this->engine_->AddOpAttr(op_name, "group", groups);
|
|
||||||
this->engine_->AddOpAttr(op_name, "axis", 1);
|
|
||||||
this->engine_->AddOpAttr(op_name, "bias_term", false);
|
|
||||||
|
|
||||||
::anakin::saber::Shape anakin_shape(weight_shape);
|
|
||||||
bool enable_int8 = boost::get<bool>(op_desc.HasAttr("enable_int8"));
|
|
||||||
|
|
||||||
if (enable_int8) {
|
|
||||||
const float int8_range = 127.;
|
|
||||||
float in_scale = boost::get<float>(op_desc.GetAttr("input_scale"));
|
|
||||||
auto weight_scale =
|
|
||||||
boost::get<std::vector<float>>(op_desc.GetAttr("weight_scale"));
|
|
||||||
PBlock<TargetT> *weight1 =
|
|
||||||
new PBlock<TargetT>(anakin_shape, ::anakin::AK_INT8);
|
|
||||||
this->engine_->RegistBlock(weight1);
|
|
||||||
float *weight_data = weight_tensor->data<float>();
|
|
||||||
std::vector<char> weight_int8;
|
|
||||||
int weight_num = weight_tensor->numel();
|
|
||||||
for (int i = 0; i < weight_tensor->numel(); i++) {
|
|
||||||
bool is_valid_int8 =
|
|
||||||
((weight_data[i] >= -128) && (weight_data[i] <= 127));
|
|
||||||
PADDLE_ENFORCE(is_valid_int8,
|
|
||||||
"We are in anakin subgraph int8 mode, the weight of conv "
|
|
||||||
"should be in range [-128, 127]");
|
|
||||||
weight_int8.push_back(static_cast<char>(weight_data[i]));
|
|
||||||
}
|
|
||||||
memcpy(static_cast<void *>(weight1->h_tensor().mutable_data()),
|
|
||||||
static_cast<void *>(weight_int8.data()), sizeof(char) * weight_num);
|
|
||||||
weight1->d_tensor().set_shape(anakin_shape);
|
|
||||||
weight1->d_tensor().copy_from(weight1->h_tensor());
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
|
|
||||||
this->engine_->Graph()->SetOpPrec(op_name, ::anakin::AK_INT8);
|
|
||||||
this->engine_->Graph()->SetWeightsScale(
|
|
||||||
op_name, {weight_scale[0] / int8_range}, false);
|
|
||||||
this->engine_->AddTensorScale(input_name, in_scale / int8_range);
|
|
||||||
} else {
|
|
||||||
auto *weight1 = pblock_from_tensor<TargetT, PrecisionT>(
|
|
||||||
*weight_tensor, weight_shape, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(conv2d, Conv2dOpConverter);
|
|
||||||
@ -1,37 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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.
|
|
||||||
|
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
class Conv2dOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
|
|
||||||
public:
|
|
||||||
Conv2dOpConverter() = default;
|
|
||||||
|
|
||||||
virtual void operator()(const framework::proto::OpDesc &op,
|
|
||||||
const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope,
|
|
||||||
bool test_mode) override;
|
|
||||||
virtual ~Conv2dOpConverter() {}
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
@ -1,115 +0,0 @@
|
|||||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
//
|
|
||||||
// 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 "paddle/fluid/inference/anakin/convert/conv2d_fusion.h"
|
|
||||||
#include <algorithm>
|
|
||||||
#include <memory>
|
|
||||||
#include <vector>
|
|
||||||
#include "paddle/fluid/inference/anakin/convert/helper.h"
|
|
||||||
|
|
||||||
using anakin::PTuple;
|
|
||||||
|
|
||||||
namespace paddle {
|
|
||||||
namespace inference {
|
|
||||||
namespace anakin {
|
|
||||||
|
|
||||||
template <typename TargetT, ::anakin::Precision PrecisionT>
|
|
||||||
void Conv2dFusionOpConverter<TargetT, PrecisionT>::operator()(
|
|
||||||
const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
|
|
||||||
const framework::Scope &scope, bool test_mode) {
|
|
||||||
framework::OpDesc op_desc(op, nullptr);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1UL);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("Filter").size(), 1UL);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Input("Bias").size(), 1UL);
|
|
||||||
PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1UL);
|
|
||||||
|
|
||||||
auto input_name = op_desc.Input("Input").front();
|
|
||||||
auto output_name = op_desc.Output("Output").front();
|
|
||||||
auto op_name = op_desc.Type() + ":" + op_desc.Output("Output").front();
