Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2299
Star 20931
Fork 5422
P
Paddle
提交 7c9aabd1 编写于 作者: X xiaolil1
浏览文件
操作

revert md reuse

上级 2b2eac47
隐藏空白更改
内联 并排
Showing with 77 addition and 77 deletion
paddle/fluid/operators/conv_mkldnn_op.cc
浏览文件 @ 7c9aabd1
...@@ -560,18 +560,18 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -560,18 +560,18 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
} }
static std::unordered_map<std::string, std::shared_ptr<mkldnn::memory::desc>> md_map; //static std::unordered_map<std::string, std::shared_ptr<mkldnn::memory::desc>> md_map;
bool md_reuse = true; //bool md_reuse = true;
auto user_src_md_key = key + "@user_src_md"; //auto user_src_md_key = key + "@user_src_md";
if (GetMdMap(md_map, user_src_md_key) == nullptr){ //if (GetMdMap(md_map, user_src_md_key) == nullptr){
md_reuse = false; //we suppose all mds are reused if the first md is in the map. // md_reuse = false; //we suppose all mds are reused if the first md is in the map.
} //}
auto user_weights_md_key = key + "@user_weights_md"; //auto user_weights_md_key = key + "@user_weights_md";
std::shared_ptr<mkldnn::memory::desc> user_src_md; std::shared_ptr<mkldnn::memory::desc> user_src_md;
std::shared_ptr<mkldnn::memory::desc> user_weights_md; std::shared_ptr<mkldnn::memory::desc> user_weights_md;
std::vector<primitive> pipeline; std::vector<primitive> pipeline;
//std::cout<<"md_reuse = "<<md_reuse<<std::endl; //std::cout<<"md_reuse = "<<md_reuse<<std::endl;
if(!md_reuse){ // if(!md_reuse){
//std::cout<<"create md.......... "<<std::endl; //std::cout<<"create md.......... "<<std::endl;
user_src_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( user_src_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
{src_tz}, paddle::framework::ToMKLDNNDataType(input->type()), input->format()))); {src_tz}, paddle::framework::ToMKLDNNDataType(input->type()), input->format())));
...@@ -579,12 +579,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -579,12 +579,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
{weights_tz}, platform::MKLDNNGetDataType<float>(), {weights_tz}, platform::MKLDNNGetDataType<float>(),
(g == 1) ? mkldnn::memory::format::oihw : mkldnn::memory::format::goihw))); (g == 1) ? mkldnn::memory::format::oihw : mkldnn::memory::format::goihw)));
SetMdMap(md_map, user_src_md_key, user_src_md); // SetMdMap(md_map, user_src_md_key, user_src_md);
SetMdMap(md_map, user_weights_md_key, user_weights_md); // SetMdMap(md_map, user_weights_md_key, user_weights_md);
} else{ // } else{
user_src_md = GetMdMap(md_map, user_src_md_key); // user_src_md = GetMdMap(md_map, user_src_md_key);
user_weights_md = GetMdMap(md_map, user_weights_md_key); // user_weights_md = GetMdMap(md_map, user_weights_md_key);
} // }
/* create memory descriptor for convolution without specified format /* create memory descriptor for convolution without specified format
* ('any') which lets a primitive (convolution in this case) choose * ('any') which lets a primitive (convolution in this case) choose
...@@ -597,16 +597,16 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -597,16 +597,16 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd; std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd;
auto bias_tz = paddle::framework::vectorize2int(bias->dims()); auto bias_tz = paddle::framework::vectorize2int(bias->dims());
auto src_md_key = key + "@src_md"; //auto src_md_key = key + "@src_md";
auto weights_md_key = key + "@weights_md_key"; //auto weights_md_key = key + "@weights_md_key";
auto dst_md_key = key + "@dst_md_key"; //auto dst_md_key = key + "@dst_md_key";
