Files
oneDNN/tests/gtests/test_eltwise.cpp
2025-01-30 09:25:17 -08:00

463 lines
20 KiB
C++

/*******************************************************************************
* Copyright 2022-2025 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/
#include "dnnl_test_common.hpp"
#include "gtest/gtest.h"
#include "oneapi/dnnl/dnnl.hpp"
namespace dnnl {
using tag = memory::format_tag;
using dt = memory::data_type;
struct eltwise_test_params_t {
dt src_dt;
dt dst_dt;
dt diff_src_dt;
tag src_tag;
tag dst_tag;
tag diff_src_tag;
memory::dims dims;
float alpha, beta;
bool expect_to_fail;
dnnl_status_t expected_status;
};
bool cuda_check_format_tag(tag atag) {
// Blocking is not supported by cuDNN
return !impl::utils::one_of(
atag, tag::aBcd8b, tag::aBcd16b, tag::aBcde8b, tag::aBcde16b);
}
template <typename... Rest>
bool cuda_check_format_tag(tag first_tag, Rest... rest_tags) {
const bool ok = cuda_check_format_tag(first_tag);
if (!ok) return ok;
return cuda_check_format_tag(rest_tags...);
}
bool hip_check_format_tag(tag atag) {
// HIP has the same limitations for `tag` as CUDA.
return cuda_check_format_tag(atag);
}
template <typename... Rest>
bool hip_check_format_tag(tag first_tag, Rest... rest_tags) {
const bool ok = hip_check_format_tag(first_tag);
if (!ok) return ok;
return hip_check_format_tag(rest_tags...);
}
template <typename... Rest>
bool generic_check_format_tag(tag first_tag, Rest... rest_tags) {
const bool ok = hip_check_format_tag(first_tag);
if (!ok) return ok;
return hip_check_format_tag(rest_tags...);
}
class eltwise_test_t : public ::testing::TestWithParam<eltwise_test_params_t> {
private:
eltwise_test_params_t p;
memory src, dst;
std::shared_ptr<eltwise_forward::primitive_desc> pd_fwd_hint;
protected:
void SetUp() override {
p = ::testing::TestWithParam<eltwise_test_params_t>::GetParam();
SKIP_IF(unsupported_data_type(p.src_dt, p.dst_dt),
"Engine does not support this data type.");
SKIP_IF_CUDA(
p.dst_dt == dt::s8, "Unsupported int8 destination data type");
SKIP_IF_HIP(p.src_dt == dt::s8, "Unsupported int8 source data type");
SKIP_IF_CUDA(!cuda_check_format_tag(p.src_tag, p.dst_tag),
"Unsupported format tag");
SKIP_IF_HIP(!hip_check_format_tag(p.src_tag, p.dst_tag),
"Unsupported format tag");
SKIP_IF_GENERIC(!generic_check_format_tag(p.src_tag, p.dst_tag),
"Unsupported format tag");
SKIP_IF_CUDA(p.src_dt != p.dst_dt && p.src_dt != dt::undef
&& p.dst_dt != dt::undef,
"Unsupported different data types for source and "
"destination");
SKIP_IF_HIP(p.src_dt != p.dst_dt && p.src_dt != dt::undef
&& p.dst_dt != dt::undef,
"Unsupported different data types for source and "
"destination");
SKIP_IF_CUDA(p.src_tag != p.dst_tag && p.src_tag != tag::any
&& p.dst_tag != tag::any,
"Unsupported different memory formats for source and "
"destination");
SKIP_IF_HIP(p.src_tag != p.dst_tag && p.src_tag != tag::any
&& p.dst_tag != tag::any,
"Unsupported different memory formats for source and "
"destination");
catch_expected_failures(
[&]() { Test(); }, p.expect_to_fail, p.expected_status);
}
