提交 94e442d4 编写于 作者: T tensor-tang

add cpp file of MKLDNNLayer

上级 9e38dafa
develop 1.8.5 2.0.1-rocm-post 2.4.1 Ligoml-patch-1 OliverLPH-patch-1 OliverLPH-patch-2 PaddlePM-patch-1 PaddlePM-patch-2 ZHUI-patch-1 add_default_att add_kylinv10 add_model_benchmark_ci add_some_yaml_config addfile all_new_design_exec ascendrc ascendrelease bugfix-eval-frame-leakgae cherry-pick-fix-customOP-random-fail cherry_undefined_var compile_windows cp_2.4_fix_numpy delete_2.0.1-rocm-post delete_add_default_att delete_all_new_design_exec delete_ascendrc delete_compile_windows delete_delete_addfile delete_disable_iterable_dataset_unittest delete_fix_dataloader_memory_leak delete_fix_imperative_dygraph_error delete_fix_retry_ci delete_fix_undefined_var delete_improve_sccache delete_incubate/lite delete_paddle_tiny_install delete_paralleltest delete_prv-disable-more-cache delete_revert-31068-fix_conv3d_windows delete_revert-31562-mean delete_revert-33630-bug-fix delete_revert-34159-add_npu_bce_logical_dev delete_revert-34910-spinlocks_for_allocator delete_revert-35069-revert-34910-spinlocks_for_allocator delete_revert-36057-dev/read_flags_in_ut dingjiaweiww-patch-1 disable_iterable_dataset_unittest dy2static enable_eager_model_test final_state_gen_python_c final_state_intermediate fix-numpy-issue fix-run-program-grad-node-mem fix_check fix_concat_slice fix_custom_device_copy_sync fix_dataloader_memory_leak fix_dlpack_for fix_imperative_dygraph_error fix_newexe_gc fix_npu_ci fix_op_flops fix_retry_ci fix_rnn_docs fix_tensor_type fix_undefined_var fix_var_stop_gradient_error fixiscan fixiscan1 fixiscan2 fixiscan3 github/fork/123malin/netifaces github/fork/123malin/tdm_abacus github/fork/AshburnLee/dev_unique github/fork/ForFishes/fix_memory_matmul github/fork/ForFishes/rm_fluid github/fork/LielinJiang/move-2.0-api github/fork/LielinJiang/visual-dl-cb github/fork/LiuChiachi/add-transformer-generate-square-subsequent-mask-api github/fork/LiuChiachi/fix-example-code-for-hapi-Model github/fork/LiuChiachi/remove-input-requirment-in-dygraph-Model github/fork/MrChengmo/fix_ps_profiler github/fork/MrChengmo/update_ps_heter github/fork/PWhiddy/patch-1 github/fork/Shixiaowei02/dev/save_load_upgrade github/fork/TCChenlong/fix_hapi github/fork/TCChenlong/fix_inden github/fork/Thunderbrook/xpu_slice github/fork/XieYunshen/disable_ut_test_parallel_executor_fetch_isolated_var github/fork/XieYunshen/disable_ut_test_parallel_executor_fetch_isolated_var_2 github/fork/XieYunshen/disable_ut_test_parallel_executor_fetch_isolated_var_3 github/fork/XieYunshen/timeout_20S_ut github/fork/ZeyuChen/remove-nltk github/fork/arlesniak/arlesniak/selective__mkldnn_flags github/fork/baiyfbupt/code_doc_mig github/fork/chalsliu/set_timeout github/fork/chen-zhiyu/develop github/fork/chenwhql/ci/try_to_find_test_buffer_shared_memory_reuse_pass_error github/fork/chenwhql/dygraph/remove_scale_loss_and_apply_collective_grads github/fork/chenwhql/saveload/add_get_inference_program github/fork/chenwhql/saveload/remove_save_load_config github/fork/cryoco/pass-compatibility-trt github/fork/danleifeng/isempty_api2.0 github/fork/frankwhzhang/api_transfer github/fork/hbwx24/error_msg/cuda_kernel_error_msg github/fork/heavengate/cherry_yolo_box github/fork/heavengate/update_yolo_box github/fork/iclementine/rnn_fix github/fork/iducn/testestse github/fork/jczaja/prv-25537-fix github/fork/jeff41404/release/1.8 github/fork/jiweibo/api_2.0 github/fork/jiweibo/fix_lite_resnet50_test github/fork/juncaipeng/fix_doc_1 github/fork/lfchener/sample_code github/fork/littletomatodonkey/fix_reg_doc github/fork/liym27/dy2stat_update_assign_to_rc20 github/fork/luotao1/profiler_ut github/fork/mapingshuo/add_wait github/fork/mapingshuo/doc_2.0 github/fork/mapingshuo/zero-0.5 github/fork/miraiwk/dev github/fork/pangyoki/add-Categorical-class-branch github/fork/pangyoki/add-multinomial-op-branch github/fork/pangyoki/fix-test_distritbution-CI github/fork/qjing666/doublegrad github/fork/qjing666/fix_hdfs_download github/fork/sandyhouse/add_gather_etc