Skip to content

P
Paddle

PaddlePaddle / Paddle
1 年多 前同步成功

通知 2302
Star 20931
Fork 5422
P
Paddle
未验证 提交 399bda2b 编写于 作者: 石 石晓伟 提交者： GitHub
浏览文件
操作

remove anakin from code, test=release/1.7 (#22421)

上级 6fec3920
展开全部 隐藏空白更改
内联 并排
Showing with 14 addition and 7529 deletion
CMakeLists.txt
浏览文件 @ 399bda2b
......@@ -149,7 +149,6 @@ if(NOT WIN32)
include(cupti)
endif()
include(anakin_subgraph)
include(flags) # set paddle compile flags
include(cudnn) # set cudnn libraries, must before configure
......
cmake/anakin_subgraph.cmake 已删除 100644 → 0
浏览文件 @ 6fec3920
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(${ANAKIN_ROOT})
add_definitions(-DPADDLE_WITH_ANAKIN)
endif()
if(ANAKIN_FOUND)
if (ANAKIN_MLU AND NOT WITH_GPU AND NOT ANAKIN_X86)
message(STATUS "Compile with anakin mlu place.")
add_definitions(-DANAKIN_MLU_PLACE)
elseif(ANAKIN_BM AND NOT WITH_GPU AND NOT ANAKIN_X86)
message(STATUS "Compile with anakin bm place.")
add_definitions(-DANAKIN_BM_PLACE)
elseif(ANAKIN_X86)
message(STATUS "Compile with anakin x86 place.")
add_definitions(-DANAKIN_X86_PLACE)
endif()
endif()
if(ANAKIN_FOUND AND WITH_GPU AND WITH_DSO)
message(STATUS "Compile with anakin subgraph.")
set(ANAKIN_SUBGRAPH ON)
endif()
cmake/configure.cmake
浏览文件 @ 399bda2b
......@@ -105,16 +105,6 @@ if(WITH_GPU)
endif()
include_directories(${TENSORRT_INCLUDE_DIR})
endif()
if(ANAKIN_FOUND)
if(${CUDA_VERSION_MAJOR} VERSION_LESS 8)
message(WARNING "Anakin needs CUDA >= 8.0 to compile. Force ANAKIN_FOUND = OFF")
set(ANAKIN_FOUND OFF CACHE STRING "Anakin is valid only when CUDA >= 8.0." FORCE)
endif()
if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
message(WARNING "Anakin needs CUDNN >= 7.0 to compile. Force ANAKIN_FOUND = OFF")
set(ANAKIN_FOUND OFF CACHE STRING "Anakin is valid only when CUDNN >= 7.0." FORCE)
endif()
endif()
elseif(WITH_AMD_GPU)
add_definitions(-DPADDLE_WITH_HIP)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__")
......
cmake/inference_lib.cmake
浏览文件 @ 399bda2b
......@@ -126,12 +126,12 @@ function(copy_part_of_thrid_party TARGET DST)
DSTS ${dst_dir}/include ${dst_dir}/lib)
endif ()
if (ANAKIN_FOUND)
set(dst_dir "${DST}/third_party/install/anakin")
if (LITE_BINARY_DIR)
set(dst_dir "${DST}/third_party/install/lite")
copy(${TARGET}
SRCS ${ANAKIN_ROOT}/*
SRCS ${LITE_BINARY_DIR}/inference_lite_lib/*
DSTS ${dst_dir})
endif ()
endif()
endfunction()
# inference library for only inference
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 399bda2b
......@@ -71,7 +71,6 @@ pass_library(identity_scale_op_clean_pass base)
pass_library(sync_batch_norm_pass base)
pass_library(runtime_context_cache_pass base)
pass_library(quant_conv2d_dequant_fuse_pass inference)
pass_library(fillconstant_elementwisemul_fuse inference)
pass_library(shuffle_channel_detect_pass inference)
pass_library(delete_quant_dequant_op_pass inference)
pass_library(simplify_with_basic_ops_pass base)
......@@ -81,10 +80,6 @@ if(WITH_GPU)
pass_library(cudnn_placement_pass base DEPS placement_pass_base)
endif()
if(ANAKIN_SUBGRAPH)
pass_library(simplify_anakin_priorbox_detection_out_pass inference)
endif()
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base DEPS placement_pass_base DIR mkldnn)
pass_library(depthwise_conv_mkldnn_pass base DIR mkldnn)
......
paddle/fluid/framework/ir/fillconstant_elementwisemul_fuse.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ 399bda2b
......@@ -1878,173 +1878,6 @@ PDNode *patterns::TransposeFlattenConcat::operator()(
return concat_out;
}
PDNode *patterns::AnakinDetectionPattern::operator()(
std::vector<PDNode *> conv_in, int times, std::string priorbox_type,
bool is_reshape) {
// The times represents the repeat times of the
// {prior_box, prior_box_loc_out, flatten, prior_box_var_out, reshape}
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;
const int kBoxCoderThirdInputOffset = times;
const int kMultiClassSecondInputNmsOffset = times + 1;
std::vector<PDNode *> nodes;
std::string op_after_priorbox = is_reshape ? "reshape2" : "flatten2";
for (int i = 0; i < times; i++) {
nodes.push_back(
pattern->NewNode(GetNodeName("prior_box" + std::to_string(i)))
->assert_is_op(priorbox_type));
nodes.push_back(pattern->NewNode(GetNodeName("box_out" + std::to_string(i)))
->assert_is_op_output(priorbox_type, "Boxes")
->assert_is_op_input(op_after_priorbox, "X")
->AsIntermediate());
nodes.push_back(
pattern->NewNode(GetNodeName("reshape1" + std::to_string(i)))
->assert_is_op(op_after_priorbox));
nodes.push_back(
pattern->NewNode(GetNodeName("reshape1_out" + std::to_string(i)))
->assert_is_op_output(op_after_priorbox)
->assert_is_op_nth_input("concat", "X", i)
->AsIntermediate());
nodes.push_back(
pattern->NewNode(GetNodeName("box_var_out" + std::to_string(i)))
->assert_is_op_output(priorbox_type, "Variances")
->assert_is_op_input(op_after_priorbox, "X")
->AsIntermediate());
nodes.push_back(
pattern->NewNode(GetNodeName("reshape2" + std::to_string(i)))
->assert_is_op(op_after_priorbox));
nodes.push_back(
pattern->NewNode(GetNodeName("reshape2_out" + std::to_string(i)))
->assert_is_op_output(op_after_priorbox)
->assert_is_op_nth_input("concat", "X", i)
->AsIntermediate());
}
auto concat_op1 = pattern->NewNode(GetNodeName("concat1"))
->assert_is_op("concat")
->assert_op_has_n_inputs("concat", times);
auto concat_out1 = pattern->NewNode(GetNodeName("concat1_out"))
->assert_is_op_output("concat")
->AsIntermediate();
auto concat_op2 = pattern->NewNode(GetNodeName("concat2"))
->assert_is_op("concat")
->assert_op_has_n_inputs("concat", times);
auto concat_out2 = pattern->NewNode(GetNodeName("concat2_out"))
->assert_is_op_output("concat")
->AsIntermediate();
auto box_coder_op = pattern->NewNode(GetNodeName("box_coder"))
->assert_is_op("box_coder")
->assert_op_has_n_inputs("box_coder", 3);
auto box_coder_out = pattern->NewNode(GetNodeName("box_coder_out"))
->assert_is_op_output("box_coder")
->AsIntermediate();
auto transpose_before_nms =
pattern->NewNode(GetNodeName("transpose_before_nms"))
->assert_is_op("transpose2");
auto transpose_before_nms_out =
pattern->NewNode(GetNodeName("transpose_before_nms_out"))
->assert_is_op_output("transpose2")
->assert_is_op_input("multiclass_nms", "Scores")
->AsIntermediate();
auto multiclass_nms_op = pattern->NewNode(GetNodeName("multiclass_nms"))
->assert_is_op("multiclass_nms")
->assert_op_has_n_inputs("multiclass_nms", 2);
auto multiclass_nms_out = pattern->NewNode(GetNodeName("multiclass_nms_out"))
->assert_is_op_output("multiclass_nms")
->AsOutput();
std::vector<PDNode *> reshape1_outs;
std::vector<PDNode *> reshape2_outs;
for (int i = 0; i < times; i++) {
conv_in[i]->AsInput();
// prior_box
nodes[i * kNumFields]->LinksFrom({conv_in[i]});
// prior_box box out
nodes[i * kNumFields + kPriorBoxLocOffset]->LinksFrom(
{nodes[i * kNumFields]});
// reshape
nodes[i * kNumFields + kReshape1Offset]->LinksFrom(
{nodes[i * kNumFields + kPriorBoxLocOffset]});
// reshape_out
nodes[i * kNumFields + kReshape1OutOffset]->LinksFrom(
{nodes[i * kNumFields + kReshape1Offset]});
nodes[i * kNumFields + kPriorBoxVarOffset]->LinksFrom(
{nodes[i * kNumFields]});
// reshape
nodes[i * kNumFields + kReshape2Offset]->LinksFrom(
{nodes[i * kNumFields + kPriorBoxVarOffset]});
// reshape_out
nodes[i * kNumFields + kReshape2OutOffset]->LinksFrom(
{nodes[i * kNumFields + kReshape2Offset]});
reshape1_outs.push_back(nodes[i * kNumFields + kReshape1OutOffset]);
reshape2_outs.push_back(nodes[i * kNumFields + kReshape2OutOffset]);
}
concat_op1->LinksFrom(reshape1_outs);
concat_op2->LinksFrom(reshape2_outs);
concat_out1->LinksFrom({concat_op1});
concat_out2->LinksFrom({concat_op2});
conv_in[kBoxCoderThirdInputOffset]->AsInput();
conv_in[kMultiClassSecondInputNmsOffset]->AsInput();
box_coder_op->LinksFrom(
{concat_out1, concat_out2, conv_in[kBoxCoderThirdInputOffset]});
box_coder_out->LinksFrom({box_coder_op});
transpose_before_nms->LinksFrom({conv_in[kMultiClassSecondInputNmsOffset]});
transpose_before_nms_out->LinksFrom({transpose_before_nms});
multiclass_nms_op->LinksFrom({box_coder_out, transpose_before_nms_out})
.LinksTo({multiclass_nms_out});
return multiclass_nms_out;
}
PDNode *patterns::FillConstantElementWiseMulFuse::operator()(
PDNode *elementwise_op_input) {
auto fill_constant =
pattern->NewNode(fill_constant_repr())->assert_is_op("fill_constant");
auto fill_constant_out = pattern->NewNode(fill_constant_out_repr())
->assert_is_op_output("fill_constant")
->assert_is_op_input("elementwise_mul", "Y")
->AsIntermediate();
auto elementwise_mul_op =
pattern->NewNode(elementwise_mul_repr())->assert_is_op("elementwise_mul");
auto elementwise_mul_out = pattern->NewNode(elementwise_mul_out_repr())
->assert_is_op_output("elementwise_mul")
->AsOutput();
fill_constant_out->LinksFrom({fill_constant});
elementwise_mul_op->LinksFrom({elementwise_op_input, fill_constant_out});
elementwise_mul_out->LinksFrom({elementwise_mul_op});
return elementwise_mul_out;
}
void patterns::QuantDequantOpFuse::operator()(PDNode *quant_op_input,
const std::string &op_type,
const std::string &weight_name,
......