void Forward(prop_kind pk, algorithm aalgorithm) {
// eltwise specific types and values
using pd_t = eltwise_forward::primitive_desc;
auto eng = get_test_engine();
auto strm = make_stream(eng);
allows_attr_t aa {};
aa.po_binary = !is_amd_gpu(eng);
auto src_md = memory::desc(p.dims, p.src_dt, p.src_tag);
auto dst_md = memory::desc(p.dims, p.dst_dt, p.dst_tag);
// default pd ctor
auto pd = pd_t();
// regular pd ctor
pd = pd_t(eng, pk, aalgorithm, src_md, dst_md, p.alpha, p.beta);
// test all pd ctors
test_fwd_pd_constructors<pd_t>(
pd, aa, pk, aalgorithm, src_md, dst_md, p.alpha, p.beta);
pd_fwd_hint = std::make_shared<pd_t>(pd);
EXPECT_ANY_THROW(eltwise_forward(pd, {}));
// default primitive ctor
auto eltwise = eltwise_forward();
// regular primitive ctor
eltwise = eltwise_forward(pd);
// check primitive kind is eltwise
ASSERT_TRUE(eltwise.get_kind() == primitive::kind::eltwise);
// query for descs from pd
const auto src_desc = pd.src_desc();
const auto dst_desc = pd.dst_desc();
// query for src_desc via exec arg
ASSERT_TRUE(pd.query_md(query::exec_arg_md, DNNL_ARG_SRC) == src_desc);
if (p.src_tag != tag::any) { ASSERT_TRUE(src_md == src_desc); }
// query for dst_desc via exec arg
ASSERT_TRUE(pd.query_md(query::exec_arg_md, DNNL_ARG_DST) == dst_desc);
if (p.dst_tag != tag::any) { ASSERT_TRUE(dst_md == dst_desc); }
// query primitive parameters
ASSERT_EQ(pd.get_prop_kind(), pk);
ASSERT_EQ(pd.get_algorithm(), aalgorithm);
ASSERT_EQ(pd.get_alpha(), p.alpha);
ASSERT_EQ(pd.get_beta(), p.beta);
// check primitive returns zero_md for all rest md
ASSERT_TRUE(pd.weights_desc().is_zero());
ASSERT_TRUE(pd.diff_src_desc().is_zero());
ASSERT_TRUE(pd.diff_dst_desc().is_zero());
ASSERT_TRUE(pd.diff_weights_desc().is_zero());
src = test::make_memory(src_desc, eng);
dst = test::make_memory(dst_desc, eng);
fill_data(p.src_dt, src, 1, 1);
// test out-place mode
eltwise.execute(strm, {{DNNL_ARG_SRC, src}, {DNNL_ARG_DST, dst}});
strm.wait();
// test in-place mode on forward
if (p.src_tag == p.dst_tag && p.src_dt == p.dst_dt) {
// TODO: add a copy of memory and result comparison with previous
// dst output.
eltwise.execute(strm, {{DNNL_ARG_SRC, src}, {DNNL_ARG_DST, src}});
strm.wait();
}
}
void Backward(algorithm aalgorithm) {
// eltwise specific types and values
using pd_t = eltwise_backward::primitive_desc;
using hint_pd_t = eltwise_forward::primitive_desc;
allows_attr_t aa {}; // doesn't support anything
auto eng = get_test_engine();
auto strm = make_stream(eng);
auto diff_src_md = memory::desc(p.dims, p.diff_src_dt, p.diff_src_tag);
auto diff_dst_md = memory::desc(p.dims, p.dst_dt, p.dst_tag);
auto src_md = memory::desc(p.dims, p.src_dt, p.src_tag);
// default pd ctor
auto pd = pd_t();
// regular pd ctor
pd = pd_t(eng, aalgorithm, diff_src_md, diff_dst_md, src_md, p.alpha,
p.beta, *pd_fwd_hint);
// test all pd ctors
test_bwd_pd_constructors<pd_t, hint_pd_t>(pd, *pd_fwd_hint, aa,
aalgorithm, diff_src_md, diff_dst_md, src_md, p.alpha, p.beta);
EXPECT_ANY_THROW(eltwise_backward(pd, {}));
// default primitive ctor
auto eltwise = eltwise_backward();