github/fork/sandyhouse/add_send_recv_alltoall_etc github/fork/sandyhouse/pipeline_exe_run github/fork/seiriosPlus/feature/large_scale_kv_save_delta github/fork/seiriosPlus/fix/paddle_errors_fix github/fork/seiriosPlus/fix/paddle_op_errors github/fork/shangzhizhou/fix_test_activation_op_random_bug github/fork/smallv0221/yxp0924 github/fork/smallv0221/yxp0925 github/fork/swtkiwi/del-matplotlib github/fork/tianshuo78520a/kunlun_test github/fork/tianshuo78520a/update_dockerfile github/fork/wanghaoshuang/bert_fuse github/fork/wanghaoshuang/label_smooth github/fork/wanghuancoder/develop_CUDASynchronize github/fork/wanghuancoder/develop_Layer_doc github/fork/wanghuancoder/develop_ParameterList_doc github/fork/wanghuancoder/develop_Sequential_doc github/fork/wanghuancoder/develop_bilinear_tensor_product github/fork/wanghuancoder/develop_coverage_build_sh github/fork/wanghuancoder/develop_in_dynamic_mode_doc github/fork/wanghuancoder/develop_unique_name_doc github/fork/wangxicoding/fleet_meta_combine github/fork/wawltor/error_message_fix_5 github/fork/willthefrog/remove_l2_norm github/fork/windstamp/momentum_op github/fork/windstamp/mv_op_5 github/fork/windstamp/normal_api github/fork/wojtuss/wojtuss/fusion_gru_quantization github/fork/wojtuss/wojtuss/quantization-with-shift github/fork/wzzju/fix_err_info github/fork/wzzju/pure_fp16 github/fork/xiemoyuan/op_error_message github/fork/xiemoyuan/optimize_error_message github/fork/yaoxuefeng6/fix_doc github/fork/yaoxuefeng6/mod_dataset_v2 github/fork/yongqiangma/lod github/fork/ysh329/fix-clip-by-norm-error github/fork/ysh329/fix-error-clip-by-value github/fork/yukavio/error_info github/fork/zhangting2020/conv_filter_grad github/fork/zhangting2020/is_compile_with_cuda github/fork/zhangting2020/place_doc github/fork/zhangting2020/program github/fork/zhhsplendid/fix_any github/fork/zhhsplendid/refine_api2 github/fork/zhhsplendid/refine_api2_test github/fork/zhhsplendid/refine_api_test_ptb_lm github/fork/zhhsplendid/refine_api_test_resnet github/fork/zhhsplendid/refine_api_test_simnet github/fork/zhiqiu/dev/refine_initializer github/fork/zhiqiu/dev/remove_inplace_argument github/fork/zlsh80826/nvinfer_plugin_var_len_cuda11 hack_event improve_sccache incuabte/new_frl incubate/frl_train_eval incubate/infrt incubate/lite incubate/new_frl incubate/new_frl_rc incubate/stride inplace_addto layer_norm make_flag_adding_easier master matmul_double_grad move_embedding_to_phi move_histogram_to_pten move_sgd_to_phi move_slice_to_pten move_temporal_shift_to_phi move_yolo_box_to_phi npu_fix_alloc numel operator_opt paddle_tiny_install paralleltest pass-compile-eval-frame preln_ernie prv-disable-more-cache prv-md-even-more prv-onednn-2.5 prv-reshape-mkldnn-ut2 pten_tensor_refactor release-deleted/2.5 release-rc/2.5 release/0.11.0 release/0.12.0 release/0.13.0 release/0.14.0 release/0.15.0 release/1.0.0 release/1.1 release/1.2 release/1.3 release/1.4 release/1.5 release/1.6 release/1.7 release/1.8 release/2.0 release/2.0-alpha release/2.0-beta release/2.0-rc release/2.0-rc1 release/2.1 release/2.2 release/2.3 release/2.3-fc-ernie-fix release/2.4 release/2.5 release/lite-0.1 release/llm_2.5 revert-24981-add_device_attr_for_regulization revert-26856-strategy_example2 revert-27520-disable_pr revert-31068-fix_conv3d_windows revert-31562-mean revert-32290-develop-hardlabel revert-33037-forci revert-33475-fix_cifar_label_dimension revert-33630-bug-fix revert-34159-add_npu_bce_logical_dev revert-34406-add_copy_from_tensor revert-34910-spinlocks_for_allocator revert-35069-revert-34910-spinlocks_for_allocator revert-36057-dev/read_flags_in_ut revert-36201-refine_fast_threaded_ssa_graph_executor revert-36985-add_license revert-37318-refactor_dygraph_to_eager revert-37926-eager_coreops_500 revert-37956-revert-37727-pylayer_support_tuple revert-38100-mingdong revert-38301-allocation_rearrange_pr revert-38703-numpy_bf16_package_reupload revert-38732-remove_useless_header_in_elementwise_mul_grad revert-38959-Reduce_Grad