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

[GPU] Convert activation's slope buffer data type for OneDNN fusions (#13444)

Browse Source
This commit is contained in:
Sergey Shlyapnikov 2022-10-14 13:58:35 +04:00 committed by GitHub
parent 0b8f1f8c00
commit 5fea4c3fc3
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 55 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
src/plugins/intel_gpu/src/graph/graph_optimizer/reorder_inputs.cpp
View File

@ -789,4 +789,31 @@ void reorder_inputs::run(program& p, layout_optimizer& lo, reorder_factory& rf)
n->set_preferred_impl_type(preferred_impl);
}
}
// WA for OneDNN PRelu activation fusions: convert activation's slope buffer to expected f32 data type
for (auto& node : p.get_processing_order()) {
if (node->get_preferred_impl_type() == impl_types::onednn) {
auto fused_prims = node->get_fused_primitives();
for (auto& fused_desc : fused_prims) {
if (!fused_desc.is_type<activation>())
continue;
auto activation_desc = fused_desc.typed_desc<activation>();
if (activation_desc->activation_function == cldnn::activation_func::relu_negative_slope &&
!activation_desc->additional_params_input.empty()) {
const auto expected_dt = data_types::f32;
const auto dep_idx = fused_desc.dep_start_idx;
const auto orig_layout = node->get_dependency(dep_idx).get_output_layout();
if (orig_layout.data_type == expected_dt)
continue;
auto new_layout = orig_layout;
new_layout.data_type = expected_dt;
auto new_input = rf.get_reorder(node->get_dependency(dep_idx).id(), orig_layout, new_layout);
if (new_input.first)
p.add_intermediate(new_input.first, *node, dep_idx);
}
}
}
}
}

28
src/plugins/intel_gpu/tests/fusions/convolution_fusion_test.cpp
View File

@ -265,6 +265,10 @@ public:
layout get_per_channel_layout(convolution_test_params& p) {
return layout{ p.default_type, p.default_format, tensor{1, p.out_shape.feature[0], 1, 1} };
}
layout get_prelu_slope_layout(convolution_test_params& p) {
return layout{ p.default_type, p.default_format, tensor{1, p.out_shape.feature[0], 1, 1} };
}
};
#endif // ENABLE_ONEDNN_FOR_GPU
@ -2914,6 +2918,30 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp32_reorder_bfyx_to_fsv32_conv_data_
}));
#ifdef ENABLE_ONEDNN_FOR_GPU
class conv_fp16_prelu_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_fp16_prelu_onednn, basic_activation_eltwise) {
auto p = GetParam();
create_topologies(
input_layout("input", get_input_layout(p)),
data("weights", get_mem(get_weights_layout(p))),
data("bias", get_mem(get_bias_layout(p))),
data("slope_data", get_mem(get_prelu_slope_layout(p))),
data("eltwise_data", get_mem(get_output_layout(p))),
convolution("conv_prim", "input", { "weights" }, { "bias" }, p.groups, p.stride, p.pad, p.dilation),
activation("activation", "conv_prim", "slope_data", activation_func::relu_negative_slope),
eltwise("eltwise", "activation", "eltwise_data", eltwise_mode::sum),
reorder("reorder_bfyx", "eltwise", p.default_format, data_types::f32)
);
tolerance = default_tolerance(p.default_type);
execute(p);
}
INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp16_prelu_onednn, ::testing::ValuesIn(std::vector<convolution_test_params>{
convolution_test_params{ CASE_CONV_FP16_1, 2, 4 },
}));
class conv_int8_eltwise_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_eltwise_onednn, u8_eltwise_sum_out) {
auto p = GetParam();