// regular primitive ctor
eltwise = eltwise_backward(pd);
// check primitive kind is eltwise
ASSERT_TRUE(eltwise.get_kind() == primitive::kind::eltwise);
// query for descs from pd
const auto diff_src_desc = pd.diff_src_desc();
const auto diff_dst_desc = pd.diff_dst_desc();
const auto src_desc = pd.src_desc();
const auto dst_desc = pd.dst_desc();
// query for diff_src_desc via exec arg
ASSERT_TRUE(pd.query_md(query::exec_arg_md, DNNL_ARG_DIFF_SRC)
== diff_src_desc);
if (p.diff_src_tag != tag::any) {
ASSERT_TRUE(diff_src_md == diff_src_desc);
}
// query for diff_dst_desc via exec arg
ASSERT_TRUE(pd.query_md(query::exec_arg_md, DNNL_ARG_DIFF_DST)
== diff_dst_desc);
if (p.dst_tag != tag::any) {
ASSERT_TRUE(diff_dst_md == diff_dst_desc);
}
// query for src_desc via exec arg
ASSERT_TRUE(pd.query_md(query::exec_arg_md, DNNL_ARG_SRC) == src_desc);
ASSERT_TRUE((p.src_tag != tag::any && src_md == src_desc)
|| (p.dst_tag != tag::any
&& pd_fwd_hint.get()->dst_desc() == dst_desc));
// query primitive parameters
ASSERT_EQ(pd.get_prop_kind(), prop_kind::backward_data);
ASSERT_EQ(pd.get_algorithm(), aalgorithm);
ASSERT_EQ(pd.get_alpha(), p.alpha);
ASSERT_EQ(pd.get_beta(), p.beta);
// check primitive returns zero_md for all rest md
ASSERT_TRUE(pd.weights_desc().is_zero());
ASSERT_TRUE(pd.dst_desc().is_zero() || pd.src_desc().is_zero());
ASSERT_TRUE(pd.diff_weights_desc().is_zero());
auto diff_src = test::make_memory(diff_src_desc, eng);
auto diff_dst = test::make_memory(diff_dst_desc, eng);
fill_data(p.diff_src_dt, diff_dst, 2, 2);
// test out-place mode
eltwise.execute(strm,
{{DNNL_ARG_SRC, src}, {DNNL_ARG_DST, dst},
{DNNL_ARG_DIFF_DST, diff_dst},
{DNNL_ARG_DIFF_SRC, diff_src}});
strm.wait();
// test in-place mode
if (p.dst_tag == p.diff_src_tag && p.dst_dt == p.diff_src_dt) {
eltwise.execute(strm,
{{DNNL_ARG_SRC, src}, {DNNL_ARG_DST, dst},
{DNNL_ARG_DIFF_DST, diff_dst},
{DNNL_ARG_DIFF_SRC, diff_dst}});
strm.wait();
}
}
void Test() {
const bool is_int8 = p.src_dt == dt::s8 || p.src_dt == dt::u8;
std::vector<prop_kind> pks = {is_int8 ? prop_kind::forward_inference
: prop_kind::forward_training};
std::vector<algorithm> algs_all = {algorithm::eltwise_relu,
algorithm::eltwise_tanh, algorithm::eltwise_elu,
algorithm::eltwise_square, algorithm::eltwise_abs,
algorithm::eltwise_sqrt, algorithm::eltwise_swish,
algorithm::eltwise_linear, algorithm::eltwise_soft_relu,
algorithm::eltwise_mish, algorithm::eltwise_logistic,
algorithm::eltwise_exp, algorithm::eltwise_gelu_tanh,
algorithm::eltwise_gelu_erf, algorithm::eltwise_log,
algorithm::eltwise_clip, algorithm::eltwise_clip_v2,
algorithm::eltwise_pow, algorithm::eltwise_round,
algorithm::eltwise_hardswish, algorithm::eltwise_hardsigmoid,
algorithm::eltwise_relu_use_dst_for_bwd,
algorithm::eltwise_tanh_use_dst_for_bwd,
algorithm::eltwise_elu_use_dst_for_bwd,
algorithm::eltwise_sqrt_use_dst_for_bwd,
algorithm::eltwise_logistic_use_dst_for_bwd,
algorithm::eltwise_exp_use_dst_for_bwd,
algorithm::eltwise_clip_v2_use_dst_for_bwd};
// TODO: generalize this function.
