Merge pull request #3337 from tensor-tang/merge
Enable mkldnn_fc for general formatrevert-3824-remove_grad_op_type
Yiqun Liu
8 years ago
committed by
GitHub
commit
2e87d747bd
@ -0,0 +1,97 @@
|
||||
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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 "mkldnn.hpp"
|
||||
|
||||
namespace paddle {
|
||||
|
||||
typedef enum {
|
||||
MKLDNN_BASE = 1, // basical info of MKLDNN
|
||||
MKLDNN_TESTS = 1, // gtest info of MKLDNN
|
||||
MKLDNN_SIZES = 2, // size info of MKLDNN
|
||||
MKLDNN_FMTS = 3, // format info of MKLDNN
|
||||
MKLDNN_ALL = 4, // show all info of MKLDNN
|
||||
} MKLDNN_LOG_LEVEL;
|
||||
|
||||
/**
|
||||
* @brief MKLDNN CPU engine.
|
||||
*
|
||||
*/
|
||||
class CPUEngine {
|
||||
public:
|
||||
static CPUEngine& Instance() {
|
||||
// Thread-safe in C++11.
|
||||
static CPUEngine myInstance;
|
||||
return myInstance;
|
||||
}
|
||||
|
||||
// Disallow copy or move
|
||||
CPUEngine(const CPUEngine&) = delete; // Copy constructor
|
||||
CPUEngine(CPUEngine&&) = delete; // Move constructor
|
||||
CPUEngine& operator=(const CPUEngine&) = delete; // Copy assignment
|
||||
CPUEngine& operator=(CPUEngine&&) = delete; // Move assignment
|
||||
|
||||
mkldnn::engine& getEngine() { return cpuEngine_; }
|
||||
|
||||
protected:
|
||||
CPUEngine() : cpuEngine_(mkldnn::engine::cpu, 0) {}
|
||||
// CPUEngine() : cpuEngine_(mkldnn::engine::cpu_lazy, 0) {}
|
||||
~CPUEngine() {}
|
||||
|
||||
private:
|
||||
mkldnn::engine cpuEngine_;
|
||||
};
|
||||
|
||||
/**
|
||||
* @brief MKLDNN Stream.
|
||||
*
|
||||
*/
|
||||
class MKLDNNStream {
|
||||
public:
|
||||
MKLDNNStream() : ready_(false) { resetState(); }
|
||||
|
||||
virtual ~MKLDNNStream() {}
|
||||
|
||||
/**
|
||||
* @brief Submit stream
|
||||
* @param prims The primitives vector
|
||||
* @param block Waiting for the stream to complete
|
||||
*/
|
||||
void submit(std::vector<mkldnn::primitive>& prims, bool block = true) {
|
||||
resetState();
|
||||
stream_->submit(prims).wait(block);
|
||||
ready_ = false;
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Reset the mkldnn stream
|
||||
*/
|
||||
void resetState() {
|
||||
if (ready_) {
|
||||
return;
|
||||
}
|
||||
// TODO(TJ): change me when mkldnn have method to reset this state
|
||||
// stream_.reset(new mkldnn::stream(mkldnn::stream::kind::lazy));
|
||||
stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
|
||||
ready_ = true;
|
||||
}
|
||||
|
||||
private:
|
||||
bool ready_;
|
||||
std::shared_ptr<mkldnn::stream> stream_;
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,80 @@
|
||||
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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 "MKLDNNLayer.h"
|
||||
#include "mkldnn.hpp"
|
||||
|
||||
namespace paddle {
|
||||
|
||||
/**
|
||||
* @brief A subclass of MKLDNNLayer fc layer.
