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

[GPU] Add condition to check deconv with b_fs_yx_fsv16 opt (#12745)

Browse Source 
Signed-off-by: Andrew Park <andrew.park@intel.com>
This commit is contained in:
Andrew Kwangwoong Park 2022-08-26 14:08:44 +09:00 committed by GitHub
parent 3bc7ce1d04
commit f55777ff1a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 6 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/plugins/intel_gpu/src/graph/program.cpp
View File

@ -1335,6 +1335,7 @@ void program::set_layout_optimizer_attributes(layout_optimizer& lo) {
size_t total_1x1_fm_conv_layers = 0;
size_t total_grouped_conv_layers = 0;
size_t opt_deconv_layers_b_fs_zyx_fsv16 = 0;
size_t opt_deconv_layers_b_fs_yx_fsv16 = 0;
size_t total_crop_layers = 0;
for (auto& node : get_processing_order()) {
@ -1370,6 +1371,8 @@ void program::set_layout_optimizer_attributes(layout_optimizer& lo) {
if (prim.type() == cldnn::deconvolution::type_id()) {
if (lo.is_format_optimized(prim.as<deconvolution>(), format::b_fs_zyx_fsv16))
opt_deconv_layers_b_fs_zyx_fsv16 += 1;
else if (lo.is_format_supported(prim.as<deconvolution>(), format::b_fs_yx_fsv16))
opt_deconv_layers_b_fs_yx_fsv16 += 1;
}
// list of layers that do not support yxfb or perform worse than bfyx
@ -1456,6 +1459,8 @@ void program::set_layout_optimizer_attributes(layout_optimizer& lo) {
// Due to fact that single winograd convolution is faster than b_fs_yx_fsv16 and
// using them together leads do redundant reorders, whole topology switch
// will be performed if at least half of layers can use b_fs_yx_fsv16.
// b_fs_yx_fsv16 deconv is faster than bfyx deconv with winograd convolution together,
// whole topology switch will be perform if at lease one layer can use b_fs_yx_fsv16.
// Crop layers are poorly optimized in fsv16 layout so whole topology stays in bfyx
// if there are many crops (2x more then b_fs_yx_fsv16 convolutions)
const float cond_denom = total_conv_layers > 0 ? 1.0f / static_cast<float>(total_conv_layers) : 1.0f;
@ -1464,7 +1469,7 @@ void program::set_layout_optimizer_attributes(layout_optimizer& lo) {
bool should_use_b_fs_yx_fsv16_conv = is_quantized_int8_model ||
(can_use_fsv16 &&
total_conv_layers > 11 &&
num_of_conv_b_fs_yx_fsv16 * cond_denom > 0.5f &&
(num_of_conv_b_fs_yx_fsv16 * cond_denom > 0.5f || opt_deconv_layers_b_fs_yx_fsv16 >= 1) &&
num_of_conv_b_fs_yx_fsv16 * 2 > total_crop_layers);
bool should_use_fs_b_yx_fsv32_conv = total_conv_layers > 11 &&