auto bias_md_key = key + "@bias_md_key"; //auto bias_md_key = key + "@bias_md_key";
std::shared_ptr<mkldnn::memory::desc> src_md; std::shared_ptr<mkldnn::memory::desc> src_md;
std::shared_ptr<mkldnn::memory::desc> weights_md; std::shared_ptr<mkldnn::memory::desc> weights_md;
std::shared_ptr<mkldnn::memory::desc> dst_md; std::shared_ptr<mkldnn::memory::desc> dst_md;
if(is_INT8){ if(is_INT8){
if(!md_reuse){ //if(!md_reuse){
src_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( src_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
src_tz, memory::data_type::u8, chosen_memory_format))); src_tz, memory::data_type::u8, chosen_memory_format)));
weights_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( weights_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
...@@ -621,25 +621,25 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -621,25 +621,25 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
if(force_fp32_output) if(force_fp32_output)
dst_dt = paddle::framework::ToMKLDNNDataType(std::type_index(typeid(float))); dst_dt = paddle::framework::ToMKLDNNDataType(std::type_index(typeid(float)));
dst_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(dst_tz, dst_dt, chosen_memory_format))); dst_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(dst_tz, dst_dt, chosen_memory_format)));
SetMdMap(md_map, src_md_key, src_md); //SetMdMap(md_map, src_md_key, src_md);
SetMdMap(md_map, weights_md_key, weights_md); //SetMdMap(md_map, weights_md_key, weights_md);
SetMdMap(md_map, dst_md_key, dst_md); //SetMdMap(md_map, dst_md_key, dst_md);
} else{ //} else{
src_md = GetMdMap(md_map, src_md_key); // src_md = GetMdMap(md_map, src_md_key);
weights_md = GetMdMap(md_map, weights_md_key); // weights_md = GetMdMap(md_map, weights_md_key);
dst_md = GetMdMap(md_map, dst_md_key); // dst_md = GetMdMap(md_map, dst_md_key);
} //}
// create a conv primitive descriptor and save it for usage in backward // create a conv primitive descriptor and save it for usage in backward
if (bias) { if (bias) {
std::shared_ptr<mkldnn::memory::desc> bias_md; std::shared_ptr<mkldnn::memory::desc> bias_md;
if(!md_reuse){ //if(!md_reuse){
bias_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( bias_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
bias_tz, memory::data_type::s32, memory::format::x))); bias_tz, memory::data_type::s32, memory::format::x)));
SetMdMap(md_map, bias_md_key, bias_md); // SetMdMap(md_map, bias_md_key, bias_md);
} else{ //} else{
bias_md = GetMdMap(md_map, bias_md_key); // bias_md = GetMdMap(md_map, bias_md_key);
} //}
conv_pd = ConvFwdPrimitiveDesc(*src_md, *weights_md, *bias_md, *dst_md, conv_pd = ConvFwdPrimitiveDesc(*src_md, *weights_md, *bias_md, *dst_md,
strides, paddings, mkldnn_engine, strides, paddings, mkldnn_engine,
...@@ -652,31 +652,31 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -652,31 +652,31 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
output_shift_scale, sum_scale[0], is_test); output_shift_scale, sum_scale[0], is_test);
} }
} else{ } else{
if(!md_reuse){ //if(!md_reuse){
src_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( src_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
src_tz, platform::MKLDNNGetDataType<float>(), chosen_memory_format))); src_tz, platform::MKLDNNGetDataType<float>(), chosen_memory_format)));
weights_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( weights_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
weights_tz, platform::MKLDNNGetDataType<float>(), chosen_memory_format))); weights_tz, platform::MKLDNNGetDataType<float>(), chosen_memory_format)));
dst_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( dst_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
