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 5ac1c95c 编写于 作者: X xiaolil1
浏览文件
操作

Revert "modify for eltwise with some useless log" 

This reverts commit fcbe4898.
上级 b26e0ec2
隐藏空白更改
内联 并排
Showing with 55 addition and 113 deletion
paddle/fluid/framework/operator.cc
浏览文件 @ 5ac1c95c
......@@ -821,7 +821,6 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
"DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)",
Type(), last_input_name, data_type, ipt_name, tmp);
data_type = tmp;
std::cout<<"data_type = "<<data_type;
last_input_name = ipt_name;
}
}
......
paddle/fluid/operators/conv_mkldnn_op.cc
浏览文件 @ 5ac1c95c
......@@ -54,7 +54,6 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler {
}
size_t GetDstMemorySize() const {
std::cout<<"dst size = "<<conv_pd_->dst_primitive_desc().get_size()<<std::endl;
return conv_pd_->dst_primitive_desc().get_size();
}
......@@ -122,9 +121,9 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler {
std::shared_ptr<mkldnn::memory> AcquireSrcMemoryFromPrimitive(
const std::shared_ptr<mkldnn::memory> user_memory_p,
std::vector<mkldnn::primitive>& pipeline, bool is_INT8=false) { // NOLINT
std::vector<mkldnn::primitive>& pipeline) { // NOLINT
auto src_pd = conv_pd_->src_primitive_desc();
auto user_pd = is_INT8? src_pd : user_memory_p->get_primitive_desc();
auto user_pd = user_memory_p->get_primitive_desc();
return this->AcquireMemory(src_pd, user_pd, user_memory_p, "@src_mem_p",
pipeline);
}
......@@ -275,7 +274,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
std::cout<<"this is conv kernel op....................."<<std::endl;
const bool is_test = ctx.Attr<bool>("is_test");
auto& dev_ctx =
......@@ -325,7 +324,7 @@ std::cout<<"this is conv kernel op....................."<<std::endl;
"dilation in convolution is not implemented yet");
const T* input_data = input->data<T>();
const float* filter_data = filter->data<float>();
const T* filter_data = filter->data<T>();
std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
std::vector<int> weights_tz =
......@@ -345,17 +344,17 @@ std::cout<<"this is conv kernel op....................."<<std::endl;
}
std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
std::vector<float> output_shift_scale;
float sum_scale = 1.0f;
std::vector<T> output_shift_scale;
T sum_scale = 1.0f;
if(is_INT8){
std::cout<<"this is conv int8 op .............."<<std::endl;
int count = is_multi_channel? (g>1? weights_tz[1]*weights_tz[0] : weights_tz[0]) : 1;
float scale_in_data = *(scale_in->data<float>());
std::vector<float> scale_weights_data(count);
T scale_in_data = *(scale_in->data<T>());
T scale_in_eltwise_data = *(scale_in_eltwise->data<T>());
std::vector<T> scale_weights_data(count);
for(int i=0; i<count; i++){
scale_weights_data[i] =*(scale_weights->data<float>() + i);
scale_weights_data[i] =*(scale_weights->data<T>() + i);
}
float scale_out_data = *(scale_out->data<float>());
T scale_out_data = *(scale_out->data<T>());
output_shift_scale.resize(count);
for(int i=0; i<count; i++){
......@@ -364,10 +363,8 @@ std::cout<<"this is conv int8 op .............."<<std::endl;
else
output_shift_scale[i] = scale_out_data / (scale_in_data * scale_weights_data[i]);
}
if(fuse_residual_conn){
float scale_in_eltwise_data = *(scale_in_eltwise->data<float>());
sum_scale = scale_out_data / scale_in_eltwise_data;
}
sum_scale = scale_out_data / scale_in_eltwise_data;
}
// Get unique name for storing MKLDNN primitives
......@@ -381,7 +378,7 @@ std::cout<<"this is conv int8 op .............."<<std::endl;
auto user_src_md = platform::MKLDNNMemDesc(
{src_tz}, platform::MKLDNNGetDataType<T>(), input->format());
auto user_weights_md = platform::MKLDNNMemDesc(
{weights_tz}, platform::MKLDNNGetDataType<float>(),
{weights_tz}, platform::MKLDNNGetDataType<T>(),
(g == 1) ? filter->format() : mkldnn::memory::format::goihw);
/* create memory descriptor for convolution without specified format
......@@ -402,28 +399,12 @@ std::cout<<"this is conv int8 op .............."<<std::endl;
auto dst_md = platform::MKLDNNMemDesc(
dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
if(is_INT8){
src_md = platform::MKLDNNMemDesc(
