Skip to content

Commit

Permalink
fixed dnnl conv forward
Browse files Browse the repository at this point in the history
  • Loading branch information
@dcslin
dcslin committed Feb 20, 2020
1 parent 06ec6ce commit 8e85fdc
Showing 1 changed file with 65 additions and 14 deletions.
79 changes: 65 additions & 14 deletions src/model/operation/convolution.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,8 @@

#include "../layer/convolution.h"
#include "./convolution.h"
#include <chrono>
#include <iostream>

namespace singa {

Expand Down Expand Up @@ -86,7 +88,7 @@ ConvHandle::ConvHandle(const Tensor &input,
// convolution forward primitive descriptor is shared between forward and
// backward process
conv_d = new dnnl::convolution_forward::desc(
dnnl::prop_kind::forward_inference, dnnl::algorithm::convolution_direct,
dnnl::prop_kind::forward, dnnl::algorithm::convolution_direct,
x_md, w_md, b_md, y_md, s_dims, p_dims, p_dims);

auto eng = input.device()->context(0)->dnnl_engine;
Expand Down Expand Up @@ -123,6 +125,7 @@ Tensor CpuConvForward(const Tensor &x, Tensor &W, Tensor &b,

#ifdef USE_DNNL
DataType dtype = x.data_type();

auto dev = x.device();

Shape shape{ch.batchsize, ch.num_filters, ch.conv_height, ch.conv_width};
Expand All @@ -132,22 +135,66 @@ Tensor CpuConvForward(const Tensor &x, Tensor &W, Tensor &b,
[&output, &x, &W, &b, &ch](Context *ctx) {

using namespace dnnl;
using tag = memory::format_tag;
using dt = memory::data_type;

auto eng = ctx->dnnl_engine;
auto x_mem = memory(ch.x_md, eng, x.block()->mutable_data());
auto w_mem = memory(ch.w_md, eng, W.block()->mutable_data());
auto b_mem = memory(ch.b_md, eng, b.block()->mutable_data());
auto y_mem = memory(ch.y_md, eng, output.block()->mutable_data());

convolution_forward(*ch.conv_pd)
.execute(ctx->dnnl_stream, {{DNNL_ARG_SRC, x_mem},
{DNNL_ARG_WEIGHTS, w_mem},
{DNNL_ARG_BIAS, b_mem},
{DNNL_ARG_DST, y_mem}});
ctx->dnnl_stream.wait();
auto dtype = dnnl::memory::data_type::f32;

auto conv_user_src_memory = memory({{ch.x_dims}, dtype, tag::nchw}, eng, x.block()->mutable_data());
auto conv_user_weights_memory = memory({{ch.w_dims}, dtype, tag::goihw}, eng, W.block()->mutable_data());
auto conv_user_bias_memory = memory({{ch.b_dims}, dtype, tag::x}, eng, b.block()->mutable_data());

auto conv_src_md = memory::desc({ch.x_dims}, dtype, tag::any);
auto conv_bias_md = memory::desc({ch.b_dims}, dtype, tag::x);
auto conv_weights_md = memory::desc({ch.w_dims}, dtype, tag::goihw);
auto conv_dst_md = memory::desc({ch.o_dims}, dtype, tag::nchw);

auto conv_desc = convolution_forward::desc(prop_kind::forward, algorithm::convolution_direct, conv_src_md, conv_weights_md, conv_bias_md, conv_dst_md, ch.s_dims, ch.p_dims, ch.p_dims);
auto conv_pd = convolution_forward::primitive_desc(conv_desc, eng);

auto conv_src_memory = conv_user_src_memory;
auto conv_weights_memory = conv_user_weights_memory;
auto conv_dst_memory = memory(conv_pd.dst_desc(), eng, output.block()->mutable_data());

auto s = ctx->dnnl_stream;

// forward performance ok
std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();

Tensor x_reo;
x_reo.ResetLike(x);
Tensor W_reo;
W_reo.ResetLike(W);

if (conv_pd.src_desc() != conv_user_src_memory.get_desc()) {
//conv_src_memory = memory(conv_pd.src_desc(), eng);
conv_src_memory = memory(conv_pd.src_desc(), eng, x_reo.block()->mutable_data());
reorder(conv_user_src_memory, conv_src_memory).execute(s,{{DNNL_ARG_FROM, conv_user_src_memory},
{DNNL_ARG_TO, conv_src_memory}});
}
if (conv_pd.weights_desc() != conv_user_weights_memory.get_desc()) {
//conv_weights_memory = memory(conv_pd.weights_desc(), eng);
conv_weights_memory = memory(conv_pd.weights_desc(), eng, W_reo.block()->mutable_data());
reorder(conv_user_weights_memory, conv_weights_memory).execute(s,{{DNNL_ARG_FROM, conv_user_weights_memory},
{DNNL_ARG_TO, conv_weights_memory}});
}

convolution_forward(conv_pd).execute(s,{{DNNL_ARG_SRC, conv_src_memory},
{DNNL_ARG_WEIGHTS, conv_weights_memory},
{DNNL_ARG_BIAS, conv_user_bias_memory},
{DNNL_ARG_DST, conv_dst_memory}});

s.wait();

std::chrono::steady_clock::time_point end = std::chrono::steady_clock::now();
std::cout << "forward Time difference = " << std::chrono::duration_cast<std::chrono::microseconds>(end - begin).count() << "[microsec]" << std::endl;

},
{x.block(), W.block(), b.block()}, {output.block()});


return output;
#else // cpp naive
Shape w_shape = W.shape();
Expand All @@ -159,9 +206,7 @@ Tensor CpuConvForward(const Tensor &x, Tensor &W, Tensor &b,

DataType dtype = x.data_type();
auto dev = x.device();
Shape shape{ch.batchsize, ch.num_filters, ch.conv_height, ch.conv_width};
Tensor output(shape, dev, dtype);

Shape shape{ch.batchsize, ch.num_filters, ch.conv_height, ch.conv_width}; Tensor output(shape, dev, dtype);
Tensor col_data(Shape{ch.col_height, ch.col_width}); // broadcasted image

float *data_col = new float[ch.col_height * ch.col_width];
Expand Down Expand Up @@ -217,12 +262,18 @@ Tensor CpuConvBackwardx(const Tensor &dy, Tensor &W, const Tensor &x,
auto conv_bwd_data_pd = convolution_backward_data::primitive_desc(
conv_bwd_data_d, eng, *ch.conv_pd);

// perf not ok
std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();

convolution_backward_data(conv_bwd_data_pd)
.execute(ctx->dnnl_stream, {{DNNL_ARG_DIFF_DST, dy_mem},
{DNNL_ARG_WEIGHTS, w_mem},
{DNNL_ARG_DIFF_SRC, dx_mem}});
ctx->dnnl_stream.wait();

std::chrono::steady_clock::time_point end = std::chrono::steady_clock::now();
std::cout << "backwardx Time difference = " << std::chrono::duration_cast<std::chrono::microseconds>(end - begin).count() << "[microsec]" << std::endl;

},
{x.block(), dy.block(), W.block()}, {dx.block()});

Expand Down

0 comments on commit 8e85fdc

Please sign in to comment.