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

[GPU] No need to add reorder after strided_slice (#19411)

Browse Source
This commit is contained in:
David Nam 2023-09-07 03:43:51 +09:00 committed by GitHub
parent bacb83f8a2
commit cb479f4a5d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 68 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
View File

@ -14,6 +14,7 @@
#include "arg_max_min_inst.h"
#include "shape_of_inst.h"
#include "condition_inst.h"
#include "strided_slice_inst.h"
#include <sstream>
#include "gemm_inst.h"
@ -234,8 +235,9 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
}
}
// Ref kernels are the main for depth_to_space and region_yolo. It can do anything.
if (next.is_type<depth_to_space>() || next.is_type<region_yolo>())
// Ref kernels are the main for depth_to_space and region_yolo and strided_slice. It can do anything.
if (next.is_type<depth_to_space>() || next.is_type<region_yolo>() ||
(next.is_type<strided_slice>() && next.get_preferred_impl_type() != cldnn::impl_types::cpu))
return true;
if (next.is_type<reorder>())

64
src/plugins/intel_gpu/tests/unit/passes/reorder_inputs_test.cpp
View File

@ -15,6 +15,10 @@
#include "gather_inst.h"
#include "border_inst.h"
#include "reshape_inst.h"
#include "strided_slice_inst.h"
#include "batch_to_space_inst.h"
#include "permute_inst.h"
#include "concatenation_inst.h"
#include "pass_manager.h"
#include "to_string_utils.h"
@ -297,6 +301,66 @@ TEST(reorder_inputs, no_add_reorder_infront_of_reshape) {
ASSERT_EQ(reshape_layout_out.get_partial_shape(), expected_out_shape);
}
TEST(reorder_inputs, no_need_of_reorder_for_strided_slice) {
tests::random_generator rg(GET_SUITE_NAME);
auto& engine = get_test_engine();
auto in_layout1 = layout{ ov::PartialShape{1080, 1920, 1, 2}, data_types::f32, format::bfyx };
auto in_layout2 = layout{ ov::PartialShape{4, 540, 960, 1, 1}, data_types::f32, format::bfzyx };
auto data_1 = engine.allocate_memory({ov::PartialShape{5}, data_types::i32, format::bfyx});
auto data_2 = engine.allocate_memory({ov::PartialShape{5}, data_types::i32, format::bfyx});
auto data_3 = engine.allocate_memory({ov::PartialShape{5}, data_types::i32, format::bfyx});
topology topology(
input_layout("input1", in_layout1),
input_layout("input2", in_layout2),
permute("permute1", input_info("input1"), {0, 1, 2, 3}),
batch_to_space("batch_to_space1",
input_info("input2"),
tensor{1, 1, 4, 1, 1},
tensor{0, 0, 1, 0, 0},
tensor{0, 0, 1, 0, 0},
tensor{1, 1080, 1920, 1, 2}),
batch_to_space("batch_to_space2",
input_info("input2"),
tensor{1, 1, 4, 1, 1},
tensor{0, 0, 1, 0, 0},
tensor{0, 0, 1, 0, 0},
tensor{1, 1080, 1920, 1, 2}),
data("data_1", data_1),
data("data_2", data_2),
data("data_3", data_3),
strided_slice("strided_slice1",
input_info("batch_to_space1"),
input_info("data_1"),
input_info("data_2"),
input_info("data_3"),
{}, {}, {}, {}, {}, {1080, 1920, 1, 2}),
strided_slice("strided_slice2",
input_info("batch_to_space2"),
input_info("data_1"),
input_info("data_2"),
input_info("data_3"),
{}, {}, {}, {}, {}, {1080, 1920, 1, 2}),
concatenation("concat", {input_info("permute1"), input_info("strided_slice1"), input_info("strided_slice2")}, 2),
permute("result", input_info("concat"), {3, 0, 1, 2})
);
ExecutionConfig config = get_test_default_config(engine);
config.set_property(ov::intel_gpu::optimize_data(true));
auto program = program::build_program(engine, topology, config, false, true);
layout_optimizer lo(true);
reorder_factory rf;
program_wrapper::apply_opt_pass<reorder_inputs>(*program, lo, rf);
ASSERT_NE(program, nullptr);
auto& result = program->get_node("result");
auto in_order = format::get_default_format(result.get_input_layout(0).get_rank()).order();
auto out_shape = result.get_output_layout(0).get_shape();
ASSERT_EQ(in_order.size(), out_shape.size());
}
// TODO Not yet implemented
//TEST(reorder_inputs, impl_forcing_conv_format_kernel) {
// auto& engine = get_test_engine();