Skip to content

[GPU] Permute f y #24319

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

Merged
p-durandin merged 12 commits into from
May 6, 2024
Merged

[GPU] Permute f y #24319

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
033cf5a
cherry pick f0abd46
OlehKravchyshyn Jul 18, 2023
eacb603
Merge branch 'openvinotoolkit:master' into permute_f_y
michal-miotk Apr 30, 2024
687a156
added JTIMES_SIMPLE
michal-miotk Apr 30, 2024
c769644
more test cases and fixing error in cl kernel
michal-miotk Apr 30, 2024
b286fcf
all test cases passes
michal-miotk Apr 30, 2024
83dfffc
temporary disabling long tests
michal-miotk May 1, 2024
33a76f8
enabled 32, 32, 256, 512, format::bfyx test
michal-miotk May 1, 2024
2b21ff3
check if fails on permute_ref
michal-miotk May 2, 2024
54b3f68
performance test compiles
michal-miotk May 2, 2024
4a03261
decreased batch and y by 2
michal-miotk May 2, 2024
965c185
deleted JTIMES_SIMPLE variable, turned on permute_f_y_axes kernel in …
michal-miotk May 2, 2024
6fd79fc
Merge branch 'openvinotoolkit:master' into permute_f_y
michal-miotk May 2, 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
6 changes: 3 additions & 3 deletions src/plugins/intel_gpu/src/kernel_selector/cl_kernels/permute_f_y_axes.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,7 +95,7 @@ KERNEL (permute_f_y_axes)(
__attribute__((opencl_unroll_hint(TILE_SIZE)))
for (int i = 0; i < TILE_SIZE; ++i) {
const int x_idx = x_begin + i;
const int f = f_begin + j;
const int f = (f_begin + j) % INPUT0_FEATURE_NUM;
const int y_idx = y_begin + bf_local;
const uint output_offset = OUTPUT_GET_INDEX(b_idx, y_idx, f, x_idx) - get_sub_group_local_id();
DT_OUTPUT_BLOCK_WRITE(output, output_offset, transpose_buf[j][bf_local][i]);
Expand Down Expand Up @@ -149,7 +149,7 @@ KERNEL (permute_f_y_axes)(
__attribute__((opencl_unroll_hint(J_TIMES)))
for (int j = 0; j < J_TIMES; ++j) {
const int j_vec = j * VEC_SIZE;
const int f = f_begin + j_vec;
const int f = (f_begin + j_vec) % INPUT0_FEATURE_NUM;;
const int y_idx = y_begin + bf_local;
const int output_idx = OUTPUT_GET_INDEX(b_idx, y_idx, f, x_idx);
WRITE_VEC(READ_VEC(0, &transpose_buf[bf_local][j_vec]), 0, &output[output_idx]);
Expand All @@ -171,7 +171,7 @@ KERNEL (permute_f_y_axes)(
}
__attribute__((opencl_unroll_hint(TILE_SIZE)))
for (int j = 0; j < TILE_SIZE; ++j) {
const int f = f_begin + j;
const int f = (f_begin + j) % INPUT0_FEATURE_NUM;
const int y_idx = y_begin + bf_local;
const uint output_offset = OUTPUT_GET_INDEX(b_idx, y_idx, f, x_idx) - get_sub_group_local_id();
DT_OUTPUT_BLOCK_WRITE(output, output_offset, transpose_buf[j][bf_local]);
Expand Down
10 changes: 3 additions & 7 deletions src/plugins/intel_gpu/src/kernel_selector/kernels/permute/permute_kernel_f_y_axes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,10 +24,7 @@ size_t GetDivisor(const size_t input_size) {
return input_size % i == 0;
};
auto result = std::find_if(begin(v), end(v), is_divided);
if (result != end(v)) {
return *result;
}
return 1;
return *result;
}

bool IsSimpleMemCopyOperation(const permute_params& params) {
Expand Down Expand Up @@ -134,7 +131,7 @@ JitConstants PermuteKernel_f_y_axes::GetJitConstants(const permute_params& param
const size_t tile_width = GetTileWidth(params);
const size_t vector_size = std::min(tile_width, static_cast<size_t>(4));
const size_t tile_size = GetTileSize(params);
const size_t j_times = tile_size / vector_size;
const size_t j_times = IsSimpleMemCopyOperation(params) ? tile_width / vector_size : tile_size / vector_size;
const size_t feature_block_size = GetFeatureBlockSize(params);
jit.AddConstant(MakeJitConstant("BLOCK_SIZE", tile_width));
jit.AddConstant(MakeJitConstant("VEC_SIZE", vector_size));
Expand Down Expand Up @@ -200,7 +197,7 @@ bool PermuteKernel_f_y_axes::Validate(const Params& p) const {