dst_tz, platform::MKLDNNGetDataType<float>(), chosen_memory_format))); dst_tz, platform::MKLDNNGetDataType<float>(), chosen_memory_format)));
SetMdMap(md_map, src_md_key, src_md); // SetMdMap(md_map, src_md_key, src_md);
SetMdMap(md_map, weights_md_key, weights_md); // SetMdMap(md_map, weights_md_key, weights_md);
SetMdMap(md_map, dst_md_key, dst_md); // SetMdMap(md_map, dst_md_key, dst_md);
} else{ //} else{
src_md = GetMdMap(md_map, src_md_key); // src_md = GetMdMap(md_map, src_md_key);
weights_md = GetMdMap(md_map, weights_md_key); // weights_md = GetMdMap(md_map, weights_md_key);
dst_md = GetMdMap(md_map, dst_md_key); // dst_md = GetMdMap(md_map, dst_md_key);
} //}
// create a conv primitive descriptor and save it for usage in backward // create a conv primitive descriptor and save it for usage in backward
if (bias) { if (bias) {
std::shared_ptr<mkldnn::memory::desc> bias_md; std::shared_ptr<mkldnn::memory::desc> bias_md;
if(!md_reuse){ //if(!md_reuse){
bias_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( bias_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
bias_tz, platform::MKLDNNGetDataType<float>(), memory::format::x))); bias_tz, platform::MKLDNNGetDataType<float>(), memory::format::x)));
SetMdMap(md_map, bias_md_key, bias_md); // SetMdMap(md_map, bias_md_key, bias_md);
} else{ //} else{
bias_md = GetMdMap(md_map, bias_md_key); // bias_md = GetMdMap(md_map, bias_md_key);
} //}
conv_pd = ConvFwdPrimitiveDesc(*src_md, *weights_md, *bias_md, *dst_md, conv_pd = ConvFwdPrimitiveDesc(*src_md, *weights_md, *bias_md, *dst_md,
strides, paddings, mkldnn_engine, strides, paddings, mkldnn_engine,
fuse_relu, fuse_residual_conn, is_test); fuse_relu, fuse_residual_conn, is_test);
...@@ -714,7 +714,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -714,7 +714,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
std::shared_ptr<mkldnn::memory> dst_memory_p; std::shared_ptr<mkldnn::memory> dst_memory_p;
bool need_s8_to_u8 = false; bool need_s8_to_u8 = false;
auto user_residual_md_key = key + "@user_residual_md"; //auto user_residual_md_key = key + "@user_residual_md";
if(fuse_residual_conn) { if(fuse_residual_conn) {
auto residual_param = ctx.Input<Tensor>("ResidualData"); auto residual_param = ctx.Input<Tensor>("ResidualData");
PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(), PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
...@@ -723,17 +723,17 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -723,17 +723,17 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto residual_dt = paddle::framework::ToMKLDNNDataType(residual_param->type()); auto residual_dt = paddle::framework::ToMKLDNNDataType(residual_param->type());
if(residual_param->format() != handler.GetDstFormat()) { if(residual_param->format() != handler.GetDstFormat()) {
std::shared_ptr<mkldnn::memory::desc> user_residual_md; std::shared_ptr<mkldnn::memory::desc> user_residual_md;
if(!md_reuse){ //if(!md_reuse){
auto residual_data_tz = auto residual_data_tz =
paddle::framework::vectorize2int(residual_param->dims()); paddle::framework::vectorize2int(residual_param->dims());
auto residual_data_type = auto residual_data_type =
paddle::framework::ToMKLDNNDataType(residual_param->type()); paddle::framework::ToMKLDNNDataType(residual_param->type());
user_residual_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( user_residual_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
residual_data_tz, residual_data_type, residual_param->format()))); residual_data_tz, residual_data_type, residual_param->format())));