src_tz, memory::data_type::u8, chosen_memory_format);
weights_md = platform::MKLDNNMemDesc(
weights_tz, memory::data_type::s8,
(g == 1) ? chosen_memory_format : mkldnn::memory::format::goihw);
dst_md = platform::MKLDNNMemDesc(
dst_tz,
fuse_relu?memory::data_type::u8:memory::data_type::s8,
chosen_memory_format);
}
// create a conv primitive descriptor and save it for usage in backward
std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd;
if (bias) {
bias_tz = paddle::framework::vectorize2int(bias->dims());
auto bias_md = platform::MKLDNNMemDesc(
bias_tz, platform::MKLDNNGetDataType<float>(), memory::format::x);
if(is_INT8){
bias_md = platform::MKLDNNMemDesc(
bias_tz, memory::data_type::s32, memory::format::x);
}
bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
if(is_INT8){
conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md,
strides, paddings, mkldnn_engine,
......@@ -455,85 +436,62 @@ std::cout<<"this is conv int8 op .............."<<std::endl;
auto user_src_memory_p =
handler.AcquireSrcMemory(user_src_md, to_void_cast<T>(input_data));
auto user_weights_memory_p = handler.AcquireWeightsMemory(
user_weights_md, to_void_cast<float>(filter_data));
user_weights_md, to_void_cast<T>(filter_data));
T* output_data = nullptr;
if (fuse_residual_conn) {
auto residual_param = ctx.Input<Tensor>("ResidualData");
auto residual_param_data = residual_param->data<T>();
PADDLE_ENFORCE(
residual_param_data != nullptr,
"Provide data if you want MKLDNN conv+elementwise_add fusion");
PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
"Output and elementwise parameter need to have the "
"same dimension sizes");
output->ShareDataWith(*residual_param);
output_data = output->mutable_data<T>(ctx.GetPlace());
} else {
output_data =
output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
}
// create reorder primitive if the input format is not the preferred one
auto src_memory_p =
handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline, is_INT8);
handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive(
user_weights_memory_p, pipeline, is_test);
if(is_INT8){
int mask_reorder = is_multi_channel? 0 : ((g!= 1) ? (1<<1)+(1<<0) : 1<<0);
int count = is_multi_channel? (g>1? weights_tz[1]*weights_tz[0] : weights_tz[0]) : 1;
std::vector<float> scale_weights_data(count);
std::vector<T> scale_weights_data(count);
for(int i=0; i<count; i++){
scale_weights_data[i] = *(scale_weights->data<T>() + i);
}
auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive(
user_weights_memory_p, pipeline, is_test, is_INT8, scale_weights_data, mask_reorder);
}
std::shared_ptr<mkldnn::memory> dst_memory_p;
if(is_INT8){
int8_t* output_data = nullptr;
if (fuse_residual_conn) {
auto residual_param = ctx.Input<Tensor>("ResidualData");
PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
"Output and elementwise parameter need to have the "
"same dimension sizes");
output->ShareDataWith(*residual_param);
output_data = output->mutable_data<int8_t>(ctx.GetPlace());
} else {
std::cout<<"conv log 1 ....................."<<std::endl;
output_data =
output->mutable_data<int8_t>(ctx.GetPlace(), handler.GetDstMemorySize());
std::cout<<"conv log 2 //////////////////////"<<std::endl;
}
dst_memory_p =
handler.AcquireDstMemoryFromPrimitive(to_void_cast<int8_t>(output_data));
std::cout<<"input fmt = "<<input->format()<<" output fmt = "<<output->format()<<" dst fmt = "<<dst_memory_p->get_primitive_desc().desc().data.format<<std::endl;
} else{
T* output_data = nullptr;
if (fuse_residual_conn) {
auto residual_param = ctx.Input<Tensor>("ResidualData");
auto residual_param_data = residual_param->data<T>();
PADDLE_ENFORCE(
residual_param_data != nullptr,
"Provide data if you want MKLDNN conv+elementwise_add fusion");
PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
"Output and elementwise parameter need to have the "
"same dimension sizes");
output->ShareDataWith(*residual_param);
output_data = output->mutable_data<T>(ctx.GetPlace());
} else {
std::cout<<"conv log 1 ....................."<<std::endl;
output_data =
output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
std::cout<<"conv log 2 //////////////////////"<<std::endl;
