Skip to content

Commit

Permalink
Added src and dst indices calculating
Browse files Browse the repository at this point in the history
  • Loading branch information
@a-sidorova
a-sidorova committed Mar 15, 2021
1 parent 35dbe02 commit 5a1b972
Show file tree
Hide file tree
Showing 2 changed files with 66 additions and 47 deletions.
98 changes: 58 additions & 40 deletions inference-engine/src/mkldnn_plugin/nodes/mkldnn_strided_slice_node.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@
#include "common/dnnl_thread.hpp"
#include "common/cpu_memcpy.h"
#include "utils/general_utils.h"
#include "common/dnnl_thread.hpp"

#include <string>
#include <algorithm>
Expand All @@ -22,6 +23,14 @@ using namespace mkldnn;
using namespace MKLDNNPlugin;
using namespace InferenceEngine;

static inline size_t parallel_init(size_t start, size_t nDims, const SizeVector& dims, SizeVector& indexes) {
for (int j = nDims - 1; j >= 0; j--) {
indexes[j] = start % dims[j];
start = start / dims[j];
}
return start;
}

MKLDNNStridedSliceNode::MKLDNNStridedSliceNode(const InferenceEngine::CNNLayerPtr& layer, const mkldnn::engine& eng, MKLDNNWeightsSharing::Ptr &cache) :
MKLDNNNode(layer, eng, cache) {}

Expand Down Expand Up @@ -365,7 +374,7 @@ void MKLDNNStridedSliceNode::createPrimitive() {

// gluing dimensions (reshaping)
std::pair<size_t, size_t> secondDim = { 0, begin.size() };
std::vector<size_t> indexes(1, 0);
SizeVector indexes(1, 0);
for (int idx = 0; idx < begin.size(); idx++) {
if (begin[idx] != 0 || end[idx] != params.srcDims[idx] - 1 || stride[idx] != 1) {
indexes.push_back(std::max(idx - 1, 0));
Expand Down Expand Up @@ -436,6 +445,51 @@ void MKLDNNStridedSliceNode::createPrimitive() {
if (params.dstDims.size() > 2)
params.lastDstDim /= newDstDims[secondDim.first];
}

// calculate src and dst indices
if (params.dstDims.size() == 1 || params.nDimsForWork != 1) {
params.nThreads = dnnl_get_max_threads();
params.srcIndices.resize(params.workAmount, 0);
params.dstIndices.resize(params.workAmount, 0);
indexes.clear();

auto getSrcIdx = [this](const SizeVector& indexes){
size_t srcIdx = 0;
for (int i = 0; i < params.nDimsForWork; ++i)
srcIdx += (begin[i] + indexes[i] * stride[i]) * params.srcStrides[i];
return srcIdx * params.dataSize;
};

for (size_t t = 0; t < params.nThreads; t++) {
size_t start = 0, end = 0;
indexes.resize(params.nDimsForWork, 0);
splitter(params.workAmount, params.nThreads, t, start, end);
parallel_init(start, params.nDimsForWork, params.dstDims, indexes);

srcIdx = getSrcIdx(indexes);
for (size_t j = start; j < end; j++) {
params.dstIndices[j] = j * params.lastDstDim;
params.srcIndices[j] = srcIdx;

bool out = false;
for (int k = params.nDimsForWork - 1; k >= 0; k--) {
indexes[k]++;
if (indexes[k] < params.dstDims[k]) {
srcIdx += stride[k] * params.srcStrides[k] * params.dataSize;
break;
} else {
indexes[k] = 0;
out = true;
}
}

if (out) {
srcIdx = getSrcIdx(indexes);
out = false;
}
}
}
}
}

void MKLDNNStridedSliceNode::execute(mkldnn::stream strm) {
Expand All @@ -445,14 +499,6 @@ void MKLDNNStridedSliceNode::execute(mkldnn::stream strm) {
stridedSlice();
}

