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

[GPU] Resolve execution failure of oneDNN reduction (#12506)

Browse Source 
+ Resolved failure of oneDNN reduction which has fused output layout.
+ Removed planar format from oneDNN reduction which selected ref kernel.
+ Modified relevant test-cases

Signed-off-by: Min, Byungil <byungil.min@intel.com>
This commit is contained in:
Min, Byungil
2022-09-16 08:59:21 +09:00
committed by GitHub
parent 96fb6ba4a6
commit eaf8364196
3 changed files with 11 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
src/plugins/intel_gpu/src/graph/impls/onednn/reduction_onednn.cpp
View File

@@ -46,14 +46,9 @@ protected:
auto input_layout = input.get_output_layout();
auto output_layout = arg.get_output_layout();
for (size_t idx = 0 ; idx < prim->axes.size(); idx++) {
if (output_layout.get_dim(prim->axes[idx]) != 1) {
// A clDNN Reduce reorders un-reduced axes of its output tensor to b-f and spatial order when keep_dims is false.
// oneDNN reduction does not allow this. So this function reverts it.
reorder_unreduced_axis_no_fusion(input_layout, output_layout, prim->axes);
break;
}
}
// A clDNN Reduce reorders un-reduced axes of its output tensor to b-f and spatial order when keep_dims is false.
// oneDNN reduction does not allow this. So this function reverts it.
reorder_unreduced_axis_no_fusion(input_layout, output_layout, prim->axes);
auto input_md = onednn::layout_to_memory_desc(input_layout);
auto output_md = onednn::layout_to_memory_desc(output_layout);
@@ -105,11 +100,6 @@ namespace detail {
attach_reduction_onednn::attach_reduction_onednn() {
implementation_map<reduce>::add(impl_types::onednn, reduction_onednn::create, {
std::make_tuple(data_types::f32, format::bfyx),
std::make_tuple(data_types::f16, format::bfyx),
std::make_tuple(data_types::u8, format::bfyx),
std::make_tuple(data_types::i8, format::bfyx),
std::make_tuple(data_types::f32, format::b_fs_yx_fsv16),
std::make_tuple(data_types::f16, format::b_fs_yx_fsv16),
std::make_tuple(data_types::u8, format::b_fs_yx_fsv16),

9
src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
View File

@@ -759,7 +759,7 @@ bool layout_optimizer::deconvolution_b_fs_yx_fsv16_opt(layout const &input_layou
return false;
}
static bool is_node_for_onednn(reduce_node const& node) {
static bool is_node_for_onednn(reduce_node const& node, format preferred_format) {
auto& input = node.input();
auto reduce_prim = node.get_primitive();
// oneDNN reduction currently does not support logical_and, logical_or, log_sum and log_sum_exp.
@@ -791,6 +791,11 @@ static bool is_node_for_onednn(reduce_node const& node) {
return false;
}
// oneDNN reduction selects ref kernel for simple formats(bfyx..) which has perf regression with a decent tensor size.
if (format::is_simple_data_format(preferred_format)) {
return false;
}
return true;
}
@@ -1543,7 +1548,7 @@ impl_types layout_optimizer::get_preferred_impl_type(program_node& node, format
if (!_optimization_attributes.use_onednn_impls)
return impl_types::ocl;
if (is_node_for_onednn(node.as<reduce>()))
if (is_node_for_onednn(node.as<reduce>(), preferred_format))
return impl_types::onednn;
else
return impl_types::ocl;

38
src/plugins/intel_gpu/tests/test_cases/reduce_gpu_test.cpp
View File

@@ -1951,21 +1951,13 @@ INSTANTIATE_TEST_SUITE_P(onednn_reduce_gpu_b_fs_yx_fsv16_i8_f32,
TestParamType_general_reduce_gpu(3, 3, 1, 1, 3, 2, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(3, 3, 1, 1, 2, 11, format::b_fs_yx_fsv16, reduce_mode::min, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(26, 12, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(1, 2, 1, 1, 5, 5, format::b_fs_yx_fsv16, reduce_mode::prod, {3, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 26, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(26, 12, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(2, 5, 1, 1, 3, 3, format::b_fs_yx_fsv16, reduce_mode::max, {3, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 26, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
TestParamType_general_reduce_gpu(17, 3, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32)
), general_reduce_gpu::PrintToStringParamName);
@@ -1977,47 +1969,19 @@ INSTANTIATE_TEST_SUITE_P(onednn_reduce_gpu_b_fs_yx_fsv16_i8_f32,
TestParamType_general_reduce_gpu(3, 3, 1, 1, 3, 3, format::b_fs_yx_fsv16, reduce_mode::l1, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(3, 3, 1, 1, 2, 11, format::b_fs_yx_fsv16, reduce_mode::min, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(26, 12, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 12, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::l1, {2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 13, 12, format::b_fs_yx_fsv16, reduce_mode::l2, {0, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 26, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 16, 11, format::b_fs_yx_fsv16, reduce_mode::l1, {1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::l2, {1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 12, 8, format::b_fs_yx_fsv16, reduce_mode::sum_square, {2}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(26, 12, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 12, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::l1, {2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 13, 9, format::b_fs_yx_fsv16, reduce_mode::l2, {0, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(2, 5, 1, 1, 3, 3, format::b_fs_yx_fsv16, reduce_mode::max, {3, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 26, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 34, 1, 1, 16, 15, format::b_fs_yx_fsv16, reduce_mode::l1, {1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 3, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 12, 8, format::b_fs_yx_fsv16, reduce_mode::sum_square, {2}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16)
), general_reduce_gpu::PrintToStringParamName);
INSTANTIATE_TEST_SUITE_P(onednn_reduce_gpu_ref_f16_f16,
onednn_reduce_gpu_f16_f16,
::testing::Values(
TestParamType_general_reduce_gpu(3, 11, 1, 1, 7, 17, format::bfyx, reduce_mode::l1, {1, 0}, "reduce_ref", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(3, 7, 1, 1, 7, 17, format::bfyx, reduce_mode::sum, {0}, "reduce_ref", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(6, 9, 1, 1, 7, 17, format::bfyx, reduce_mode::l2, {1}, "reduce_ref", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(5, 5, 1, 1, 17, 17, format::bfyx, reduce_mode::mean, {2}, "reduce_ref", false, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 12, 4, format::bfyx, reduce_mode::mean, {3, 2, 1, 0}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 4, 1, 1, 12, 12, format::bfyx, reduce_mode::mean, {0, 2, 1}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(4, 3, 1, 1, 7, 12, format::bfyx, reduce_mode::sum, {3, 1, 0}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(16, 16, 1, 1, 4, 11, format::bfyx, reduce_mode::mean, {2, 1}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(2, 16, 1, 1, 3, 6, format::bfyx, reduce_mode::mean, {1, 2}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(7, 4, 1, 1, 3, 6, format::bfyx, reduce_mode::mean, {3}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(7, 32, 1, 1, 3, 6, format::bfyx, reduce_mode::mean, {0}, "reduce_ref", true, data_types::f16, false, data_types::f16),
TestParamType_general_reduce_gpu(17, 4, 1, 1, 12, 15, format::bfyx, reduce_mode::mean, {0}, "reduce_ref", true, data_types::f16, false, data_types::f16)
TestParamType_general_reduce_gpu(17, 3, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::f16, false, data_types::f16)
), general_reduce_gpu::PrintToStringParamName);
#endif // ENABLE_ONEDNN_FOR_GPU