}
dst_memory_p =
handler.AcquireDstMemoryFromPrimitive(to_void_cast<T>(output_data));
}
auto dst_memory_p =
handler.AcquireDstMemoryFromPrimitive(to_void_cast<T>(output_data));
// create convolution op primitive
std::shared_ptr<mkldnn::convolution_forward> conv_p;
if (bias) {
const float* bias_data = bias->data<float>();
const T* bias_data = bias->data<T>();
auto user_bias_md = platform::MKLDNNMemDesc(
{bias_tz}, platform::MKLDNNGetDataType<float>(), memory::format::x);
{bias_tz}, platform::MKLDNNGetDataType<T>(), memory::format::x);
auto user_bias_memory_p =
handler.AcquireBiasMemory(user_bias_md, to_void_cast<float>(bias_data));
handler.AcquireBiasMemory(user_bias_md, to_void_cast<T>(bias_data));
auto bias_memory_p =
handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline);
if(is_INT8){
int mask_reorder = is_multi_channel? 0 : 1<<0;
int count = is_multi_channel? (g>1? weights_tz[1]*weights_tz[0] : weights_tz[0]) : 1;
std::vector<float> scale_bias_data(count);
std::vector<T> scale_bias_data(count);
for(int i=0; i<count; i++){
scale_bias_data[i] = (*scale_in->data<float>()) * (*(scale_weights->data<float>() + i));
scale_bias_data[i] = (*scale_in->data<T>()) * (*(scale_weights->data<T>() + i));
}
auto bias_memory_p =
handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline, is_INT8, scale_bias_data, mask_reorder);
......@@ -545,19 +503,17 @@ std::cout<<"input fmt = "<<input->format()<<" output fmt = "<<output->format()<
dst_memory_p);
}
// push primitive to stream and wait until it's executed
pipeline.push_back(*conv_p);
stream(stream::kind::eager).submit(pipeline).wait();
output->set_layout(DataLayout::kMKLDNN);
output->set_format(GetMKLDNNFormat(*dst_memory_p));
std::cout<<"input fmt = "<<input->format()<<" output fmt = "<<output->format()<<" dst fmt = "<<dst_memory_p->get_primitive_desc().desc().data.format<<std::endl;
}
private:
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<T> output_shift_scale, T sum_scale) const {
mkldnn::primitive_attr conv_attr;
mkldnn::post_ops post_operations;
// Fusion with Elementwise layer relies on adding a sum post-operation with
......@@ -612,7 +568,7 @@ std::cout<<"input fmt = "<<input->format()<<" output fmt = "<<output->format()<
const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
const bool fuse_residual_conn,
const std::vector<float> output_shift_scale, const float sum_scale) const {
const std::vector<T> output_shift_scale, const T sum_scale) const {
memory::dims stride_dims = {strides[0], strides[1]};
memory::dims padding_dims = {paddings[0], paddings[1]};
......@@ -661,7 +617,7 @@ std::cout<<"input fmt = "<<input->format()<<" output fmt = "<<output->format()<
const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
const bool fuse_residual_conn,
const std::vector<float> output_shift_scale, const float sum_scale) const {
const std::vector<T> output_shift_scale, const T sum_scale) const {
memory::dims stride_dims = {strides[0], strides[1]};
memory::dims padding_dims = {paddings[0], paddings[1]};
......@@ -885,8 +841,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
namespace ops = paddle::operators;
REGISTER_OP_KERNEL(conv2d, MKLDNN, ::paddle::platform::CPUPlace,
ops::ConvMKLDNNOpKernel<float>,
ops::ConvMKLDNNOpKernel<int8_t>);
ops::ConvMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(conv2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
ops::ConvMKLDNNGradOpKernel<float>);
paddle/fluid/operators/conv_op.cc
浏览文件 @ 5ac1c95c
......@@ -94,10 +94,10 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
auto input_data_type =
framework::ToDataType(ctx.Input<Tensor>("Input")->type());
//auto filter_data_type =
// framework::ToDataType(ctx.Input<Tensor>("Filter")->type());
//PADDLE_ENFORCE_EQ(input_data_type, filter_data_type,
// "input and filter data type should be consistent");
auto filter_data_type =
framework::ToDataType(ctx.Input<Tensor>("Filter")->type());
PADDLE_ENFORCE_EQ(input_data_type, filter_data_type,
"input and filter data type should be consistent");
if (input_data_type == framework::proto::VarType::FP16) {
PADDLE_ENFORCE_EQ(library, framework::LibraryType::kCUDNN,
......