SetMdMap(md_map, user_residual_md_key, user_residual_md); //SetMdMap(md_map, user_residual_md_key, user_residual_md);
} else{ //} else{
user_residual_md = GetMdMap(md_map, user_residual_md_key); // user_residual_md = GetMdMap(md_map, user_residual_md_key);
} //}
if(is_INT8){ if(is_INT8){
PADDLE_ENFORCE( PADDLE_ENFORCE(
force_fp32_output == false, force_fp32_output == false,
...@@ -818,17 +818,17 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -818,17 +818,17 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
// create convolution op primitive // create convolution op primitive
std::shared_ptr<mkldnn::convolution_forward> conv_p; std::shared_ptr<mkldnn::convolution_forward> conv_p;
//auto scale_bias_key = key + "@scale_bias"; //auto scale_bias_key = key + "@scale_bias";
auto user_bias_md_key = key + "@user_bias_md"; //auto user_bias_md_key = key + "@user_bias_md";
if (bias) { if (bias) {
const float* bias_data = bias->data<float>(); const float* bias_data = bias->data<float>();
std::shared_ptr<mkldnn::memory::desc> user_bias_md; std::shared_ptr<mkldnn::memory::desc> user_bias_md;
if(!md_reuse){ //if(!md_reuse){
user_bias_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc( user_bias_md.reset(new mkldnn::memory::desc(platform::MKLDNNMemDesc(
{bias_tz}, platform::MKLDNNGetDataType<float>(), memory::format::x))); {bias_tz}, platform::MKLDNNGetDataType<float>(), memory::format::x)));
SetMdMap(md_map, user_bias_md_key, user_bias_md); // SetMdMap(md_map, user_bias_md_key, user_bias_md);
} else{ //} else{
user_bias_md = GetMdMap(md_map, user_bias_md_key); // user_bias_md = GetMdMap(md_map, user_bias_md_key);
} //}
auto user_bias_memory_p = auto user_bias_memory_p =
handler.AcquireBiasMemory(*user_bias_md, to_void_cast<float>(bias_data)); handler.AcquireBiasMemory(*user_bias_md, to_void_cast<float>(bias_data));
std::shared_ptr<mkldnn::memory> bias_memory_p; std::shared_ptr<mkldnn::memory> bias_memory_p;
...@@ -898,25 +898,25 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -898,25 +898,25 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
return {{0.0f}}; return {{0.0f}};
} }
void SetMdMap(std::unordered_map<std::string, std::shared_ptr<mkldnn::memory::desc>> &md_map, //void SetMdMap(std::unordered_map<std::string, std::shared_ptr<mkldnn::memory::desc>> &md_map,
const std::string& name, std::shared_ptr<mkldnn::memory::desc> mds) const { // const std::string& name, std::shared_ptr<mkldnn::memory::desc> mds) const {
auto it = md_map.find(name); // auto it = md_map.find(name);
if (it == md_map.end()) { // if (it == md_map.end()) {
md_map[name] = mds; // create new blob // md_map[name] = mds; // create new blob
} else { // } else {
(*it).second = mds; // set data to existing blob // (*it).second = mds; // set data to existing blob
} // }
return; // return;
} //}
std::shared_ptr<mkldnn::memory::desc> GetMdMap(std::unordered_map<std::string, std::shared_ptr<mkldnn::memory::desc>> md_map, //std::shared_ptr<mkldnn::memory::desc> GetMdMap(std::unordered_map<std::string, std::shared_ptr<mkldnn::memory::desc>> md_map,
const std::string& name) const { // const std::string& name) const {
auto it = md_map.find(name); // auto it = md_map.find(name);
if (it != md_map.end()) { // if (it != md_map.end()) {
return (*it).second; // return (*it).second;
} // }
return nullptr; // return nullptr;
} //}
mkldnn::primitive_attr CreatePostOps(bool fuse_relu, bool fuse_residual_conn, mkldnn::primitive_attr CreatePostOps(bool fuse_relu, bool fuse_residual_conn,
const std::vector<float> output_shift_scale, float sum_scale) const { const std::vector<float> output_shift_scale, float sum_scale) const {
......
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