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

[GPU] Skip remove reorder in case input padding is different from output padding (#21619)

Browse Source
This commit is contained in:
Kelvin Choi
2023-12-15 09:25:46 +09:00
committed by GitHub
parent c01f50e3ff
commit 6f0db3330a
2 changed files with 41 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/plugins/intel_gpu/src/graph/graph_optimizer/remove_redundant_reorders.cpp
View File

@@ -164,7 +164,8 @@ void remove_redundant_reorders::run(program& p) {
bool remove_current = r_node.is_simple_reorder() &&
r_dep_node.get_users().size() == 1 &&
!r_dep_node.is_output() &&
!r_node.get_primitive()->has_surface_input();
!r_node.get_primitive()->has_surface_input() &&
r_node.get_input_layout().data_padding == r_node.get_output_layout().data_padding;
if (remove_dep) {
// for chains like

39
src/plugins/intel_gpu/tests/unit/test_cases/reorder_gpu_test.cpp
View File

@@ -1643,6 +1643,45 @@ TEST(reorder_gpu_opt, basic_remove_redundant)
ASSERT_TRUE(outputs.at("r2").get_memory()->get_layout().format == format::yxfb);
}
TEST(reorder_gpu_opt, remove_redundant_reorder_reorder_with_padding)
{
auto& engine = get_test_engine();
memory::ptr in = engine.allocate_memory({ data_types::f16, format::bfyx, tensor{ 1, 4, 4, 1 } });
layout r2_output(data_types::f32, format::b_fs_yx_fsv16, { 1, 4, 4, 1 });
memory::ptr scale_mem = engine.allocate_memory({ data_types::f32, format::bfyx, tensor{ 1, 1, 1, 1 } });
set_values(scale_mem, { 2.0f });
topology tpl{
input_layout("in", in->get_layout()),
data("scale_data", scale_mem),
eltwise("eltwise", { input_info("in"), input_info("scale_data") }, eltwise_mode::prod),
reorder("r1", input_info("eltwise"), format::bfyx, data_types::f32, std::vector<float>{0, 0, 0, 1}),
reorder("r2", input_info("r1"), r2_output.with_padding(padding({ 0, 0, 1, 1 }, 0.f))),
eltwise("output", { input_info("r2"), input_info("scale_data") }, eltwise_mode::prod)
};
ExecutionConfig config = get_test_default_config(engine);
config.set_property(ov::intel_gpu::optimize_data(true));
network net(engine, tpl, config);
net.set_input_data("in", in);
auto outputs = net.execute();
auto executed_primitives = net.get_executed_primitives();
ASSERT_EQ(executed_primitives.count("r1"), 1);
// r2 would be removed, but the padding value should be remained at the input primitive of r2.
std::vector<int32_t> gt = {0, 0, 1, 1};
auto r1_output_data_padding = net.get_primitive("r1")->get_output_layout().data_padding;
auto upper_padding = r1_output_data_padding.upper_size().sizes();
auto lower_padding = r1_output_data_padding.lower_size().sizes();
for (int32_t i = 0 ; i < 4; i++) {
ASSERT_EQ(upper_padding[i], gt[i]);
ASSERT_EQ(lower_padding[i], gt[i]);
}
}
TEST(reorder_gpu_opt, remove_redundant_activation_fuse)
{
auto& engine = get_test_engine();