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] Skip Depth To Space fusing when dynamic shape and skip broadcastable check in select typed_primitive_inst() when new shape infer #23270

Merged
merged 3 commits into from
Mar 7, 2024
Merged

[GPU] Skip Depth To Space fusing when dynamic shape and skip broadcastable check in select typed_primitive_inst() when new shape infer #23270

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
db1e693
skip dts fusing when dynamic shape and skip broadcastable check in ty…
wilson-seok Mar 5, 2024
13d7b06
fix unintended comment
wilson-seok Mar 5, 2024
5384cd9
remove debugging code
wilson-seok Mar 7, 2024
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
2 changes: 2 additions & 0 deletions src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_primitive_fusing.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -536,6 +536,8 @@ void prepare_primitive_fusing::fuse_simple_primitives(program &p) {
bool input_conv = node.get_dependency(0).is_type<convolution>();
bool out_eltw = node.get_users().front()->is_type<eltwise>();
if (input_conv && out_eltw) {
if (node.is_dynamic())
return false;
auto& eltw = static_cast<const eltwise&>(*node.get_users().front()->get_primitive());
auto& conv = node.get_dependency(0).as<convolution>();
auto eltw_mode = eltw.mode == eltwise_mode::sum;
Expand Down
88 changes: 47 additions & 41 deletions src/plugins/intel_gpu/src/graph/select.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,49 +92,55 @@ select_inst::typed_primitive_inst(network& network, select_node const& node) : p
3,
"");

if (node.get_primitive()->broadcast_spec.m_type == ov::op::AutoBroadcastType::NONE) {
CLDNN_ERROR_LAYOUT_MISMATCH(node.id(),
"Positive input layout",
deps[1].first->get_output_layout(),
"Negative input layout",
deps[2].first->get_output_layout(),
"");

CLDNN_ERROR_NOT_EQUAL(node.id(),
"Mask size",
deps[0].first->get_output_layout().get_tensor(),
"Positive input format",
deps[1].first->get_output_layout().get_tensor(),
"");
} else if (node.get_primitive()->broadcast_spec.m_type == ov::op::AutoBroadcastType::NUMPY) {
CLDNN_ERROR_DATA_TYPES_MISMATCH(node.id(),
"Positive input data type",
deps[1].first->get_output_layout().data_type,
"Negative input data type",
deps[2].first->get_output_layout().data_type,
"");

auto dep1_size = deps[1].first->get_output_layout().get_tensor();
auto dep2_size = deps[2].first->get_output_layout().get_tensor();
cldnn::tensor output_tensor = tensor::max(dep1_size, dep2_size);
// Cond input0 also can be broadcasted.
auto dep0_size = deps[0].first->get_output_layout().get_tensor();
output_tensor = tensor::max(dep0_size, output_tensor);

auto max_dim_count = output_tensor.raw.size();

for (size_t i = 0; i < deps.size(); i++) {
for (size_t d = 0; d < max_dim_count; d++) {
auto current_dim = deps[i].first->get_output_layout().get_tensor().raw[d];

CLDNN_ERROR_BOOL(node.id(),
"Sizes equal or broadcast is possible",
!(current_dim == output_tensor.raw[d] || current_dim == 1),
"Invalid input shapes");
bool allow_new_shape_infer = network.get_program()->get_config().get_property(ov::intel_gpu::allow_new_shape_infer);
// Broadcast check is performed in ngraph shape infer of select when allow_new_shape_infer=true
if (!allow_new_shape_infer) {
if (node.get_primitive()->broadcast_spec.m_type == ov::op::AutoBroadcastType::NONE) {
CLDNN_ERROR_LAYOUT_MISMATCH(node.id(),
"Positive input layout",
deps[1].first->get_output_layout(),
"Negative input layout",
deps[2].first->get_output_layout(),
"");

CLDNN_ERROR_NOT_EQUAL(node.id(),
"Mask size",
deps[0].first->get_output_layout().get_tensor(),
"Positive input format",
deps[1].first->get_output_layout().get_tensor(),
"");
} else if (node.get_primitive()->broadcast_spec.m_type == ov::op::AutoBroadcastType::NUMPY) {
CLDNN_ERROR_DATA_TYPES_MISMATCH(node.id(),
"Positive input data type",
deps[1].first->get_output_layout().data_type,
"Negative input data type",
deps[2].first->get_output_layout().data_type,
"");

auto dep1_size = deps[1].first->get_output_layout().get_tensor();
auto dep2_size = deps[2].first->get_output_layout().get_tensor();
cldnn::tensor output_tensor = tensor::max(dep1_size, dep2_size);
// Cond input0 also can be broadcasted.
auto dep0_size = deps[0].first->get_output_layout().get_tensor();
output_tensor = tensor::max(dep0_size, output_tensor);

auto max_dim_count = output_tensor.raw.size();

for (size_t i = 0; i < deps.size(); i++) {
for (size_t d = 0; d < max_dim_count; d++) {
auto current_dim = deps[i].first->get_output_layout().get_tensor().raw[d];

if (!(current_dim == output_tensor.raw[d] || current_dim == 1))
std::cout << "error!" << std::endl;
Copy link
Contributor

Choose a reason for hiding this comment

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

please remove this line

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Ooops!.My mistake! Updated! Thanks!

