Skip to content

Second fix to #34554 #36290

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lidanqing-vv merged 1 commit into from
Oct 13, 2021
Merged

Second fix to #34554 #36290

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
49 changes: 25 additions & 24 deletions paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,7 +78,8 @@ class ConvMKLDNNHandlerT
mkldnn::convolution_backward_weights>(
dev_ctx, mkldnn_engine, cpu_place,
platform::CreateKey(dev_ctx, framework::vectorize(input->dims()),
unique_name)) {
unique_name)),
is_test_(ctx.Attr<bool>("is_test")) {
if (!this->isCached()) {
PADDLE_ENFORCE_EQ(
input->layout(), framework::DataLayout::kMKLDNN,
Expand Down Expand Up @@ -159,7 +160,6 @@ class ConvMKLDNNHandlerT
framework::slice_ddim(filter_dims, 2, filter_dims.size());

const auto ksize = framework::vectorize(filter_data_dims);
const bool is_test = ctx.Attr<bool>("is_test");

auto strides_temp = ctx.Attr<std::vector<int>>("strides");
std::vector<int64_t> strides(begin(strides_temp), end(strides_temp));
Expand Down Expand Up @@ -214,9 +214,8 @@ class ConvMKLDNNHandlerT

const auto dst_md = platform::MKLDNNMemDesc(
dst_tz, platform::MKLDNNGetDataType<T_out>(), chosen_memory_format);
const auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference
: mkldnn::prop_kind::forward_training;

const auto fwd_prop_kind = is_test_ ? mkldnn::prop_kind::forward_inference
: mkldnn::prop_kind::forward_training;
float sum_scale = 1.0f;
std::vector<float> output_shift_scale;
if (platform::is_int8<T>())
Expand Down Expand Up @@ -261,7 +260,8 @@ class ConvMKLDNNHandlerT
mkldnn::convolution_backward_weights>(
dev_ctx, dev_ctx.GetEngine(), cpu_place,
platform::CreateKey(dev_ctx, framework::vectorize(in->dims()),
unique_name)) {
unique_name)),
is_test_(false) {
if (!this->isBwdCached()) {
PADDLE_ENFORCE_EQ(
in->layout(), framework::DataLayout::kMKLDNN,
Expand Down Expand Up @@ -291,7 +291,7 @@ class ConvMKLDNNHandlerT
"Wrong format set for output_grad tensor"));

PADDLE_ENFORCE_EQ(
ctx.Attr<bool>("is_test"), false,
is_test_, false,
platform::errors::InvalidArgument(
"is_test attribute should be set to False in training phase."));

Expand Down Expand Up @@ -557,26 +557,26 @@ class ConvMKLDNNHandlerT
framework::vectorize(in_mem->dims()),
platform::MKLDNNGetDataType<T>(), in_mem->format());
return this->AcquireMemoryWithReorder(
user_mem_md, mem_md, platform::to_void_cast<T>(in_mem_data), key_mem);
user_mem_md, mem_md, platform::to_void_cast<T>(in_mem_data), key_mem,
is_test_);
} else {
const std::string target_key_suffix{key_mem_target};
const auto target_mem_p = this->AcquireMemory(target_key_suffix);
user_mem_p->set_data_handle(platform::to_void_cast<T>(in_mem_data));
if (user_mem_p != target_mem_p) {
this->AcquireReorder(user_mem_p, target_mem_p, key_mem);
this->AcquireReorder(user_mem_p, target_mem_p);
}
return target_mem_p;
}
}

std::shared_ptr<mkldnn::memory> AcquireWeightsMemoryWithReorder(
const framework::Tensor* filter, const int groups, const bool is_conv3d,
const bool is_test, const std::vector<float>& scale_data = {1.0f},
int mask = 0) {
const std::vector<float>& scale_data = {1.0f}, int mask = 0) {
// This is workaround to make execution faster, delete
// if statement after including md inside Tensor
auto weights_mem_p = this->AcquireMemory("@weights_mem_p_target");
if (is_test && weights_mem_p) {
if (is_test_ && weights_mem_p) {
return weights_mem_p;
} else {
const K* filter_data = filter->data<K>();
Expand All @@ -589,16 +589,16 @@ class ConvMKLDNNHandlerT

return this->AcquireMemoryWithReorder(
user_src_md, this->fwd_pd_->weights_desc(),
platform::to_void_cast<K>(filter_data), "@weights_mem_p", is_test, {},
scale_data, mask);
platform::to_void_cast<K>(filter_data), "@weights_mem_p", is_test_,
{}, scale_data, mask);
}
}

