Skip to content

GitLab

Commit c8c016dd authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8

parents e8ca3daf 4e731776
Show whitespace changes
Inline Side-by-side
Showing with 1511 additions and 383 deletions
test/ck_tile/grouped_gemm/CMakeLists.txt 0 → 100644
View file @ c8c016dd
# Currently ck_tile is only built on gfx9
if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_ck_tile_grouped_gemm test_grouped_gemm.cpp)
endif()
test/ck_tile/grouped_gemm/test_grouped_gemm.cpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include "gtest/gtest.h"
#include "ck_tile/host.hpp"
#include "test_grouped_gemm_util.hpp"
using F16 = ck_tile::half_t;
using F32 = float;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
// clang-format off
using KernelTypes = ::testing::Types<
// ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType, CDataType
std::tuple< Row, Row, Row, F16, F16, F32, F16>,
//std::tuple< Col, Row, Row, F16, F16, F32, F16>,
std::tuple< Row, Col, Row, F16, F16, F32, F16>//,
//std::tuple< Col, Col, Row, F16, F16, F32, F16>
>;
// clang-format on
TYPED_TEST_SUITE(TestCkTileGroupedGemm, KernelTypes);
#include "test_grouped_gemm_ut_cases.inc"
test/ck_tile/grouped_gemm/test_grouped_gemm_ut_cases.inc 0 → 100644
View file @ c8c016dd
#pragma once
TYPED_TEST(TestCkTileGroupedGemm, Basic)
{
const int group_count = 16;
std::vector<int> Ms;
std::vector<int> Ns;
std::vector<int> Ks;
std::vector<int> stride_As;
std::vector<int> stride_Bs;
std::vector<int> stride_Cs;
for(int i = 0; i < group_count; i++)
{
Ms.push_back(256 + 256 * i);
Ns.push_back(128 + 128 * i);
Ks.push_back(128 + 64 * i);
stride_As.push_back(Ks[i]);
stride_Bs.push_back(Ks[i]);
stride_Cs.push_back(Ns[i]);
}
this->Run(Ms, Ns, Ks, stride_As, stride_Bs, stride_Cs, group_count);
}
test/ck_tile/grouped_gemm/test_grouped_gemm_util.hpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <sstream>
#include <gtest/gtest.h>
#include "ck_tile/core.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp"
template <typename Tuple>
class TestCkTileGroupedGemm : public ::testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CLayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using AccDataType = std::tuple_element_t<5, Tuple>;
using CDataType = std::tuple_element_t<6, Tuple>;
struct GroupedGemKernelParam
{
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const bool kTilePermute = false;
static const ck_tile::index_t kOutputRank = 2;
static const int kBlockPerCu = 1;
static const ck_tile::index_t M_Tile = 128;
static const ck_tile::index_t N_Tile = 128;
static const ck_tile::index_t K_Tile = 32;
static const ck_tile::index_t M_Warp = 2;
static const ck_tile::index_t N_Warp = 2;
static const ck_tile::index_t K_Warp = 1;
static const ck_tile::index_t M_Warp_Tile = 32;
static const ck_tile::index_t N_Warp_Tile = 32;
static const ck_tile::index_t K_Warp_Tile = 8;
};
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<GroupedGemKernelParam::M_Tile,
GroupedGemKernelParam::N_Tile,
GroupedGemKernelParam::K_Tile>,
ck_tile::sequence<GroupedGemKernelParam::M_Warp,
GroupedGemKernelParam::N_Warp,
GroupedGemKernelParam::K_Warp>,
ck_tile::sequence<GroupedGemKernelParam::M_Warp_Tile,
GroupedGemKernelParam::N_Warp_Tile,
GroupedGemKernelParam::K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
template <typename CLayout>
using GemmEpilogue =
std::conditional_t<std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>,
ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
GroupedGemKernelParam::kPadM,
GroupedGemKernelParam::kPadN,
GroupedGemKernelParam::kTilePermute,
GroupedGemKernelParam::kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType,
CDataType,
GroupedGemKernelParam::kPadM,
GroupedGemKernelParam::kPadN>>>;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmTraits = ck_tile::TileGemmTraits<GroupedGemKernelParam::kPadM,
GroupedGemKernelParam::kPadN,
GroupedGemKernelParam::kPadK,
ALayout,
BLayout,
CLayout>;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenGemmShape,
CodegenGemmTraits<ALayout, BLayout, CLayout>>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>,
CodegenGemmPolicy>;
template <typename ALayout, typename BLayout, typename CLayout>
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,
CodegenGemmPipeline<ALayout, BLayout, CLayout>,
GemmEpilogue<CLayout>>;
using grouped_gemm_kargs = ck_tile::GroupedGemmHostArgs;
std::size_t GetWorkspaceSize(const std::vector<grouped_gemm_kargs>& gemm_descs)
{
return Kernel<std::nullptr_t, std::nullptr_t, std::nullptr_t>::GetWorkSpaceSize(gemm_descs);
}
template <typename ALayout, typename BLayout, typename CLayout>
void invoke_grouped_gemm(const std::vector<grouped_gemm_kargs>& gemm_descs,
const ck_tile::stream_config& s,
void* p_workspace_)
{
using GroupedGemmKernel = Kernel<ALayout, BLayout, CLayout>;
