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

[GPU] Fix permute reorder fusing bfzyx (#20896)

Browse Source 
* fix accuracy issue of permute reorder fusing

* add func test

* remove debug code

* use unit test

* simplify condition

* add additionall condition
This commit is contained in:
Wilson Seok
2023-11-15 09:33:12 +00:00
committed by GitHub
parent aa1fcbbad1
commit 0180078c70
2 changed files with 36 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
+5
View File
@@ -414,6 +414,7 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
}
bool layout_optimizer::can_fuse_reorder_to_prev(program_node& prev, reorder_node& node, format fmt_prev, format fmt_next) {
bool allow_new_shape_infer = node.get_program().get_config().get_property(ov::intel_gpu::allow_new_shape_infer);
// Because mvn and concatenation kernel can work cross-layout, if reorder only performs type conversion,
// fusing reorder to the previous node can be done even if it is a dynamic shape case
if ((prev.is_type<mvn>() || prev.is_type<concatenation>()) &&
@@ -476,6 +477,10 @@ bool layout_optimizer::can_fuse_reorder_to_prev(program_node& prev, reorder_node
if (fmt_prev.dimension() > 6 || fmt_next.dimension() > 6)
return false;
// Skip reorder fusing to permute when allow_new_shape_infer is True and input and output rank is different
if (allow_new_shape_infer && (fmt_prev.dimension() != fmt_next.dimension()))
return false;
return true;
}
src/plugins/intel_gpu/tests/unit/passes/remove_redundant_reorders_tests.cpp
+31
View File
@@ -219,6 +219,37 @@ TEST(remove_redundant_reorders, not_to_fuse_permute) {
network network(engine, topology, config);
}
TEST(remove_redundant_reorders, not_to_fuse_permute_new_shape_infer) {
auto& engine = get_test_engine();
auto input1 = engine.allocate_memory({data_types::f16, format::bfyx, {4, 64, 512, 512}});
auto input2 = engine.allocate_memory({data_types::f16, format::bfzyx, {1, 4, 64, 512, 512}});
layout output_layout_fp16( data_types::f16, format::bfzyx, { 4, 512, 64, 512 } );
topology topology;
topology.add(input_layout("input1", input1->get_layout()));
topology.add(input_layout("input2", input2->get_layout()));
topology.add(permute("permute", input_info("input1"), {0, 2, 3, 1}));
topology.add(reorder("reorder1", input_info("permute"), output_layout_fp16));
topology.add(reshape("reshape", input_info("reorder1"), false, {}, ov::PartialShape{1, 4, 512, 512, 64}));
topology.add(concatenation("concat", {input_info("reshape"), input_info("input2")}, 4));
ExecutionConfig config = get_test_default_config(engine);
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("input1", input1);
network.set_input_data("input2", input2);
network.execute();
auto prog = network.get_program();
ASSERT_NE(prog, nullptr);
auto& permute_node = prog->get_node("permute");
auto permute_layout = permute_node.get_output_layout();
ASSERT_EQ(permute_layout.format.value, format::bfyx);
}
TEST(remove_redundant_reorders, remove_fused) {
auto& engine = get_test_engine();
layout output_layout_fp16( data_types::f16, format::bfyx, { 1, 3, 2, 2 } );