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

[GPU] Forcing to use clDNN FC on small batch size (#11715)

Browse Source 
+ forced to use clDNN FC due to perf drop

Signed-off-by: byungilm <byungil.min@intel.com>
This commit is contained in:
Min, Byungil
2022-06-07 16:25:14 +09:00
committed by GitHub
parent f9afe07c9d
commit 49942c2f80
1 changed files with 51 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
View File

@@ -818,6 +818,44 @@ static bool is_node_for_onednn(deconvolution_node const& node) {
return onednn_valid_dt && onednn_valid_params && spatial_dims_num <= 3;
}
static bool is_node_for_onednn(program_node& node, fully_connected_node const& fc_node) {
bool is_suitable_for_onednn = true;
auto out_layout = node.get_output_layout();
for (auto& fo : node.get_fused_primitives()) {
if (fo.node->is_type<eltwise>()) {
// FC checkings
auto in_layout = node.get_dependency(fo.dep_start_idx).get_output_layout();
auto in_dt = in_layout.data_type;
auto out_dt = out_layout.data_type;
// if it is not eltwise sum and input is full tensor
if ((out_layout.count() == in_layout.count()) && in_dt != out_dt
&& (data_type_traits::is_floating_point(in_dt) || data_type_traits::is_floating_point(out_dt))
&& onednn_add_fusing_helpers::is_full_tensor(in_layout)) {
return false;
}
// WA: onednn sum/binary_add post-op are not supported due to perf drop.
auto add_type = onednn_add_fusing_helpers::get_add_fusing_type(node, fo);
if (add_type == add_fusing_type::sum || add_type == add_fusing_type::binary_per_tensor || add_type == add_fusing_type::binary_per_oc) {
return false;
}
}
}
auto fc_prim = fc_node.get_primitive();
size_t rank = cldnn::format::dimension(out_layout.format);
auto size = out_layout.size;
// OneDnn doesn't support spatial dimensions for output
for (int i = 0; i < rank - 2 - (fc_prim->input_size == 3 ? 1 : 0); i++) {
if (size.spatial[i] != 1) {
return false;
}
}
return is_suitable_for_onednn;
}
bool layout_optimizer::needs_all_usr_onednn_small_ic_to_blocked(const program_node& node) {
bool all_users_match = true;
for (auto usr : node.get_users()) {
@@ -1538,33 +1576,20 @@ impl_types layout_optimizer::get_preferred_impl_type(program_node& node, format
impl_candidate = impl_types::ocl;
}
for (auto& fo : node.get_fused_primitives()) {
if (fo.node->is_type<eltwise>()) {
// FC checkings
if (node.is_type<fully_connected>()) {
auto in_layout = node.get_dependency(fo.dep_start_idx).get_output_layout();
auto out_layout = node.get_output_layout();
auto in_dt = in_layout.data_type;
auto out_dt = out_layout.data_type;
// if it is not eltwise sum and input is full tensor
if ((out_layout.count() == in_layout.count()) && in_dt != out_dt
&& (data_type_traits::is_floating_point(in_dt) || data_type_traits::is_floating_point(out_dt))
&& onednn_add_fusing_helpers::is_full_tensor(in_layout)) {
impl_candidate = impl_types::ocl;
break;
}
if (node.is_type<fully_connected>()) {
if (!is_node_for_onednn(node, node.as<fully_connected>()))
impl_candidate = impl_types::ocl;
// WA: onednn sum/binary_add post-op are not supported due to perf drop.
auto add_type = onednn_add_fusing_helpers::get_add_fusing_type(node, fo);
if (add_type == add_fusing_type::sum || add_type == add_fusing_type::binary_per_tensor || add_type == add_fusing_type::binary_per_oc) {
impl_candidate = impl_types::ocl;
break;
}
// Gemm checkings
// TODO: investigate why currently onednn gemm has some "sum" post-op restrictions
// which don't correlate with fc checkings in the code above
// Temprorary WA: disable onednn gemm with sum post-op inside
} else {
// WA : Use cldnn FC due to perf drop of small batch size until onednn FC improve perf
if (node.get_output_layout().batch() < 32)
impl_candidate = impl_types::ocl;
} else {
for (auto& fo : node.get_fused_primitives()) {
if (fo.node->is_type<eltwise>()) {
// Gemm checkings
// TODO: investigate why currently onednn gemm has some "sum" post-op restrictions
// which don't correlate with fc checkings in the code above
// Temprorary WA: disable onednn gemm with sum post-op inside
auto& e_node = fo.node->as<eltwise>();
if (e_node.get_primitive()->mode == eltwise_mode::sum) {
impl_candidate = impl_types::ocl;
@@ -1572,21 +1597,7 @@ impl_types layout_optimizer::get_preferred_impl_type(program_node& node, format
}
}
}
}
if (node.is_type<fully_connected>()) {
auto fc_prim = node.as<fully_connected>().get_primitive();
auto out_layout = node.get_output_layout();
size_t rank = cldnn::format::dimension(out_layout.format);
auto size = out_layout.size;
// OneDnn doesn't support spatial dimensions for output
for (int i = 0; i < rank - 2 - (fc_prim->input_size == 3 ? 1 : 0); i++) {
if (size.spatial[i] != 1) {
impl_candidate = impl_types::ocl;
break;
}
}
} else {
impl_candidate = impl_types::ocl;
auto gemm_prim = node.as<gemm>().get_primitive();
auto in0_l = node.get_dependency(0).get_output_layout();