IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GNA] Fix large eltwise split in case split dimension is less than alignment (#9933)

Browse Source 
* [GNA] Fix large eltwise split in case split dimension is less than alignment

* Update src/plugins/intel_gna/optimizer/gna_pass_manager.cpp

Co-authored-by: Krzysztof Bruniecki <krzysztof.bruniecki@intel.com>

Co-authored-by: Krzysztof Bruniecki <krzysztof.bruniecki@intel.com>
This commit is contained in:
Szymon Irzabek 2022-01-27 20:09:32 +01:00 committed by GitHub
parent d4779bb351
commit 069b3839ce
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 18 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
src/plugins/intel_gna/backend/gna_limitations.hpp
View File

@ -23,6 +23,7 @@ constexpr uint32_t convFiltersNumDivider = 4;
constexpr uint32_t convFilterSizeDivider = 8;
constexpr uint32_t convFilterMaxSize = 768;
constexpr uint32_t convEachKernelByteAlignment = 16;
constexpr uint32_t inputByteAlignment = 64;
constexpr uint32_t noOfInputsDivisor = 8;
constexpr uint32_t noOfInputsLowPrecDivisor = 16;

3
src/plugins/intel_gna/layers/gna_split_layer.hpp
View File

@ -7,6 +7,7 @@
#include <vector>
#include <legacy/ie_layers.h>
#include "backend/gna_limitations.hpp"
namespace GNAPluginNS {
// Split, Slice
@ -47,7 +48,7 @@ public:
};
// @brief Returns sizes of split outputs to split the input tensor to aligned parts not greater than the specified size
static std::vector<uint32_t> GetAlignedSplitSizes(uint32_t totalSize, uint32_t maxSplitSize, uint32_t alignment = 64) {
static std::vector<uint32_t> GetAlignedSplitSizes(uint32_t totalSize, uint32_t maxSplitSize, uint32_t alignment = GNALimitations::inputByteAlignment) {
std::vector<uint32_t> splitSizes;
uint32_t maxAlignedSplitSize = maxSplitSize - maxSplitSize % alignment;
uint32_t usedSize = 0;

15
src/plugins/intel_gna/optimizer/gna_pass_manager.cpp
View File

@ -1451,10 +1451,21 @@ void EltwiseSplitOverChannelsPass::run() {
IE_ASSERT(firstValuableDim != std::end(oDims));
auto splittedElementsSize = *firstValuableDim;
auto splittedDimIx = std::distance(std::begin(oDims), firstValuableDim);
auto alignment = GNALimitations::inputByteAlignment;
// Split output size should be multiple by 64 to avoid align filters insertion,
// but we need to check if our input size to split exceeds 64; if not we can always
// split if the remaining size is aligned
if (splittedElementsSize <= 64) {
if ((totalElementsSize / splittedElementsSize) % alignment == 0) {
alignment = 1;
} else {
THROW_GNA_LAYER_EXCEPTION(l) << "splitting didn't succeed\n";
}
}
// Split output size should be multiple by 64 to avoid align filters insertion
auto splitSizes = GetAlignedSplitSizes(splittedElementsSize,
GNALimitations::bufferMaxSize * splittedElementsSize / totalElementsSize);
GNALimitations::bufferMaxSize * splittedElementsSize / totalElementsSize, alignment);
pass_trace() << "transforming " << LAYER_NAME(l) << " by splitting it to multiple eltwise operations\n";
auto quantized = InferenceEngine::getInjectedData<QuantizedLayerParams>(l);

3
src/tests/functional/plugin/gna/pass_tests/eltwise_split_over_channels_pass.cpp
View File

@ -78,7 +78,8 @@ const std::vector<std::map<std::string, std::string>> configs = {
const std::vector<std::vector<size_t>> inputShape = {
{1, 67000},
{1, 500000},
{1, 936, 513}
{1, 936, 513},
{1, 64, 64, 64}
};
INSTANTIATE_TEST_SUITE_P(smoke_EltwiseSplitOverChennels, EltwiseSplitOverChannelsPassTest,