revert-39143-adjust_empty revert-39227-move_trace_op_to_pten revert-39268-dev/remove_concat_fluid_kernel revert-40170-support_partial_grad revert-41056-revert-40727-move_some_activaion_to_phi revert-41065-revert-40993-mv_ele_floordiv_pow revert-41068-revert-40790-phi_new revert-41944-smaller_inference_api_test revert-42149-do-not-reset-default-stream-for-stream-safe-cuda-allocator revert-43155-fix_ut_tempfile revert-43882-revert-41944-smaller_inference_api_test revert-45808-phi/simplify_size_op revert-46827-deform_comment revert-47325-remove_cudnn_hardcode revert-47645-add_npu_storage_dims revert-48815-set_free_when_no_cache_hit_default_value_true revert-49499-test_ninja_on_ci revert-49654-prim_api_gen revert-49673-modify_get_single_cov revert-49763-fix_static_composite_gen revert-50158-fix_found_inf_bug_for_custom_optimizer revert-50188-refine_optimizer_create_accumulators revert-50335-fix_optminizer_set_auxiliary_var_bug revert-51676-flag_delete revert-51850-fix_softmaxce_dev revert-52175-dev_peak_memory revert-52186-deve revert-52523-test_py38 revert-52912-develop revert-53248-set_cmake_policy revert-54029-fix_windows_compile_bug revert-54068-support_translating_op_attribute revert-54214-modify_cmake_dependencies revert-54370-offline_pslib revert-54391-fix_cmake_md5error revert-54411-fix_cpp17_compile revert-54466-offline_pslib revert-54480-cmake-rocksdb revert-55568-fix_BF16_bug1 revert-56328-new_ir_support_vector_type_place_transfer revert-56366-fix_openssl_bug revert-56545-revert-56366-fix_openssl_bug revert-56620-fix_new_ir_ocr_bug revert-56925-check_inputs_grad_semantic revert-57005-refine_stride_flag rocm_dev_0217 sd_conv_linear_autocast semi-auto/rule-base support-0D-sort support_weight_transpose test_benchmark_ci test_feature_precision_test_c test_for_Filtetfiles test_model_benchmark test_model_benchmark_ci zhiqiu-patch-1 v2.5.1 v2.5.0 v2.5.0-rc1 v2.5.0-rc0 v2.4.2 v2.4.1 v2.4.0 v2.4.0-rc0 v2.3.2 v2.3.1 v2.3.0 v2.3.0-rc0 v2.2.2 v2.2.1 v2.2.0 v2.2.0-rc0 v2.2.0-bak0 v2.1.3 v2.1.2 v2.1.1 v2.1.0 v2.1.0-rc0 v2.0.2 v2.0.1 v2.0.0 v2.0.0-rc1 v2.0.0-rc0 v2.0.0-beta0 v2.0.0-alpha0 v1.8.5 v1.8.4 v1.8.3 v1.8.2 v1.8.1 v1.8.0 v1.7.2 v1.7.1 v1.7.0 v1.6.3 v1.6.2 v1.6.1 v1.6.0 v1.6.0-rc0 v1.5.2 v1.5.1 v1.5.0 v1.4.1 v1.4.0 v1.3.2 v1.3.1 v1.3.0 v1.2.1 v1.2.0 v1.1.0 v1.0.2 v1.0.1 v1.0.0 v1.0.0-rc0 v0.15.0 v0.15.0-rc0 v0.14.0 v0.13.0 v0.12.0 v0.11.1a2 v0.11.1a1 v0.11.0 lite-v0.1
6 合并请求!11636[IMPORTANT] MKLDNN layout: Support for sum operator,!8482Release/0.11.0,!8190Release/0.11.0,!8189Release/0.11.0,!6633给线性回归的get-started代码加上了预测的示例~~,!4953refine the mkldnn logic
/* 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 "MKLDNNLayer.h"
using namespace mkldnn; // NOLINT
typedef memory::format format;
namespace paddle {
bool MKLDNNLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
<< "Please set WITH_MKLDNN=ON "
<< "and set use_mkldnn=True";
CHECK(!useGpu_) << "Do not support GPU yet";
// set device id before Layer::init
setDevice(MKLDNN_DEVICE);
// change param device to MKLDNN device
setParamsDevice(MKLDNN_DEVICE, parameterMap);
if (!Layer::init(layerMap, parameterMap)) {
return false;
}
setOutputMap();
checkCPUOutputsNumber();
stream_.reset(new MKLDNNStream());
engine_ = CPUEngine::Instance().getEngine();
return true;
}
void MKLDNNLayer::forward(PassType passType) {
passType_ = passType;
{
REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
CHECK(!inputLayers_.empty());
copySeqInfoToOutputs();
size_t elemenCnt = inputLayers_[0]->getOutputValue()->getElementCnt();
if (inputElemenCnt_ != elemenCnt) {
VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
// reset when input total sizes changed, not only the batchsize
inputElemenCnt_ = elemenCnt;
pipelineFwd_.clear();
reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
// all cpu device output grad or value share output's
shareCPUDevice();
resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
// MKLDNNLayer output value should be MKLDNNMatrix
// so external output value is necessary.