|
||||
*
|
||||
* The config file api is mkldnn_fc
|
||||
*/
|
||||
class MKLDNNFcLayer : public MKLDNNLayer {
|
||||
protected:
|
||||
// input layer size, can not be change after init
|
||||
size_t iLayerSize_; // == ic * ih * iw
|
||||
|
||||
// if has already init the weight
|
||||
bool hasInitedWgt_;
|
||||
|
||||
// if input layer has image size info (ih>1 && iw>1)
|
||||
bool hasSpatial_;
|
||||
|
||||
// fc weight and bias
|
||||
std::unique_ptr<Weight> weight_;
|
||||
std::unique_ptr<Weight> biases_;
|
||||
|
||||
public:
|
||||
explicit MKLDNNFcLayer(const LayerConfig& config)
|
||||
: MKLDNNLayer(config), hasInitedWgt_(false), hasSpatial_(true) {}
|
||||
|
||||
~MKLDNNFcLayer() {}
|
||||
|
||||
bool init(const LayerMap& layerMap,
|
||||
const ParameterMap& parameterMap) override;
|
||||
|
||||
void convertWeightsFromPaddle() override;
|
||||
|
||||
void convertWeightsToPaddle() override;
|
||||
|
||||
void forward(PassType passType) override;
|
||||
|
||||
void backward(const UpdateCallback& callback) override;
|
||||
|
||||
protected:
|
||||
/**
|
||||
* reshape the input image sizes
|
||||
* and reset output buffer size
|
||||
* and reset mkldnn forward
|
||||
*/
|
||||
void reshape();
|
||||
|
||||
/**
|
||||
* reset the forward primitve and memory
|
||||
* only would be called when input size changes
|
||||
*/
|
||||
void resetFwd();
|
||||
|
||||
/**
|
||||
* reset the backward primitve and memory for mkldnn fc
|
||||
* only would be called when needed
|
||||
*/
|
||||
void resetBwd();
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,132 @@
|
||||
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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 <vector>
|
||||
#include "Layer.h"
|
||||
#include "MKLDNNBase.h"
|
||||
#include "mkldnn.hpp"
|
||||
|
||||
DECLARE_bool(use_mkldnn);
|
||||
DECLARE_bool(use_mkldnn_wgt);
|
||||
|
||||
namespace paddle {
|
||||
|
||||
class MKLDNNLayer;
|
||||
typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
|
||||
|
||||
/**
|
||||
* @brief Base class of MKLDNNlayer.
|
||||
*
|
||||
*/
|
||||
class MKLDNNLayer : public Layer {
|
||||
protected:
|
||||
// batch size
|
||||
int bs_;
|
||||
// input image channel, height and width
|
||||
int ic_, ih_, iw_;
|
||||
// output image channel, height and width
|
||||
int oc_, oh_, ow_;
|
||||
|
||||
// backward also need reset after reset forward handle
|
||||
bool needResetBwd_;
|
||||
|
||||
// mkldnn engine, stream and primivtives
|
||||
mkldnn::engine engine_;
|
||||
std::shared_ptr<MKLDNNStream> stream_;
|
||||
std::shared_ptr<mkldnn::primitive> fwd_;
|
||||
std::shared_ptr<mkldnn::primitive> bwdWgt_;
|
||||
std::shared_ptr<mkldnn::primitive> bwdData_;
|
||||
std::vector<mkldnn::primitive> pipelineFwd_;
|
||||
std::vector<mkldnn::primitive> pipelineBwd_;
|
||||
|
||||
// TODO(TJ): change below memory as MKLDNNMatrixPtr type
|
||||
std::shared_ptr<mkldnn::memory> inVal_;
|
||||
std::shared_ptr<mkldnn::memory> inGrad_;
|
||||
std::shared_ptr<mkldnn::memory> outVal_;
|
||||
std::shared_ptr<mkldnn::memory> outGrad_;
|
||||
std::shared_ptr<mkldnn::memory> wgtVal_;
|
||||
std::shared_ptr<mkldnn::memory> wgtGrad_;
|
||||
std::shared_ptr<mkldnn::memory> biasVal_;
|
||||
std::shared_ptr<mkldnn::memory> biasGrad_;
|
||||
|
||||
public:
|
||||
explicit MKLDNNLayer(const LayerConfig& config)
|
||||
: Layer(config),
|
||||
bs_(0),
|
||||
ic_(0),
|
||||
ih_(0),
|
||||
iw_(0),
|
||||
oc_(0),
|
||||
oh_(0),
|
||||
ow_(0),
|
||||
needResetBwd_(true),
|
||||
engine_(mkldnn::engine::cpu, 0),
|
||||
stream_(nullptr),
|
||||
fwd_(nullptr),
|
||||
bwdWgt_(nullptr),
|
||||
bwdData_(nullptr) {}
|
||||
|
||||
~MKLDNNLayer() {}
|
||||
|
||||
virtual bool init(const LayerMap& layerMap,
|
||||
const ParameterMap& parameterMap) {
|
||||
if (!Layer::init(layerMap, parameterMap)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
|
||||
<< "Please set WITH_MKLDNN=ON "
|
||||
<< "and set use_mkldnn=True";
|
||||
stream_.reset(new MKLDNNStream());
|
||||
engine_ = CPUEngine::Instance().getEngine();
|
||||
|
||||
