MNN/source/backend/cpu/OneDNNConvolution.cpp

194 lines
8.6 KiB
C++

#ifdef MNN_USE_ONEDNN
#include "OneDNNConvolution.hpp"
#include "CPUConvolution.hpp"
#include "dnnl.hpp"
using namespace dnnl;
using tag = memory::format_tag;
using dt = memory::data_type;
namespace MNN {
namespace OneDNN {
class OneDNNConvolution : public Execution {
public:
OneDNNConvolution(const Convolution2DCommon *common, Backend *b, const float *originWeight,
size_t originWeightSize, const float *bias, size_t biasSize) : Execution(b) {
mCommon = common;
const auto convCommon = common;
const auto kw = convCommon->kernelX();
const auto kh = convCommon->kernelY();
auto ic = convCommon->inputCount();
const auto oc = convCommon->outputCount();
const auto strideX = convCommon->strideX();
const auto strideY = convCommon->strideY();
if (0 == ic) {
ic = originWeightSize / oc / kw / kh;
}
eng = engine(engine::kind::cpu, 0);
stm = stream(eng);
memory::dims conv_weights_tz = {oc, ic, kh, kw};
memory::dims conv_bias_tz = {oc};
memory::dims conv_strides = {strideX, strideY};
int defaultOw = 10;
int defaultOh = 10;
memory::dims conv_src_tz = {1, ic, mCommon->strideY() * (defaultOh - 1) + (kh - 1) * mCommon->dilateY() + 1, (kw - 1) * mCommon->dilateX() + 1 + mCommon->strideX() * (defaultOw - 1)};
memory::dims conv_dst_tz = {1, oc, defaultOh, defaultOw};
memory::dims conv_padding = {0, 0};
if (mCommon->relu()) {
post_ops ops;
ops.append_eltwise(1.0f, algorithm::eltwise_relu, 0.0f, 0.0f);
conv_attr.set_post_ops(ops);
}
if (mCommon->relu6()) {
post_ops ops;
ops.append_eltwise(1.0f, algorithm::eltwise_clip, 0.0f, 6.0f);
conv_attr.set_post_ops(ops);
}
auto user_weights_md = memory::desc({conv_weights_tz}, dt::f32, tag::oihw);
auto conv_src_md = memory::desc({conv_src_tz}, dt::f32, tag::any);
auto conv_weights_md = memory::desc({conv_weights_tz}, dt::f32, tag::any);
auto conv_bias_md = memory::desc({conv_bias_tz}, dt::f32, tag::a);
auto conv_dst_md = memory::desc({conv_dst_tz}, dt::f32, tag::any);
auto conv_desc = convolution_forward::desc(prop_kind::forward_inference,
algorithm::convolution_auto, conv_src_md, conv_weights_md, conv_bias_md,
conv_dst_md, conv_strides, conv_padding, conv_padding);
auto conv_pd = convolution_forward::primitive_desc(conv_desc, conv_attr, eng);
const auto* weightSrc = originWeight;
mWeight.reset(Tensor::createDevice<int8_t>({(int)conv_pd.weights_desc().get_size()}));
auto res = b->onAcquireBuffer(mWeight.get(), Backend::STATIC);
if (!res) {
mValid = false;
return;
}
auto user_weights = memory(user_weights_md, eng, (float*)weightSrc);
conv_weights = memory(conv_pd.weights_desc(), eng, mWeight->host<float>());
auto r_pd = reorder::primitive_desc(user_weights, conv_weights);
reorder(r_pd).execute(stm, user_weights, conv_weights);
conv_bias = memory(conv_bias_md, eng);
{
auto ptr = conv_bias.map_data();
::memcpy(ptr, bias, biasSize * sizeof(float));
conv_bias.unmap_data(ptr);
}
}
virtual ~OneDNNConvolution() {
if (nullptr != mWeight) {
backend()->onReleaseBuffer(mWeight.get(), Backend::STATIC);
}
}
virtual ErrorCode onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) override {
const auto convCommon = mCommon;
const auto kw = convCommon->kernelX();
const auto kh = convCommon->kernelY();
const auto ic = inputs[0]->channel();
const auto oc = convCommon->outputCount();
const auto strideX = convCommon->strideX();
const auto strideY = convCommon->strideY();
const auto ih = inputs[0]->height();
const auto iw = inputs[0]->width();
const auto oh = outputs[0]->height();
