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

[IE][VPU][Myriad]: NMS layer optimization; support big buffers (#2839) (#2839)

Browse Source 
* Myriad: new firmware version (1445)
This commit is contained in:
Andrey Sokolov 2020-10-30 13:30:21 +03:00 committed by GitHub
parent 6845af5880
commit 4ce6e3414c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 45 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
inference-engine/cmake/vpu_dependencies.cmake
View File

@ -19,7 +19,7 @@ set(VPU_SUPPORTED_FIRMWARES usb-ma2450 usb-ma2x8x pcie-ma248x)
# Default packages
#
set(FIRMWARE_PACKAGE_VERSION 1440)
set(FIRMWARE_PACKAGE_VERSION 1445)
set(VPU_CLC_MA2X8X_VERSION "movi-cltools-20.09.1")
#

1
inference-engine/src/vpu/graph_transformer/include/vpu/middleend/allocator/structs.hpp
View File

@ -21,6 +21,7 @@ namespace vpu {
const int DDR_MAX_SIZE = 512 * 1024 * 1024;
const int CMX_SLICE_SIZE = 128 * 1024;
const int DATA_ALIGNMENT = 64;
const int CMX_SHAVE_BUFFER_SIZE = 100 * 1024;
//
// Allocator Structs

44
inference-engine/src/vpu/graph_transformer/src/stages/static_shape_nms.cpp
View File

@ -3,7 +3,7 @@
//
#include <vpu/frontend/frontend.hpp>
#include <vpu/compile_env.hpp>
#include <precision_utils.h>
#include <memory>
@ -54,8 +54,10 @@ private:
void serializeParamsImpl(BlobSerializer& serializer) const override {
const auto center_point_box = attrs().get<bool>("center_point_box");
const auto use_ddr_buffer = !tempBuffers().empty();
serializer.append(static_cast<int32_t>(center_point_box));
serializer.append(static_cast<int32_t>(use_ddr_buffer));
}
void serializeDataImpl(BlobSerializer& serializer) const override {
@ -74,11 +76,35 @@ private:
input5->serializeBuffer(serializer);
outputData->serializeBuffer(serializer);
outputDims->serializeBuffer(serializer);
if (!tempBuffers().empty())
tempBuffer(0)->serializeBuffer(serializer);
}
};
bool isCMXEnough(int cmxSize, int numSlices, std::vector<int> bufferSizes) {
int curOffset = 0;
int curSlice = 0;
const auto buffer_allocate = [&curOffset, &curSlice, &numSlices, cmxSize](int numBytes) {
if (curOffset + numBytes < cmxSize) {
curOffset += numBytes;
} else if (curSlice < numSlices && numBytes < cmxSize) {
curSlice++;
curOffset = numBytes;
} else {
return false;
}
return true;
};
return std::all_of(bufferSizes.begin(), bufferSizes.end(), buffer_allocate);
}
} // namespace
void FrontEnd::parseStaticShapeNMS(const Model& model, const ie::CNNLayerPtr& layer, const DataVector& inputs, const DataVector& outputs) const {
VPU_THROW_UNLESS(inputs.size() == 6,
"StaticShapeNMS with name {} parsing failed, expected number of inputs: 6, but {} provided",
@ -119,6 +145,22 @@ void FrontEnd::parseStaticShapeNMS(const Model& model, const ie::CNNLayerPtr& la
auto stage = model->addNewStage<StaticShapeNMS>(layer->name, StageType::StaticShapeNMS, layer, usedInputs, DataVector{outIndices, outShape});
stage->attrs().set<bool>("center_point_box", centerPointBox);
const auto inputDims0 = inputs[0]->desc().dims();
const auto perm = DimsOrder::fromNumDims(inputDims0.size()).toPermutation();
const auto spatDim = inputDims0[perm[1]];
const int ddrBufferSize0 = 2 * sizeof(int16_t) * 4 * spatDim;
const int ddrBufferSize1 = 2 * sizeof(int16_t) * spatDim;
const int ddrBufferSize2 = 2 * sizeof(int32_t) * spatDim;
const int ddrBufferSize = ddrBufferSize0 + ddrBufferSize1 + ddrBufferSize2 + 2 * vpu::DATA_ALIGNMENT;
const auto& env = CompileEnv::get();
const auto numSlices = env.resources.numSHAVEs;
const int cmxTempBufferSize = 4 * sizeof(int32_t) * 256;
if (!isCMXEnough(CMX_SHAVE_BUFFER_SIZE, numSlices, {ddrBufferSize0, ddrBufferSize1, ddrBufferSize2, cmxTempBufferSize}))
model->addTempBuffer(stage, DataDesc({ddrBufferSize}));
}
} // namespace vpu