// then external input value is not necessary,
// since input may be mkldnn internal buffer.
CHECK(extOutVal_) << "external output value is necessary";
output_.value = std::dynamic_pointer_cast<Matrix>(extOutVal_);
CHECK(inVal_ && outVal_) << "internal memories are necessary";
if (cvtInVal_) {
pipelineFwd_.insert(pipelineFwd_.begin(), *cvtInVal_);
}
if (cvtOutVal_) {
pipelineFwd_.push_back(*cvtOutVal_);
}
convertWeightsFromPaddle();
printSizeInfo();
printValueFormat();
needResetBwd_ = true;
}
if (inputLayers_[0]->getType() == "data") {
// Update input value data when input layer is "data" type,
// since the input value data address might be changed.
CHECK(extInVal_);
extInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
if (!outputOnlyMKLDNN_) {
clearGrads();
}
stream_->submit(pipelineFwd_);
}
{
REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
forwardActivation();
}
}
void MKLDNNLayer::backward(const UpdateCallback& callback) {
if (needResetBwd_) {
VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
pipelineBwd_.clear();
pipelineMergeGrad_.clear();
mergeGrad_ = nullptr;
resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
// external output grad is not necessary
// since output may be mkldnn internal buffer or merge them directly.
CHECK(outGrad_) << "internal output grad is necessary";
if (cvtOutGrad_) {
pipelineBwd_.insert(pipelineBwd_.begin(), *cvtOutGrad_);
}
if (cvtInGrad_) {
pipelineBwd_.push_back(*cvtInGrad_);
}
printGradFormat();
needResetBwd_ = false;
}
// merge grad must before backward activation
if (mergeGrad_) {
REGISTER_TIMER_INFO("MergeBpGrad", getName().c_str());
stream_->submit(pipelineMergeGrad_);
}
{
REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
backwardActivation();
}
{
REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
stream_->submit(pipelineBwd_);
}
{
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
updateWeights(callback);
}
}
void MKLDNNLayer::reshapeInput(int& batchsize, int& height, int& width) {
const Argument& input = inputLayers_[0]->getOutput();
batchsize = input.getBatchSize();
int h = input.getFrameHeight();
int w = input.getFrameWidth();
if (h != 0) {
height = h;
}
if (w != 0) {
width = w;
}
}
void MKLDNNLayer::reshapeOutput(size_t height, size_t width) {
output_.setFrameHeight(height);
output_.setFrameWidth(width);
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
outputOtherDevice_[i].setFrameHeight(height);
outputOtherDevice_[i].setFrameWidth(width);
}
}
void MKLDNNLayer::resetWithMatrix(MKLDNNMatrixPtr& dnn,
const MatrixPtr& mat,
memory::primitive_desc pd) {
dnn = nullptr;
if (mat == nullptr) {
return;
}
dnn = MKLDNNMatrix::create(pd, mat);
}
void MKLDNNLayer::resetInValue(
MKLDNNMatrixPtr& in, const std::shared_ptr<memory::primitive_desc>& intPD) {
cvtInVal_ = nullptr;
extInVal_ = nullptr;
in = nullptr;
CHECK_GT(bs_ * ic_ * ih_ * iw_, 0);
auto extPD = MKLDNNMatrix::createPrimitiveDesc(
{bs_, ic_, ih_, iw_}, format::nchw, engine_);
const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
in = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
CHECK_EQ(inputIsOnlyMKLDNN(), in != nullptr);
if (in == nullptr || in->getFormat() == format::nc) {
in = MKLDNNMatrix::create(extPD, inMat);
}
extInVal_ = isPaddleFormat(in->getFormat()) ? in : nullptr;
if (in->getFormat() == format::nc) {
CHECK(ih_ == 1 && iw_ == 1);
}
if (nullptr == intPD || in->getPrimitiveDesc() == *intPD) {
return;
}
// need create reorder
in = MKLDNNMatrix::create(*intPD);
extInVal_ = extInVal_ ? extInVal_ : MKLDNNMatrix::create(extPD, inMat);
cvtInVal_ = MKLDNNMatrix::createReorder(extInVal_, in);
CHECK(cvtInVal_) << "should not be emptry";
}
void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out,
memory::primitive_desc intPD) {
cvtOutVal_ = nullptr;
out = MKLDNNMatrix::create(intPD, output_.value);
extOutVal_ = out;
if (outputIsOnlyMKLDNN() || isPaddleFormat(extOutVal_->getFormat())) {
return;
}
// need create reorder
CHECK_GT(bs_ * oc_ * oh_ * ow_, 0);
extOutVal_ = MKLDNNMatrix::create(
memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_, output_.value);
out = MKLDNNMatrix::create(intPD);
cvtOutVal_ = MKLDNNMatrix::createReorder(out, extOutVal_);
CHECK(cvtOutVal_) << "should not be empty";
}
void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
memory::primitive_desc intPD) {
cvtInGrad_ = nullptr;
extInGrad_ = nullptr;
in = nullptr;
LayerPtr& input = inputLayers_[0];
if (input->getOutputGrad() == nullptr) {
// no need input grad
return;
}
CHECK(inputIsOnlyMKLDNN() || input->getOutputMapSize() <= 1)
<< "only support input is MKLDNN layer or only have one output layer";
// when input is a mkldnn branch node,
// this layer will save input grad to a internal buffer,
// and the mkldnn input layer will merge them to actual prev->output_.grad