const auto is_swapping_f_with_y = [](const std::vector<uint16_t>& order) {
// Target transform: Swap feature with y
// OV order: 0 2 1 3 => bfyx -> byfx
// IE order: 0 2 1 3 => bfyx -> byfx
// cldnn order: 0 3 2 1 => bfxy -> byxf
if (order.size() != 4) {
return false;
Expand Down Expand Up @@ -244,5 +241,4 @@ bool PermuteKernel_f_y_axes::Validate(const Params& p) const {
KernelsPriority PermuteKernel_f_y_axes::GetKernelsPriority(const Params& /*params*/) const {
return FORCE_PRIORITY_3;
}

} // namespace kernel_selector
2 changes: 1 addition & 1 deletion src/plugins/intel_gpu/src/kernel_selector/kernels/permute/permute_kernel_selector.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ permute_kernel_selector::permute_kernel_selector() {
Attach<PermuteKernel_tile_8x8_4x4>();
Attach<PermuteKernel_tile_8x8_4x4_fsv>();
Attach<PermuteKernel_bfzyx_to_bfyxz>();
// Attach<PermuteKernel_f_y_axes>();
Attach<PermuteKernel_f_y_axes>();
}

KernelsData permute_kernel_selector::GetBestKernels(const Params& params) const {
Expand Down
96 changes: 79 additions & 17 deletions src/plugins/intel_gpu/tests/unit/test_cases/permute_gpu_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2248,18 +2248,38 @@ TEST_P(permute_tile_fsv_5d, i64_cached) {

class permute_f_y_axes_tile: public TiledPermuteTest {};

// Test cases are disabled because permute_f_y_axes_tile kernel itself is disabled for accuracy issue
// INSTANTIATE_TEST_SUITE_P(smoke_permute_f_y_axes_tile,
// permute_f_y_axes_tile,
// ::testing::ValuesIn(std::vector<TiledPermuteParam>{
// {{1, 4, 8, 1}, format::bfyx}, // permute_f_y_axes
// {{1, 64, 32, 1}, format::bfyx}, // permute_f_y_axes
// {{1, 32, 256, 512}, format::b_fs_yx_fsv32}, // THREE_DIM_TRANSPOSE
// {{1, 32, 256, 512}, format::bfyx}, // PERMUTE_SIMPLE_MEM_COPY
// {{1, 256, 256, 1}, format::b_fs_yx_fsv32}, // permute_f_y_axes
// {{1, 32, 16, 4}, format::b_fs_yx_fsv16}, // THREE_DIM_TRANSPOSE
// }),
// TiledPermuteTest::PrintToStringParamName);
INSTANTIATE_TEST_SUITE_P(smoke_permute_f_y_axes_tile,
permute_f_y_axes_tile,
::testing::ValuesIn(std::vector<TiledPermuteParam>{
{{1, 4, 8, 1}, format::bfyx}, // permute_f_y_axes
{{1, 64, 32, 1}, format::bfyx}, // permute_f_y_axes
{{1, 32, 256, 512}, format::b_fs_yx_fsv32}, // THREE_DIM_TRANSPOSE
{{1, 32, 256, 512}, format::bfyx}, // PERMUTE_SIMPLE_MEM_COPY
{{1, 256, 256, 1}, format::b_fs_yx_fsv32}, // permute_f_y_axes
{{1, 32, 16, 4}, format::b_fs_yx_fsv16}, // THREE_DIM_TRANSPOSE
//4 batch version
{{4, 4, 4, 1}, format::bfyx}, // permute_f_y_axes
//32 batch version
{{32, 4, 8, 1}, format::bfyx}, // permute_f_y_axes
{{32, 64, 32, 1}, format::bfyx}, // permute_f_y_axes
{{32, 128, 196, 1}, format::bfyx}, // permute_f_y_axes
{{32, 196, 4, 16}, format::bfyx}, // permute_f_y_axes
{{32, 196, 4, 32}, format::bfyx}, // permute_f_y_axes
{{32, 4, 196, 32}, format::bfyx}, // permute_f_y_axes
{{32, 196, 128, 1}, format::bfyx}, // permute_f_y_axes
{{32, 196, 8, 16}, format::bfyx}, // permute_f_y_axes
{{16, 32, 128, 512}, format::b_fs_yx_fsv32}, // THREE_DIM_TRANSPOSE
{{16, 32, 128, 512}, format::bfyx}, // PERMUTE_SIMPLE_MEM_COPY
{{32, 256, 256, 1}, format::b_fs_yx_fsv32}, // permute_f_y_axes
{{32, 32, 16, 4}, format::b_fs_yx_fsv16}, // THREE_DIM_TRANSPOSE
{{32, 16, 16, 16}, format::bfyx},
{{32, 16, 8, 16}, format::bfyx},
{{32, 16, 16, 64}, format::bfyx},
{{32, 16, 8, 32}, format::bfyx},
{{32, 8, 16, 32}, format::bfyx},
{{32, 196, 8, 64}, format::bfyx}, // permute_f_y_axes
}),
TiledPermuteTest::PrintToStringParamName);