|
|
||||||
this->engine_->AddOp(op_name, "Convolution", {input_name}, {output_name});
|
|
||||||
|
|
||||||
auto *filter_v = scope.FindVar(op_desc.Input("Filter").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(filter_v);
|
|
||||||
|
|
||||||
auto weight_tensor = tensor_from_var(*filter_v, platform::CPUPlace());
|
|
||||||
auto weight_shape = framework::vectorize<int>(weight_tensor->dims());
|
|
||||||
|
|
||||||
auto *b_v = scope.FindVar(op_desc.Input("Bias").front());
|
|
||||||
PADDLE_ENFORCE_NOT_NULL(b_v);
|
|
||||||
|
|
||||||
PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
|
|
||||||
const int filter_h = weight_tensor->dims()[2];
|
|
||||||
const int filter_w = weight_tensor->dims()[3];
|
|
||||||
auto filter_num = weight_tensor->dims()[0];
|
|
||||||
this->engine_->template AddOpAttr<int>(op_name, "filter_num", filter_num);
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "kernel_size",
|
|
||||||
{filter_h, filter_w});
|
|
||||||
auto strides = boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
|
|
||||||
auto paddings = boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
|
|
||||||
auto dilations = boost::get<std::vector<int>>(op_desc.GetAttr("dilations"));
|
|
||||||
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dilation_rate",
|
|
||||||
dilations);
|
|
||||||
const int groups = boost::get<int>(op_desc.GetAttr("groups"));
|
|
||||||
this->engine_->AddOpAttr(op_name, "group", groups);
|
|
||||||
this->engine_->AddOpAttr(op_name, "axis", 1);
|
|
||||||
this->engine_->AddOpAttr(op_name, "bias_term", true);
|
|
||||||
|
|
||||||
::anakin::saber::Shape anakin_shape(weight_shape);
|
|
||||||
bool enable_int8 = boost::get<bool>(op_desc.HasAttr("enable_int8"));
|
|
||||||
if (enable_int8) {
|
|
||||||
const float int8_range = 127.;
|
|
||||||
float in_scale = boost::get<float>(op_desc.GetAttr("input_scale"));
|
|
||||||
auto weight_scale =
|
|
||||||
boost::get<std::vector<float>>(op_desc.GetAttr("weight_scale"));
|
|
||||||
PBlock<TargetT> *weight1 =
|
|
||||||
new PBlock<TargetT>(anakin_shape, ::anakin::AK_INT8);
|
|
||||||
this->engine_->RegistBlock(weight1);
|
|
||||||
float *weight_data = weight_tensor->data<float>();
|
|
||||||
std::vector<char> weight_int8;
|
|
||||||
int weight_num = weight_tensor->numel();
|
|
||||||
for (int i = 0; i < weight_tensor->numel(); i++) {
|
|
||||||
bool is_valid_int8 =
|
|
||||||
((weight_data[i] >= -128) && (weight_data[i] <= 127));
|
|
||||||
PADDLE_ENFORCE(is_valid_int8,
|
|
||||||
"We are in anakin subgraph int8 mode, the weight of conv "
|
|
||||||
"should be in range [-128, 127]");
|
|
||||||
weight_int8.push_back(static_cast<char>(weight_data[i]));
|
|
||||||
}
|
|
||||||
memcpy(static_cast<void *>(weight1->h_tensor().mutable_data()),
|
|
||||||
static_cast<void *>(weight_int8.data()), sizeof(char) * weight_num);
|
|
||||||
weight1->d_tensor().set_shape(anakin_shape);
|
|
||||||
weight1->d_tensor().copy_from(weight1->h_tensor());
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
|
|
||||||
this->engine_->Graph()->SetOpPrec(op_name, ::anakin::AK_INT8);
|
|
||||||
this->engine_->Graph()->SetWeightsScale(
|
|
||||||
op_name, {weight_scale[0] / int8_range}, false);
|
|
||||||
this->engine_->AddTensorScale(input_name, in_scale / int8_range);
|
|
||||||
} else {
|
|
||||||
auto weight_tensor = tensor_from_var(*filter_v, platform::CPUPlace());
|
|
||||||
auto weight_shape = framework::vectorize<int>(weight_tensor->dims());
|
|
||||||
auto *weight1 = pblock_from_tensor<TargetT, PrecisionT>(
|
|
||||||
*weight_tensor, weight_shape, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
|
|
||||||
auto weight2 = pblock_from_var<TargetT, PrecisionT>(*b_v, this->engine_);
|
|
||||||
this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace anakin
|
|
||||||
} // namespace inference
|
|
||||||
} // namespace paddle
|
|
||||||
|
|
||||||
REGISTER_ANAKIN_OP_CONVERTER(conv2d_fusion, Conv2dFusionOpConverter);
|
|
||||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue