提交 0aa01929 编写于 作者: K Krzysztof Binias

Add backward

上级 0cc25a40
...@@ -15,6 +15,7 @@ ...@@ -15,6 +15,7 @@
#include "mkldnn.hpp" #include "mkldnn.hpp"
#include "paddle/fluid/operators/activation_op.h" #include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/mkldnn_activation_op.h" #include "paddle/fluid/operators/mkldnn_activation_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
namespace paddle { namespace paddle {
namespace operators { namespace operators {
...@@ -25,9 +26,14 @@ using paddle::platform::MKLDNNDeviceContext; ...@@ -25,9 +26,14 @@ using paddle::platform::MKLDNNDeviceContext;
namespace { namespace {
std::string gethash(const mkldnn::memory::dims &operand_dims, std::string gethash(const mkldnn::memory::dims &operand_dims,
const mkldnn::algorithm algorithm) { const mkldnn::algorithm algorithm) {
return std::string(std::to_string(operand_dims[0]) + "-" + auto dim2str = [](const mkldnn::memory::dims &operand_dims) {
std::to_string(operand_dims[1]) + "-" + std::string dstr = "";
std::to_string(algorithm)); for (size_t i = 0; i < operand_dims.size(); ++i) {
dstr += std::to_string(operand_dims[i]) + "-";
}
return dstr;
};
return dim2str(operand_dims) + std::to_string(algorithm);
} }
template <typename T, typename ExecContext> template <typename T, typename ExecContext>
...@@ -44,7 +50,7 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, ...@@ -44,7 +50,7 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
const auto *src_data = src->template data<T>(); const auto *src_data = src->template data<T>();
auto *dst = ctx.template Output<Tensor>("Out"); auto *dst = ctx.template Output<Tensor>("Out");
const T *dst_data = dst->template mutable_data<T>(ctx.GetPlace()); T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());
// get memory dim // get memory dim
PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4, PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4,
...@@ -52,15 +58,14 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, ...@@ -52,15 +58,14 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
std::vector<int> src_tz = framework::vectorize2int(src->dims()); std::vector<int> src_tz = framework::vectorize2int(src->dims());
const std::string key = gethash(src_tz, algorithm); const std::string key = gethash(src_tz, algorithm);
const std::string key_src_mem = key + "@eltwise_src_mem"; const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
const std::string key_dst_mem = key + "@eltwise_dst_mem"; const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem";
const std::string key_fwd = key + "@eltwise_fwd"; const std::string key_fwd = key + "@eltwise_fwd";
std::shared_ptr<void> p_src_mem = dev_ctx.GetBlob(key_src_mem); auto p_fwd = std::static_pointer_cast<mkldnn::eltwise_forward>(
std::shared_ptr<void> p_dst_mem = dev_ctx.GetBlob(key_dst_mem); dev_ctx.GetBlob(key_fwd));
std::shared_ptr<void> p_fwd = dev_ctx.GetBlob(key_fwd);
if (p_src_mem == nullptr || p_dst_mem == nullptr || p_fwd == nullptr) { if (p_fwd == nullptr) {
// create memory description // create memory description
auto data_md = src_tz.size() == 2 auto data_md = src_tz.size() == 2
? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
...@@ -69,35 +74,40 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, ...@@ -69,35 +74,40 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
mkldnn::memory::format::nchw); mkldnn::memory::format::nchw);
// create memory primitives // create memory primitives
p_src_mem = std::make_shared<mkldnn::memory>( auto p_src_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
mkldnn::memory({data_md, mkldnn_engine}, {data_md, mkldnn_engine}, platform::to_void_cast(src_data)));
static_cast<void *>(const_cast<float *>(src_data))));
dev_ctx.SetBlob(key_src_mem, p_src_mem); dev_ctx.SetBlob(key_src_mem, p_src_mem);
p_dst_mem = std::make_shared<mkldnn::memory>( auto p_dst_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
mkldnn::memory({data_md, mkldnn_engine}, {data_md, mkldnn_engine}, platform::to_void_cast(dst_data)));
static_cast<void *>(const_cast<float *>(dst_data))));
dev_ctx.SetBlob(key_dst_mem, p_dst_mem); dev_ctx.SetBlob(key_dst_mem, p_dst_mem);
auto fwd_desc = mkldnn::eltwise_forward::desc( auto fwd_desc = mkldnn::eltwise_forward::desc(
mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
auto p_fwd_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>( auto p_fwd_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
fwd_desc, mkldnn_engine); fwd_desc, mkldnn_engine);
const std::string key_fwd_pd = key + "eltwise_fwd_pd";
dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd);
p_fwd = std::make_shared<mkldnn::eltwise_forward>( p_fwd = std::make_shared<mkldnn::eltwise_forward>(
*(p_fwd_pd.get()), *(static_cast<mkldnn::memory *>(p_src_mem.get())), *p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get()));
*(static_cast<mkldnn::memory *>(p_dst_mem.get())));
dev_ctx.SetBlob(key_fwd, p_fwd); dev_ctx.SetBlob(key_fwd, p_fwd);
} else { } else {
std::static_pointer_cast<mkldnn::memory>(p_src_mem)->set_data_handle( // primitives already exist
reinterpret_cast<void *>(const_cast<T *>(src_data))); auto p_src_mem =
std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
std::static_pointer_cast<mkldnn::memory>(p_dst_mem)->set_data_handle( PADDLE_ENFORCE(p_src_mem != nullptr,
reinterpret_cast<void *>(const_cast<T *>(dst_data))); "Fail to find eltwise p_src_mem in device context.");
