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

[GPU] Prevent concat different formats for onednn #28668

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
+160 −0
Open

[GPU] Prevent concat different formats for onednn #28668

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 8 additions & 0 deletions src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,9 +53,14 @@ struct ConcatenationImplementationManager : public ImplementationManager {
if (out_layout.data_padding)
return false;

std::vector<format::type> all_dep_types;

bool any_dep_is_onednn = false;
for (const auto& dep : node.get_dependencies()) {
const auto& in_layout = dep.first->get_output_layout(false, dep.second);

all_dep_types.push_back(in_layout.format.value);

if (!one_of(in_layout.data_type, supported_types))
return false;

Expand All @@ -69,6 +74,9 @@ struct ConcatenationImplementationManager : public ImplementationManager {
any_dep_is_onednn = true;
}

if (std::adjacent_find(all_dep_types.begin(), all_dep_types.end(), std::not_equal_to<>() ) != all_dep_types.end())
return false;

if (!any_dep_is_onednn && format::is_simple_data_format(out_layout.format))
return false;

Expand Down
152 changes: 152 additions & 0 deletions src/plugins/intel_gpu/tests/unit/test_cases/concatenation_gpu_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1725,6 +1725,158 @@ INSTANTIATE_TEST_SUITE_P(smoke,
),
concat_gpu::PrintToStringParamName);


template <typename Type>
struct concat_gpu_4d_implicit_mix_types_onednn : public concat_gpu {
public:
cldnn::memory::ptr run_concat_network(std::vector<std::vector<std::vector<std::vector<std::vector<Type>>>>> input, format::type fmt, ExecutionConfig config) {
auto data_type = ov::element::from<Type>();
auto& engine = get_test_engine();
const size_t batch_num = testing::get<0>(GetParam());
const std::vector<size_t> in_features = testing::get<1>(GetParam());
const size_t input_y = testing::get<2>(GetParam());
const size_t input_x = testing::get<3>(GetParam());
size_t output_f = 0;
for (auto& f : in_features)
output_f += f;

topology topology;

std::vector<memory::ptr> in_memory;
std::vector<primitive_id> input_ids;
std::vector<input_info> pooling_ids;

for (size_t i = 0; i < in_features.size(); i++) {
auto size = tensor(static_cast<int32_t>(batch_num),
static_cast<int32_t>(in_features[i]),
static_cast<int32_t>(input_x),
static_cast<int32_t>(input_y));
auto data = input[i];
auto in_lay = layout(data_type, fmt, size);
auto data_flat = std::vector<Type>(in_lay.get_linear_size(), 0);

for (size_t bi = 0; bi < batch_num; ++bi) {
for (size_t fi = 0; fi < in_features[i]; ++fi) {
for (size_t yi = 0; yi < input_y; ++yi) {
for (size_t xi = 0; xi < input_x; ++xi) {
auto coords = tensor(batch(bi), feature(fi), spatial(xi, yi, 0, 0));
auto in_offset = in_lay.get_linear_offset(coords);
data_flat[in_offset] = data[bi][fi][yi][xi];
}
}
}
}

auto in_mem = engine.allocate_memory(in_lay);
set_values(in_mem, data_flat);
in_memory.push_back(in_mem);

topology.add(input_layout("input" + std::to_string(i), in_lay));
topology.add(pooling("pool" + std::to_string(i), input_info("input" + std::to_string(i)), pooling_mode::max, {1, 1}, {1, 1}));

input_ids.push_back("input" + std::to_string(i));
pooling_ids.push_back(input_info("pool" + std::to_string(i)));
}

topology.add(concatenation("concat", pooling_ids, 1));
auto weights_lay = cldnn::layout(data_type, cldnn::format::bfyx, tensor(batch(output_f), feature(output_f)));
auto weights_mem = engine.allocate_memory(weights_lay);
auto& stream = get_test_stream();
weights_mem->fill(stream);
stream.finish();
{
cldnn::mem_lock<Type> weights_ptr(weights_mem, stream);
for (size_t fi = 0; fi < output_f; ++fi) {
auto coords = tensor(batch(fi), feature(fi), spatial(0, 0, 0, 0));
auto offset = weights_lay.get_linear_offset(coords);
weights_ptr[offset] = static_cast<Type>(1.f);
}
}

