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

[GPU] Minor fixes for dynamic model (#16075)

Browse Source 
Signed-off-by: Andrew Park <andrew.park@intel.com>
This commit is contained in:
Andrew Kwangwoong Park
2023-03-06 20:50:38 +09:00
committed by GitHub
parent 54d3641baa
commit 4ce35fd851
4 changed files with 42 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_primitive_fusing.cpp
View File

@@ -70,7 +70,7 @@ void prepare_primitive_fusing::remove_redundant_reshape(program &p) {
for (auto prev : node.get_dependencies()) {
if (!prev.first->is_type<reshape>())
return;
if (prev.first->get_users().size() > 1)
if (prev.first->get_users().size() > 1 || prev.first->get_dependencies().size() > 1)
return;
if (prev.first->as<reshape>().input().get_output_layout() == node.get_output_layout()) {
p.add_optimized_primitive_info(prev.first->id());

3
src/plugins/intel_gpu/src/graph/primitive_inst.cpp
View File

@@ -153,6 +153,9 @@ void primitive_inst::update_shape() {
auto idx = _deps[i].second;
auto new_shape = _deps[i].first->_impl_params->get_output_layout(idx);
if (_impl_params->get_input_layout(i) != new_shape) {
GPU_DEBUG_TRACE_DETAIL << id() << ": update shape dep: " << _deps[i].first->id()
<< " was: " << _impl_params->get_input_layout(i).to_short_string()
<< " now: " << new_shape.to_short_string() << std::endl;
_impl_params->input_layouts[i] = new_shape;
input_shape_changed = true;
}

2
src/plugins/intel_gpu/src/kernel_selector/kernels/gather/gather_kernel_ref.cpp
View File

@@ -297,7 +297,7 @@ KernelsData GatherKernelRef::GetKernelsData(const Params& params, const optional
2,
GetFusedPrimitiveInputsCount(params),
1,
newParams.outputs[0].is_dynamic());
newParams.has_dynamic_tensors());
return {kd};
}

37
src/plugins/intel_gpu/tests/passes/prepare_primitive_fusing_test.cpp
View File

@@ -11,6 +11,8 @@
#include "data_inst.h"
#include "eltwise_inst.h"
#include "reduce_inst.h"
#include "reshape_inst.h"
#include "gemm_inst.h"
#include "pass_manager.h"
#include "to_string_utils.h"
@@ -403,3 +405,38 @@ TEST(prepare_primitive_fusing, fuse_eltwise_to_fc_dyn_illegal_2) {
ASSERT_EQ(lock[2], 93);
ASSERT_EQ(lock[3], 94);
}
TEST(prepare_primitive_fusing, dont_remove_only_dep_reshape) {
// Topology:
// input -> reshape(w/ 2nd non-const input) -> reshape(w/ 2nd const input) -> gemm
//
// Expectation:
// If only the input size of depedency reshape is not 1 among the sequence of reshapes
// The current reshape alone should not be removed, and removing redundant reshapes is skipped
auto& engine = get_test_engine();
auto in_layout = layout{ ov::PartialShape::dynamic(4), data_types::f32, format::bfyx };
auto pattern_layout = layout{ ov::PartialShape{ 4 }, data_types::i64, format::bfyx };
std::vector<int64_t> output_pattern { 0, 1, -1, 0 };
topology topology;
topology.add(input_layout("input1", in_layout));
topology.add(input_layout("pattern1", pattern_layout));
topology.add(input_layout("input2", in_layout));
topology.add(reshape("reshape1", input_info("input1"), input_info("pattern1"), true, ov::PartialShape::dynamic(4)));
topology.add(reshape("reshape2", input_info("reshape1"), true, output_pattern, ov::PartialShape::dynamic(4)));
topology.add(gemm("gemm", { input_info("reshape2"), input_info("input2") }, data_types::f32, false, false));
ExecutionConfig config;
config.set_property(ov::intel_gpu::optimize_data(true));
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, "reshape2"));
}