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

[GPU] Enable BroadcastTransition before CommonOptimizations (#21198)

Browse Source
This commit is contained in:
Vladimir Paramuzov 2023-11-21 19:54:04 +04:00 committed by GitHub
parent d1f71ee233
commit eb94c80804
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 5 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp
View File

@ -52,6 +52,8 @@
#include "transformations/fp16_compression/convert_compression_only_to_legacy.hpp" #include "transformations/fp16_compression/convert_compression_only_to_legacy.hpp"
#include "transformations/fp16_compression/mark_decompression_convert_constant_folding.hpp" #include "transformations/fp16_compression/mark_decompression_convert_constant_folding.hpp"
#include "transformations/common_optimizations/common_optimizations.hpp" #include "transformations/common_optimizations/common_optimizations.hpp"
#include "transformations/common_optimizations/broadcast_elementwise_fusion.hpp"
#include "transformations/common_optimizations/broadcast_transition.hpp"
#include "transformations/common_optimizations/lin_op_sequence_fusion.hpp" #include "transformations/common_optimizations/lin_op_sequence_fusion.hpp"
#include "transformations/common_optimizations/weights_dequantize_to_fake_quantize.hpp" #include "transformations/common_optimizations/weights_dequantize_to_fake_quantize.hpp"
#include "transformations/common_optimizations/convert_quantize_dequantize.hpp" #include "transformations/common_optimizations/convert_quantize_dequantize.hpp"
@ -230,6 +232,9 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
manager.register_pass<ov::pass::MVNFusion>(); manager.register_pass<ov::pass::MVNFusion>();
// decompose MVNs that sre not supported in GPU, so that they will be marked as precision sensitive in ConvertPrecision // decompose MVNs that sre not supported in GPU, so that they will be marked as precision sensitive in ConvertPrecision
manager.register_pass<ov::pass::MVN6Decomposition>(); manager.register_pass<ov::pass::MVN6Decomposition>();
// Run these broadcast optimizations earlier to ensure that those are executed before NopElimination/ConstantFolding
manager.register_pass<ov::pass::BroadcastElementwiseFusion>();
manager.register_pass<ov::pass::BroadcastTransition>();
manager.register_pass<ov::pass::KeepConstantsPrecisionAndAddConverts>(); manager.register_pass<ov::pass::KeepConstantsPrecisionAndAddConverts>();
pass_config->set_callback<ov::pass::KeepConstantsPrecisionAndAddConverts>( pass_config->set_callback<ov::pass::KeepConstantsPrecisionAndAddConverts>(