if (p.src_dt != dt::f32) {
auto it = algs_all.begin();
while (true) {
if (*it == algorithm::eltwise_round) {
algs_all.erase(it);
break;
}
it++;
if (it == algs_all.end()) break;
}
}
std::vector<algorithm> algs_int8
= {algorithm::eltwise_relu, algorithm::eltwise_linear};
const auto &algs = is_int8 ? algs_int8 : algs_all;
for_(auto pk : pks)
for (auto alg : algs) {
SKIP_FOR_LOOP_CUDA(is_fwd(pk)
&& !impl::utils::one_of(alg,
algorithm::eltwise_relu,
algorithm::eltwise_tanh,
algorithm::eltwise_elu,
algorithm::eltwise_logistic),
"Unsupported algorithm type for CUDA");
SKIP_FOR_LOOP_CUDA(alg == algorithm::eltwise_relu && p.alpha != 0.f,
"Unsupported combination of algorithm type and alpha "
"parameter for CUDA");
SKIP_FOR_LOOP_HIP(is_fwd(pk)
&& !impl::utils::one_of(alg,
algorithm::eltwise_relu,
algorithm::eltwise_tanh,
algorithm::eltwise_elu,
algorithm::eltwise_logistic,
algorithm::eltwise_soft_relu,
algorithm::eltwise_abs),
"Unsupported algorithm type for HIP");
SKIP_FOR_LOOP_HIP(
alg == algorithm::eltwise_soft_relu && p.alpha != 1.f,
"Unsupported combination of algorithm type and alpha "
"parameter for HIP");
Forward(pk, alg);
bool to_continue = pk != prop_kind::forward_training
|| p.diff_src_dt == dt::undef
|| alg == algorithm::eltwise_round;
if (to_continue) continue;
SKIP_FOR_LOOP_CUDA(
!impl::utils::one_of(alg, algorithm::eltwise_relu),
"Unsupported algorithm type for CUDA");
SKIP_FOR_LOOP_HIP(!impl::utils::one_of(alg, algorithm::eltwise_relu,
algorithm::eltwise_soft_relu),
"Unsupported algorithm type for HIP");
SKIP_IF(unsupported_data_type(p.diff_src_dt),
"Engine does not support this data type.");
SKIP_IF_CUDA(!cuda_check_format_tag(p.diff_src_tag),
"Unsupported format tag");
SKIP_IF_HIP(!hip_check_format_tag(p.diff_src_tag),
"Unsupported format tag");
SKIP_IF_CUDA(p.src_dt != p.diff_src_dt && p.src_dt != dt::undef
&& p.diff_src_dt != dt::undef,
"Unsupported different data types for diff_source and "
"diff_destination");
SKIP_IF_HIP(p.src_dt != p.diff_src_dt && p.src_dt != dt::undef
&& p.diff_src_dt != dt::undef,
"Unsupported different data types for diff_source and "
"diff_destination");
SKIP_IF_CUDA(p.src_tag != p.diff_src_tag && p.src_tag != tag::any
&& p.diff_src_tag != tag::any,
"Unsupported different memory formats for diff_source "
"and "
"diff_destination");
SKIP_IF_HIP(p.src_tag != p.diff_src_tag && p.src_tag != tag::any
&& p.diff_src_tag != tag::any,
"Unsupported different memory formats for diff_source "
"and diff_destination");
Backward(alg);
}
}
bool is_fwd(prop_kind pk) const {
return pk == prop_kind::forward_training
|| pk == prop_kind::forward_inference;
}
};
using tp = eltwise_test_params_t;
TEST_P(eltwise_test_t, TestsEltwise) {}
INSTANTIATE_TEST_SUITE_P(Test_Eltwise_EF, eltwise_test_t,
::testing::Values(
// Negative dims
tp {dt::f32, dt::f32, dt::undef, tag::nchw, tag::nchw,
tag::undef, {2, -2, 128, 256}, 1.f, 2.f, true,
dnnl_invalid_arguments},