const MatrixPtr& inMat =
input->getOutputMapSize() <= 1 ? input->getOutputGrad() : nullptr;
in = MKLDNNMatrix::create(intPD, inMat);
Argument& arg = input->getOutput(this->getName());
arg.grad = std::dynamic_pointer_cast<Matrix>(in);
CHECK(inVal_ != nullptr && inVal_->getPrimitiveDesc() == intPD)
<< "should have internal input value and primitive desc must equal";
if (inputIsOnlyMKLDNN()) {
return;
}
extInGrad_ = in;
if (isPaddleFormat(extInGrad_->getFormat())) {
return;
}
// need create reorder
CHECK(extInVal_ != nullptr && isPaddleFormat(extInVal_->getFormat()))
<< "should have external input value and the format must be nchw(nc)";
extInGrad_ = MKLDNNMatrix::create(extInVal_->getPrimitiveDesc(), inMat);
CHECK(inVal_ != nullptr && inVal_->getPrimitiveDesc() == intPD)
<< "should have internal input value and primitive desc must equal";
in = MKLDNNMatrix::create(intPD);
cvtInGrad_ = MKLDNNMatrix::createReorder(in, extInGrad_);
CHECK(cvtInGrad_);
}
void MKLDNNLayer::resetOutGrad(MKLDNNMatrixPtr& out,
memory::primitive_desc intPD) {
cvtOutGrad_ = nullptr;
extOutGrad_ = nullptr;
out = nullptr;
MatrixPtr& outMat = output_.grad;
out = MKLDNNMatrix::create(intPD, outMat);
resetMergeGrad(out);
if (outputIsOnlyMKLDNN()) {
return;
}
CHECK_LE(outputMap_.size(), 1U) << "do not support mixed with cpu device";
extOutGrad_ = out;
if (isPaddleFormat(extOutGrad_->getFormat())) {
return;
}
// need create reorder
CHECK(extOutVal_ != nullptr && isPaddleFormat(extOutVal_->getFormat()))
<< "should have external output value and the format must be nchw(nc)";
extOutGrad_ = MKLDNNMatrix::create(extOutVal_->getPrimitiveDesc(), outMat);
CHECK(outVal_ != nullptr && outVal_->getPrimitiveDesc() == intPD)
<< "should have internal output value and primitive desc must equal";
out = MKLDNNMatrix::create(intPD);
cvtOutGrad_ = MKLDNNMatrix::createReorder(extOutGrad_, out);
CHECK(cvtOutGrad_);
}
void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) {
mergeGrad_ = nullptr;
pipelineMergeGrad_.clear();
if (outputMap_.size() <= 1 || !outputIsOnlyMKLDNN()) {
// do not merge when output is not all MKLDNN or only one output
return;
}
CHECK(out) << "should have reset internal ouput grad";
std::vector<double> scales(outputMap_.size(), 1.0);
std::vector<memory::primitive_desc> srcPDs;
std::vector<primitive::at> srcs;
for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {
MKLDNNMatrixPtr src =
std::dynamic_pointer_cast<MKLDNNMatrix>(it->second->grad);
VLOG(MKLDNN_BASE) << getName() << " has output grad " << it->first;
CHECK(src) << "should be MKLDNNMatrix";
auto srcDims = src->getDims();
auto dstDims = out->getDims();
CHECK_EQ(srcDims.size(), dstDims.size());
for (size_t i = 0; i < srcDims.size(); ++i) {
CHECK_EQ(srcDims[i], dstDims[i]);
}
srcPDs.push_back(src->getPrimitiveDesc());
srcs.push_back(*src);
}
// TODO(TJ): remove me when mkldnn sum support different formats
for (size_t i = 1; i < srcPDs.size(); ++i) {
CHECK(srcPDs[0] == srcPDs[i]);
}
tmpOutGrad_ = out;
tmpCvt_ = nullptr;
if (out->getPrimitiveDesc() != srcPDs[0]) {
tmpOutGrad_ = MKLDNNMatrix::create(srcPDs[0]);
tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out);
CHECK(tmpCvt_);
pipelineMergeGrad_.push_back(*tmpCvt_);
}
auto sumPD =
sum::primitive_desc(tmpOutGrad_->getMemoryDesc(), scales, srcPDs);
mergeGrad_.reset(new sum(sumPD, srcs, *tmpOutGrad_));
pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_);
}
} // namespace paddle
......@@ -119,119 +119,9 @@ public:
~MKLDNNLayer() {}
virtual bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
<< "Please set WITH_MKLDNN=ON "
<< "and set use_mkldnn=True";
CHECK(!useGpu_) << "Do not support GPU yet";
// set device id before Layer::init
setDevice(MKLDNN_DEVICE);
// change param device to MKLDNN device
setParamsDevice(MKLDNN_DEVICE, parameterMap);
if (!Layer::init(layerMap, parameterMap)) {
return false;
}
setOutputMap();
checkCPUOutputsNumber();
stream_.reset(new MKLDNNStream());
engine_ = CPUEngine::Instance().getEngine();
return true;
}
void forward(PassType passType) override {
passType_ = passType;
{
REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
CHECK(!inputLayers_.empty());
copySeqInfoToOutputs();
size_t elemenCnt = inputLayers_[0]->getOutputValue()->getElementCnt();
if (inputElemenCnt_ != elemenCnt) {
VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
// reset when input total sizes changed, not only the batchsize
inputElemenCnt_ = elemenCnt;
pipelineFwd_.clear();
reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
// all cpu device output grad or value share output's
shareCPUDevice();
resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
// MKLDNNLayer output value should be MKLDNNMatrix
// so external output value is necessary.
