Skip to content

[GPU] Enable uint8 weights for oneDNN-based convolution #32650

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
Nov 4, 2025
+69 −3
Merged

[GPU] Enable uint8 weights for oneDNN-based convolution #32650

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
6ec8c65
Enable uint8 weights for oneDNN-based convolution
mklimenk Oct 31, 2025
54638e5
Add uint8 conv test
mklimenk Nov 3, 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
5 changes: 2 additions & 3 deletions src/plugins/intel_gpu/src/graph/impls/onednn/convolution_onednn.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,11 +87,10 @@ struct ConvolutionImplementationManager : public ImplementationManager {
return false;

bool f16_conv = everyone_is(data_types::f16, in_dt, wei_dt) && one_of(out_dt, {data_types::f16, data_types::f32, data_types::u8, data_types::i8});
bool u8s8_conv = one_of(in_dt, {data_types::i8, data_types::u8}) &&
wei_dt == data_types::i8 &&
bool int8_conv = one_of(in_dt, {data_types::i8, data_types::u8}) && one_of(wei_dt, {data_types::i8, data_types::u8}) &&
one_of(out_dt, {data_types::i32, data_types::f16, data_types::f32, data_types::u8, data_types::i8});

if (!f16_conv && !u8s8_conv)
if (!f16_conv && !int8_conv)
Copy link
Contributor

@isanghao isanghao Nov 3, 2025

Choose a reason for hiding this comment

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

could you add a unittest for this case?

Copy link
Contributor Author

@mklimenk mklimenk Nov 3, 2025

Choose a reason for hiding this comment

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

Added in 54638e5

Copy link
Contributor

@michal-miotk michal-miotk Nov 3, 2025

Choose a reason for hiding this comment

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

this int8_conv is little confusing, what about 8bit_conv or sth similar?

Copy link
Contributor

@michal-miotk michal-miotk Nov 3, 2025

Choose a reason for hiding this comment

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

hmm it is good name

return false;

if (!is_supported_post_ops(conv_node))
Expand Down
67 changes: 67 additions & 0 deletions src/plugins/intel_gpu/tests/unit/test_cases/convolution_gpu_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4404,6 +4404,73 @@ TEST(convolution_int8_fw_gpu, quantized_convolution_u8s8f32_symmetric) {
}
}

TEST(convolution_int8_fw_gpu, quantized_convolution_u8u8f32_symmetric) {
auto& engine = get_test_engine();

auto input = engine.allocate_memory({ data_types::u8, format::bfyx, { 1, 1, 5, 4 } });
auto weights = engine.allocate_memory({ data_types::u8, format::bfyx, { 2, 1, 3, 3 } });
auto biases = engine.allocate_memory({ data_types::f32, format::bfyx, { 1, 2, 1, 1 } });

set_values<uint8_t>(input, { 1, 2, 3, 4, 5,
2, 2, 3, 4, 6,
3, 3, 3, 5, 1,
1, 1, 1, 1, 1 });
set_values<uint8_t>(weights, { 1, 2, 1,
2, 1, 2,
9, 7, 1,

9, 0, 4,
1, 3, 2,
0, 2, 5 });
set_values(biases, { -1.0f, -8.0f });

VVVF<float> output_vec = {
{
{ 70.0f, 100.0f, 25.0f },
{ 16.0f, 18.0f, 2.0f }
},
{
{ 48.0f, 81.0f, 43.0f },
{ 37.0f, 29.0f, 2.0f }
} };

topology topology(
input_layout("input", input->get_layout()),
data("weights", weights),
data("biases", biases),
convolution("conv", input_info("input"), "weights", "biases", 1, { 2, 2 }, { 1, 1 }, { 0, 0 }, { 1, 2 }, false),
reorder("out", input_info("conv"), format::bfyx, data_types::f32));

ExecutionConfig config = get_test_default_config(engine);
config.set_property(ov::intel_gpu::optimize_data(true));
network network(engine, topology, config);
network.set_input_data("input", input);

auto outputs = network.execute();
ASSERT_EQ(outputs.begin()->first, "out");

auto output_memory = outputs.at("out").get_memory();
cldnn::mem_lock<float> output_ptr(output_memory, get_test_stream());

auto output_layout = output_memory->get_layout();
int y_size = output_layout.spatial(1);
int x_size = output_layout.spatial(0);
int f_size = output_layout.feature();
int b_size = output_layout.batch();
ASSERT_EQ(output_layout.format, format::bfyx);
ASSERT_EQ(y_size, 2);
ASSERT_EQ(x_size, 3);
ASSERT_EQ(f_size, 2);
ASSERT_EQ(b_size, 1);
for (int f = 0; f < f_size; f++)
for (int y = 0; y < y_size; ++y) {
for (int x = 0; x < x_size; ++x) {
ASSERT_NEAR(output_vec[f][y][x], ((float)output_ptr[f * y_size * x_size + y * x_size + x]), 1e-5f) <<
" x="<<x << " y=" << y << " f=" << f;
}
}
}

TEST(convolution_int8_fw_gpu, quantized_convolution_u8s8f32_asymmetric_weight_and_activations) {
auto& engine = get_test_engine();

Expand Down
Loading