Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Resolve accuracy issue from fusing prelu #29754

Merged
merged 6 commits into from
Mar 31, 2025
+33 −11
Merged

Resolve accuracy issue from fusing prelu #29754

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
765016d
[GPU] Resolve acc issue by fusing prelu
byungilm Mar 26, 2025
759091a
[GPU] Fix conv fusing logic
byungilm Mar 26, 2025
108a26e
[GPU] Fix comment for prelu fusing
byungilm Mar 26, 2025
1f81115
[GPU] clear CI failure form prelu fusing
byungilm Mar 26, 2025
e613d7f
[GPU] Modified test-code for prelu fusing issue
byungilm Mar 27, 2025
8703f66
[GPU] Apply comment for prelu fusing
byungilm Mar 27, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_primitive_fusing.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -742,6 +742,15 @@ void prepare_primitive_fusing::fuse_simple_primitives(program &p) {
// prelu fusion is not implemented in oneDNN3.1 (CVS-108233)
return;
}

// Fusing prelu to multi batch onednn conv caused an accuracy issue. Blocked fusing of the case.
auto input_layout = input.get_output_layout();
if (input.is_type<convolution>() && (lo.get_preferred_impl_type(input, format::any /*dummy*/) == impl_types::onednn) &&
activation_func == cldnn::activation_func::relu_negative_slope &&
input_layout.is_static() && input_layout.batch() > 1) {
return;
}

// Activation should not be fused if oneDNN does NOT support it
if (lo.is_primitive_implemented_for_onednn(input)) {
#ifdef ENABLE_ONEDNN_FOR_GPU
Expand Down
35 changes: 24 additions & 11 deletions src/plugins/intel_gpu/tests/unit/fusions/convolution_fusion_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -527,6 +527,14 @@ class ConvActivationTestOnednn : public BaseFusingTest<conv_activation_onednn_te
#define CASE_CONV_ELTW_i8_4 { 1, 16, 1, 4 }, { 1, 16, 1, 2 }, { 1, 16, 1, 1 }, { 16, 16, 1, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::i8, format::b_fs_yx_fsv16, data_types::i8, format::os_is_yx_osv16_isv16, data_types::f32, format::bfyx
#define CASE_CONV_ELTW_i8_5 { 1, 16, 1, 4, 1 }, { 1, 16, 1, 2, 1 }, { 1, 16, 2, 1, 1 }, { 16, 16, 1, 3, 1 }, { 1, 1, 1 }, { 0, 0, 0 }, { 1, 1, 1 }, 1, data_types::i8, format::bfzyx, data_types::i8, format::oizyx, data_types::f32, format::bfzyx

#define CASE_MUL_BATCH_CONV_FP16_1 { 20, 15, 4, 5 }, { 20, 30, 2, 3 }, { 30, 15, 3, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::f16, format::bfyx, data_types::f16, format::bfyx, data_types::f16, format::bfyx
#define CASE_MUL_BATCH_CONV_FP16_2 { 32, 16, 5, 5 }, { 32, 32, 3, 3 }, { 32, 16, 3, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::f16, format::bs_fs_yx_bsv16_fsv16, data_types::f16, format::os_is_yx_osv16_isv16, data_types::f16, format::bfyx
#define CASE_MUL_BATCH_CONV_U8S8_1 { 32, 15, 4, 5 }, { 32, 30, 2, 3 }, { 30, 15, 3, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::u8, format::bfyx, data_types::i8, format::bfyx, data_types::f32, format::bfyx
#define CASE_MUL_BATCH_CONV_U8S8_2 { 20, 15, 5, 5 }, { 20, 30, 3, 3 }, { 30, 15, 3, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::u8, format::bs_fs_yx_bsv16_fsv16, data_types::i8, format::os_is_yx_osv16_isv16, data_types::f32, format::bfyx
#define CASE_MUL_BATCH_CONV_S8S8_1 { 32, 15, 4, 5 }, { 32, 30, 2, 3 }, { 30, 15, 3, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::i8, format::bfyx, data_types::i8, format::bfyx, data_types::f32, format::bfyx
#define CASE_MUL_BATCH_CONV_S8S8_2 { 20, 15, 5, 5 }, { 20, 30, 3, 3 }, { 30, 15, 3, 3 }, { 1, 1 }, { 0, 0 }, { 1, 1 }, 1, data_types::i8, format::bs_fs_yx_bsv16_fsv16, data_types::i8, format::os_is_yx_osv16_isv16, data_types::f32, format::bfyx
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

please remove unused macros.



