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

[GPU] Remove WA to fallback to cldnn (#14147)

Browse Source
This commit is contained in:
Jade Cho
2022-11-28 12:23:32 +09:00
committed by GitHub
parent 003e25b6fe
commit 6bc7cd73cd
3 changed files with 7 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -102,7 +102,10 @@ bool concat_in_place_optimization::match(concatenation_node& node) {
else
continue;
}
is_onednn_impl = true;
// Optimized-out input node is no longer onednn impl.
if (!input->can_be_optimized())
is_onednn_impl = true;
}
}

26
src/plugins/intel_gpu/src/graph/graph_optimizer/reorder_inputs.cpp
View File

@@ -40,10 +40,7 @@ std::map<program_node*, format::type> get_preferred_formats(program& p, layout_o
#ifdef ENABLE_ONEDNN_FOR_GPU
size_t onednn_impls_counter = 0;
size_t all_impls_counter = 0;
const float onednn_min_threshold = 0.09f;
bool should_update_fmt_map = false;
// Calculate onednn kernels number and all kernels number inside the network
for (auto n : p.get_processing_order()) {
if (!n->is_in_data_flow())
@@ -57,22 +54,9 @@ std::map<program_node*, format::type> get_preferred_formats(program& p, layout_o
if (impl == impl_types::onednn)
onednn_impls_counter++;
all_impls_counter++;
}
float onednn_usage_ratio = all_impls_counter ? static_cast<float>(onednn_impls_counter) / static_cast<float>(all_impls_counter) : 0.f;
GPU_DEBUG_IF(debug_config->verbose >= 1) {
GPU_DEBUG_COUT << "----------------------------------------------" << std::endl;
GPU_DEBUG_COUT << "Onednn kernels number: " << onednn_impls_counter << " from " << all_impls_counter
<< " (" << onednn_usage_ratio * 100.f << "%)" << std::endl;
GPU_DEBUG_COUT << "Onednn usage threshold: " << onednn_min_threshold * 100.f << "%" << std::endl;
}
// Reverted to cldnn way for cases when onednn kernels number inside the whole network is extremely low =>
// improvements from onednn usage less than losses due to unoptimized formats for cldnn kernels, extra reorders, etc.
if (onednn_usage_ratio < onednn_min_threshold && lo.get_optimization_attributes().use_onednn_impls) {
if (onednn_impls_counter < 1 && lo.get_optimization_attributes().use_onednn_impls) {
should_update_fmt_map = true;
lo.set_optimization_attribute(layout_optimizer::optimization_attributes_type::use_onednn_impls, 0);
GPU_DEBUG_IF(debug_config->verbose >= 1) {
@@ -80,14 +64,8 @@ std::map<program_node*, format::type> get_preferred_formats(program& p, layout_o
}
}
GPU_DEBUG_IF(debug_config->verbose >= 1) {
GPU_DEBUG_COUT << "----------------------------------------------" << std::endl;
}
#endif // ENABLE_ONEDNN_FOR_GPU
#ifdef ENABLE_ONEDNN_FOR_GPU
if (should_update_fmt_map)
#endif
#endif // ENABLE_ONEDNN_FOR_GPU
{
for (auto n : p.get_processing_order()) {
if (!n->is_in_data_flow())

2
src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
View File

@@ -322,7 +322,7 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
prev.is_input() && (prev_dt == data_types::u8 || prev_dt == data_types::i8))
return true;
if (!use_onednn_impls) {
if (!use_onednn_impls || next.get_preferred_impl_type() == impl_types::ocl) {
if (next.is_type<convolution>() &&
(fmt_prev == format::bfyx || fmt_prev == format::bs_fs_yx_bsv4_fsv2) &&
((fmt_next == format::fs_b_yx_fsv32 && next.as<convolution>().get_primitive()->groups == 1) ||