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ 399bda2b
......@@ -1093,37 +1093,6 @@ struct TransposeFlattenConcat : public PatternBase {
}
};
struct AnakinDetectionPattern : public PatternBase {
AnakinDetectionPattern(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "anakin_detect_pattern") {}
PDNode* operator()(std::vector<PDNode*> conv_inputs, int times,
std::string priorbox_type, bool is_reshape);
std::string GetNodeName(const std::string& op_type) {
return PDNodeName(name_scope_, repr_, id_, op_type);
}
PDNode* GetPDNode(const std::string& op_type) {
return pattern->RetrieveNode(GetNodeName(op_type));
}
};
struct FillConstantElementWiseMulFuse : public PatternBase {
FillConstantElementWiseMulFuse(PDPattern* pattern,
const std::string& name_scope)
: PatternBase(pattern, name_scope,
"anakin_fillconstant_elementwisemul_fuse") {}
PDNode* operator()(PDNode* elementwise_op_input);
// declare operator node's name
PATTERN_DECL_NODE(fill_constant);
PATTERN_DECL_NODE(fill_constant_out);
PATTERN_DECL_NODE(elementwise_mul);
PATTERN_DECL_NODE(elementwise_mul_out);
};
struct QuantDequantOpFuse : public PatternBase {
QuantDequantOpFuse(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "quant_dequant_fuse") {}
......
paddle/fluid/framework/ir/simplify_anakin_priorbox_detection_out_pass.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/framework/ir/simplify_anakin_priorbox_detection_out_pass.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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
paddle/fluid/framework/ir/subgraph_detector.h
浏览文件 @ 399bda2b
......@@ -65,7 +65,7 @@ class SubGraphFuser {
using NodeInsideSubgraphTeller = SubgraphDetector::NodeInsideSubgraphTeller;
SubGraphFuser(Graph *graph, const NodeInsideSubgraphTeller &teller,
int min_subgraph_size, std::string name = "anakin_engine")
int min_subgraph_size, std::string name = "tensorrt_engine")
: graph_(graph),
node_inside_subgraph_teller_(teller),
min_subgraph_size_{min_subgraph_size},
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ 399bda2b
......@@ -30,10 +30,6 @@ if (TENSORRT_FOUND)
add_subdirectory(tensorrt)
endif()
if (ANAKIN_SUBGRAPH)
add_subdirectory(anakin)
endif()
if (WITH_LITE)
add_subdirectory(lite)
endif()
......@@ -68,9 +64,6 @@ if(NOT APPLE)
set_target_properties(paddle_fluid PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
endif()
if(ANAKIN_FOUND)
set(ANAKIN_SHARED_INFERENCE_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/api/api_anakin_engine.cc)
endif()
set(SHARED_INFERENCE_SRCS
io.cc
${CMAKE_CURRENT_SOURCE_DIR}/../framework/data_feed.cc
......@@ -80,8 +73,7 @@ set(SHARED_INFERENCE_SRCS
${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc
${mkldnn_quantizer_src_file}
${ANAKIN_SHARED_INFERENCE_SRCS})
${mkldnn_quantizer_src_file})
# Create shared inference library defaultly
cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
......
paddle/fluid/inference/anakin/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ 6fec3920
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)
paddle/fluid/inference/anakin/convert/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ 6fec3920
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)
paddle/fluid/inference/anakin/convert/activation.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/inference/anakin/convert/activation.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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
paddle/fluid/inference/anakin/convert/affine_channel.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/inference/anakin/convert/affine_channel.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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
paddle/fluid/inference/anakin/convert/batch_norm.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/inference/anakin/convert/batch_norm.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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
paddle/fluid/inference/anakin/convert/concat.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/inference/anakin/convert/concat.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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
paddle/fluid/inference/anakin/convert/conv2d.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/inference/anakin/convert/conv2d.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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
paddle/fluid/inference/anakin/convert/conv2d_fusion.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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);
paddle/fluid/inference/anakin/convert/conv2d_fusion.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 Conv2dFusionOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
Conv2dFusionOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~Conv2dFusionOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/density_prior_box.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/density_prior_box.h"
#include <algorithm>
#include <map>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void DensityPriorBoxOpConverter<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);
auto input_name = op_desc.Input("Input").front();
auto image_name = op_desc.Input("Image").front();
auto output_name = op_desc.Output("Boxes").front();
auto op_type = op_desc.Type();
auto op_name = op_type + ":" + op_desc.Output("Boxes").front();
// only for density_prior_box
std::vector<float> fixed_sizes = {};
std::vector<float> fixed_ratios = {};
std::vector<int> densities = {};
std::vector<float> min_sizes = {};
std::vector<float> max_sizes = {};
std::vector<float> aspect_ratios = {};
bool is_clip = false;
bool is_flip = false;
if (op_type == "density_prior_box") {
fixed_sizes =
boost::get<std::vector<float>>(op_desc.GetAttr("fixed_sizes"));
fixed_ratios =
boost::get<std::vector<float>>(op_desc.GetAttr("fixed_ratios"));
densities = boost::get<std::vector<int>>(op_desc.GetAttr("densities"));
is_clip = boost::get<bool>(op_desc.GetAttr("clip"));
} else if (op_type == "prior_box") {
min_sizes = boost::get<std::vector<float>>(op_desc.GetAttr("min_sizes"));
max_sizes = boost::get<std::vector<float>>(op_desc.GetAttr("max_sizes"));
aspect_ratios =
boost::get<std::vector<float>>(op_desc.GetAttr("aspect_ratios"));
is_clip = boost::get<bool>(op_desc.GetAttr("clip"));
is_flip = boost::get<bool>(op_desc.GetAttr("flip"));
}
std::vector<float> dens;
for (auto& ele : densities) {
dens.push_back(static_cast<float>(ele));
}
auto variances = boost::get<std::vector<float>>(op_desc.GetAttr("variances"));
// lack img_h, img_w
auto step_h = boost::get<float>(op_desc.GetAttr("step_h"));
auto step_w = boost::get<float>(op_desc.GetAttr("step_w"));
auto offset = boost::get<float>(op_desc.GetAttr("offset"));
PTuple<std::string> t_order;
t_order.push_back("MIN");
t_order.push_back("COM");
t_order.push_back("MAX");
std::vector<float> temp_v = {};
this->engine_->AddOp(op_name, "PriorBox", {input_name, image_name},
{output_name});
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "min_size",
min_sizes);
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "max_size",
max_sizes);
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "aspect_ratio",
aspect_ratios);
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "fixed_size",
fixed_sizes);
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "fixed_ratio",
fixed_ratios);
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "density", dens);
this->engine_->AddOpAttr(op_name, "is_flip", is_flip);
this->engine_->AddOpAttr(op_name, "is_clip", is_clip);
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "variance",
variances);
this->engine_->AddOpAttr(op_name, "img_h", static_cast<int>(0));
this->engine_->AddOpAttr(op_name, "img_w", static_cast<int>(0));
this->engine_->AddOpAttr(op_name, "step_h", step_h);
this->engine_->AddOpAttr(op_name, "step_w", step_w);
this->engine_->AddOpAttr(op_name, "offset", offset);
this->engine_->template AddOpAttr<PTuple<std::string>>(op_name, "order",
t_order);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(density_prior_box, DensityPriorBoxOpConverter);
REGISTER_ANAKIN_OP_CONVERTER(prior_box, DensityPriorBoxOpConverter);
paddle/fluid/inference/anakin/convert/density_prior_box.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 DensityPriorBoxOpConverter
: public AnakinOpConverter<TargetT, PrecisionT> {
public:
DensityPriorBoxOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~DensityPriorBoxOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/detection_out.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/detection_out.h"
#include <algorithm>
#include <map>
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void DetectionOutOpConverter<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);
auto target_name = op_desc.Input("TargetBox").front();
auto prior_box_name = op_desc.Input("PriorBox").front();
auto scores_name = op_desc.Input("Scores").front();
auto output_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
auto code_type = boost::get<std::string>(op_desc.GetAttr("code_type"));
auto background_label = boost::get<int>(op_desc.GetAttr("background_label"));
auto score_threshold = boost::get<float>(op_desc.GetAttr("score_threshold"));
auto nms_top_k = boost::get<int>(op_desc.GetAttr("nms_top_k"));
auto nms_threshold = boost::get<float>(op_desc.GetAttr("nms_threshold"));
auto nms_eta = boost::get<float>(op_desc.GetAttr("nms_eta"));
auto keep_top_k = boost::get<int>(op_desc.GetAttr("keep_top_k"));
std::string anakin_code_type;
if (code_type == "decode_center_size") {
anakin_code_type = "CENTER_SIZE";
} else if (code_type == "encode_center_size") {
PADDLE_THROW(
"Not support encode_center_size code_type in DetectionOut of anakin");
}
this->engine_->AddOp(op_name, "DetectionOutput",
{target_name, scores_name, prior_box_name},
{output_name});
this->engine_->AddOpAttr(op_name, "share_location", true);