std::shared_ptr<mkldnn::memory> AcquireBiasMemoryWithReorder(
const framework::Tensor* bias, const bool is_test,
const framework::Tensor* bias,
const std::vector<float>& scale_data = {1.0f}, int mask = 0) {
auto bias_mem_p = this->AcquireMemory("@bias_mem_p_target");
if (is_test && bias_mem_p) {
if (is_test_ && bias_mem_p) {
return bias_mem_p;
} else {
const K* bias_data = bias->data<K>();
Expand All @@ -608,7 +608,7 @@ class ConvMKLDNNHandlerT

return this->AcquireMemoryWithReorder(
user_bias_md, this->fwd_pd_->bias_desc(),
platform::to_void_cast<K>(bias_data), "@bias_mem_p", is_test, {},
platform::to_void_cast<K>(bias_data), "@bias_mem_p", is_test_, {},
scale_data, mask);
}
}
Expand Down Expand Up @@ -641,7 +641,7 @@ class ConvMKLDNNHandlerT
platform::GetMKLDNNFormat(this->fwd_pd_->dst_desc())) {
auto residual_memory_p = this->AcquireResidualMemory(residual_param);
dst_memory_p = this->template AcquireDstMemory<T_out>(output);
this->AcquireReorder(residual_memory_p, dst_memory_p, "@residual_dst");
this->AcquireReorder(residual_memory_p, dst_memory_p);
} else {
// Changing ShareDataWith to TensorCopy results in performance drop
// on ResNet architectures
Expand All @@ -651,6 +651,9 @@ class ConvMKLDNNHandlerT
}
return dst_memory_p;
}

private:
const bool is_test_;
};

} // anonymous namespace
Expand Down Expand Up @@ -695,7 +698,6 @@ class ConvMKLDNNOpKernel : public framework::OpKernel<T> {
ctx.template device_context<platform::MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine();

const bool is_test = ctx.Attr<bool>("is_test");
const bool is_conv3d = ctx.Attr<std::vector<int>>("strides").size() == 3U;
const bool fuse_residual_conn = ctx.Attr<bool>("fuse_residual_connection");

Expand All @@ -712,7 +714,7 @@ class ConvMKLDNNOpKernel : public framework::OpKernel<T> {
auto src_memory_p = handler.AcquireSrcMemoryWithReorder(input);

auto weights_memory_p = handler.AcquireWeightsMemoryWithReorder(
filter, ctx.Attr<int>("groups"), is_conv3d, is_test);
filter, ctx.Attr<int>("groups"), is_conv3d);

std::shared_ptr<dnnl::memory> dst_memory_p;
if (fuse_residual_conn) {
Expand All @@ -731,7 +733,7 @@ class ConvMKLDNNOpKernel : public framework::OpKernel<T> {
{MKLDNN_ARG_DST, *dst_memory_p}};

if (bias) {
auto bias_memory_p = handler.AcquireBiasMemoryWithReorder(bias, is_test);
auto bias_memory_p = handler.AcquireBiasMemoryWithReorder(bias);
args.insert({MKLDNN_ARG_BIAS, *bias_memory_p});
}

Expand Down Expand Up @@ -783,11 +785,10 @@ class ConvMKLDNNOpKernel : public framework::OpKernel<T> {
ctx.Attr<std::vector<float>>("Scale_weights");
const bool is_multi_channel = scale_weights_data.size() > 1;
const int& groups = ctx.Attr<int>("groups");
const bool& is_test = ctx.Attr<bool>("is_test");
int mask_reorder =
is_multi_channel ? ((groups != 1) ? (1 << 1) + (1 << 0) : 1 << 0) : 0;
auto weights_memory_p = handler.AcquireWeightsMemoryWithReorder(
filter, groups, false, is_test, scale_weights_data, mask_reorder);
filter, groups, false, scale_weights_data, mask_reorder);

std::shared_ptr<dnnl::memory> dst_memory_p;
if (fuse_residual_conn) {
Expand Down Expand Up @@ -822,7 +823,7 @@ class ConvMKLDNNOpKernel : public framework::OpKernel<T> {
handler.get_int8_bias_scales(ctx);

auto bias_memory_p = handler.AcquireBiasMemoryWithReorder(
bias, is_test, scale_bias_data, mask_reorder);
bias, scale_bias_data, mask_reorder);
args.insert({MKLDNN_ARG_BIAS, *bias_memory_p});
}