const auto ow = outputs[0]->width();
auto pads = ConvolutionCommon::convolutionPadFull(inputs[0], outputs[0], mCommon);
memory::dims conv_src_tz = {inputs[0]->batch(), ic, ih, iw};
memory::dims conv_weights_tz = {oc, ic, kh, kw};
memory::dims conv_bias_tz = {oc};
memory::dims conv_dst_tz = {outputs[0]->batch(), oc, oh, ow};
memory::dims conv_strides = {strideX, strideY};
auto user_src_md = memory::desc({conv_src_tz}, dt::f32, tag::nChw4c);
auto user_weights_md = memory::desc({conv_weights_tz}, dt::f32, tag::oihw);
auto user_dst_md = memory::desc({conv_dst_tz}, dt::f32, tag::nChw4c);
auto conv_src_md = memory::desc({conv_src_tz}, dt::f32, tag::any);
auto conv_dst_md = memory::desc({conv_dst_tz}, dt::f32, tag::any);
user_src = memory(user_src_md, eng, inputs[0]->host<float>());
user_dst = memory(user_dst_md, eng, outputs[0]->host<float>());
mSrcTemp = nullptr;
mDstTemp = nullptr;
// Fix weight desc and bias desc
auto conv_desc = convolution_forward::desc(prop_kind::forward_inference,
algorithm::convolution_auto, conv_src_md, conv_weights.get_desc(), conv_bias.get_desc(),
conv_dst_md, conv_strides, {std::get<1>(pads), std::get<0>(pads)}, {std::get<3>(pads), std::get<2>(pads)});
auto conv_pd = convolution_forward::primitive_desc(conv_desc, conv_attr, eng);
conv = convolution_forward(conv_pd);
mSrcTemp = nullptr;
mDstTemp = nullptr;
if (conv_pd.src_desc() != user_src.get_desc()) {
auto needSize = conv_pd.src_desc().get_size();
mSrcTemp.reset(Tensor::createDevice<int8_t>({(int)needSize}));
auto res = backend()->onAcquireBuffer(mSrcTemp.get(), Backend::DYNAMIC);
if (!res) {
return OUT_OF_MEMORY;
}
conv_src = memory(conv_pd.src_desc(), eng, mSrcTemp->host<float>());
}
if (conv_pd.dst_desc() != user_dst.get_desc()) {
auto needSize = conv_pd.dst_desc().get_size();
mDstTemp.reset(Tensor::createDevice<int8_t>({(int)needSize}));
auto res = backend()->onAcquireBuffer(mDstTemp.get(), Backend::DYNAMIC);
if (!res) {
return OUT_OF_MEMORY;
}
conv_dst = memory(conv_pd.dst_desc(), eng, mDstTemp->host<float>());
}
if (nullptr != mSrcTemp) {
backend()->onReleaseBuffer(mSrcTemp.get(), Backend::DYNAMIC);
}
if (nullptr != mDstTemp) {
backend()->onReleaseBuffer(mDstTemp.get(), Backend::DYNAMIC);
} return NO_ERROR;
}
virtual ErrorCode onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) override {
memory conv_src_temp = user_src;
if (nullptr != mSrcTemp) {
auto r_pd = reorder::primitive_desc(user_src, conv_src);
reorder(r_pd).execute(stm, user_src, conv_src);
conv_src_temp = conv_src;
}
memory conv_dst_temp = user_dst;
if (nullptr != mDstTemp) {
conv_dst_temp = conv_dst;
}
conv.execute(stm, {{DNNL_ARG_SRC, conv_src_temp},
{DNNL_ARG_WEIGHTS, conv_weights},
{DNNL_ARG_BIAS, conv_bias},
{DNNL_ARG_DST, conv_dst_temp}});
if (nullptr != mDstTemp) {
auto r_pd = reorder::primitive_desc(conv_dst, user_dst);
reorder(r_pd).execute(stm, conv_dst, user_dst);
}
return NO_ERROR;
}
private:
engine eng;
stream stm;
convolution_forward conv;
memory conv_weights;
memory conv_bias;
primitive_attr conv_attr;
std::shared_ptr<Tensor> mWeight;
std::shared_ptr<Tensor> mSrcTemp;
std::shared_ptr<Tensor> mDstTemp;
memory user_src;
memory user_dst;
memory conv_src;
memory conv_dst;
const Convolution2DCommon* mCommon;
};
Execution* createConvolution(const Convolution2DCommon *common, Backend *b, const float *originWeight,
size_t originWeightSize, const float *bias, size_t biasSize) {
return new OneDNNConvolution(common, b, originWeight, originWeightSize, bias, biasSize);
}
}
};
#endif