this->engine_->AddOpAttr(op_name, "variance_encode_in_target", false);
this->engine_->AddOpAttr(op_name, "class_num", static_cast<int>(0));
this->engine_->AddOpAttr(op_name, "background_id", background_label);
this->engine_->AddOpAttr(op_name, "keep_top_k", keep_top_k);
this->engine_->AddOpAttr(op_name, "code_type", anakin_code_type);
this->engine_->AddOpAttr(op_name, "conf_thresh", score_threshold);
this->engine_->AddOpAttr(op_name, "nms_top_k", nms_top_k);
this->engine_->AddOpAttr(op_name, "nms_thresh", nms_threshold);
this->engine_->AddOpAttr(op_name, "nms_eta", nms_eta);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(detection_out, DetectionOutOpConverter);
paddle/fluid/inference/anakin/convert/detection_out.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 DetectionOutOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
DetectionOutOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~DetectionOutOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/dropout.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/dropout.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 DropoutOpConverter<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("Mask").size(), 1);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
auto x_name = op_desc.Input("X").front();
auto out_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Scale", {x_name}, {out_name});
auto dropout_prob = boost::get<float>(op_desc.GetAttr("dropout_prob"));
auto factor = 1 - dropout_prob;
auto *weight1 = pblock_from_vector<TargetT, PrecisionT>(
std::vector<float>({factor}), this->engine_);
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
this->engine_->AddOpAttr(op_name, "axis", 0);
this->engine_->AddOpAttr(op_name, "num_axes", 0);
this->engine_->AddOpAttr(op_name, "bias_term", false);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(dropout, DropoutOpConverter);
paddle/fluid/inference/anakin/convert/dropout.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 DropoutOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
DropoutOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~DropoutOpConverter() {}
private:
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/elementwise.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/elementwise.h"
#include <algorithm>
#include <string>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void ElementwiseAddOpConverter<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.Input("Y").size(), 1);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
auto x_name = op_desc.Input("X").front();
auto y_name = op_desc.Input("Y").front();
auto out_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Eltwise", {x_name, y_name}, {out_name});
std::string elementwise_type = "Add";
this->engine_->template AddOpAttr<std::string>(op_name, "type",
elementwise_type);
std::vector<float> coeff = {1.0, 1.0};
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
}
template <typename TargetT, ::anakin::Precision PrecisionT>
void ElementwiseMulOpConverter<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.Input("Y").size(), 1);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
auto x_name = op_desc.Input("X").front();
auto y_name = op_desc.Input("Y").front();
auto out_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Eltwise", {x_name, y_name}, {out_name});
std::string elementwise_type = "Mul";
this->engine_->template AddOpAttr<std::string>(op_name, "type",
elementwise_type);
std::vector<float> coeff = {1.0, 1.0};
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(elementwise_add, ElementwiseAddOpConverter);
REGISTER_ANAKIN_OP_CONVERTER(elementwise_mul, ElementwiseMulOpConverter);
paddle/fluid/inference/anakin/convert/elementwise.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 ElementwiseAddOpConverter
: public AnakinOpConverter<TargetT, PrecisionT> {
public:
ElementwiseAddOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~ElementwiseAddOpConverter() {}
private:
};
template <typename TargetT, ::anakin::Precision PrecisionT>
class ElementwiseMulOpConverter
: public AnakinOpConverter<TargetT, PrecisionT> {
public:
ElementwiseMulOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~ElementwiseMulOpConverter() {}
private:
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/fc.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/fc.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 FcBaseOpConverter<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);
auto input_names = op_desc.InputNames();
bool with_bias = input_names.size() >= 3;
std::string w_name = "Y";
std::string i_name = "X";
if (with_bias) {
w_name = "W";
i_name = "Input";
}
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
// get weights
auto *y_v = scope.FindVar(op_desc.Input(w_name).front());
PADDLE_ENFORCE_NOT_NULL(y_v);
auto weight_tensor = tensor_from_var(*y_v, platform::CPUPlace());
auto weight_shape = framework::vectorize<int>(weight_tensor->dims());
int out_dim = weight_shape[1];
const int w_m = weight_shape[0];
const int w_k = weight_shape[1];
auto input_name = op_desc.Input(i_name).front();
auto output_name = op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Dense", {input_name}, {output_name});
this->engine_->AddOpAttr(op_name, "bias_term", with_bias);
this->engine_->AddOpAttr(op_name, "axis", 1);
this->engine_->AddOpAttr(op_name, "out_dim", out_dim);
auto *weight_data = weight_tensor->data<float>();
PADDLE_ENFORCE(w_m * w_k == weight_tensor->numel());
std::vector<float> trans_weight_data(weight_tensor->numel());
for (int i = 0; i < w_m; i++) {
for (int j = 0; j < w_k; j++) {
trans_weight_data[i + j * w_m] = weight_data[i * w_k + j];
}
}
int weight_num = weight_tensor->numel();
bool enable_int8 = boost::get<bool>(op_desc.HasAttr("enable_int8"));
if (enable_int8) {
if (weight_shape.size() < 4UL) {
weight_shape.insert(weight_shape.begin(), 4UL - weight_shape.size(), 1);
}
::anakin::saber::Shape anakin_shape(weight_shape);
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);
std::vector<char> weight_int8;
for (int i = 0; i < weight_num; i++) {
bool is_valid_int8 =
((trans_weight_data[i] >= -128) && (trans_weight_data[i] <= 127));
PADDLE_ENFORCE(is_valid_int8,
"We are in anakin subgraph int8 mode, the weight of fc "
"should be in range [-128, 127]");
weight_int8.push_back(static_cast<char>(trans_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_vector<TargetT, PrecisionT>(trans_weight_data,
this->engine_);
this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
}
// get bias
if (with_bias) {
auto *b_v = scope.FindVar(op_desc.Input("Bias").front());
PADDLE_ENFORCE_NOT_NULL(b_v);
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(mul, MulOpConverter);
REGISTER_ANAKIN_OP_CONVERTER(fc, FcOpConverter);
paddle/fluid/inference/anakin/convert/fc.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 FcBaseOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
FcBaseOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~FcBaseOpConverter() {}
};
// with bias
template <typename TargetT, ::anakin::Precision PrecisionT>
class FcOpConverter : public FcBaseOpConverter<TargetT, PrecisionT> {
public:
FcOpConverter() = default;
};
// without bias
template <typename TargetT, ::anakin::Precision PrecisionT>
class MulOpConverter : public FcBaseOpConverter<TargetT, PrecisionT> {
public:
MulOpConverter() = default;
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/flatten.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/flatten.h"
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void FlattenOpConverter<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(), 1UL);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1UL);
auto input = op_desc.Input("X").front();
auto output = op_desc.Output("Out").front();
int axis = boost::get<int>(op_desc.GetAttr("axis"));
PADDLE_ENFORCE(axis == 1,
"the anakin flatten op converter now only support aixs == 1.");
std::vector<int> out_dims = {0, -1, 1, 1};
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Reshape", {input}, {output});
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dims", out_dims);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(flatten, FlattenOpConverter);
paddle/fluid/inference/anakin/convert/flatten.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 FlattenOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
FlattenOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~FlattenOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/helper.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/helper.h"
namespace paddle {
namespace inference {
namespace anakin {
std::unique_ptr<framework::LoDTensor> tensor_from_var(
const framework::Variable& var, const platform::Place& place) {
auto& src = var.Get<framework::LoDTensor>();
std::unique_ptr<framework::LoDTensor> dst(new framework::LoDTensor());
dst->Resize(src.dims());
TensorCopySync((src), place, dst.get());
return dst;
}
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/helper.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 <algorithm>
#include <map>
#include <memory>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/inference/anakin/engine.h"
#include "framework/core/net/net.h"
#include "framework/core/types.h"
#include "framework/graph/graph.h"
#include "framework/graph/graph_global_mem.h"
#include "saber/saber_types.h"
using anakin::saber::Shape;
using anakin::AK_FLOAT;
using anakin::AK_INT8;
using anakin::PBlock;
namespace paddle {
namespace inference {
namespace anakin {
std::unique_ptr<framework::LoDTensor> tensor_from_var(
const framework::Variable& var, const platform::Place& place);
template <typename TargetT, ::anakin::Precision PrecisionT>
PBlock<TargetT>* pblock_from_tensor(const framework::LoDTensor& tensor,
std::vector<int> shape_vec,
AnakinEngine<TargetT, PrecisionT>* engine) {
while (shape_vec.size() < 4) {
shape_vec.insert(shape_vec.begin(), 1);
}
Shape shape(shape_vec);
PBlock<TargetT>* weight = new PBlock<TargetT>(shape, AK_FLOAT);