std::vector<input_info> concat_ids;

concat_ids.push_back(input_info("input1"));
concat_ids.push_back(input_info("pool_final"));

topology.add(data("weights" , weights_mem));
topology.add(convolution("conv", input_info("concat"), "weights", "", 1, {1, 1}, {1, 1}, {0, 0}, {0, 0}, false));
topology.add(pooling("pool_final", input_info("conv"), pooling_mode::max, {1, 1}, {1, 1}));
topology.add(concatenation("concat_final", concat_ids, 1));
topology.add(reorder("reorder", input_info("concat_final"), layout(data_type, format::byxf, {(int32_t)batch_num, (int32_t)output_f, (int32_t)input_y, (int32_t)input_x})));

network concat_network(engine, topology, config);
for (size_t i = 0; i < in_features.size(); i++) {
concat_network.set_input_data(input_ids[i], in_memory[i]);
}
auto outputs = concat_network.execute();

bool concat_opt_enabled = config.get_property(ov::intel_gpu::optimize_data);
bool concat_opt_result = std::static_pointer_cast<concatenation_inst>(concat_network.get_primitive("concat_final"))->node->can_be_optimized();
EXPECT_EQ(concat_opt_enabled, concat_opt_result);

return outputs.at("reorder").get_memory();
}

std::vector<std::vector<std::vector<std::vector<std::vector<Type>>>>> generate_input() {
const size_t batch_num = testing::get<0>(GetParam());
const std::vector<size_t> in_features = testing::get<1>(GetParam());
const size_t input_y = testing::get<2>(GetParam());
const size_t input_x = testing::get<3>(GetParam());

std::vector<std::vector<std::vector<std::vector<std::vector<Type>>>>> input(in_features.size());
for (size_t i = 0; i < in_features.size(); ++i) {
input[i] = rg.generate_random_4d<Type>(batch_num, in_features[i], input_y, input_x, -10, 10);
}
return input;
}

void test(format::type fmt) {
auto& engine = get_test_engine();
auto& stream = get_test_stream();
if (!engine.get_device_info().supports_immad) {
// This case is only for device that uses onednn.
return;
}
auto input = generate_input();

// implicit concat
ExecutionConfig config1 = get_test_default_config(engine);
config1.set_property(ov::intel_gpu::optimize_data(true));
ov::intel_gpu::ImplementationDesc impl = { format::bfyx, std::string(""), impl_types::onednn };
config1.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ {"conv", impl} }));

auto out_mem1 = run_concat_network(input, fmt, config1);
cldnn::mem_lock<Type> out_ptr1(out_mem1, stream);

// explicit concat
ExecutionConfig config2 = get_test_default_config(engine);
config2.set_property(ov::intel_gpu::optimize_data(false));
auto out_mem2 = run_concat_network(input, fmt, config2);
cldnn::mem_lock<Type> out_ptr2(out_mem2, stream);

ASSERT_EQ(out_ptr1.size(), out_ptr2.size());
size_t diff_count = 0;
for (size_t i = 0; i < out_ptr1.size(); ++i) {
if (out_ptr1[i] != out_ptr2[i]) diff_count++;
}
ASSERT_EQ(diff_count, 0);
}
};

using concat_implicit_gpu_onednn_4d_mix_i8 = concat_gpu_4d_implicit_mix_types_onednn<ov::float16>;

TEST_P(concat_implicit_gpu_onednn_4d_mix_i8, input_order_opt_b_fs_yx_fsv32) {
ASSERT_NO_FATAL_FAILURE(test(format::b_fs_yx_fsv32));
}

INSTANTIATE_TEST_SUITE_P(smoke,
concat_implicit_gpu_onednn_4d_mix_i8,
::testing::Values(
TestParamType_concat(1, { 8, 32 }, 2, 2, false)
),
concat_gpu::PrintToStringParamName);


template <typename Type>
struct concat_gpu_4d_explicit : public concat_gpu {
public:
Expand Down
Loading