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

Fix prepare padding which was not handling group size properly (#16977)

Browse Source
This commit is contained in:
Taylor Yeonbok Lee 2023-04-16 21:42:03 -07:00 committed by GitHub
parent 31398bb3eb
commit c96a5c4b70
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 45 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_padding.cpp
View File

@ -51,7 +51,8 @@ void prepare_padding::run(program& p) {
if (!prim->with_output_size) if (!prim->with_output_size)
continue; continue;
auto weights_layout = prim_node.weights().get_output_layout();
tensor filter_size = weights_layout.convert_to_weights_layout(prim->grouped_weights_shape).get_tensor();
auto format = node->get_output_layout().format; auto format = node->get_output_layout().format;
if (format == format::b_fs_zyx_fsv16 || if (format == format::b_fs_zyx_fsv16 ||
format == format::bs_fs_zyx_bsv16_fsv16 || format == format::bs_fs_zyx_bsv16_fsv16 ||
@ -60,8 +61,6 @@ void prepare_padding::run(program& p) {
format == format::b_fs_zyx_fsv32) format == format::b_fs_zyx_fsv32)
continue; continue;
auto filter_size = prim_node.weights().get_output_layout().get_tensor();
auto needed_padding = calc_sliding_window_needed_input_padding(prim_node.input().get_output_layout(), auto needed_padding = calc_sliding_window_needed_input_padding(prim_node.input().get_output_layout(),
prim->output_size, prim->output_size,
filter_size, filter_size,

43
src/plugins/intel_gpu/tests/passes/prepare_padding_test.cpp Normal file
View File

@ -0,0 +1,43 @@
// 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/program.hpp"
#include "data_inst.h"
#include "convolution_inst.h"
#include "pass_manager.h"
#include "program_wrapper.h"
#include <memory>
using namespace cldnn;
using namespace ::tests;
TEST(prepare_padding, groupconv_with_output) {
auto& engine = get_test_engine();
auto in_layout = layout{data_types::f16, format::bfyx, tensor{1, 18, 135, 76}};
auto weight_layout = layout{data_types::f16, format::bfzyx, tensor{1, 18, 3, 3, 18}};
auto weights_data = generate_random_5d<FLOAT16>(1, 18, 18, 3, 3, -1, 1);
auto weights_mem = engine.allocate_memory({ data_types::f16, format::bfzyx, weight_layout.get_tensor()});
set_values(weights_mem, weights_data);
auto output_size = tensor{1, 18, 135, 76};
ov::CoordinateDiff pad = {0, 0};
topology topo;
topo.add(input_layout("input", in_layout));
topo.add(data("weight", weights_mem));
topo.add(convolution("conv", input_info("input"), { "weight" }, {}, 1, {1, 1}, {0, 0}, {1, 1}, output_size, data_types::f16, true));
topo.add(reorder("reorder", input_info("conv"), format::bfyx, data_types::f32));
ExecutionConfig config = get_test_default_config(engine);
config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
auto prog = program::build_program(engine, topo, config, false, true);
reorder_factory rf;
program_wrapper::apply_opt_pass<prepare_padding>(*prog, true);
const auto& node = prog->get_node("reorder_input_conv");
auto params = node.get_kernel_impl_params();
ASSERT_EQ(params->get_output_layout().data_padding.upper_size().spatial[2], 0);
}