Expand Down
33 changes: 17 additions & 16 deletions paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,10 +51,10 @@ class ConvTransposeMKLDNNHandlerT
: platform::MKLDNNHandlerT<T, mkldnn::deconvolution_forward>(
dev_ctx, mkldnn_engine, cpu_place,
platform::CreateKey(dev_ctx, framework::vectorize(input->dims()),
unique_name)) {
unique_name)),
is_test_(ctx.Attr<bool>("is_test")) {
if (!this->isCached()) {
const bool is_test = ctx.Attr<bool>("is_test");
PADDLE_ENFORCE_EQ(is_test, true,
PADDLE_ENFORCE_EQ(is_test_, true,
platform::errors::InvalidArgument(
"ConvTransposeMKLDNN works only for inference. "
"The attribute \'is_test\' value should be set to "
Expand Down Expand Up @@ -169,8 +169,8 @@ class ConvTransposeMKLDNNHandlerT

const mkldnn::primitive_attr conv_trans_attr =
CreatePostOps(fuse_activation, fuse_alpha, fuse_beta);
auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference
: mkldnn::prop_kind::forward_training;
auto fwd_prop_kind = is_test_ ? mkldnn::prop_kind::forward_inference
: mkldnn::prop_kind::forward_training;
if (bias) {
std::vector<int64_t> bias_tz = framework::vectorize(bias->dims());
const auto bias_md =
Expand Down Expand Up @@ -231,18 +231,18 @@ class ConvTransposeMKLDNNHandlerT
const auto target_src_mem_p = this->AcquireMemory(target_key_suffix);
user_src_mem_p->set_data_handle(platform::to_void_cast<T>(input_data));
if (user_src_mem_p != target_src_mem_p) {
this->AcquireReorder(user_src_mem_p, target_src_mem_p, "@src_mem_p");
this->AcquireReorder(user_src_mem_p, target_src_mem_p);
}
return target_src_mem_p;
}
}

std::shared_ptr<mkldnn::memory> AcquireWeightsMemoryWithReorder(
const framework::Tensor* filter, const int& groups, const bool& is_test) {
const framework::Tensor* filter, const int& groups) {
// This is workaround to make execution faster, delete
// if statement after including md inside Tensor
auto weights_mem_p = this->AcquireMemory("@weights_mem_p_target");
if (is_test && weights_mem_p) {
if (is_test_ && weights_mem_p) {
return weights_mem_p;
} else {
const K* filter_data = filter->data<K>();
Expand Down Expand Up @@ -277,15 +277,15 @@ class ConvTransposeMKLDNNHandlerT

return this->template AcquireMemoryWithReorder<K>(
user_src_md, this->fwd_pd_->weights_desc(),
platform::to_void_cast<K>(filter_data), "@weights_mem_p", is_test,
platform::to_void_cast<K>(filter_data), "@weights_mem_p", is_test_,
iohw2oihw_reorder);
}
}

std::shared_ptr<mkldnn::memory> AcquireBiasMemoryWithReorder(
const framework::Tensor* bias, const bool& is_test) {
const framework::Tensor* bias) {
auto bias_mem_p = this->AcquireMemory("@bias_mem_p_target");
if (is_test && bias_mem_p) {
if (is_test_ && bias_mem_p) {
return bias_mem_p;
} else {
const K* bias_data = bias->data<K>();
Expand All @@ -294,9 +294,12 @@ class ConvTransposeMKLDNNHandlerT
MKLDNNMemoryFormat::x);
return this->AcquireMemoryWithReorder(
user_bias_md, this->fwd_pd_->bias_desc(),
platform::to_void_cast<K>(bias_data), "@bias_mem_p", is_test);
platform::to_void_cast<K>(bias_data), "@bias_mem_p", is_test_);
}
}

private:
const bool is_test_;
};

template <typename T, typename K>
Expand Down Expand Up @@ -325,8 +328,6 @@ class ConvTransposeMKLDNNOpKernel : public framework::OpKernel<T> {
ctx.template device_context<platform::MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine();

const bool is_test = ctx.Attr<bool>("is_test");

const auto* input = ctx.Input<Tensor>("Input");
const auto* filter = ctx.Input<Tensor>("Filter");
const auto* bias =
Expand All @@ -340,7 +341,7 @@ class ConvTransposeMKLDNNOpKernel : public framework::OpKernel<T> {
output, unique_name);
auto src_memory_p = handler.AcquireSrcMemoryWithReorder(input);
auto weights_memory_p = handler.AcquireWeightsMemoryWithReorder(
filter, ctx.Attr<int>("groups"), is_test);
filter, ctx.Attr<int>("groups"));

std::shared_ptr<dnnl::memory> dst_memory_p =
handler.template AcquireDstMemory<T_out>(output);
Expand All @@ -352,7 +353,7 @@ class ConvTransposeMKLDNNOpKernel : public framework::OpKernel<T> {
{MKLDNN_ARG_DST, *dst_memory_p}};

if (bias) {
auto bias_memory_p = handler.AcquireBiasMemoryWithReorder(bias, is_test);
auto bias_memory_p = handler.AcquireBiasMemoryWithReorder(bias);
args.insert({MKLDNN_ARG_BIAS, *bias_memory_p});
}
auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
Expand Down
Loading