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

[GPU] Fix output format not changing at runtime (#15887)

Browse Source 
* [GPU] Fix output format not changing at runtime

Signed-off-by: Andrew Park <andrew.park@intel.com>

* Add remove_redundant_reorders pass TC for ov_gpu_unit_tests

Signed-off-by: Andrew Park <andrew.park@intel.com>

---------

Signed-off-by: Andrew Park <andrew.park@intel.com>
This commit is contained in:
Andrew Kwangwoong Park
2023-02-25 07:26:54 +09:00
committed by GitHub
parent 30939f5021
commit 46e8aad4bb
2 changed files with 74 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/plugins/intel_gpu/src/graph/graph_optimizer/remove_redundant_reorders.cpp
View File

@@ -169,6 +169,18 @@ void remove_redundant_reorders::run(program& p) {
!r_node.get_primitive()->has_surface_input();
if (remove_dep) {
// for chains like
// b_fs_yx_fsv16 -> reorder(ofmt:bfyx) -> bfyx -> reorder(ofmt:any) -> bfyx
// if output_format of current node is format::any, input format of the dependency node is propagated as it is
// b_fs_yx_fsv16 -> reorder(ofmt:any) -> b_fs_yx_fsv16
// so output format of dependency node must be stored in output_format of current node
// b_fs_yx_fsv16 -> reorder(ofmt:bfyx) -> bfyx
auto output_layout = r_dep_node.get_output_layout();
auto prim = std::const_pointer_cast<reorder>(r_node.get_primitive());
if (prim->output_format == format::any)
prim->output_format = output_layout.format;
LOG_NODE_REMOVAL(r_dep_node.id());
r_dep_node.can_be_optimized(true);
p.add_optimized_primitive_info(r_dep_node.id());
p.extract_and_remove(r_dep_node);

62
src/plugins/intel_gpu/tests/passes/remove_redundant_reorders_tests.cpp Normal file
View File

@@ -0,0 +1,62 @@
// Copyright (C) 2023 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include "test_utils.h"
#include "intel_gpu/runtime/engine.hpp"
#include "intel_gpu/graph/network.hpp"
#include "intel_gpu/graph/program.hpp"
#include "data_inst.h"
#include "convolution_inst.h"
#include "reorder_inst.h"
#include "softmax_inst.h"
#include "pass_manager.h"
#include "to_string_utils.h"
#include "program_wrapper.h"
#include <memory>
using namespace cldnn;
using namespace ::tests;
TEST(remove_redundant_reorders, remove_dep_dynamic) {
// Topology:
// convolution -> reorder -> softmax
//
// Expectation:
// The preferred format of convolution should be selected as b_fs_yx_fsv16 (reorder_inputs)
// A new reorder that converts to bfyx should be inserted after convolution (reorder_inputs)
// In reorders, output format of dependency reorder should be saved as output_format of orginial reorder (remove_redundant_reorders)
auto& engine = get_test_engine();
auto input_layout_dynamic = layout{ov::PartialShape{1, 3, ov::Dimension::dynamic(), ov::Dimension::dynamic()},
data_types::f16, format::bfyx};
auto input = engine.allocate_memory({ data_types::f32, format::bfyx, { 1, 3, 224, 224 } });
auto weights = engine.allocate_memory({ data_types::f16, format::bfyx, { 64, 3, 7, 7 } });
topology topology;
topology.add(data("weights", weights));
topology.add(input_layout("input", input_layout_dynamic));
topology.add(convolution("conv", input_info("input"), { "weights" }));
topology.add(reorder("reorder", input_info("conv"), format::any, data_types::f32));
topology.add(softmax("softmax", input_info("reorder"), 1));
ExecutionConfig config;
config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
config.set_property(ov::intel_gpu::optimize_data(true));
network network(engine, topology, config);
network.set_input_data("input", input);
network.execute();
auto prog = network.get_program();
ASSERT_NE(prog, nullptr);
auto& softmax_node = prog->get_node("softmax");
auto softmax_layout = softmax_node.get_output_layout();
ASSERT_EQ(softmax_layout.format.value, format::bfyx);
}