engine->RegistBlock(weight);
float* cpu_data = static_cast<float*>(weight->h_tensor().mutable_data());
std::copy_n(tensor.data<float>(), tensor.numel(), cpu_data);
weight->d_tensor().set_shape(shape);
weight->d_tensor().copy_from(weight->h_tensor());
return weight;
}
template <typename TargetT, ::anakin::Precision PrecisionT>
PBlock<TargetT>* pblock_from_vector(const std::vector<float>& vec,
std::vector<int> shape_vec,
AnakinEngine<TargetT, PrecisionT>* engine) {
while (shape_vec.size() < 4) {
shape_vec.insert(shape_vec.begin(), 1);
}
Shape shape(shape_vec);
PBlock<TargetT>* weight = new PBlock<TargetT>(shape, AK_FLOAT);
engine->RegistBlock(weight);
auto* weight_data = static_cast<float*>(weight->h_tensor().mutable_data());
std::copy(std::begin(vec), std::end(vec), weight_data);
weight->d_tensor().set_shape(shape);
weight->d_tensor().copy_from(weight->h_tensor());
return weight;
}
template <typename TargetT, ::anakin::Precision PrecisionT>
PBlock<TargetT>* pblock_from_vector(const std::vector<float>& vec,
AnakinEngine<TargetT, PrecisionT>* engine) {
int size = vec.size();
return pblock_from_vector<TargetT, PrecisionT>(
vec, std::vector<int>({1, 1, 1, size}), engine);
}
template <typename TargetT, ::anakin::Precision PrecisionT>
PBlock<TargetT>* pblock_from_var(const framework::Variable& var,
AnakinEngine<TargetT, PrecisionT>* engine) {
auto tensor = tensor_from_var(var, platform::CPUPlace());
auto shape = framework::vectorize<int>(tensor->dims());
return pblock_from_tensor<TargetT, PrecisionT>(*tensor, shape, engine);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/im2sequence.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/im2sequence.h"
#include <algorithm>
#include <string>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void Im2SequenceConverter<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("Y").size(), 0);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
auto x_name = op_desc.Input("X").front();
auto out_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Im2Sequence", {x_name}, {out_name});
std::vector<int> dilations = {1, 1};
auto paddings = boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
auto strides = boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
auto kernels = boost::get<std::vector<int>>(op_desc.GetAttr("kernels"));
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "paddings", paddings);
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "window_size",
kernels);
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dilations",
dilations);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(im2sequence, Im2SequenceConverter);
paddle/fluid/inference/anakin/convert/im2sequence.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 Im2SequenceConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
Im2SequenceConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~Im2SequenceConverter() {}
private:
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/op_converter.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "framework/core/types.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/anakin/engine.h"
#include "paddle/fluid/inference/utils/singleton.h"
#include "saber/saber_types.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
class AnakinOpConverter {
using AnakinEngineT = AnakinEngine<TargetT, PrecisionT>;
public:
AnakinOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope, bool test_mode) {}
void ConvertOp(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const std::unordered_set<std::string> &parameters,
const framework::Scope &scope, AnakinEngineT *engine,
bool test_mode = false) {
framework::OpDesc op_desc(op, nullptr);
std::string op_type = op_desc.Type();
AnakinOpConverter *it = nullptr;
if (op_type == "depthwise_conv2d") op_type = "conv2d";
if (op_type == "reshape2") op_type = "reshape";
if (op_type == "transpose2") op_type = "transpose";
if (op_type == "flatten2") op_type = "flatten";
if (!it) {
it = Registry<AnakinOpConverter>::Global().Lookup(op_type);
}
PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", op_type);
it->SetEngine(engine);
(*it)(op, block_desc, scope, test_mode);
}
void ConvertBlock(framework::BlockDesc *block_desc,
const std::unordered_set<std::string> &parameters,
const framework::Scope &scope, AnakinEngineT *engine) {
std::unique_lock<std::mutex> lock(mutex_);
framework::proto::BlockDesc *block = block_desc->Proto();
for (auto i = 0; i < block->ops_size(); i++) {
auto &op = block->ops(i);
ConvertOp(op, *block_desc, parameters, scope, engine);
}
}
// The scope here should be inited with the parameter vars.
void ConvertBlockToAnakinEngine(
framework::BlockDesc *block_desc, framework::Scope *scope,
const std::vector<std::string> &inputs,
const std::unordered_set<std::string> &parameters,
const std::vector<std::string> &outputs, AnakinEngineT *engine) {
ConvertBlock(block_desc, parameters, *scope, engine);
// if the max_batch size
int max_batch_size = engine->GetMaxBatchSize();
PADDLE_ENFORCE(max_batch_size > 0,
"the max_batch_size setted from config->EnableAnakinEngine "
"must largger than 0");
// If the user does not specify this variable, we use the input shape from
// the block_desc.
auto max_input_shape = engine->GetMaxInputShape();
std::map<std::string, std::vector<int>> temp_max_input_shape;
// Register outputs with anakin using the RegistVar interface before Freeze.
// Note that RegistVar's parameters can only be outputs, not inputs.
for (auto &output : outputs) {
engine->Graph()->RegistVar(output);
}
engine->Freeze();
// Add scale for tensor in int8 mode.
auto tensor_scales = engine->GetTensorScales();
for (auto &item : tensor_scales) {
engine->Graph()->SetVarScale(item.first, item.second);
}
for (auto &input : inputs) {
if (parameters.count(input)) continue;
std::vector<int> input_shape;
input_shape.resize(4);
input_shape[0] = max_batch_size;
if (max_input_shape.count(input)) {
PADDLE_ENFORCE(max_input_shape[input].size() == 4,
"the dimensions of max_input_shape setted from "
"config->EnableAnakinEngine must be 4");
for (int i = 1; i < 4; i++) {
input_shape[i] = max_input_shape[input][i];
}
} else {
auto *var = block_desc->FindVar(input);
PADDLE_ENFORCE(var, "no variable called %s", input);
auto var_shape = var->GetShape();
std::cout << "input :" << input << std::endl;
PADDLE_ENFORCE(var_shape.size() == 4);
for (size_t i = 1; i < var_shape.size(); i++) {
input_shape[i] = var_shape[i];
}
}
temp_max_input_shape[input] = input_shape;
engine->SetInputShape(input, input_shape);
}
engine->SetMaxInputShape(temp_max_input_shape);
engine->Optimize();
engine->InitNet();
}
void SetEngine(AnakinEngineT *engine) { engine_ = engine; }
virtual ~AnakinOpConverter() {}
protected:
bool test_mode_;
AnakinEngineT *engine_{nullptr};
private:
std::unordered_map<std::string, AnakinOpConverter<TargetT, PrecisionT> *>
converters_;
framework::Scope *scope_{nullptr};
std::mutex mutex_;
};
template class AnakinOpConverter<::anakin::saber::NV,
::anakin::Precision::FP32>;
template class AnakinOpConverter<::anakin::saber::NV,
::anakin::Precision::INT8>;
#ifdef ANAKIN_X86_PLACE
template class AnakinOpConverter<::anakin::saber::X86,
::anakin::Precision::FP32>;
template class AnakinOpConverter<::anakin::saber::X86,
::anakin::Precision::INT8>;
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
#define REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, \
place_type__, place_class__, \
precision_type__, precision_class__) \
struct anakin_##op_type__##_##place_type__##_##precision_type__##_converter \
: public ::paddle::framework::Registrar { \
anakin_##op_type__##_##place_type__##_##precision_type__##_converter() { \
LOG(INFO) << "register convert " << #op_type__ << " "; \
::paddle::inference::Registry< \
::paddle::inference::anakin::AnakinOpConverter< \
place_class__, precision_class__>>::Global() \
.Register<Converter__>(#op_type__); \
} \
}; \
anakin_##op_type__##_##place_type__##_##precision_type__##_converter \
anakin_##op_type__##_##place_type__##_##precision_type__##_converter__; \
int Touch_anakin_##op_type__##_##place_type__##_##precision_type__() { \
anakin_##op_type__##_##place_type__##_##precision_type__##_converter__ \
.Touch(); \
return 0; \
}
#define WRAP(...) __VA_ARGS__
#define REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__, \
precision_type__) \
REGISTER_ANAKIN_OP_CONVERTER_BASE( \
op_type__, \
::paddle::inference::anakin::Converter__<WRAP( \
::anakin::saber::NV, ::anakin::Precision::precision_type__)>, \
CUDA, ::anakin::saber::NV, precision_type__, \
::anakin::Precision::precision_type__)
#define REGISTER_CPU_ANAKIN_OP_CONVERTER(op_type__, Converter__, \
precision_type__) \
REGISTER_ANAKIN_OP_CONVERTER_BASE( \
op_type__, \
::paddle::inference::anakin::Converter__<WRAP( \
::anakin::saber::X86, ::anakin::Precision::precision_type__)>, \
CPU, ::anakin::saber::X86, precision_type__, \
::anakin::Precision::precision_type__)
#if defined(PADDLE_WITH_CUDA) && defined(ANAKIN_X86_PLACE)
#define REGISTER_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__, FP32); \
REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__, INT8); \
REGISTER_CPU_ANAKIN_OP_CONVERTER(op_type__, Converter__, FP32); \
REGISTER_CPU_ANAKIN_OP_CONVERTER(op_type__, Converter__, INT8)
#elif defined(PADDLE_WITH_CUDA)
#define REGISTER_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__, FP32); \
REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__, INT8)
#endif
#define USE_ANAKIN_CONVERTER_BASE(op_type__, place_type__, precision_type__) \
extern int Touch_anakin_##op_type__##_##place_type__##_##precision_type__(); \
int use_converter_anakin_##op_type__##_##place_type__##_##precision_type__ \
UNUSED = \
Touch_anakin_##op_type__##_##place_type__##_##precision_type__();
#if defined(PADDLE_WITH_CUDA) && defined(ANAKIN_X86_PLACE)
#define USE_ANAKIN_CONVERTER(op_type__) \
USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA, FP32) \