// then external input value is not necessary,
// since input may be mkldnn internal buffer.
CHECK(extOutVal_) << "external output value is necessary";
output_.value = std::dynamic_pointer_cast<Matrix>(extOutVal_);
CHECK(inVal_ && outVal_) << "internal memories are necessary";
if (cvtInVal_) {
pipelineFwd_.insert(pipelineFwd_.begin(), *cvtInVal_);
}
if (cvtOutVal_) {
pipelineFwd_.push_back(*cvtOutVal_);
}
convertWeightsFromPaddle();
printValueFormat();
needResetBwd_ = true;
}
if (inputLayers_[0]->getType() == "data") {
// Update input value data when input layer is "data" type,
// since the input value data address might be changed.
CHECK(extInVal_);
extInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
if (!outputOnlyMKLDNN_) {
clearGrads();
}
stream_->submit(pipelineFwd_);
}
{
REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
forwardActivation();
}
}
void backward(const UpdateCallback& callback) override {
if (needResetBwd_) {
VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
pipelineBwd_.clear();
pipelineMergeGrad_.clear();
mergeGrad_ = nullptr;
resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
// external output grad is not necessary
// since output may be mkldnn internal buffer or merge them directly.
CHECK(outGrad_) << "internal output grad is necessary";
if (cvtOutGrad_) {
pipelineBwd_.insert(pipelineBwd_.begin(), *cvtOutGrad_);
}
if (cvtInGrad_) {
pipelineBwd_.push_back(*cvtInGrad_);
}
printGradFormat();
needResetBwd_ = false;
}
// merge grad must before backward activation
if (mergeGrad_) {
REGISTER_TIMER_INFO("MergeBpGrad", getName().c_str());
stream_->submit(pipelineMergeGrad_);
}
{
REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
backwardActivation();
}
{
REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
stream_->submit(pipelineBwd_);
}
{
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
updateWeights(callback);
}
}
virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
void forward(PassType passType) override;
void backward(const UpdateCallback& callback) override;
/**
* reshape the input image sizes
......@@ -287,30 +177,12 @@ protected:
/**
* reshape the input image sizes and input batchsize
*/
virtual void reshapeInput(int& batchsize, int& height, int& width) {
const Argument& input = inputLayers_[0]->getOutput();
batchsize = input.getBatchSize();
int h = input.getFrameHeight();
int w = input.getFrameWidth();
if (h != 0) {
height = h;
}
if (w != 0) {
width = w;
}
}
void reshapeInput(int& batchsize, int& height, int& width);
/**
* reshape output image sizes
*/
virtual void reshapeOutput(size_t height, size_t width) {
output_.setFrameHeight(height);
output_.setFrameWidth(width);
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
outputOtherDevice_[i].setFrameHeight(height);
outputOtherDevice_[i].setFrameWidth(width);
}
}
void reshapeOutput(size_t height, size_t width);
/**
* reset MKLDNNMatrix from Matrix and internal primitive desc.
......@@ -318,13 +190,7 @@ protected:
*/
void resetWithMatrix(MKLDNNMatrixPtr& dnn,
const MatrixPtr& mat,
mkldnn::memory::primitive_desc pd) {
dnn = nullptr;
if (mat == nullptr) {
return;
}
dnn = MKLDNNMatrix::create(pd, mat);
}
mkldnn::memory::primitive_desc pd);
/**
* reset input value from input MKLDNNMatrix and internal primitive desc.