CLDNN_ERROR_BOOL(node.id(),
"Sizes equal or broadcast is possible",
!(current_dim == output_tensor.raw[d] || current_dim == 1),
"Invalid input shapes");
}
}
} else {
CLDNN_ERROR_MESSAGE(node.id(), "Unsupported broadcast_type: " + std::to_string(static_cast<int>(node.get_primitive()->broadcast_spec.m_type)));
}
} else {
CLDNN_ERROR_MESSAGE(node.id(), "Unsupported broadcast_type: " + std::to_string(static_cast<int>(node.get_primitive()->broadcast_spec.m_type)));
}
}
} // namespace cldnn
5 changes: 5 additions & 0 deletions src/plugins/intel_gpu/tests/functional/single_layer_tests/dynamic/select.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,6 +120,11 @@ const std::vector<std::vector<InputShape>> inShapesDynamicNumpy = {
{ { -1, -1, -1}, {{ 4, 5, 6}} },
{ { -1, -1}, {{ 5, 6}} }
},
{
{ { -1}, {{ 130048}} },
{ { -1, -1}, {{ 2, 130048}} },
{ { -1, -1}, {{ 2, 130048}} }
},
};

const auto numpyCases = ::testing::Combine(
Expand Down
40 changes: 40 additions & 0 deletions src/plugins/intel_gpu/tests/unit/passes/prepare_primitive_fusing_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
#include "fully_connected_inst.h"
#include "gemm_inst.h"
#include "convolution_inst.h"
#include "depth_to_space_inst.h"
#include "pass_manager.h"
#include "to_string_utils.h"

Expand Down Expand Up @@ -648,4 +649,43 @@ TEST(prepare_primitive_fusing, can_profiling_data_when_fuse_illegal) {
auto output = net.execute();
for (auto& iter : output)
ASSERT_NE(iter.second.get_event(), nullptr);
}

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

auto in_layout = layout{ ov::PartialShape{-1, -1, -1, -1}, data_types::f32, format::bfyx };
auto weight_layout = layout{ ov::PartialShape{32, 32, 3, 3}, data_types::f32, format::bfyx};
auto weight_mem = engine.allocate_memory(weight_layout);
auto weight_data = rg.generate_random_4d<ov::float16>(32, 32, 3, 3, -1, 1);
set_values(weight_mem, weight_data);
auto scale_layout = layout{ ov::PartialShape{1, 2, 1, 1}, data_types::f32, format::bfyx };
auto scale_mem = engine.allocate_memory(scale_layout);
auto elt_layout = layout{ ov::PartialShape{1, 2, 32, 32}, data_types::f32, format::bfyx };
auto elt_mem = engine.allocate_memory(elt_layout);

topology topology;

topology.add(data("weights", weight_mem));
topology.add(input_layout("input", in_layout));
topology.add(convolution("conv", input_info("input"), "weights", "", 1, {1, 1}, {1, 1}, {0, 0}, {0, 0}, false));
topology.add(depth_to_space("depth_to_space", input_info("conv"), 4, depth_to_space_mode::blocks_first));
topology.add(data("scale1_data", scale_mem));
topology.add(eltwise("scale1", { input_info("depth_to_space"), input_info("scale1_data") }, eltwise_mode::prod, data_types::f32));
topology.add(activation("actv1", input_info("scale1"), activation_func::relu));
topology.add(data("eltw_data", elt_mem));
topology.add(eltwise("eltw", { input_info("actv1"), input_info("eltw_data") }, eltwise_mode::sum, data_types::f32));
topology.add(reorder("reorder_bfyx", input_info("eltw"), format::bfyx, data_types::f32));

ExecutionConfig config = get_test_default_config(engine);
config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
auto prog = program::build_program(engine, topology, config, false, true);

layout_optimizer lo(true);

program_wrapper::apply_opt_pass<prepare_primitive_fusing>(*prog, lo);

ASSERT_NE(prog, nullptr);
ASSERT_TRUE(has_node(*prog, "scale1"));
}
Loading