USE_ANAKIN_CONVERTER_BASE(op_type__, CPU, FP32)
#define USE_INT8_ANAKIN_CONVERTER(op_type__) \
USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA, INT8) \
USE_ANAKIN_CONVERTER_BASE(op_type__, CPU, INT8)
#elif defined(PADDLE_WITH_CUDA)
#define USE_ANAKIN_CONVERTER(op_type__) \
USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA, FP32)
#define USE_INT8_ANAKIN_CONVERTER(op_type__) \
USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA, INT8)
#endif
paddle/fluid/inference/anakin/convert/pool2d.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/pool2d.h"
#include <algorithm>
#include <string>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void Pool2dOpConverter<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 x_name = op_desc.Input("X").front();
auto y_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
bool global_pooling = boost::get<bool>(op_desc.GetAttr("global_pooling"));
std::string pool_type =
boost::get<std::string>(op_desc.GetAttr("pooling_type"));
std::vector<int> ksize =
boost::get<std::vector<int>>(op_desc.GetAttr("ksize"));
std::vector<int> strides =
boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
std::vector<int> paddings =
boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
bool ceil_mode = boost::get<bool>(op_desc.GetAttr("ceil_mode"));
std::string anakin_pool_type;
if (pool_type == "max") {
anakin_pool_type = "MAX";
} else if (pool_type == "avg") {
if (paddings[0] || paddings[1]) {
anakin_pool_type = "AVGEXC";
} else {
anakin_pool_type = "AVG";
}
} else {
PADDLE_THROW("TensorRT unsupported pooling type!");
}
this->engine_->AddOp(op_name, "Pooling", {x_name}, {y_name});
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "pool_size", ksize);
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
this->engine_->AddOpAttr(op_name, "method", anakin_pool_type);
this->engine_->AddOpAttr(op_name, "global_pooling", global_pooling);
this->engine_->AddOpAttr(op_name, "cmp_out_shape_floor_as_conv", !ceil_mode);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(pool2d, Pool2dOpConverter);
paddle/fluid/inference/anakin/convert/pool2d.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 Pool2dOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
Pool2dOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~Pool2dOpConverter() {}
private:
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/relu.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/relu.h"
#include <algorithm>
#include <map>
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void ReluOpConverter<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, "ReLU", {input_name}, {output_name});
this->engine_->AddOpAttr(op_name, "alpha", 0);
}
template <typename TargetT, ::anakin::Precision PrecisionT>
void LeakyReluOpConverter<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();
float alpha = boost::get<float>(op_desc.GetAttr("alpha"));
this->engine_->AddOp(op_name, "ReLU", {input_name}, {output_name});
this->engine_->AddOpAttr(op_name, "alpha", alpha);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(relu, ReluOpConverter);
REGISTER_ANAKIN_OP_CONVERTER(leaky_relu, LeakyReluOpConverter);
paddle/fluid/inference/anakin/convert/relu.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 ReluOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
ReluOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~ReluOpConverter() {}
};
template <typename TargetT, ::anakin::Precision PrecisionT>
class LeakyReluOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
LeakyReluOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~LeakyReluOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/reshape.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/reshape.h"
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void ReshapeOpConverter<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(), 1UL);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1UL);
auto input = op_desc.Input("X").front();
auto output = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Reshape", {input}, {output});
auto shape = boost::get<std::vector<int>>(op_desc.GetAttr("shape"));
if (shape.size() < 4) {
shape.insert(shape.end(), 4 - shape.size(), 1);
}
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dims", shape);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(reshape, ReshapeOpConverter);
paddle/fluid/inference/anakin/convert/reshape.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 ReshapeOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
ReshapeOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~ReshapeOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/roi_align.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/roi_align.h"
#include <algorithm>
#include <map>
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void RoiAlignOpConverter<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.Input("ROIs").size(), 1);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
auto input_x_name = op_desc.Input("X").front();
auto input_rois_name = op_desc.Input("ROIs").front();
auto output_name = op_desc.Output("Out").front();
auto spatial_scale = boost::get<float>(op_desc.GetAttr("spatial_scale"));
auto pooled_height = boost::get<int>(op_desc.GetAttr("pooled_height"));
auto pooled_width = boost::get<int>(op_desc.GetAttr("pooled_width"));
auto sampling_ratio = boost::get<int>(op_desc.GetAttr("sampling_ratio"));
this->engine_->AddOp(op_name, "RoiAlign", {input_x_name, input_rois_name},
{output_name});
this->engine_->AddOpAttr(op_name, "spatial_scale", spatial_scale);
this->engine_->AddOpAttr(op_name, "pooled_height", pooled_height);
this->engine_->AddOpAttr(op_name, "pooled_width", pooled_width);
this->engine_->AddOpAttr(op_name, "sampling_ratio", sampling_ratio);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(roi_align, RoiAlignOpConverter);
paddle/fluid/inference/anakin/convert/roi_align.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 RoiAlignOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
RoiAlignOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~RoiAlignOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/scale.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/scale.h"
#include <algorithm>
#include <map>
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void ScaleOpConverter<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();
float scale = boost::get<float>(op_desc.GetAttr("scale"));
float bias = boost::get<float>(op_desc.GetAttr("bias"));
float bias_after_scale =
boost::get<bool>(op_desc.GetAttr("bias_after_scale"));
PADDLE_ENFORCE(bias_after_scale,
"The anakin scale layer only support bias after scale now.");
this->engine_->AddOp(op_name, "Power", {input_name}, {output_name});
this->engine_->AddOpAttr(op_name, "shift", bias);
this->engine_->AddOpAttr(op_name, "scale", scale);
this->engine_->AddOpAttr(op_name, "power", static_cast<float>(1.0));
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(scale, ScaleOpConverter);
paddle/fluid/inference/anakin/convert/scale.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 ScaleOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
ScaleOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~ScaleOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/shuffle_channel.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/shuffle_channel.h"
#include <algorithm>
#include <string>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void ShuffleChannelOpConverter<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 input = op_desc.Input("X").front();
auto output = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "ShuffleChannel", {input}, {output});
auto group = boost::get<int>(op_desc.GetAttr("group"));
this->engine_->AddOpAttr(op_name, "group", group);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(shuffle_channel, ShuffleChannelOpConverter);
paddle/fluid/inference/anakin/convert/shuffle_channel.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 ShuffleChannelOpConverter
: public AnakinOpConverter<TargetT, PrecisionT> {
public:
ShuffleChannelOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~ShuffleChannelOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/softmax.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/softmax.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void SoftMaxOpConverter<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(), 1UL);
auto input = op_desc.Input("X").front();
auto output = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
auto input_var_desc = block_desc.FindVar(input);
PADDLE_ENFORCE(input_var_desc,
"Cant find %s variable When runing Anakin Softmax converter.",
input);
auto input_shape_in_fluid = input_var_desc->GetShape();
size_t input_dims = input_shape_in_fluid.size();
this->engine_->AddOp(op_name, "Softmax", {input}, {output});
this->engine_->AddOpAttr(op_name, "axis", static_cast<int>(input_dims - 1));
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(softmax, SoftMaxOpConverter);
paddle/fluid/inference/anakin/convert/softmax.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 SoftMaxOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
SoftMaxOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~SoftMaxOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/split.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/split.h"
#include <algorithm>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void SplitOpConverter<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);
auto input_name = op_desc.Input("X").front();
auto y_names = op_desc.Output("Out");
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
int axis = boost::get<int>(op_desc.GetAttr("axis"));
std::vector<int> output_lengths =
boost::get<std::vector<int>>(op_desc.GetAttr("sections"));
int split_num = output_lengths.size();
PADDLE_ENFORCE(split_num > 1,
"anakin split op converter: the split num should > 1");
int num_sum = 0;
std::vector<int> slice_point;
