Skip to content

Commit

Permalink
[GPU] Fix to enable fc 5d
Browse files Browse the repository at this point in the history
  • Loading branch information
@steve-y
steve-y committed Oct 8, 2024
1 parent 0c36373 commit 70afb74
Show file tree
Hide file tree
Showing 2 changed files with 19 additions and 25 deletions.
24 changes: 9 additions & 15 deletions src/plugins/intel_gpu/src/graph/fully_connected.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -115,38 +115,32 @@ layout fully_connected_inst::calc_output_layout(fully_connected_node const& node
};

int64_t feature = input_pshape[std::min(desc->input_size, static_cast<size_t>(4)) - 1].get_length();

if (desc->input_size == 3) {
feature = std::max({input_layout.spatial(0), input_layout.spatial(1), input_layout.spatial(2)});
}

if (weights_pshape.size() != 2) {
weights_layout.set_partial_shape(reshape_to_2d(weights_pshape, feature));
}

auto output_size = tensor();

// If immad is supported, spatial dimensions are reshaped to 2d in order to select oneDnn impl,
// because oneDnn doesn't support spatial dimensions for output.
if (supports_immad) {
if (desc->input_size == 3) {
feature = std::max({input_layout.spatial(0), input_layout.spatial(1), input_layout.spatial(2)});
}

if (desc->input_size > 3) {
input_layout.set_partial_shape(reshape_to_2d(input_pshape, feature));
}
if (weights_pshape.size() != 2) {
weights_layout.set_partial_shape(reshape_to_2d(weights_pshape, feature));
}

output_size = tensor(input_layout.batch(), weights_layout.batch(), 1, 1);
if (desc->input_size == 3) {
output_size = tensor(input_layout.batch(), input_layout.feature(), 1, weights_layout.batch());
}
} else {
feature = input_pshape[std::min(desc->input_size, static_cast<size_t>(5)) - 1].get_length();
if (desc->input_size == 3) {
feature = std::max({input_layout.spatial(0), input_layout.spatial(1), input_layout.spatial(2)});
}

if (desc->input_size > 5) {
input_layout.set_partial_shape(reshape_to_2d(input_pshape, feature));
}
if (weights_pshape.size() != 2) {
weights_layout.set_partial_shape(reshape_to_2d(weights_pshape, feature));
}

output_size = tensor(input_layout.batch(), weights_layout.batch(), 1, 1);
if (desc->input_size == 3) {
Expand Down
20 changes: 10 additions & 10 deletions src/plugins/intel_gpu/tests/unit/test_cases/fully_connected_gpu_test.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -461,17 +461,17 @@ TEST(fully_connected_gpu, no_biases_5d_input) {
return;

// Input : 1x8x8x8x12
// Weights: 48x12x1x1
// Weights: 48x12
// Output : 1x8x8x8x48

const int32_t input_b = 1, input_f = 8, input_z = 8, input_y = 8, input_x = 12, // size of the whole input buffer
weight_b = 48, weight_f = 12, weight_y = 1, weight_x = 1; // size of the whole weights buffer
weight_b = 48, weight_f = 12; // size of the whole weights buffer

auto input_prim = engine.allocate_memory({ data_types::f32, format::bfzyx, { input_b, input_f, input_x, input_y, input_z } });
auto weights_prim = engine.allocate_memory({ data_types::f32, format::bfyx, { weight_b, weight_f, weight_x, weight_y } });
auto weights_prim = engine.allocate_memory({ { weight_b, weight_f }, data_types::f32, format::bfyx });

std::vector<float> input_data(input_b * input_f * input_z * input_y * input_x, 0);
std::vector<float> weights_data(weight_b * weight_f * weight_y * weight_x, 0);
std::vector<float> weights_data(weight_b * weight_f, 0);

set_values(input_prim, std::move(input_data));
set_values(weights_prim, std::move(weights_data));
Expand Down Expand Up @@ -501,17 +501,17 @@ TEST(fully_connected_gpu, no_biases_5d_input_immad) {
return;

// Input : 1x8x8x8x12
// Weights: 48x12x1x1
// Weights: 48x12
// Output : 512x48x1x1

const int32_t input_b = 1, input_f = 8, input_z = 8, input_y = 8, input_x = 12, // size of the whole input buffer
weight_b = 48, weight_f = 12, weight_y = 1, weight_x = 1; // size of the whole weights buffer
weight_b = 48, weight_f = 12; // size of the whole weights buffer

auto input_prim = engine.allocate_memory({ data_types::f32, format::bfzyx, { input_b, input_f, input_x, input_y, input_z } });
auto weights_prim = engine.allocate_memory({ data_types::f32, format::bfyx, { weight_b, weight_f, weight_x, weight_y } });
auto weights_prim = engine.allocate_memory({ { weight_b, weight_f }, data_types::f32, format::bfyx });

std::vector<float> input_data(input_b * input_f * input_z * input_y * input_x, 0);
std::vector<float> weights_data(weight_b * weight_f * weight_y * weight_x, 0);
std::vector<float> weights_data(weight_b * weight_f, 0);

set_values(input_prim, std::move(input_data));
set_values(weights_prim, std::move(weights_data));
Expand All @@ -536,8 +536,8 @@ TEST(fully_connected_gpu, no_biases_5d_input_immad) {
auto outputs = network.execute();
ASSERT_EQ(outputs.begin()->second.get_layout().batch(), input_f*input_z*input_y);
ASSERT_EQ(outputs.begin()->second.get_layout().feature(), weight_b);
ASSERT_EQ(outputs.begin()->second.get_layout().spatial(1), weight_y);
ASSERT_EQ(outputs.begin()->second.get_layout().spatial(0), weight_x);
ASSERT_EQ(outputs.begin()->second.get_layout().spatial(1), 1);
ASSERT_EQ(outputs.begin()->second.get_layout().spatial(0), 1);
}

TEST(fully_connected_gpu, xb_f32_batch_1) {
Expand Down

0 comments on commit 70afb74

Please sign in to comment.