IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Fix onednn implicit concat issue with reorder as input. (#18180)

Browse Source 
* [GPU] Fix onednn implicit concat issue with reorder as input.

Fix for missed memory offset handling in onednn reorder.

Signed-off-by: hyunback <hyunback.kim@intel.com>
This commit is contained in:
hyunback kim 2023-06-23 19:46:50 +09:00 committed by GitHub
parent 54e969012d
commit 3c378eb7ac
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 69 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/plugins/intel_gpu/src/graph/impls/onednn/reorder_onednn.cpp
View File

@ -32,12 +32,14 @@ protected:
int input_idx = DNNL_ARG_FROM;
for (size_t i = 0; i < instance.inputs_memory_count(); i++) {
auto& input = instance.input_memory(i);
args.insert({input_idx++, input.get_onednn_memory(_pd.src_desc())});
auto offset = onednn::get_f_offset(instance.get_input_layout(i), _pd.dnnl::primitive_desc_base::src_desc(i));
args.insert({input_idx++, input.get_onednn_memory(_pd.dnnl::primitive_desc_base::src_desc(static_cast<uint8_t>(i)), offset)});
}
{
auto& output = instance.output_memory();
args.insert({DNNL_ARG_TO, output.get_onednn_memory(_pd.dst_desc())});
auto offset = onednn::get_f_offset(instance.get_output_layout(), _pd.dnnl::primitive_desc_base::dst_desc(0));
args.insert({DNNL_ARG_DST, output.get_onednn_memory(_pd.dnnl::primitive_desc_base::dst_desc(0), offset)});
}
return args;

65
src/plugins/intel_gpu/tests/unit/passes/prepare_buffer_fusing_test.cpp
View File

@ -363,3 +363,68 @@ TEST(prepare_buffer_fusing, crop_b_axis) {
}
}
}
#ifdef ENABLE_ONEDNN_FOR_GPU
TEST(prepare_buffer_fusing, in_place_onednn_concat_static) {
auto& engine = get_test_engine();
if (!engine.get_device_info().supports_immad)
return;
auto in_layout1 = layout{ ov::PartialShape{1, 1, 4, 2}, data_types::f32, format::bfyx };
auto in_layout2 = layout{ ov::PartialShape{1, 2, 4, 2}, data_types::f32, format::bfyx };
topology topology;
topology.add(input_layout("input1", in_layout1));
topology.add(input_layout("input2", in_layout2));
topology.add(reorder("reorder1", input_info("input1"), format::bfyx, data_types::f16));
topology.add(reorder("reorder2", input_info("input2"), format::bfyx, data_types::f16));
topology.add(concatenation("concat", { input_info("reorder1"), input_info("reorder2") }, 1));
topology.add(reorder("output", input_info("concat"), format::bfyx, data_types::f32));
ExecutionConfig config = get_test_default_config(engine);
config.set_property(ov::intel_gpu::optimize_data(true));
config.set_property(ov::intel_gpu::allow_new_shape_infer(false));
network network(engine, topology, config);
auto input_memory1 = engine.allocate_memory(in_layout1);
auto input_memory2 = engine.allocate_memory(in_layout2);
set_values<float>(input_memory1,
{1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f});
set_values<float>(input_memory2,
{11.f, 22.f, 33.f, 44.f, 55.f, 66.f, 77.f, 88.f,
111.f, 222.f, 333.f, 444.f, 555.f, 666.f, 777.f, 888.f});
network.set_input_data("input1", input_memory1);
network.set_input_data("input2", input_memory2);
std::vector<float> ref_output = {
1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f,
11.f, 22.f, 33.f, 44.f, 55.f, 66.f, 77.f, 88.f,
111.f, 222.f, 333.f, 444.f, 555.f, 666.f, 777.f, 888.f};
std::map<cldnn::primitive_id, cldnn::network_output> output;
EXPECT_NO_THROW(output = network.execute());
auto out_l = network.get_output_layout("output");
auto out_mem = output.at("output").get_memory();
cldnn::mem_lock<float> output_ptr(out_mem, get_test_stream());
cldnn::mem_lock<float> input1_ptr(input_memory1, get_test_stream());
cldnn::mem_lock<float> input2_ptr(input_memory2, get_test_stream());
const auto& concat_node_n = network.get_primitive("concat")->get_node();
auto concat_mem = network.get_primitive("concat")->output_memory_ptr();
auto reorder1_mem = network.get_primitive("reorder1")->output_memory_ptr();
auto reorder2_mem = network.get_primitive("reorder2")->output_memory_ptr();
ASSERT_EQ(concat_mem.get(), reorder1_mem.get());
ASSERT_EQ(concat_mem.get(), reorder2_mem.get());
ASSERT_TRUE(concat_node_n.can_be_optimized());
for (size_t x = 0; x < out_l.count(); ++x) {
ASSERT_EQ(ref_output[x], output_ptr[x]);
}
}
#endif