......@@ -332,99 +198,20 @@ protected:
*/
void resetInValue(
MKLDNNMatrixPtr& in,
const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr) {
cvtInVal_ = nullptr;
extInVal_ = nullptr;
in = nullptr;
CHECK_GT(bs_ * ic_ * ih_ * iw_, 0);
auto extPD = MKLDNNMatrix::createPrimitiveDesc(
{bs_, ic_, ih_, iw_}, mkldnn::memory::format::nchw, engine_);
const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
in = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
CHECK_EQ(inputIsOnlyMKLDNN(), in != nullptr);
if (in == nullptr || in->getFormat() == mkldnn::memory::format::nc) {
in = MKLDNNMatrix::create(extPD, inMat);
}
extInVal_ = isPaddleFormat(in->getFormat()) ? in : nullptr;
if (in->getFormat() == mkldnn::memory::format::nc) {
CHECK(ih_ == 1 && iw_ == 1);
}
if (nullptr == intPD || in->getPrimitiveDesc() == *intPD) {
return;
}
// need create reorder
in = MKLDNNMatrix::create(*intPD);
extInVal_ = extInVal_ ? extInVal_ : MKLDNNMatrix::create(extPD, inMat);
cvtInVal_ = MKLDNNMatrix::createReorder(extInVal_, in);
CHECK(cvtInVal_) << "should not be emptry";
}
const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr);
/**
* reset output value from internal primitive desc.
* reset both internal and external buffer and create reorder if necessary.
*/
void resetOutValue(MKLDNNMatrixPtr& out,
mkldnn::memory::primitive_desc intPD) {
cvtOutVal_ = nullptr;
out = MKLDNNMatrix::create(intPD, output_.value);
extOutVal_ = out;
if (outputIsOnlyMKLDNN() || isPaddleFormat(extOutVal_->getFormat())) {
return;
}
// need create reorder
CHECK_GT(bs_ * oc_ * oh_ * ow_, 0);
extOutVal_ = MKLDNNMatrix::create(mkldnn::memory::dims{bs_, oc_, oh_, ow_},
mkldnn::memory::format::nchw,
engine_,
output_.value);
out = MKLDNNMatrix::create(intPD);
cvtOutVal_ = MKLDNNMatrix::createReorder(out, extOutVal_);
CHECK(cvtOutVal_) << "should not be empty";
}
mkldnn::memory::primitive_desc intPD);
/**
* reset input grad from internal primitive desc.
* reset both internal and external buffer and create reorder if necessary.
*/
void resetInGrad(MKLDNNMatrixPtr& in, mkldnn::memory::primitive_desc intPD) {
cvtInGrad_ = nullptr;
extInGrad_ = nullptr;
in = nullptr;
LayerPtr& input = inputLayers_[0];
if (input->getOutputGrad() == nullptr) {
// no need input grad
return;
}
CHECK(inputIsOnlyMKLDNN() || input->getOutputMapSize() <= 1)
<< "only support input is MKLDNN layer or only have one output layer";
// when input is a mkldnn branch node,
// this layer will save input grad to a internal buffer,
// and the mkldnn input layer will merge them to actual prev->output_.grad
const MatrixPtr& inMat =
input->getOutputMapSize() <= 1 ? input->getOutputGrad() : nullptr;
in = MKLDNNMatrix::create(intPD, inMat);
Argument& arg = input->getOutput(this->getName());
arg.grad = std::dynamic_pointer_cast<Matrix>(in);
CHECK(inVal_ != nullptr && inVal_->getPrimitiveDesc() == intPD)
<< "should have internal input value and primitive desc must equal";
if (inputIsOnlyMKLDNN()) {
return;
}
extInGrad_ = in;
if (isPaddleFormat(extInGrad_->getFormat())) {
return;
}
// need create reorder
CHECK(extInVal_ != nullptr && isPaddleFormat(extInVal_->getFormat()))
<< "should have external input value and the format must be nchw(nc)";
extInGrad_ = MKLDNNMatrix::create(extInVal_->getPrimitiveDesc(), inMat);
CHECK(inVal_ != nullptr && inVal_->getPrimitiveDesc() == intPD)
<< "should have internal input value and primitive desc must equal";
in = MKLDNNMatrix::create(intPD);
cvtInGrad_ = MKLDNNMatrix::createReorder(in, extInGrad_);
CHECK(cvtInGrad_);
}
void resetInGrad(MKLDNNMatrixPtr& in, mkldnn::memory::primitive_desc intPD);
/**
* reset output grad from internal primitive desc.
......@@ -434,81 +221,59 @@ protected:
* it could not be mixed with cpu device,
* since it can not get memory desc from cpu device.
*/
void resetOutGrad(MKLDNNMatrixPtr& out,
mkldnn::memory::primitive_desc intPD) {
cvtOutGrad_ = nullptr;
extOutGrad_ = nullptr;
out = nullptr;
MatrixPtr& outMat = output_.grad;
out = MKLDNNMatrix::create(intPD, outMat);
resetMergeGrad(out);
if (outputIsOnlyMKLDNN()) {
return;
}
CHECK_LE(outputMap_.size(), 1U) << "do not support mixed with cpu device";
extOutGrad_ = out;
if (isPaddleFormat(extOutGrad_->getFormat())) {
return;
}
// need create reorder
CHECK(extOutVal_ != nullptr && isPaddleFormat(extOutVal_->getFormat()))
<< "should have external output value and the format must be nchw(nc)";
extOutGrad_ = MKLDNNMatrix::create(extOutVal_->getPrimitiveDesc(), outMat);
CHECK(outVal_ != nullptr && outVal_->getPrimitiveDesc() == intPD)
<< "should have internal output value and primitive desc must equal";
out = MKLDNNMatrix::create(intPD);
cvtOutGrad_ = MKLDNNMatrix::createReorder(extOutGrad_, out);
CHECK(cvtOutGrad_);
}
void resetOutGrad(MKLDNNMatrixPtr& out, mkldnn::memory::primitive_desc intPD);
/**
* reset the merge grad primitive if necessary.