// TODO(TJ): deivecId
|
||||
return true;
|
||||
}
|
||||
|
||||
/**
|
||||
* convert weight from paddle format to mkldnn format
|
||||
* weight_ will be override
|
||||
*/
|
||||
virtual void convertWeightsFromPaddle() {}
|
||||
|
||||
/**
|
||||
* convert mkldnn weight to paddle format
|
||||
* weight_ will be override
|
||||
*/
|
||||
virtual void convertWeightsToPaddle() {}
|
||||
|
||||
/**
|
||||
* print info about sizes
|
||||
*/
|
||||
virtual void printSizeInfo() {
|
||||
VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_
|
||||
<< ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_
|
||||
<< ", oh: " << oh_ << ", ow: " << ow_;
|
||||
}
|
||||
|
||||
// TODO(TJ): move to MkldnnMatrix
|
||||
// create memory desc
|
||||
inline mkldnn::memory::desc createMD(
|
||||
mkldnn::memory::dims dims,
|
||||
mkldnn::memory::format fmt,
|
||||
mkldnn::memory::data_type type = mkldnn::memory::data_type::f32) {
|
||||
// TODO(TJ): isFmtSuppoted(fmt)
|
||||
return mkldnn::memory::desc(dims, type, fmt);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,120 @@
|
||||
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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 <string>
|
||||
#include <vector>
|
||||
#include "LayerGradUtil.h"
|
||||
#include "paddle/gserver/layers/MKLDNNBase.h"
|
||||
|
||||
namespace paddle {
|
||||
|
||||
/**
|
||||
* @brief test the functionality of Mkldnnlayers
|
||||
* refer to paddle original function
|
||||
*/
|
||||
class MKLDNNTester {
|
||||
enum {
|
||||
DNN = 0, // MKLDNN layer
|
||||
REF = 1, // Reference layer
|
||||
NUM = 2, // Number of total
|
||||
};
|
||||
|
||||
protected:
|
||||
std::vector<TestConfig> configs_;
|
||||
vector<string> layerNames_;
|
||||
vector<vector<DataLayerPtr>> dataLayers_;
|
||||
vector<vector<Argument>> datas_;
|
||||
vector<LayerMap> layerMaps_;
|
||||
vector<vector<ParameterPtr>> parameters_;
|
||||
vector<LayerPtr> testLayers_;
|
||||
LayerPtr dnnLayer_, refLayer_;
|
||||
|
||||
/// run some iterations, all the result should pass
|
||||
size_t iter_;
|
||||
/// whether to print out the details
|
||||
bool log_;
|
||||
/// vlog level to print the matrix details datas
|
||||
int lvl_;
|
||||
/// epsilon
|
||||
float eps_;
|
||||
/// input image size, default 1
|
||||
size_t ih_, iw_;
|
||||
|
||||
public:
|
||||
explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) {
|
||||
iter_ = iter;
|
||||
eps_ = epsilon;
|
||||
log_ = false;
|
||||
lvl_ = MKLDNN_ALL;
|
||||
}
|
||||
|
||||
~MKLDNNTester() {}
|
||||
|
||||
public:
|
||||
void run(const TestConfig& dnn,
|
||||
const TestConfig& ref,
|
||||
size_t batchSize,
|
||||
size_t inputImgH = 1,
|
||||
size_t inputImgW = 1,
|
||||
size_t iter = 3,
|
||||
float epsilon = 1e-4,
|
||||
bool log = false,
|
||||
int level = MKLDNN_ALL);
|
||||
void setLogLevel(int lvl) { lvl_ = lvl; }
|
||||
|
||||
private:
|
||||
void reset(const TestConfig& dnn, const TestConfig& ref, size_t batchSize);
|
||||
void setInputImgSize();
|
||||
void runOnce();
|
||||
|
||||
void randomWgtDatas();
|
||||
void randomBotDatas();
|
||||
void randomTopDiffs();
|
||||
|
||||
void checkForward();
|
||||
void checkBackwardData();
|
||||
void checkBackwardWgts();
|
||||
|
||||
void clearWgtDiffs();
|
||||
void clearBotDiffs();
|
||||
void clearBotDiffs(int n); // clear specific layer
|
||||
void clearTopDatas();
|
||||
|
||||
void printTopDatas();
|
||||
void printMatrix(const MatrixPtr& m);
|
||||
void printVector(const VectorPtr& v);
|
||||
|
||||
void saveWgt(const vector<ParameterPtr>& from, vector<VectorPtr>& to);
|
||||
void restoreWgt(const vector<VectorPtr>& from, vector<ParameterPtr>& to);
|
||||
|
||||
double compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2);
|
||||
double compareVector(const VectorPtr& v1, const VectorPtr& v2);
|
||||
|
||||
/**
|
||||
* Get delta percent
|
||||
* if many(>failRate) wrong(abs(dnn-ref)/abs(ref)>thres) points return the
|
||||
* max(diff/ref)
|
||||
* else return sum(abs(a-b)) / sum(abs(b))
|
||||
* The return value should smaller than eps when passing.