for (int i = 0; i < split_num - 1; i++) {
num_sum += output_lengths[i];
slice_point.push_back(num_sum);
}
this->engine_->AddOp(op_name, "Slice", {input_name}, y_names);
this->engine_->AddOpAttr(op_name, "axis", axis);
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "slice_point",
slice_point);
// slice_dim is useless in anakin
this->engine_->AddOpAttr(op_name, "slice_dim", 4);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(split, SplitOpConverter);
paddle/fluid/inference/anakin/convert/split.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 SplitOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
SplitOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~SplitOpConverter() {}
private:
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/sum.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/sum.h"
#include <algorithm>
#include <string>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void SumOpConverter<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(), 2);
PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
auto input_names = op_desc.Input("X");
auto out_name = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
std::vector<float> coeff = {1, 1};
std::string elementwise_type = "Add";
this->engine_->AddOp(op_name, "Eltwise", input_names, {out_name});
this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
this->engine_->template AddOpAttr<std::string>(op_name, "type",
elementwise_type);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(sum, SumOpConverter);
paddle/fluid/inference/anakin/convert/sum.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 SumOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
SumOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~SumOpConverter() {}
private:
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/convert/test_activation_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/activation.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
static void test_activation_op(const std::string& op_type,
const platform::DeviceContext& context,
bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("act-X", {10, 6, 1, 1});
validator.DeclOutputVar("act-Out", {10, 6, 1, 1});
framework::OpDesc desc;
desc.SetType(op_type);
desc.SetInput("X", {"act-X"});
desc.SetOutput("Out", {"act-Out"});
if (op_type == "swish") {
desc.SetAttr("beta", 1.0f);
}
if (op_type == "relu6") {
desc.SetAttr("threshold", 6.0f);
}
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(5);
}
#ifdef PADDLE_WITH_CUDA
TEST(sigm_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_activation_op<::anakin::saber::NV>("sigmoid", ctx, true);
}
TEST(tanh_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_activation_op<::anakin::saber::NV>("tanh", ctx, true);
}
TEST(relu6_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_activation_op<::anakin::saber::NV>("relu6", ctx, true);
}
TEST(swish_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_activation_op<::anakin::saber::NV>("swish", ctx, true);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(sigmoid);
USE_OP(tanh);
USE_OP(relu6);
USE_OP(swish);
USE_ANAKIN_CONVERTER(sigmoid);
USE_ANAKIN_CONVERTER(tanh);
USE_ANAKIN_CONVERTER(relu6);
USE_ANAKIN_CONVERTER(swish);
paddle/fluid/inference/anakin/convert/test_affine_channel_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/affine_channel.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_affine_channel_op(const platform::DeviceContext& context,
bool use_gpu) {
// Declare the difference between the inputs.
std::unordered_set<std::string> parameters({"scale", "bias"});
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("x", {1, 3, 5, 2});
validator.DeclOutputVar("out", {1, 3, 5, 2});
validator.DeclParamVar("scale", {3});
validator.DeclParamVar("bias", {3});
// Prepare Op descriptions.
framework::OpDesc desc;
desc.SetType("affine_channel");
desc.SetInput("X", {"x"});
desc.SetInput("Bias", {"bias"});
desc.SetInput("Scale", {"scale"});
desc.SetOutput("Out", {"out"});
// Layout must be explicitly specified here as NCHW.
desc.SetAttr("data_layout", std::string("NCHW"));
validator.SetOp(*desc.Proto());
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(affine_channel_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_affine_channel_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(affine_channel_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_affine_channel_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(affine_channel);
USE_ANAKIN_CONVERTER(affine_channel);
paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_batchnorm_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters(
{"batch_norm_scale", "batch_norm_bias", "batch_norm_mean",
"batch_norm_variance"});
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
std::vector<int> param_shape{2};
validator.DeclInputVar("batch_norm_X", {1, 2, 5, 5});
validator.DeclParamVar("batch_norm_scale", param_shape);
validator.DeclParamVar("batch_norm_bias", param_shape);
validator.DeclParamVar("batch_norm_mean", param_shape);
validator.DeclParamVar("batch_norm_variance", param_shape);
validator.DeclOutputVar("batch_norm_Y", {1, 2, 5, 5});
validator.DeclOutputVar("batch_norm_save_mean", param_shape);
validator.DeclOutputVar("batch_norm_save_variance", param_shape);
// Prepare Op description
framework::OpDesc desc;
desc.SetType("batch_norm");
desc.SetInput("X", {"batch_norm_X"});
desc.SetInput("Scale", {"batch_norm_scale"});
desc.SetInput("Bias", {"batch_norm_bias"});
desc.SetInput("Mean", {"batch_norm_mean"});
desc.SetInput("Variance", {"batch_norm_variance"});
desc.SetOutput("Y", {"batch_norm_Y"});
desc.SetOutput("MeanOut", {"batch_norm_mean"});
desc.SetOutput("VarianceOut", {"batch_norm_variance"});
desc.SetOutput("SavedMean", {"batch_norm_save_mean"});
desc.SetOutput("SavedVariance", {"batch_norm_save_variance"});
float eps = 1e-5f;
bool is_test = true;
desc.SetAttr("epsilon", eps);
desc.SetAttr("is_test", is_test);
validator.SetOp(*desc.Proto());
std::unordered_set<std::string> neglected_output = {
"batch_norm_save_mean", "batch_norm_save_variance", "batch_norm_mean",
"batch_norm_variance"};
validator.Execute(1, neglected_output);
}
#ifdef PADDLE_WITH_CUDA
TEST(batch_norm_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_batchnorm_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(batch_norm_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_batchnorm_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(batch_norm);
USE_ANAKIN_CONVERTER(batch_norm);
paddle/fluid/inference/anakin/convert/test_concat_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/concat.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_concat_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters({""});
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("concat_x1", {1, 2, 1, 1});
validator.DeclInputVar("concat_x2", {1, 3, 1, 1});
validator.DeclInputVar("concat_x3", {1, 1, 1, 1});
validator.DeclOutputVar("concat_out", {1, 6, 1, 1});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("concat");
desc.SetInput("X", {"concat_x1", "concat_x2", "concat_x3"});
desc.SetOutput("Out", {"concat_out"});
int axis = 1;
desc.SetAttr("axis", axis);
validator.SetOp(*desc.Proto());
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(concat_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_concat_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(concat_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_concat_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(concat);
USE_ANAKIN_CONVERTER(concat);
paddle/fluid/inference/anakin/convert/test_conv2d_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/conv2d.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_conv2d_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters({"conv2d-Y"});
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("conv2d-X", {1, 3, 3, 3});
validator.DeclParamVar("conv2d-Y", {4, 3, 1, 1});
validator.DeclOutputVar("conv2d-Out", {1, 4, 3, 3});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("conv2d");
desc.SetInput("Input", {"conv2d-X"});
desc.SetInput("Filter", {"conv2d-Y"});
desc.SetOutput("Output", {"conv2d-Out"});
const std::vector<int> strides({1, 1});
const std::vector<int> paddings({0, 0});
const std::vector<int> dilations({1, 1});
const int groups = 1;
desc.SetAttr("strides", strides);
desc.SetAttr("paddings", paddings);
desc.SetAttr("dilations", dilations);
desc.SetAttr("groups", groups);
validator.SetOp(*desc.Proto());
validator.Execute(3);
}
#ifdef PADDLE_WITH_CUDA
TEST(conv2d_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_conv2d_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(conv2d_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_conv2d_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(conv2d);
USE_ANAKIN_CONVERTER(conv2d);
paddle/fluid/inference/anakin/convert/test_dropout_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/dropout.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_dropout_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("x", {1, 1, 2, 2});
validator.DeclOutputVar("out", {1, 1, 2, 2});
validator.DeclOutputVar("mask", {1, 1, 2, 2});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("dropout");
desc.SetInput("X", {"x"});
desc.SetOutput("Out", {"out"});
desc.SetOutput("Mask", {"mask"});
float dropout_prob = 0.5;
desc.SetAttr("dropout_prob", dropout_prob);
desc.SetAttr("is_test", true);
validator.SetOp(*desc.Proto());
std::unordered_set<std::string> neglected_output = {"mask"};
validator.Execute(1, neglected_output);
}
#ifdef PADDLE_WITH_CUDA
TEST(dropout_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_dropout_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(dropout_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_dropout_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(dropout);
USE_ANAKIN_CONVERTER(dropout);