// Tag for src on forward is not specified
tp {dt::f32, dt::f32, dt::undef, tag::any, tag::nchw,
tag::undef, {2, 2, 128, 256}, 1.f, 2.f, true,
dnnl_invalid_arguments},
// Tag for src on backward is not specified
tp {dt::f32, dt::f32, dt::f32, tag::any, tag::nchw, tag::nchw,
{2, 2, 128, 256}, 1.f, 2.f, true,
dnnl_invalid_arguments},
// Data type for src is not specified
tp {dt::undef, dt::f32, dt::undef, tag::nchw, tag::nchw,
tag::undef, {2, 2, 128, 256}, 1.f, 2.f, true,
dnnl_invalid_arguments},
// Different data types are not supported
tp {dt::f32, dt::bf16, dt::undef, tag::nchw, tag::nchw,
tag::undef, {2, 2, 128, 256}, 1.f, 2.f, true,
dnnl_unimplemented},
// Different data types are not supported
tp {dt::f32, dt::bf16, dt::f32, tag::nchw, tag::nchw, tag::nchw,
{2, 2, 128, 256}, 1.f, 2.f, true, dnnl_unimplemented},
// Different memory formats are not supported
tp {dt::f32, dt::f32, dt::undef, tag::nchw, tag::nhwc,
tag::undef, {2, 2, 128, 256}, 1.f, 2.f, true,
dnnl_unimplemented},
// Different memory formats are not supported
tp {dt::f32, dt::f32, dt::f32, tag::nchw, tag::nhwc, tag::nchw,
{2, 2, 128, 256}, 1.f, 2.f, true, dnnl_unimplemented}));
static auto all_cases = [](dt src_dt, dt dst_dt, dt diff_src_dt) {
return ::testing::Values(tp {src_dt, dst_dt, diff_src_dt, tag::nwc,
tag::nwc, tag::nwc, {2, 16, 10}, 0.f, 0.f},
tp {src_dt, dst_dt, diff_src_dt, tag::ncw, tag::ncw, tag::ncw,
{2, 64, 27}, 1.f, 2.f},
tp {src_dt, dst_dt, diff_src_dt, tag::nhwc, tag::nhwc, tag::nhwc,
{2, 16, 10, 8}, 0.f, 0.9f},
tp {src_dt, dst_dt, diff_src_dt, tag::nchw, tag::nchw, tag::nchw,
{2, 64, 27, 27}, 1.f, 2.f},
tp {src_dt, dst_dt, diff_src_dt, tag::nChw8c, tag::nChw8c,
tag::nChw8c, {2, 16, 16, 8}, 0.1f, 0.9f},
tp {src_dt, dst_dt, diff_src_dt, tag::nChw16c, tag::nChw16c,
tag::nChw16c, {2, 16, 4, 4}, 0.f, 0.f},
tp {src_dt, dst_dt, diff_src_dt, tag::ncdhw, tag::ncdhw, tag::ncdhw,
{2, 64, 7, 7, 7}, 1.f, 1.f},
tp {src_dt, dst_dt, diff_src_dt, tag::ncdhw, tag::ncdhw, tag::ncdhw,
{10, 10, 10, 10, 10}, 0.f, 0.f},
tp {src_dt, dst_dt, diff_src_dt, tag::nCdhw16c, tag::nCdhw16c,
tag::nCdhw16c, {4, 15, 2, 2, 2}, 0.1f, 0.2f});
};
#define EXPAND_DTS(src, dst, diff_src) \
memory::data_type::src, memory::data_type::dst, memory::data_type::diff_src
#define INST_TEST_CASE(name, suite, ...) \
INSTANTIATE_TEST_SUITE_P(name, eltwise_test_t, suite(__VA_ARGS__));
#define CPU_INST_TEST_CASE(name, suite, ...) \
CPU_INSTANTIATE_TEST_SUITE_P(name, eltwise_test_t, suite(__VA_ARGS__));
#define GPU_INST_TEST_CASE(name, suite, ...) \
GPU_INSTANTIATE_TEST_SUITE_P(name, eltwise_test_t, suite(__VA_ARGS__));
INST_TEST_CASE(EltwiseSimpleF32, all_cases, EXPAND_DTS(f32, f32, f32));
INST_TEST_CASE(EltwiseSimpleBF16, all_cases, EXPAND_DTS(bf16, bf16, bf16));
INST_TEST_CASE(EltwiseSimpleF16, all_cases, EXPAND_DTS(f16, f16, undef));
INST_TEST_CASE(EltwiseSimpleU8, all_cases, EXPAND_DTS(u8, u8, undef));
} // namespace dnnl