Skip to content

[GPU] Fix MVN fusion accuracy issue #30416

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

Merged
merged 2 commits into from
Jun 19, 2025
Merged

[GPU] Fix MVN fusion accuracy issue #30416

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dd8d533
[GPU] Fix the out of fused ops index issue for mvn
ahnyoung-paul Jun 11, 2025
979e324
Fix CI failures
ahnyoung-paul Jun 13, 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
14 changes: 13 additions & 1 deletion src/plugins/intel_gpu/src/kernel_selector/kernels/mvn/mvn_kernel_bfyx_opt.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,7 +135,19 @@ JitConstants MVNKernelBfyxOpt::GetJitConstants(const mvn_params& params, MVNKern
"((in_data_set_idx + iteration_in_data_set_offset) % OUTPUT_SIZE_X)" };
}
}
auto conf = FusedOpsConfiguration("", idx_order, "result", activation_dt, 1, LoadType::LT_UNALIGNED, BoundaryCheck::DISABLED);

auto boundary_check = BoundaryCheck::DISABLED;
if (params.has_dynamic_tensors()) {
boundary_check = BoundaryCheck::ENABLED;
} else {
for (const auto& fused_op : params.fused_ops) {
if (!fused_op.output_tensor.SameDims(params.outputs[0])) {
boundary_check = BoundaryCheck::ENABLED;
break;
}
}
}
auto conf = FusedOpsConfiguration("", idx_order, "result", activation_dt, 1, LoadType::LT_UNALIGNED, boundary_check);
jit.Merge(MakeFusedOpsJitConstants(params, { conf }));
}

Expand Down
71 changes: 71 additions & 0 deletions src/plugins/intel_gpu/tests/unit/test_cases/mvn_gpu_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -973,4 +973,75 @@ TEST_P(mvn_random_test, random_cached) {
TEST_P(mvn_random_test_bsv32, random_cached) {
this->execute(GetParam(), true);
}

TEST(mvn_bfyx_opt_fused_ops, basic_fused) {
tests::random_generator rg(GET_SUITE_NAME);
auto& engine = get_test_engine();

auto input_dyn_layout = layout{ov::PartialShape{-1,-1,2048}, data_types::f16, format::bfyx};
auto input0_layout = layout{ov::PartialShape{2,990,2048}, data_types::f16, format::bfyx};
auto input1_layout = layout{ov::PartialShape{2,1,2048}, data_types::f16, format::bfyx};

topology topo;
topo.add(input_layout("input0", input_dyn_layout));
topo.add(input_layout("input1", input_dyn_layout));
topo.add(input_layout("input2", input_dyn_layout));
topo.add(mvn("mvn", input_info("input0"), true, 1e-10f, true, {2}));
topo.add(eltwise("mul", { input_info("mvn"), input_info("input1") }, eltwise_mode::prod, data_types::f16));
topo.add(eltwise("add", { input_info("mul"), input_info("input2") }, eltwise_mode::sum, data_types::f16));
topo.add(reorder("reorder", input_info("add"), format::bfyx, data_types::f16));

auto input0 = engine.allocate_memory(input0_layout);
auto input1 = engine.allocate_memory(input1_layout);
auto input2 = engine.allocate_memory(input1_layout);

std::vector<ov::float16> input0_data = rg.generate_random_1d<ov::float16>(input0->get_layout().count(), -1, 1, 128);
std::vector<ov::float16> input1_data = rg.generate_random_1d<ov::float16>(input1->get_layout().count(), -1, 1, 128);
std::vector<ov::float16> input2_data = rg.generate_random_1d<ov::float16>(input2->get_layout().count(), -1, 1, 128);
set_values(input0, input0_data);
set_values(input1, input1_data);
set_values(input2, input2_data);

ExecutionConfig config_fused = get_test_default_config(engine);
config_fused.set_property(ov::intel_gpu::optimize_data(true));
config_fused.set_property(ov::intel_gpu::allow_new_shape_infer(true));
config_fused.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ {"mvn", {format::type::bfyx, "mvn_gpu_bfyx_opt"}} }));

network network_fused(engine, topo, config_fused);
network_fused.set_input_data("input0", input0);
network_fused.set_input_data("input1", input1);
network_fused.set_input_data("input2", input2);

auto outputs_fused = network_fused.execute();
ASSERT_EQ(outputs_fused.size(), size_t(1));
ASSERT_EQ(outputs_fused.begin()->first, "reorder");
auto output_fused = outputs_fused.begin()->second.get_memory();

ExecutionConfig config_unfused = get_test_default_config(engine);
config_unfused.set_property(ov::intel_gpu::optimize_data(false));
config_unfused.set_property(ov::intel_gpu::allow_new_shape_infer(true));

network network_unfused(engine, topo, config_unfused);
network_unfused.set_input_data("input0", input0);
network_unfused.set_input_data("input1", input1);
network_unfused.set_input_data("input2", input2);

auto outputs_unfused = network_unfused.execute();
ASSERT_EQ(outputs_unfused.size(), size_t(1));
ASSERT_EQ(outputs_unfused.begin()->first, "reorder");
auto output_unfused = outputs_unfused.begin()->second.get_memory();

ASSERT_NE(network_fused.get_executed_primitive_ids().size(), network_unfused.get_executed_primitive_ids().size());
ASSERT_EQ(output_fused->get_layout(), output_unfused->get_layout());

cldnn::mem_lock<ov::float16, mem_lock_type::read> buff_fused(output_fused, get_test_stream());
cldnn::mem_lock<ov::float16, mem_lock_type::read> buff_unfused(output_unfused, get_test_stream());
float tolerance = 0.002f;
for (size_t i = 0; i < output_fused->get_layout().count(); ++i) {
ASSERT_NEAR(buff_fused[i], buff_unfused[i], tolerance) << " at index: " << i
<< " fused: " << static_cast<float>(buff_fused[i])
<< " unfused: " << static_cast<float>(buff_unfused[i]);
}
}

#endif
Loading