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

[GPU] Prevent Conv's input data type changing at reorder_inputs pass (#19042) (#19245)

Browse Source
This commit is contained in:
Sergey Shlyapnikov
2023-08-17 17:57:14 +04:00
committed by GitHub
parent 0262662050
commit b7415f5c3b
3 changed files with 75 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/plugins/intel_gpu/src/graph/graph_optimizer/add_required_reorders.cpp
View File

@@ -23,11 +23,14 @@ If not than required reorder is added to the network.
/*
Add a reorder in between node and usr
*/
void add_required_reorders::add_reorder(program& p, program_node* node, program_node* usr) {
void add_required_reorders::add_reorder(program& p, program_node* node, program_node* usr, bool keep_original_dt) {
layout reorder_layout = node->get_output_layout();
reorder_layout.format = usr->get_output_layout().format;
reorder_layout.data_type = usr->get_output_layout().data_type;
if (keep_original_dt)
reorder_layout.data_type = node->get_output_layout().data_type;
auto new_reorder = std::make_shared<reorder>(node->id() + "_reorder_" + usr->id(), node->id(), reorder_layout);
auto& new_reorder_node = p.get_or_create(new_reorder);
new_reorder_node.set_output_layout(reorder_layout, false);
@@ -309,8 +312,16 @@ void add_required_reorders::run(program& p) {
continue;
}
if (usr->get_output_layout() != node.first->get_output_layout())
add_reorder(p, node.first, usr);
if (usr->get_output_layout() != node.first->get_output_layout()) {
// Preserve original data type to prevent Convolution input data type from changing
// in the following sequence: Node(U8, unsupported format) -> Conv(FP16, bfyx).
// Without this condition, inserted reorder will change Conv's input to FP16, instead of
// expected U8 format.
bool keep_original_dt = false;
if (usr->is_type<convolution>())
keep_original_dt = true;
add_reorder(p, node.first, usr, keep_original_dt);
}
}
}
}

2
src/plugins/intel_gpu/src/graph/include/pass_manager.h
View File

@@ -57,7 +57,7 @@ public:
private:
void run(program& p) override;
void add_reorder(program& p, program_node* node, program_node* usr);
void add_reorder(program& p, program_node* node, program_node* usr, bool keep_original_dt = false);
};
class add_reshape_to_primitives : public base_pass {

60
src/plugins/intel_gpu/tests/passes/add_required_reorders_test.cpp Normal file
View File

@@ -0,0 +1,60 @@
// 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 "concatenation_inst.h"
#include "crop_inst.h"
#include "data_inst.h"
#include "eltwise_inst.h"
#include "fully_connected_inst.h"
#include "gather_inst.h"
#include "pass_manager.h"
#include "permute_inst.h"
#include "reshape_inst.h"
#include "shape_of_inst.h"
#include "convolution_inst.h"
#include "dft_inst.h"
#include "to_string_utils.h"
#include "program_wrapper.h"
#include <memory>
using namespace cldnn;
using namespace ::tests;
TEST(add_required_reorders, prevent_input_dt_changing_for_convs) {
auto& engine = get_test_engine();
int input_b = 1, input_f = 16, input_y = 3, input_x = 3;
int output_b = input_b, output_f = 16, output_y = 3, output_x = 3;
auto input_mem = engine.allocate_memory({ {input_b, input_f, input_y, input_x}, data_types::u8, format::bs_fs_yx_bsv16_fsv32 });
auto input2_mem = engine.allocate_memory({ {input_b, input_f, input_y, input_x}, data_types::u8, format::bs_fs_yx_bsv16_fsv32 });
auto weights_mem = engine.allocate_memory({ {16, 16, 1, 1}, data_types::i8, format::bfyx });
auto input = input_layout("input", input_mem->get_layout());
auto input_const = data("input_const", input2_mem);
auto weights = data("weights", weights_mem);
auto eltwise1 = eltwise("eltwise1", input_info("input"), input_info("input_const"), eltwise_mode::sum);
auto conv1 = convolution("conv1", input_info("eltwise1"), {"weights"}, {}, 1);
auto output_reorder = reorder("reorder", input_info("conv1"), { data_types::f32, format::bfyx, { output_b, output_f, output_y, output_x } });
topology topology_test(input, input_const, eltwise1, weights, conv1, output_reorder);
ExecutionConfig config_test = get_test_default_config(engine);
ov::intel_gpu::ImplementationDesc conv1_impl_test = { format::bfyx, "", impl_types::ocl };
config_test.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ { "conv1", conv1_impl_test } }));
auto prog = program::build_program(engine, topology_test, config_test, false, true);
program_wrapper::apply_opt_pass<add_required_reorders>(*prog);
ASSERT_NE(prog, nullptr);
ASSERT_TRUE(prog->has_node("conv1"));
ASSERT_EQ(prog->get_node("conv1").get_dependency(0).get_output_layout(false).data_type, data_types::u8);
}