/* ----------------------------------------------------------------------------------------------------- */
/* ---------------------------------------- FP32 convolution cases ------------------------------------- */
/* ----------------------------------------------------------------------------------------------------- */
Expand Down Expand Up @@ -1755,7 +1763,7 @@ TEST_P(conv_fp32_activation_eltwise_in_u8_fp32, basic) {
data("bias", get_mem(get_per_channel_layout(p))),
data("eltwise_data", get_mem(layout{ p.out_shape, data_types::i8, p.input_format })),
convolution("conv_prim", input_info("input"), "weights", "bias", p.groups, p.stride, p.dilation, p.pad, p.pad, format::is_grouped(get_weights_layout(p).format)),
activation("activation", input_info("conv_prim"), activation_func::relu_negative_slope),
activation("activation", input_info("conv_prim"), activation_func::relu),
eltwise("sum", { input_info("activation"), input_info("eltwise_data") }, eltwise_mode::sum, data_types::f32),
reorder("reorder_bfyx", input_info("sum"), p.default_format, data_types::f32)
);
Expand Down Expand Up @@ -3113,6 +3121,11 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, conv_fp16_prelu_onednn, ::testing::ValuesI
convolution_test_params{ CASE_CONV_FP16_1, 2, 2, 4 },
}));

INSTANTIATE_TEST_SUITE_P(fusings_gpu_multi_batch, conv_fp16_prelu_onednn, ::testing::ValuesIn(std::vector<convolution_test_params>{
convolution_test_params{ CASE_MUL_BATCH_CONV_FP16_1, 4, 4, 4 },
convolution_test_params{ CASE_MUL_BATCH_CONV_FP16_2, 4, 4, 4 },
}));

class conv_int8_eltwise_onednn : public WeightsPrimitiveFusingTestOneDNN {};
TEST_P(conv_int8_eltwise_onednn, u8_eltwise_sum_out) {
auto p = GetParam();
Expand Down Expand Up @@ -3807,18 +3820,18 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, post_ops_optimizations_onednn_eltw_any_sum
convolution_test_params{ CASE_CONV_S8S8_3, 2, 2, 5 },

// cases with batch = 16
convolution_test_params{ CASE_CONV_U8S8_10, 2, 2, 5 },
convolution_test_params{ CASE_CONV_S8S8_10, 2, 2, 5 },
convolution_test_params{ CASE_CONV_U8S8_10, 2, 4, 5 },
convolution_test_params{ CASE_CONV_S8S8_10, 2, 4, 5 },

// cases with batch = 32
convolution_test_params{ CASE_CONV_U8S8_11, 2, 2, 5 },
convolution_test_params{ CASE_CONV_U8S8_12, 2, 2, 5 },
convolution_test_params{ CASE_CONV_U8S8_13, 2, 2, 5 },
convolution_test_params{ CASE_CONV_U8S8_14, 2, 2, 5 },
convolution_test_params{ CASE_CONV_S8S8_12, 2, 2, 5 },
convolution_test_params{ CASE_CONV_S8S8_13, 2, 2, 5 },
convolution_test_params{ CASE_CONV_S8S8_14, 2, 2, 5 },
convolution_test_params{ CASE_CONV_S8S8_15, 2, 2, 5 },
convolution_test_params{ CASE_CONV_U8S8_11, 2, 4, 5 },
convolution_test_params{ CASE_CONV_U8S8_12, 2, 4, 5 },
convolution_test_params{ CASE_CONV_U8S8_13, 2, 4, 5 },
convolution_test_params{ CASE_CONV_U8S8_14, 2, 4, 5 },
convolution_test_params{ CASE_CONV_S8S8_12, 2, 4, 5 },
convolution_test_params{ CASE_CONV_S8S8_13, 2, 4, 5 },
convolution_test_params{ CASE_CONV_S8S8_14, 2, 4, 5 },
convolution_test_params{ CASE_CONV_S8S8_15, 2, 4, 5 },
}));

// Input range uses in 2 cases: not per-tensor output range or out_lo > out_hi
Expand Down
Loading