未验证 提交 3581c075 编写于 作者: J jakpiase 提交者: GitHub

disabled unnecessary int reorders profiling (#39498)

上级 2bc91cc5
...@@ -71,9 +71,6 @@ static Tensor FoldFirstAndLastDims(const MKLDNNDeviceContext& dev_ctx, ...@@ -71,9 +71,6 @@ static Tensor FoldFirstAndLastDims(const MKLDNNDeviceContext& dev_ctx,
auto reorder_p = reorder_handler.AcquireReorder(reorder_src_memory_p, auto reorder_p = reorder_handler.AcquireReorder(reorder_src_memory_p,
reorder_dst_memory_p); reorder_dst_memory_p);
paddle::platform::RecordEvent record_reorder(
"int_reorder", paddle::platform::EventRole::kUniqueOp);
auto& astream = MKLDNNDeviceContext::tls().get_stream(); auto& astream = MKLDNNDeviceContext::tls().get_stream();
reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p); reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p);
astream.wait(); astream.wait();
......
...@@ -106,12 +106,8 @@ class QuantOpKernel : public framework::OpKernel<T> { ...@@ -106,12 +106,8 @@ class QuantOpKernel : public framework::OpKernel<T> {
reorder_p = std::shared_ptr<reorder>(new reorder(*reorder_pd)); reorder_p = std::shared_ptr<reorder>(new reorder(*reorder_pd));
auto& astream = platform::MKLDNNDeviceContext::tls().get_stream(); auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
{ reorder_p->execute(astream, *src_memory, *dst_memory);
platform::RecordEvent record_reorder("int_reorder", astream.wait();
platform::EventRole::kUniqueOp);
reorder_p->execute(astream, *src_memory, *dst_memory);
astream.wait();
}
output->set_layout(DataLayout::kMKLDNN); output->set_layout(DataLayout::kMKLDNN);
output->set_format(GetMKLDNNFormat(*dst_memory)); output->set_format(GetMKLDNNFormat(*dst_memory));
......
...@@ -138,12 +138,9 @@ class ReQuantOpKernel : public framework::OpKernel<T> { ...@@ -138,12 +138,9 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
} }
auto& astream = platform::MKLDNNDeviceContext::tls().get_stream(); auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
{
platform::RecordEvent record_reorder("int_reorder", reorder_p->execute(astream, *src_memory, *dst_memory);
platform::EventRole::kUniqueOp); astream.wait();
reorder_p->execute(astream, *src_memory, *dst_memory);
astream.wait();
}
output->set_layout(framework::DataLayout::kMKLDNN); output->set_layout(framework::DataLayout::kMKLDNN);
output->set_format(platform::GetMKLDNNFormat(*dst_memory)); output->set_format(platform::GetMKLDNNFormat(*dst_memory));
......
...@@ -174,12 +174,9 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -174,12 +174,9 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
output, in_out.format(), ctx.GetPlace()); output, in_out.format(), ctx.GetPlace());
auto reorder_p = reorder_handler.AcquireReorder(target_mem, dst_mem); auto reorder_p = reorder_handler.AcquireReorder(target_mem, dst_mem);
{
platform::RecordEvent record_reorder("int_reorder", reorder_p->execute(astream, *dst_mem, *target_mem);
platform::EventRole::kUniqueOp); astream.wait();
reorder_p->execute(astream, *dst_mem, *target_mem);
astream.wait();
}
} }
output->set_layout(framework::DataLayout::kMKLDNN); output->set_layout(framework::DataLayout::kMKLDNN);
output->set_format(platform::GetMKLDNNFormat(*dst_mem)); output->set_format(platform::GetMKLDNNFormat(*dst_mem));
......
...@@ -85,9 +85,6 @@ class ReduceMKLDNNKernel : public framework::OpKernel<T> { ...@@ -85,9 +85,6 @@ class ReduceMKLDNNKernel : public framework::OpKernel<T> {
auto reorder_p = reorder_handler.AcquireReorder(reorder_src_memory_p, auto reorder_p = reorder_handler.AcquireReorder(reorder_src_memory_p,
reorder_dst_memory_p); reorder_dst_memory_p);
platform::RecordEvent record_reorder("int_reorder",
platform::EventRole::kUniqueOp);
reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p); reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p);
astream.wait(); astream.wait();
......
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