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

[GPU][DG2] Fix deconv unit tests (#14176)

Browse Source 
* Quantize clamp optimization(prepare_quantization.cpp:227) for cldnn is not proper to onednn.
  * cast intermediate variable to output type after ~
  * cldnn: every fused quantization.
  * onednn: all post-ops are processed.
* A little perf regression expected in onednn because clamp is relatively heavier operation than type casting.
* Remove deconv workaround from PR #14139
This commit is contained in:
Felix Dohyun Kim 2022-12-21 11:39:57 +09:00 committed by GitHub
parent 55879bdbee
commit fcd33063be
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 1 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -1185,7 +1185,7 @@ void program_node::init_onednn_primitive_attributes() {
// 4. clamp
{
if (q_param->has_clamp) {
if (q_param->has_clamp || idx < cldnn_post_ops.size() - 1) {
float out_lo = q_param->has_min_clamp ? q_param->out_lo : data_type_traits::min<float>(out_dt);
float out_hi = q_param->has_max_clamp ? q_param->out_hi : data_type_traits::max<float>(out_dt);
post_ops.append_eltwise(1.0f, dnnl::algorithm::eltwise_clip, out_lo, out_hi);

26
src/plugins/intel_gpu/tests/fusions/deconvolution_fusion_test.cpp
View File

@ -475,18 +475,6 @@ TEST_P(deconv_actv_eltw_actv, basic) {
activation("act2", input_info("eltw"), activation_func::relu),
reorder("out", input_info("act2"), p.default_format, data_types::f32)
);
// Waiting MFDNN-8947
std::set<deconv_test_params> onednn_skip = {
deconv_test_params{CASE_DECONV_U8S8_6, 2, 5},
deconv_test_params{CASE_DECONV_U8S8_7, 2, 5},
deconv_test_params{CASE_DECONV_S8S8_6, 2, 5},
deconv_test_params{CASE_DECONV_S8S8_7, 2, 5},
};
if (engine.get_device_info().supports_immad && onednn_skip.count(p)) {
std::cout << "SKIP" << std::endl;
return;
}
// Need much higher tolerance because of deconvolution -> convolution optimization
tolerance = 1.f;
execute(p);
@ -702,20 +690,6 @@ TEST_P(deconv_scale_actv_quant_u8_eltw_scale_actv_quant_i8, basic) {
input_info("out2_lo"), input_info("out2_hi"), 255, data_types::i8),
reorder("out", input_info("quant2"), p.default_format, data_types::f32)
);
// Waiting MFDNN-8947
std::set<deconv_test_params> onednn_skip = {
deconv_test_params{CASE_DECONV_FP16_2, 2, 9},
deconv_test_params{CASE_DECONV_FP16_3, 2, 9},
deconv_test_params{CASE_DECONV_FP16_6, 2, 9},
deconv_test_params{CASE_DECONV_FP16_7, 2, 9},
deconv_test_params{CASE_DECONV_S8S8_6, 2, 9},
deconv_test_params{CASE_DECONV_S8S8_7, 2, 9},
};
if (engine.get_device_info().supports_immad && onednn_skip.count(p)) {
std::cout << "SKIP" << std::endl;
return;
}
tolerance = 2.1f;
execute(p);
}