|
||||
*/
|
||||
double getDelta(const real* d1,
|
||||
const real* d2,
|
||||
size_t len,
|
||||
const float failRate = 1e-3,
|
||||
const float thres = 0.1);
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,76 @@
|
||||
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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 <string>
|
||||
#include <vector>
|
||||
#include "MKLDNNTester.h"
|
||||
#include "ModelConfig.pb.h"
|
||||
|
||||
using namespace paddle; // NOLINT
|
||||
|
||||
DECLARE_bool(thread_local_rand_use_global_seed);
|
||||
DECLARE_bool(use_gpu);
|
||||
DECLARE_bool(use_mkldnn);
|
||||
|
||||
struct testFCDesc {
|
||||
int bs;
|
||||
int ic;
|
||||
int oc;
|
||||
int ih, iw; // oh == ow == 1
|
||||
};
|
||||
|
||||
void testFcLayer(const testFCDesc& pm) {
|
||||
const std::string compareTypes[] = {"mkldnn_fc", "fc"};
|
||||
TestConfig cfg;
|
||||
cfg.layerConfig.set_type(compareTypes[0]);
|
||||
cfg.layerConfig.set_size(pm.oc);
|
||||
cfg.inputDefs.push_back(
|
||||
{INPUT_DATA,
|
||||
"layer_0",
|
||||
/* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw),
|
||||
/* size of weight= */ size_t(pm.oc * pm.ic * pm.ih * pm.iw)});
|
||||
cfg.layerConfig.add_inputs();
|
||||
|
||||
MKLDNNTester tester;
|
||||
for (auto biasSize : {pm.oc, 0}) {
|
||||
cfg.biasSize = biasSize;
|
||||
TestConfig ref = cfg;
|
||||
ref.layerConfig.set_type(compareTypes[1]);
|
||||
for (auto bs : {pm.bs, 1}) {
|
||||
tester.run(cfg, ref, bs, pm.ih, pm.iw);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
TEST(MKLDNNLayer, FcLayer) {
|
||||
testFcLayer({/*bs*/ 2, /*ic*/ 2, /*oc*/ 3, /*ih*/ 1, /*iw*/ 1});
|
||||
testFcLayer({/*bs*/ 3, /*ic*/ 7, /*oc*/ 19, /*ih*/ 1, /*iw*/ 1});
|
||||
testFcLayer({/*bs*/ 8, /*ic*/ 16, /*oc*/ 32, /*ih*/ 13, /*iw*/ 13});
|
||||
testFcLayer({/*bs*/ 4, /*ic*/ 12, /*oc*/ 18, /*ih*/ 13, /*iw*/ 11});
|
||||
testFcLayer({/*bs*/ 2, /*ic*/ 64, /*oc*/ 32, /*ih*/ 16, /*iw*/ 16});
|
||||
testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16});
|
||||
}
|
||||
|
||||
// TODO(TJ): add branch test
|
||||
|
||||
int main(int argc, char** argv) {
|
||||
testing::InitGoogleTest(&argc, argv);
|
||||
FLAGS_use_gpu = false;
|
||||
FLAGS_use_mkldnn = true;
|
||||
initMain(argc, argv);
|
||||
FLAGS_thread_local_rand_use_global_seed = true;
|
||||
srand(1);
|
||||
return RUN_ALL_TESTS();
|
||||
}
|
||||
Loading…
Reference in new issue