TEST_P(permute_f_y_axes_tile, combined) {
auto p = GetParam();
Expand All @@ -2273,14 +2293,53 @@ TEST_P(permute_f_y_axes_tile, combined) {

struct TiledPerformancePermuteTest : TiledPermuteTest
{
static double get_exectime(const std::map<cldnn::primitive_id, cldnn::network_output>& outputs,
const std::string& primitive_id)
{
using namespace std::chrono;
std::shared_ptr<event> e = outputs.at(primitive_id).get_event();
e->wait(); // should ensure execution completion, if not segfault will occur
double avg_time = 0.0;
auto intervals = e->get_profiling_info();
for (const auto& q : intervals)
{
if (q.stage != instrumentation::profiling_stage::executing) {
continue;
}
avg_time = duration_cast<duration<double, microseconds::period>>(q.value->value()).count();
break;
}
return avg_time;
}

static void print_all_perf(std::map<primitive_id, network_output> outputs)
{
std::cout << "Print last run time" << std::endl;
using namespace std::chrono;
for( const auto &n : outputs ) {
std::shared_ptr<event> e = n.second.get_event();
auto intervals = e->get_profiling_info();
double time = 0.0;
for (const auto& q : intervals)
{
if (q.stage == instrumentation::profiling_stage::executing) {
continue;
}
time = duration_cast<duration<double, microseconds::period>>(q.value->value()).count();
break;
}
std::cout << n.first << ":" << time << std::endl;
}
std::cout << std::endl;
}

template<data_types Data_Type>
void execute_perf_test(const std::vector<cldnn::tensor::value_type>& sizes, cldnn::format format_fsv,
const std::string & kernel_name, std::vector<uint16_t> permute_order)
{
auto& engine = get_test_engine();
// convert ov::float16 to ov::float16
using type_ = typename ov::element_type_traits<Data_Type>::value_type;
using type = typename std::conditional<std::is_same<type_, ov::float16>::value, ov::float16, type_>::type;
// convert half_t to FLOAT16
using type = typename ov::element_type_traits<Data_Type>::value_type;

std::vector<cldnn::tensor::value_type> internal_sizes(sizes);
std::swap(internal_sizes.at(2), internal_sizes.back());
Expand Down Expand Up @@ -2342,11 +2401,11 @@ struct TiledPerformancePermuteTest : TiledPermuteTest
double exectime_opt = 0.f;
for (int i = 0; i < r; ++i) {
output_permute_opt = network_tile.execute();
auto t_opt = get_profiling_exectime(output_permute_opt, "output");
auto t_opt = get_exectime(output_permute_opt, "output");
exectime_opt += t_opt;

output_permute_ref = network_ref.execute();
auto t_ref = get_profiling_exectime(output_permute_ref, "output");
auto t_ref = get_exectime(output_permute_ref, "output");
exectime_ref += t_ref;
}
exectime_ref /= r;
Expand All @@ -2365,14 +2424,17 @@ struct TiledPerformancePermuteTest : TiledPermuteTest
<< frm_str << " " << input_type << " " << exectime_opt << std::endl;

}

};


// No need to run performance tests on CI
TEST_P(TiledPerformancePermuteTest, DISABLED_f32) {
auto p = GetParam();
execute_perf_test<cldnn::data_types::f32>(p.sizes, p.format_fsv, "permute_f_y_axes", {0, 2, 1, 3});
}


INSTANTIATE_TEST_SUITE_P(, TiledPerformancePermuteTest,
::testing::ValuesIn(std::vector<TiledPermuteParam> {
// b_fs_zy_fsv16
Expand Down