paddle/fluid/inference/anakin/convert/test_elementwise_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/elementwise.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
static void test_elementwise_op(const std::string& op_type,
const platform::DeviceContext& context,
bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("x", {1, 1, 2, 2});
validator.DeclInputVar("y", {1, 1, 2, 2});
validator.DeclOutputVar("out", {1, 1, 2, 2});
// Prepare Op description
framework::OpDesc desc;
desc.SetType(op_type);
desc.SetInput("X", {"x"});
desc.SetInput("Y", {"y"});
desc.SetOutput("Out", {"out"});
int axis = -1;
desc.SetAttr("axis", axis);
validator.SetOp(*desc.Proto());
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(elementwise_op, native_add_gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_elementwise_op<::anakin::saber::NV>("elementwise_add", ctx, true);
}
TEST(elementwise_op, native_mul_gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_elementwise_op<::anakin::saber::NV>("elementwise_mul", ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(elementwise_op, native_add_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_elementwise_op<::anakin::saber::X86>("elementwise_add", ctx, false);
}
TEST(elementwise_op, native_mul_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_elementwise_op<::anakin::saber::X86>("elementwise_mul", ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(elementwise_add);
USE_OP(elementwise_mul);
USE_ANAKIN_CONVERTER(elementwise_add);
USE_ANAKIN_CONVERTER(elementwise_mul);
paddle/fluid/inference/anakin/convert/test_fc_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_mul_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters({"mul_y"});
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("mul_x", {1, 1, 2, 2});
validator.DeclParamVar("mul_y", {4, 2});
validator.DeclOutputVar("mul_out", {1, 2});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("mul");
desc.SetInput("X", {"mul_x"});
desc.SetInput("Y", {"mul_y"});
desc.SetOutput("Out", {"mul_out"});
validator.SetOp(*desc.Proto());
validator.Execute(10);
}
#ifdef PADDLE_WITH_CUDA
TEST(mul_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_mul_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(mul_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_mul_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(mul);
USE_ANAKIN_CONVERTER(fc);
paddle/fluid/inference/anakin/convert/test_flatten_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_flatten_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("flatten-X", {3, 10, 10, 4});
validator.DeclOutputVar("flatten-Out", {3, 400, 1, 1});
framework::OpDesc desc;
desc.SetType("flatten");
desc.SetInput("X", {"flatten-X"});
desc.SetOutput("Out", {"flatten-Out"});
desc.SetAttr("axis", 1);
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(5);
}
#ifdef PADDLE_WITH_CUDA
TEST(flatten_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_flatten_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(flatten_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_flatten_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(reshape);
USE_OP_ITSELF(flatten);
USE_ANAKIN_CONVERTER(flatten);
paddle/fluid/inference/anakin/convert/test_im2sequence_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/im2sequence.h"
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
TEST(im2sequence_op, native) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation validator(parameters, &scope);
std::vector<int> kernels = {6, 1};
std::vector<int> strides = {1, 1};
std::vector<int> paddings = {0, 0, 0, 0};
validator.DeclInputVar("x", {1, 1, 2, 2});
validator.DeclOutputVar("out", {1, 1 * kernels[0] * kernels[1]});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("im2sequence");
desc.SetInput("X", {"x"});
desc.SetOutput("Out", {"out"});
desc.SetAttr("kernels", kernels);
desc.SetAttr("strides", strides);
desc.SetAttr("paddings", paddings);
validator.SetOp(*desc.Proto());
validator.Execute(1);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(im2sequence);
USE_ANAKIN_CONVERTER(im2sequence);
paddle/fluid/inference/anakin/convert/test_pool2d_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_pool2d(const platform::DeviceContext& context, bool use_gpu,
bool global_pooling, bool ceil_mode,
std::string pool_type = "max") {
framework::Scope scope;
std::unordered_set<std::string> parameters;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
// The ITensor's Dims should not contain the batch size.
// So, the ITensor's Dims of input and output should be C * H * W.
validator.DeclInputVar("pool2d_x", {1, 3, 6, 7});
if (global_pooling)
validator.DeclOutputVar("pool2d_out", {1, 3, 1, 1});
else if (ceil_mode)
validator.DeclOutputVar("pool2d_out", {1, 3, 3, 4});
else
validator.DeclOutputVar("pool2d_out", {1, 3, 3, 3});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("pool2d");
desc.SetInput("X", {"pool2d_x"});
desc.SetOutput("Out", {"pool2d_out"});
std::vector<int> ksize({2, 2});
std::vector<int> strides({2, 2});
std::vector<int> paddings({0, 0});
std::string pooling_t = pool_type;
desc.SetAttr("pooling_type", pooling_t);
desc.SetAttr("ksize", ksize);
desc.SetAttr("strides", strides);
desc.SetAttr("paddings", paddings);
desc.SetAttr("global_pooling", global_pooling);
desc.SetAttr("ceil_mode", ceil_mode);
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(Pool2dOpConverter, normal) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_pool2d<::anakin::saber::NV>(ctx, true, false, false);
}
TEST(Pool2dOpConverter, test_global_pooling) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_pool2d<::anakin::saber::NV>(ctx, true, true, false);
}
TEST(Pool2dOpConverter, max_ceil_test) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_pool2d<::anakin::saber::NV>(ctx, true, false, true);
}
TEST(Pool2dOpConverter, avg_ceil_test) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_pool2d<::anakin::saber::NV>(ctx, true, false, true, "avg");
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(Pool2dOpConverter, normal_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_pool2d<::anakin::saber::X86>(ctx, false, false, false);
}
TEST(Pool2dOpConverter, test_global_pooling_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_pool2d<::anakin::saber::X86>(ctx, false, true, false);
}
TEST(Pool2dOpConverter, max_ceil_test_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_pool2d<::anakin::saber::X86>(ctx, false, false, true);
}
TEST(Pool2dOpConverter, avg_ceil_test_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_pool2d<::anakin::saber::X86>(ctx, false, false, true, "avg");
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(pool2d);
USE_ANAKIN_CONVERTER(pool2d);
paddle/fluid/inference/anakin/convert/test_relu_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/relu.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
static void test_activation_op(const std::string& op_type,
const platform::DeviceContext& context,
bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("act-X", {10, 6, 1, 1});
validator.DeclOutputVar("act-Out", {10, 6, 1, 1});
framework::OpDesc desc;
desc.SetType(op_type);
desc.SetInput("X", {"act-X"});
desc.SetOutput("Out", {"act-Out"});
if (op_type == "leaky_relu") {
desc.SetAttr("alpha", 0.1f);
}
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(5);
}
#ifdef PADDLE_WITH_CUDA
TEST(relu_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_activation_op<::anakin::saber::NV>("relu", ctx, true);
}
TEST(leaky_relu_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_activation_op<::anakin::saber::NV>("leaky_relu", ctx, true);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(relu);
USE_OP(leaky_relu);
USE_ANAKIN_CONVERTER(relu);
USE_ANAKIN_CONVERTER(leaky_relu);
paddle/fluid/inference/anakin/convert/test_reshape_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_reshape1_op(const platform::DeviceContext& context, bool use_gpu) {
framework::Scope scope;
std::unordered_set<std::string> parameters;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
// validator.DeclInputVar("reshape-X", {2, 3, 3, 1});
// validator.DeclOutputVar("reshape-Out", {3, 2, 1, 3});
validator.DeclInputVar("reshape-X", {1, 2, 4, 1});
validator.DeclOutputVar("reshape-Out", {1, 8, 1, 1});
framework::OpDesc desc;
desc.SetType("reshape");
desc.SetInput("X", {"reshape-X"});
desc.SetOutput("Out", {"reshape-Out"});
// desc.SetAttr("shape", std::vector<int>({3, 2, 1, 3}));
desc.SetAttr("shape", std::vector<int>({1, 8, 1, 1}));
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(1);
}
template <typename TargetT>
void test_reshape2_op(const platform::DeviceContext& context, bool use_gpu) {
framework::Scope scope;
std::unordered_set<std::string> parameters;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("reshape-X", {1, 2, 4});
validator.DeclOutputVar("reshape-Out", {1, 4, 2});
framework::OpDesc desc;
desc.SetType("reshape");
desc.SetInput("X", {"reshape-X"});
desc.SetOutput("Out", {"reshape-Out"});
// desc.SetAttr("shape", std::vector<int>({3, 2, 1, 3}));
desc.SetAttr("shape", std::vector<int>({0, -1, 2}));
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(reshape1_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_reshape1_op<::anakin::saber::NV>(ctx, true);
}
TEST(reshape2_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_reshape2_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(reshape1_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_reshape2_op<::anakin::saber::X86>(ctx, false);
}
TEST(reshape2_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_reshape2_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(reshape);
USE_ANAKIN_CONVERTER(reshape);
paddle/fluid/inference/anakin/convert/test_softmax_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_softmax_op(const platform::DeviceContext& context, bool use_gpu) {
framework::Scope scope;
std::unordered_set<std::string> parameters;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("softmax-X", {1, 10, 2});