* note: do not support the grads are mixed with cpu device,
* since it can not get memory desc from cpu device.
*/
virtual void resetMergeGrad(MKLDNNMatrixPtr& out) {
mergeGrad_ = nullptr;
pipelineMergeGrad_.clear();
if (outputMap_.size() <= 1 || !outputIsOnlyMKLDNN()) {
// do not merge when output is not all MKLDNN or only one output
return;
}
CHECK(out) << "should have reset internal ouput grad";
std::vector<double> scales(outputMap_.size(), 1.0);
std::vector<mkldnn::memory::primitive_desc> srcPDs;
std::vector<mkldnn::primitive::at> srcs;
for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {
MKLDNNMatrixPtr src =
std::dynamic_pointer_cast<MKLDNNMatrix>(it->second->grad);
VLOG(MKLDNN_BASE) << getName() << " has output grad " << it->first;
CHECK(src) << "should be MKLDNNMatrix";
auto srcDims = src->getDims();
auto dstDims = out->getDims();
CHECK_EQ(srcDims.size(), dstDims.size());
for (size_t i = 0; i < srcDims.size(); ++i) {
CHECK_EQ(srcDims[i], dstDims[i]);
}
srcPDs.push_back(src->getPrimitiveDesc());
srcs.push_back(*src);
}
void resetMergeGrad(MKLDNNMatrixPtr& out);
protected:
/**
* Set deviceId of this layer.
*/
void setDevice(int id) { deviceId_ = id; }
// TODO(TJ): remove me when mkldnn sum support different formats
for (size_t i = 1; i < srcPDs.size(); ++i) {
CHECK(srcPDs[0] == srcPDs[i]);
/**
* check the format is nchw or nc,
* which is supported by Paddle default memory layout
*/
bool isPaddleFormat(mkldnn::memory::format fmt) {
if (fmt == mkldnn::memory::format::nchw ||
fmt == mkldnn::memory::format::nc) {
return true;
} else {
return false;
}
tmpOutGrad_ = out;
tmpCvt_ = nullptr;
if (out->getPrimitiveDesc() != srcPDs[0]) {
tmpOutGrad_ = MKLDNNMatrix::create(srcPDs[0]);
tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out);
CHECK(tmpCvt_);
pipelineMergeGrad_.push_back(*tmpCvt_);
}
/**
* If input only has MKLDNN device.
* Otherwise, only support the previous layer using CPU device.
*/
bool inputIsOnlyMKLDNN(int index = 0) {
int prevDevice = getPrev(index)->getDeviceId();
if (prevDevice == MKLDNN_DEVICE) {
return true;
} else {
CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet";
return false;
}
}
auto sumPD = mkldnn::sum::primitive_desc(
tmpOutGrad_->getMemoryDesc(), scales, srcPDs);
mergeGrad_.reset(new mkldnn::sum(sumPD, srcs, *tmpOutGrad_));
pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_);
/**
* If output only has MKLDNN device.
* Otherwise, other devices should only using CPU device.
*/
bool outputIsOnlyMKLDNN() {
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE)
<< "Only support other device is CPU yet";
}
outputOnlyMKLDNN_ = outputOtherDevice_.size() == 0;
return outputOnlyMKLDNN_;
}
/**
......@@ -568,54 +333,7 @@ protected:
}
}
protected:
/**
* If input only has MKLDNN device.
* Otherwise, only support the previous layer using CPU device.
*/
bool inputIsOnlyMKLDNN(int index = 0) {
int prevDevice = getPrev(index)->getDeviceId();
if (prevDevice == MKLDNN_DEVICE) {
return true;
} else {
// do not support GPU yet
CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet";
return false;
}
}
/**
* If output only has MKLDNN device.
* Otherwise, other devices should only using CPU device.
*/
bool outputIsOnlyMKLDNN() {
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE)
<< "Only support other device is CPU yet";
}
outputOnlyMKLDNN_ = outputOtherDevice_.size() == 0;
return outputOnlyMKLDNN_;
}
/**
* Set deviceId of this layer.
*/
void setDevice(int id) { deviceId_ = id; }
private:
/**
* check the format is nchw or nc,
* which is supported by Paddle default memory layout
*/
bool isPaddleFormat(mkldnn::memory::format fmt) {
if (fmt == mkldnn::memory::format::nchw ||
fmt == mkldnn::memory::format::nc) {
return true;
} else {
return false;
}
}
/**
* clear all grad
*/
......
Markdown is supported
0% .
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
反馈
建议
客服 返回
顶部