static inline size_t parallel_init(size_t start, size_t nDims, const SizeVector& dims, SizeVector& indexes) {
for (int j = nDims - 1; j >= 0; j--) {
indexes[j] = start % dims[j];
start = start / dims[j];
}
return start;
}

void MKLDNNStridedSliceNode::stridedSliceV() {
const uint8_t* srcData = reinterpret_cast<const uint8_t*>(this->getParentEdgeAt(DATA_ID)->getMemoryPtr()->GetPtr()) +
(begin[0] * params.srcStrides[0] + begin[1] * params.srcStrides[1]) * params.dataSize;
Expand All @@ -479,43 +525,15 @@ void MKLDNNStridedSliceNode::stridedSlice() {
(stride.back() == 1 && stride.size() > 1 ? begin[params.nDimsForWork] * params.srcStrides[params.nDimsForWork] * params.dataSize : 0);
uint8_t* dstData = reinterpret_cast<uint8_t*>(this->getChildEdgeAt(0)->getMemoryPtr()->GetPtr());

parallel_nt(0, [&](const int ithr, const int nthr) {
parallel_nt(params.nThreads, [&](const int ithr, const int nthr) {
size_t start = 0, end = 0;
SizeVector indexes(params.nDimsForWork, 0);
splitter(params.workAmount, nthr, ithr, start, end);
parallel_init(start, params.nDimsForWork, params.dstDims, indexes);

size_t srcIdx;
getSrcIdx(srcIdx, indexes);

for (size_t iwork = start, dstIdx = start * params.lastDstDim, i = 1; iwork < end; ++iwork, dstIdx += params.lastDstDim) {
cpu_memcpy(&dstData[dstIdx], &srcData[srcIdx], params.lastDstDim);

for (int j = params.nDimsForWork - 1; j >= 0; j--) {
indexes[j]++;
if (indexes[j] < params.dstDims[j]) {
srcIdx += stride[j] * params.srcStrides[j] * params.dataSize;
break;
} else {
indexes[j] = i = 0;
}
}

if (!i) {
getSrcIdx(srcIdx, indexes);
i++;
}
}
for (size_t iwork = start; iwork < end; ++iwork)
cpu_memcpy(&dstData[params.dstIndices[iwork]], &srcData[params.srcIndices[iwork]], params.lastDstDim);
});
}

inline void MKLDNNStridedSliceNode::getSrcIdx(size_t& srcIdx, const SizeVector& indexes) {
srcIdx = 0;
for (int i = 0; i < params.nDimsForWork; ++i)
srcIdx += (begin[i] + indexes[i] * stride[i]) * params.srcStrides[i];
srcIdx *= params.dataSize;
}

bool MKLDNNStridedSliceNode::created() const {
return getType() == StridedSlice;
}
Expand Down
15 changes: 8 additions & 7 deletions inference-engine/src/mkldnn_plugin/nodes/mkldnn_strided_slice_node.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,8 +29,6 @@ class MKLDNNStridedSliceNode : public MKLDNNNode {
void stridedSliceV();
void stridedSlice();

inline void getSrcIdx(size_t& srcIdx, const InferenceEngine::SizeVector& indexes);

const size_t DATA_ID = 0;
const size_t BEGIN_ID = 1;
const size_t END_ID = 2;
Expand All @@ -51,11 +49,14 @@ class MKLDNNStridedSliceNode : public MKLDNNNode {
InferenceEngine::SizeVector dstDims;
InferenceEngine::SizeVector srcStrides;
InferenceEngine::SizeVector dstStrides;
size_t nDimsForWork;
size_t workAmount;
size_t lastDstDim;
size_t dataSize;
bool equalDims;
InferenceEngine::SizeVector srcIndices;
InferenceEngine::SizeVector dstIndices;
size_t nThreads = 0;
size_t nDimsForWork = 0;
size_t workAmount = 0;
size_t lastDstDim = 0;
size_t dataSize = 0;
bool equalDims = false;
} params;
};

Expand Down

0 comments on commit 5a1b972

Please sign in to comment.