auto p_dst_mem =
std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_dst_mem));
PADDLE_ENFORCE(p_dst_mem != nullptr,
"Fail to find eltwise p_src_mem in device context.");
p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data));
p_dst_mem->set_data_handle(dst_data);
} }
// push primitive to stream and wait until it's executed // push primitive to stream and wait until it's executed
std::vector<mkldnn::primitive> pipeline = { std::vector<mkldnn::primitive> pipeline = {*(p_fwd.get())};
*(static_cast<mkldnn::eltwise_forward::primitive *>(p_fwd.get()))};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
} }
...@@ -121,47 +131,64 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, ...@@ -121,47 +131,64 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
std::vector<int> src_tz = framework::vectorize2int(out->dims()); std::vector<int> src_tz = framework::vectorize2int(out->dims());
const std::string key = gethash(src_tz, algorithm); const std::string key = gethash(src_tz, algorithm);
const std::string key_src_mem = key + "@eltwise_src_mem";
const std::string key_dst_mem = key + "@eltwise_dst_mem";
const std::string key_fwd = key + "@eltwise_fwd";
// create memory description const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem";
auto data_md = src_tz.size() == 2 const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem";
? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, const std::string key_grad = key + "@eltwise_grad";
mkldnn::memory::format::nc)
: platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, auto p_grad = std::static_pointer_cast<mkldnn::eltwise_forward::primitive>(
mkldnn::memory::format::nchw); dev_ctx.GetBlob(key_grad));
// retrieve source memory from device context if (p_grad == nullptr) {
const std::shared_ptr<void> src_mem = dev_ctx.GetBlob(key_src_mem); // create memory description
auto *p_src_mem = static_cast<mkldnn::memory *>(src_mem.get()); auto data_md = src_tz.size() == 2
? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
// create memory primitives mkldnn::memory::format::nc)
auto diff_src_memory = : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
mkldnn::memory({data_md, mkldnn_engine}, mkldnn::memory::format::nchw);
static_cast<void *>(const_cast<float *>(diff_src)));
auto diff_dst_memory = // create memory primitives
mkldnn::memory({data_md, mkldnn_engine}, std::shared_ptr<void> p_diff_src_mem =
static_cast<void *>(const_cast<float *>(diff_dst))); std::make_shared<mkldnn::memory>(mkldnn::memory(
{data_md, mkldnn_engine}, platform::to_void_cast(diff_src)));
auto backward_desc = dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem);
mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta); std::shared_ptr<void> p_diff_dst_mem =
std::make_shared<mkldnn::memory>(mkldnn::memory(
// retrieve eltwise primitive desc from device context {data_md, mkldnn_engine}, platform::to_void_cast(diff_dst)));
const std::shared_ptr<void> forward_pd = dev_ctx.GetBlob(key_fwd); dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem);
PADDLE_ENFORCE(forward_pd != nullptr,
"Fail to find eltwise_pd in device context"); auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md,
auto *p_forward_pd = alpha, beta);
static_cast<mkldnn::eltwise_forward::primitive_desc *>(forward_pd.get());
const std::string key_fwd_pd = key + "eltwise_fwd_pd";
auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( auto *p_fwd_pd = static_cast<mkldnn::eltwise_forward::primitive_desc *>(
backward_desc, mkldnn_engine, *p_forward_pd); dev_ctx.GetBlob(key_fwd_pd).get());
auto eltwise_bwd = mkldnn::eltwise_backward(eltwise_bwd_prim_desc, *p_src_mem, auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc(
diff_dst_memory, diff_src_memory); bwd_desc, mkldnn_engine, *p_fwd_pd);
const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
const std::shared_ptr<void> p_src_mem = dev_ctx.GetBlob(key_src_mem);
p_grad = std::make_shared<mkldnn::eltwise_backward>(
eltwise_bwd_prim_desc, *static_cast<mkldnn::memory *>(p_src_mem.get()),
*(static_cast<mkldnn::memory *>(p_diff_dst_mem.get())),
*(static_cast<mkldnn::memory *>(p_diff_src_mem.get())));
} else {
// primitives already exist
auto p_diff_src_mem = std::static_pointer_cast<mkldnn::memory>(
dev_ctx.GetBlob(key_diff_src_mem));
auto p_diff_dst_mem = std::static_pointer_cast<mkldnn::memory>(
dev_ctx.GetBlob(key_diff_dst_mem));
p_diff_src_mem->set_data_handle(
platform::to_void_reinterpret_cast(diff_src));
p_diff_dst_mem->set_data_handle(
platform::to_void_reinterpret_cast(diff_dst));
}
// push primitive to stream and wait until it's executed // push primitive to stream and wait until it's executed
std::vector<mkldnn::primitive> pipeline = {eltwise_bwd}; std::vector<mkldnn::primitive> pipeline = {*(p_grad.get())};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
} }
} // anonymous namespace } // anonymous namespace
......
...@@ -38,6 +38,11 @@ void* to_void_cast(const Type* t) { ...@@ -38,6 +38,11 @@ void* to_void_cast(const Type* t) {
return static_cast<void*>(const_cast<Type*>(t)); return static_cast<void*>(const_cast<Type*>(t));
} }
template <typename Type>
void* to_void_reinterpret_cast(const Type* t) {
return reinterpret_cast<void*>(const_cast<Type*>(t));
}
template <class Type> template <class Type>
using tf_desc = typename Type::desc; using tf_desc = typename Type::desc;
......
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