validator.DeclOutputVar("softmax-Out", {1, 10, 2});
framework::OpDesc desc;
desc.SetType("softmax");
desc.SetInput("X", {"softmax-X"});
desc.SetOutput("Out", {"softmax-Out"});
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(softmax_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_softmax_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(relu_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_softmax_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(softmax);
USE_ANAKIN_CONVERTER(softmax);
paddle/fluid/inference/anakin/convert/test_split_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/split.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, int Axis>
void AnakinSliceTest(const platform::DeviceContext &context, bool use_gpu,
const std::vector<int> &in_shape,
const std::vector<int> &sections) {
std::unordered_set<std::string> parameters({""});
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("split_input", in_shape);
std::vector<std::string> output_vars;
for (size_t i = 0; i < sections.size(); ++i) {
auto out_shape = in_shape;
out_shape[Axis] = sections[i];
std::string output_name = "split_out" + std::to_string(i);
validator.DeclOutputVar(output_name, out_shape);
output_vars.push_back(output_name);
}
// Prepare Op description
framework::OpDesc desc;
desc.SetType("split");
desc.SetInput("X", {"split_input"});
desc.SetOutput("Out", output_vars);
desc.SetAttr("axis", Axis);
desc.SetAttr("num", 0);
desc.SetAttr("sections", sections);
validator.SetOp(*desc.Proto());
validator.Execute(1);
}
// batch = 0, axis = 1, same shape
TEST(split_op, test_same_shape_axis1_batch1) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
AnakinSliceTest<::anakin::saber::NV, 1>(ctx, true, {1, 4, 2, 2}, {2, 2});
}
// batch = 0, axis = 1, different shape
TEST(split_op, test_different_shape_axis1_batch1) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
AnakinSliceTest<::anakin::saber::NV, 1>(ctx, true, {1, 3, 2, 2}, {2, 1});
}
// batch = 0, axis = 2, same shape
TEST(split_op, test_same_shape_axis2_batch1) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
AnakinSliceTest<::anakin::saber::NV, 2>(ctx, true, {1, 3, 4, 2}, {2, 2});
}
// batch = 0, axis = 2, different shape
TEST(split_op, test_different_shape_axis2_batch1) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
AnakinSliceTest<::anakin::saber::NV, 2>(ctx, true, {1, 3, 3, 2}, {2, 1});
}
// batch = 0, axis = 3, same shape
TEST(split_op, test_same_shape_axis3_batch1) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
AnakinSliceTest<::anakin::saber::NV, 3>(ctx, true, {1, 3, 2, 4}, {2, 2});
}
// batch = 0, axis = 3, different shape
TEST(split_op, test_different_shape_axis3_batch1) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
AnakinSliceTest<::anakin::saber::NV, 3>(ctx, true, {1, 3, 2, 3}, {2, 1});
}
#ifdef ANAKIN_X86_PLACE
TEST(split_op, test_different_shape_axis1_batch1_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
AnakinSliceTest<::anakin::saber::X86, 1>(ctx, false, {1, 3, 2, 3}, {2, 1});
}
TEST(split_op, test_different_shape_axis2_batch1_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
AnakinSliceTest<::anakin::saber::X86, 2>(ctx, false, {1, 3, 4, 2}, {2, 2});
}
TEST(split_op, test_different_shape_axis3_batch1_cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
AnakinSliceTest<::anakin::saber::X86, 3>(ctx, false, {1, 3, 2, 4}, {2, 2});
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(split);
USE_ANAKIN_CONVERTER(split);
paddle/fluid/inference/anakin/convert/test_sum_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/sum.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
#include "paddle/fluid/operators/sum_op.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
static void test_sum_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("sum_x1", {1, 2, 1, 2});
validator.DeclInputVar("sum_x2", {1, 2, 1, 2});
validator.DeclOutputVar("sum_out", {1, 2, 1, 2});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("sum");
desc.SetInput("X", {"sum_x1", "sum_x2"});
desc.SetOutput("Out", {"sum_out"});
validator.SetOp(*desc.Proto());
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(sum_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_sum_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(sum_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_sum_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(sum);
USE_ANAKIN_CONVERTER(sum);
paddle/fluid/inference/anakin/convert/test_transpose_op.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
/* 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 <gtest/gtest.h>
#include "paddle/fluid/inference/anakin/convert/op_converter.h"
#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT>
void test_transpose1_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("transpose-X", {2, 3, 4, 5});
validator.DeclOutputVar("transpose-Out", {4, 2, 5, 3});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("transpose");
desc.SetInput("X", {"transpose-X"});
desc.SetOutput("Out", {"transpose-Out"});
desc.SetAttr("axis", std::vector<int>({2, 0, 3, 1}));
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(3);
}
template <typename TargetT>
void test_transpose2_op(const platform::DeviceContext& context, bool use_gpu) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
AnakinConvertValidation<TargetT, ::anakin::Precision::FP32> validator(
parameters, &scope, context, use_gpu);
validator.DeclInputVar("transpose-X", {3, 4, 5});
validator.DeclOutputVar("transpose-Out", {3, 5, 4});
// Prepare Op description
framework::OpDesc desc;
desc.SetType("transpose");
desc.SetInput("X", {"transpose-X"});
desc.SetOutput("Out", {"transpose-Out"});
desc.SetAttr("axis", std::vector<int>({0, 2, 1}));
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(1);
}
#ifdef PADDLE_WITH_CUDA
TEST(transpose1_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_transpose1_op<::anakin::saber::NV>(ctx, true);
}
TEST(transpose2_op, gpu) {
platform::CUDAPlace gpu_place(0);
platform::CUDADeviceContext ctx(gpu_place);
test_transpose2_op<::anakin::saber::NV>(ctx, true);
}
#endif
#ifdef ANAKIN_X86_PLACE
TEST(transpose1_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_transpose2_op<::anakin::saber::X86>(ctx, false);
}
TEST(transpose2_op, cpu) {
platform::CPUPlace cpu_place;
platform::CPUDeviceContext ctx(cpu_place);
test_transpose2_op<::anakin::saber::X86>(ctx, false);
}
#endif
} // namespace anakin
} // namespace inference
} // namespace paddle
USE_OP(transpose);
USE_ANAKIN_CONVERTER(transpose);
paddle/fluid/inference/anakin/convert/transpose.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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/transpose.h"
#include <algorithm>
#include <string>
#include <vector>
using anakin::PTuple;
namespace paddle {
namespace inference {
namespace anakin {
template <typename TargetT, ::anakin::Precision PrecisionT>
void TransposeOpConverter<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 input = op_desc.Input("X").front();
auto output = op_desc.Output("Out").front();
auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
this->engine_->AddOp(op_name, "Permute", {input}, {output});
auto axis = boost::get<std::vector<int>>(op_desc.GetAttr("axis"));
size_t axis_size = axis.size();
while (axis.size() < 4) {
axis.push_back(axis_size);
axis_size += 1;
}
this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dims", axis);
}
} // namespace anakin
} // namespace inference
} // namespace paddle
REGISTER_ANAKIN_OP_CONVERTER(transpose, TransposeOpConverter);
paddle/fluid/inference/anakin/convert/transpose.h 已删除 100644 → 0
浏览文件 @ 6fec3920
// 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 TransposeOpConverter : public AnakinOpConverter<TargetT, PrecisionT> {
public:
TransposeOpConverter() = default;
virtual void operator()(const framework::proto::OpDesc &op,
const framework::BlockDesc &block_desc,
const framework::Scope &scope,
bool test_mode) override;
virtual ~TransposeOpConverter() {}
};
} // namespace anakin
} // namespace inference
} // namespace paddle
paddle/fluid/inference/anakin/engine.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
此差异已折叠。
paddle/fluid/inference/anakin/engine.h 已删除 100644 → 0
浏览文件 @ 6fec3920
此差异已折叠。
paddle/fluid/inference/anakin/op_teller.cc 已删除 100644 → 0
浏览文件 @ 6fec3920
此差异已折叠。
paddle/fluid/inference/anakin/op_teller.h 已删除 100644 → 0
浏览文件 @ 6fec3920
此差异已折叠。
paddle/fluid/inference/analysis/argument.h
浏览文件 @ 399bda2b
......@@ -59,7 +59,6 @@ struct Argument {
using unique_ptr_t = std::unique_ptr<void, std::function<void(void*)>>;
using fusion_statis_t = std::unordered_map<std::string, int>;
using anakin_max_shape_t = std::map<std::string, std::vector<int>>;
bool Has(const std::string& key) const { return valid_fields_.count(key); }
// If we set the model using config.SetModelBuffer,
......@@ -184,19 +183,6 @@ struct Argument {
bool);
DECL_ARGUMENT_FIELD(tensorrt_use_calib_mode, TensorRtUseCalibMode, bool);
DECL_ARGUMENT_FIELD(anakin_max_input_shape, AnakinMaxInputShape,
anakin_max_shape_t);
DECL_ARGUMENT_FIELD(anakin_max_batch_size, AnakinMaxBatchSize, int);
DECL_ARGUMENT_FIELD(anakin_min_subgraph_size, AnakinMinSubgraphSize, int);
DECL_ARGUMENT_FIELD(anakin_precision_mode, AnakinPrecisionMode,
AnalysisConfig::Precision);
DECL_ARGUMENT_FIELD(anakin_auto_config_layout, AnakinAutoConfigLayout, bool);
DECL_ARGUMENT_FIELD(use_anakin, UseAnakin, bool);
DECL_ARGUMENT_FIELD(anakin_passes_filter, AnakinPassesFilter,
std::vector<std::string>);
DECL_ARGUMENT_FIELD(anakin_ops_filter, AnakinOpsFilter,
std::vector<std::string>);
DECL_ARGUMENT_FIELD(lite_passes_filter, LitePassesFilter,
std::vector<std::string>);
DECL_ARGUMENT_FIELD(lite_ops_filter, LiteOpsFilter, std::vector<std::string>);
......
paddle/fluid/inference/capi/c_api.h
浏览文件 @ 399bda2b
此差异已折叠。
paddle/fluid/operators/anakin/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ 6fec3920
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或