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

[GPU] Fixed matmul handling for some shapes (#6642)

Browse Source
This commit is contained in:
Vladimir Paramuzov 2021-07-16 17:05:52 +03:00 committed by GitHub
parent a92a737ba0
commit 22fddb4315
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 60 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
inference-engine/src/cldnn_engine/cldnn_engine.cpp
View File

@ -83,6 +83,7 @@
#include "gpu/gpu_config.hpp"
#include "cldnn/runtime/device_query.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
#ifdef __linux__
# include <dlfcn.h>
@ -436,6 +437,11 @@ InferenceEngine::CNNNetwork clDNNEngine::CloneAndTransformNetwork(const Inferenc
manager.run_passes(nGraphFunc);
}
}
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(!debug_config->dump_graphs.empty()) {
clonedNetwork.serialize(debug_config->dump_graphs + "/transformed_func.xml");
}
return clonedNetwork;
}

39
inference-engine/src/cldnn_engine/ops/matmul.cpp
View File

@ -62,6 +62,8 @@ void CreateMatMulOp(Program& p, const std::shared_ptr<ngraph::op::v0::MatMul>& o
bool is_fc = IsNodeOnConstPath(op->get_input_node_shared_ptr(1));
is_fc &= std::count_if(shape_b.begin(), shape_b.end(), [](size_t x) { return x != 1; }) <= 2;
// TODO: This conditions can be relaxed with proper handling in FC path
is_fc &= shape_b.size() > 1 && shape_a.size() > 1;
if (is_fc) {
ngraph::Shape shape_a_aligned, shape_b_aligned;
@ -73,10 +75,10 @@ void CreateMatMulOp(Program& p, const std::shared_ptr<ngraph::op::v0::MatMul>& o
auto inputName = inputPrimitives[0];
auto weightsName = inputPrimitives[1];
// Weights normalization
if (!op->get_transpose_b()) {
ngraph::Shape output_shape = shape_b;
std::vector<uint16_t> transpose_order(output_shape.size());
std::vector<uint16_t> transpose_order(shape_b.size());
std::iota(transpose_order.begin(), transpose_order.end(), 0);
std::swap(*(transpose_order.end() - 1), *(transpose_order.end() - 2));
@ -95,8 +97,7 @@ void CreateMatMulOp(Program& p, const std::shared_ptr<ngraph::op::v0::MatMul>& o
// Input normalization
if (op->get_transpose_a()) {
ngraph::Shape output_shape = shape_a;
std::vector<uint16_t> transpose_order(output_shape.size());
std::vector<uint16_t> transpose_order(shape_a.size());
std::iota(transpose_order.begin(), transpose_order.end(), 0);
std::swap(*(transpose_order.end() - 1), *(transpose_order.end() - 2));
@ -131,16 +132,20 @@ void CreateMatMulOp(Program& p, const std::shared_ptr<ngraph::op::v0::MatMul>& o
if (reshape_fc) {
inputName = reshape_to_2d(shape_a, inputName, shape_a.back(), "_cldnn_reshape_in");
}
if (shape_b.size() != 2) {
weightsName = reshape_to_2d(shape_b, weightsName, K, "_cldnn_reshape_weights");
}
auto input_rank = reshape_fc ? 2 : shape_a.size();
auto fcPrim = cldnn::fully_connected(layerName,
inputName,
weightsName,
"",
DataTypeFromPrecision(op->get_output_element_type(0)),
cldnn::padding(),
op->get_output_shape(0).size());
input_rank);
p.AddPrimitive(fcPrim);
@ -196,7 +201,29 @@ void CreateMatMulOp(Program& p, const std::shared_ptr<ngraph::op::v0::MatMul>& o
auto reshapeName = layerName + "_cldnn_in" + std::to_string(i) + "_reshape";
// Extend input dimensions by prepending ones
inputDims.insert(inputDims.begin(), outDimsN - inputDimsN, 1ul);
if (inputDimsN == 1) {
// One-dimensional tensors unsqueezing is applied for each input independently.
// The axes inserted in this step are not included in the output shape.
// * If rank of the **first** input is equal to 1, it is always unsqueezed to 2D tensor **row vector** (regardless of `transpose_a`)
// by adding axes with size 1 at ROW_INDEX_DIM, to the **left** of the shape. For example `[S]` will be reshaped to `[1, S]`.
// * If rank of the **second** input is equal to 1, it is always unsqueezed to 2D tensor **column vector** (regardless of `transpose_b`)
// by adding axes with size 1 at COL_INDEX_DIM, to the **right** of the shape. For example `[S]` will be reshaped to `[S, 1]`.
bool transpose = false;
if (i == 0) {
transpose = op->get_transpose_a();
inputDims.insert(inputDims.begin(), 1);
} else {
transpose = op->get_transpose_b();
inputDims.insert(inputDims.end(), 1);
}
// Specs says that shapes must be unsqueezed regardless of tranpose flag, but primitive implementation always respects transposes
// so we have to swap dimensions correspondingly to have consistent shapes.
if (transpose) {
std::swap(inputDims[0], inputDims[1]);
}
}
if (inputDimsN < outDimsN)
inputDims.insert(inputDims.begin(), outDimsN - inputDimsN, 1ul);
auto targetShape = gemmSpecificTensor(inputDims);

22
inference-engine/tests/functional/plugin/gpu/shared_tests_instances/single_layer_tests/mat_mul.cpp
View File

@ -16,7 +16,27 @@ const std::vector<InferenceEngine::Precision> inputPrecisions = {
};
const std::vector<ShapeRelatedParams> shapeRelatedParams = {
{ { {1, 4, 5, 6}, false }, { {1, 4, 6, 4}, false } }
{ { {1, 4, 5, 6}, false }, { {1, 4, 6, 4}, false } },
{ { {1, 16, 128}, false }, { {1, 64, 128}, true } },
{ { {4, 5, 6}, false }, { {6, 3}, false } },
{ { {9, 9, 9}, false }, { {9, 9}, false } },
{ { {1, 2, 3}, false }, { {1, 1, 3, 2}, false } },
{ { {1, 3, 2, 4}, false }, { {2, 1, 4, 2}, false } },
{ { {2, 1, 2, 4}, false }, { {1, 3, 4, 2}, false } },
{ { {3, 2, 4}, false }, { {2, 1, 4, 2}, false } },
{ { {2, 1, 4, 2}, false }, { {3, 2, 4}, false } },
{ { {2, 1, 2, 3}, true }, { {3, 2, 4}, false } },
{ { {2, 1, 3, 2}, false }, { {3, 4, 2}, true } },
{ { {2, 1, 2, 3}, true }, { {3, 4, 2}, true } },
{ { {3}, false }, { {2, 2, 3, 1}, false } },
{ { {2, 2, 1, 3}, false }, { {3}, false } },
{ { {1, 5}, false }, { {5, 1}, false } },
{ { {5, 1}, true }, { {5, 1}, false } },
{ { {1, 5}, false }, { {1, 5}, true } },
{ { {1, 5}, false }, { {5}, false } },
{ { {5}, false }, { {5, 1}, false } },
{ { {5}, false }, { {5}, false } },
{ { {5}, true }, { {5}, true } }
};
std::vector<ngraph::helpers::InputLayerType> secondaryInputTypes = {