paddle/fluid/operators/dequantize_op.cc
浏览文件 @ 5ac1c95c
......@@ -40,7 +40,7 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
auto* input = ctx.Input<Tensor>("Input");
auto* scale = ctx.Input<Tensor>("Scale");
auto* output = ctx.Output<Tensor>("Output");
std::cout<<"this is dequant op ***********"<<std::endl;
auto& dev_ctx =
ctx.template device_context<platform::MKLDNNDeviceContext>();
const auto& engine = dev_ctx.GetEngine();
......
paddle/fluid/operators/quantize_op.cc
浏览文件 @ 5ac1c95c
......@@ -37,7 +37,7 @@ class QuantOpKernel : public framework::OpKernel<T> {
auto* input = ctx.Input<Tensor>("Input");
auto* scale = ctx.Input<Tensor>("Scale");
auto* output = ctx.Output<Tensor>("Output");
std::cout<<"this is quantize op!!!!!!!!!!!!!!"<<std::endl;
auto& dev_ctx =
ctx.template device_context<platform::MKLDNNDeviceContext>();
const auto& engine = dev_ctx.GetEngine();
......@@ -68,12 +68,7 @@ std::cout<<"this is quantize op!!!!!!!!!!!!!!"<<std::endl;
auto reorder_pd = std::shared_ptr<reorder::primitive_desc>(
new reorder::primitive_desc(dst_pd, src_pd, attri));
auto reorder_p= std::shared_ptr<reorder>(new reorder(*reorder_pd, *src_memory_p, dst_memory));
pipeline.push_back(*reorder_p);
stream(stream::kind::eager).submit(pipeline).wait();
output->set_layout(DataLayout::kMKLDNN);
output->set_format(GetMKLDNNFormat(dst_memory));
}
};
......
paddle/fluid/platform/mkldnn_helper.h
浏览文件 @ 5ac1c95c
......@@ -153,11 +153,8 @@ class MKLDNNHandler {
std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
"Fail to find mem primitive in device context");
//mem_p = nullptr;
if (mem_p == nullptr) {
mem_p = std::make_shared<mkldnn::memory>(mdp, ptr);
std::cout<<"mem_p == null"<<std::endl;
//std::cout<<"mdp fmt = "<<mdp.desc().data.format<<" mem_p fmt = "<<mem_p->get_primitive_desc().desc().data.format<<std::endl;
dev_ctx_.SetBlob(local_key, mem_p);
} else {
mem_p->set_data_handle(ptr);
......@@ -165,7 +162,6 @@ std::cout<<"mem_p == null"<<std::endl;
// should be reused or none of them. So we check consistency
is_reusing_ = true;
}
std::cout<<"mdp fmt = "<<mdp.desc().data.format<<" mem_p fmt = "<<mem_p->get_primitive_desc().desc().data.format<<std::endl;
return mem_p;
}
......@@ -178,9 +174,7 @@ std::cout<<"mdp fmt = "<<mdp.desc().data.format<<" mem_p fmt = "<<mem_p->get_p
std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
"Fail to find mem primitive in device context");
//mem_p = nullptr;
if (mem_p == nullptr) {
std::cout<<"mem_p == null"<<std::endl;
mem_p = std::make_shared<mkldnn::memory>(
mkldnn::memory::primitive_desc{md, engine_}, ptr);
dev_ctx_.SetBlob(local_key, mem_p);
......@@ -190,7 +184,6 @@ std::cout<<"mem_p == null"<<std::endl;
// should be reused or none of them. So we check consistency
is_reusing_ = true;
}
std::cout<<"md fmt = "<<md.data.format<<" mem_p fmt = "<<mem_p->get_primitive_desc().desc().data.format<<std::endl;
return mem_p;
}
......
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