Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

[mkldnn]Mkldnn bn opt backport from master to 1.7x #18009

Merged
merged 4 commits into from
Apr 15, 2020
Merged

[mkldnn]Mkldnn bn opt backport from master to 1.7x #18009

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
59a88f2
optimize for backward batchnorm
rongzha1 Mar 25, 2020
ad37ffe
using memcpy instead of 'for' loop
rongzha1 Mar 25, 2020
b07cd0c
rm unnecessary pointer cast and add const for some variable
rongzha1 Mar 25, 2020
79c1788
trigger CI
rongzha1 Apr 13, 2020
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
50 changes: 27 additions & 23 deletions src/operator/nn/mkldnn/mkldnn_batch_norm-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,9 +180,10 @@ void MKLDNNBatchNormForward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,
CHECK(weight_mem.get_desc().get_size() == channels_ * sizeof(float) * 2);
float* weight_ptr = gamma.data().dptr<float>();
float* bias_ptr = beta.data().dptr<float>();
const size_t copy_size = sizeof(weight_buf[0]) * channels_;
if (!param.fix_gamma) {
memcpy(weight_buf, weight_ptr, sizeof(weight_buf[0]) * channels_);
memcpy(&weight_buf[channels_], bias_ptr, sizeof(weight_buf[0]) * channels_);
memcpy(weight_buf, weight_ptr, copy_size);
memcpy(&weight_buf[channels_], bias_ptr, copy_size);
} else if (IsBNWriting(req[batchnorm::kGamma])) {
for (int i = 0; i < channels_; i++) {
weight_buf[i] = 1.0f;
Expand Down Expand Up @@ -332,17 +333,18 @@ void MKLDNNBatchNormBackward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,
const NDArray &beta = in_data[batchnorm::kBeta];
DType *weight_buf = reinterpret_cast<DType *>(bwd.GetWeight().get_data_handle());
nnvm::dim_t channels_ = data.shape()[1];
for (int i = 0; i < channels_; i++) {
if (!param.fix_gamma)
weight_buf[i] = (gamma.data().dptr<DType>())[i]; // weight
else
DType *weight_ptr = gamma.data().dptr<DType>();
DType* bias_ptr = beta.data().dptr<DType>();
const size_t copy_size = sizeof(DType) * channels_;
if (!param.fix_gamma) {
memcpy(weight_buf, weight_ptr, copy_size);
memcpy(&weight_buf[channels_], bias_ptr, copy_size);
} else {
for (int i = 0; i < channels_; i++) {
weight_buf[i] = static_cast<DType>(1.0f);
}
memcpy(&weight_buf[channels_], bias_ptr, copy_size);
}

for (int i = 0; i < channels_; i++) {
weight_buf[channels_ + i] = (beta.data().dptr<DType>())[i]; // bias
}

mkldnn_args_map_t net_args;
net_args[MKLDNN_ARG_SRC] = *data_mem;
net_args[MKLDNN_ARG_DIFF_SRC] = *gradi_mem;
Expand All @@ -352,10 +354,10 @@ void MKLDNNBatchNormBackward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,

// training but no input mean and variance
if (ctx.is_train && !param.use_global_stats) {
DType* moving_mean_ptr = reinterpret_cast<DType *>(moving_mean.data().dptr<DType>());
DType* moving_var_ptr = reinterpret_cast<DType *>(moving_var.data().dptr<DType>());
DType* out_mean_ptr = reinterpret_cast<DType *>(out_mean.data().dptr<DType>());
DType* out_var_ptr = reinterpret_cast<DType *>(out_var.data().dptr<DType>());
DType* moving_mean_ptr = moving_mean.data().dptr<DType>();
DType* moving_var_ptr = moving_var.data().dptr<DType>();
DType* out_mean_ptr = out_mean.data().dptr<DType>();
DType* out_var_ptr = out_var.data().dptr<DType>();
mkldnn::memory var_mem(bwd.pd.variance_desc(), CpuEngine::Get()->get_engine());
DType *tmp_var_ptr = reinterpret_cast<DType *>(var_mem.get_data_handle());

Expand All @@ -381,15 +383,17 @@ void MKLDNNBatchNormBackward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,

// copy data from gradw_mem to in_grad[1] and in_grad[2]
DType *gw_buf = reinterpret_cast<DType *>(bwd.GetGradw().get_data_handle());
for (int i = 0; i < channels_; i++) {
if (!param.fix_gamma)
(in_grad[1].data().dptr<DType>())[i] = gw_buf[i];
else
(in_grad[1].data().dptr<DType>())[i] = 0.0f;
}
DType *w_grad_1 = in_grad[1].data().dptr<DType>();
DType *w_grad_2 = in_grad[2].data().dptr<DType>();

for (int i = 0; i < channels_; i++) {
(in_grad[2].data().dptr<DType>())[i] = gw_buf[i + channels_];
if (!param.fix_gamma) {
memcpy(w_grad_1, gw_buf, copy_size);
memcpy(w_grad_2, &gw_buf[channels_], copy_size);
} else {
for (int i = 0; i < channels_; i++) {
(in_grad[1].data().dptr<DType>())[i] = 0.0f;
}
memcpy(w_grad_2, &gw_buf[channels_], copy_size);
}
} else {
LOG(FATAL) << "MKLDNN batch normalization backward: should not reach here ...";
Expand Down