auto arguments = GroupedGemmKernel::MakeKargs(gemm_descs);
const dim3 grids = GroupedGemmKernel::GridSize(gemm_descs);
constexpr dim3 blocks = GroupedGemmKernel::BlockSize();
ck_tile::hip_check_error(hipMemcpyWithStream(
p_workspace_,
arguments.data(),
arguments.size() * sizeof(typename GroupedGemmKernel::GemmTransKernelArg),
hipMemcpyHostToDevice,
s.stream_id_));
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GroupedGemKernelParam::kBlockPerCu>(
GroupedGemmKernel{},
grids,
blocks,
0,
ck_tile::cast_pointer_to_constant_address_space(p_workspace_),
gemm_descs.size()));
}
public:
void Run(const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
std::vector<int>& stride_As,
std::vector<int>& stride_Bs,
std::vector<int>& stride_Cs,
const int group_count = 16)
{
using namespace ck_tile::literals;
auto f_host_tensor_descriptor = [](std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return ck_tile::HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return ck_tile::HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(stride == 0)
{
if constexpr(std::is_same_v<decltype(layout),
ck_tile::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
std::vector<ck_tile::HostTensor<ADataType>> a_m_k_tensors;
std::vector<ck_tile::HostTensor<BDataType>> b_k_n_tensors;
std::vector<ck_tile::HostTensor<CDataType>> c_m_n_tensors;
a_m_k_tensors.reserve(group_count);
b_k_n_tensors.reserve(group_count);
c_m_n_tensors.reserve(group_count);
std::vector<std::unique_ptr<ck_tile::DeviceMem>> a_m_k_dev_buf;
std::vector<std::unique_ptr<ck_tile::DeviceMem>> b_k_n_dev_buf;
std::vector<std::unique_ptr<ck_tile::DeviceMem>> c_m_n_dev_buf;
a_m_k_dev_buf.reserve(group_count);
b_k_n_dev_buf.reserve(group_count);
c_m_n_dev_buf.reserve(group_count);
std::vector<grouped_gemm_kargs> gemm_descs;
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; ++i)
{
const ck_tile::index_t M = Ms[i];
const ck_tile::index_t N = Ns[i];
const ck_tile::index_t K = Ks[i];
stride_As[i] = f_get_default_stride(M, N, stride_As[i], ALayout{});
stride_Bs[i] = f_get_default_stride(K, N, stride_Bs[i], BLayout{});
stride_Cs[i] = f_get_default_stride(M, N, stride_Cs[i], CLayout{});
a_m_k_tensors.push_back(ck_tile::HostTensor<ADataType>(
f_host_tensor_descriptor(M, K, stride_As[i], ALayout{})));
b_k_n_tensors.push_back(ck_tile::HostTensor<BDataType>(
f_host_tensor_descriptor(K, N, stride_Bs[i], BLayout{})));
c_m_n_tensors.push_back(ck_tile::HostTensor<CDataType>(
f_host_tensor_descriptor(M, N, stride_Cs[i], CLayout{})));
std::cout << "gemm[" << i << "]"
<< " a_m_k: " << a_m_k_tensors[i].mDesc
<< " b_k_n: " << b_k_n_tensors[i].mDesc
<< " c_m_n: " << c_m_n_tensors[i].mDesc << std::endl;
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k_tensors[i]);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n_tensors[i]);
a_m_k_dev_buf.push_back(std::make_unique<ck_tile::DeviceMem>(
a_m_k_tensors[i].get_element_space_size_in_bytes()));
b_k_n_dev_buf.push_back(std::make_unique<ck_tile::DeviceMem>(
b_k_n_tensors[i].get_element_space_size_in_bytes()));
c_m_n_dev_buf.push_back(std::make_unique<ck_tile::DeviceMem>(
c_m_n_tensors[i].get_element_space_size_in_bytes()));
a_m_k_dev_buf[i]->ToDevice(a_m_k_tensors[i].data());
b_k_n_dev_buf[i]->ToDevice(b_k_n_tensors[i].data());
c_m_n_dev_buf[i]->SetZero();
c_m_n_tensors[i].SetZero();
const void* p_a = a_m_k_dev_buf[i]->GetDeviceBuffer();
const void* p_b = b_k_n_dev_buf[i]->GetDeviceBuffer();
void* p_c = c_m_n_dev_buf[i]->GetDeviceBuffer();
gemm_descs.push_back(
{p_a, p_b, p_c, M, N, K, stride_As[i], stride_Bs[i], stride_Cs[i]});
}
ck_tile::DeviceMem gemm_workspace;
gemm_workspace.Realloc(GetWorkspaceSize(gemm_descs));
invoke_grouped_gemm<ALayout, BLayout, CLayout>(
gemm_descs, ck_tile::stream_config{nullptr, false}, gemm_workspace.GetDeviceBuffer());
for(int i = 0; i < group_count; i++)
{
c_m_n_dev_buf[i]->FromDevice(c_m_n_tensors[i].data());
}
bool pass{true};
for(int i = 0; i < group_count; ++i)
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
f_host_tensor_descriptor(Ms[i], Ns[i], stride_Cs[i], CLayout{}));
c_m_n_host_ref.SetZero();
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k_tensors[i], b_k_n_tensors[i], c_m_n_host_ref);
pass &= ck_tile::check_err(c_m_n_tensors[i], c_m_n_host_ref);
}
EXPECT_TRUE(pass);
}
};
test/data_type/CMakeLists.txt
View file @ c8c016dd
......@@ -9,13 +9,38 @@ if (USE_BITINT_EXTENSION_INT4)
endif()
endif()
add_gtest_executable(test_fp8 test_fp8.cpp)
if(result EQUAL 0)
target_link_libraries(test_fp8 PRIVATE utility)
add_custom_target(test_fp8)
if (CK_USE_OCP_FP8)
add_gtest_executable(test_fp8_ocp test_fp8_ocp.cpp)
if(result EQUAL 0)
target_link_libraries(test_fp8_ocp PRIVATE utility)
endif()
add_gtest_executable(test_bf8_ocp test_bf8_ocp.cpp)
if(result EQUAL 0)
target_link_libraries(test_bf8_ocp PRIVATE utility)
endif()
add_dependencies(test_fp8 test_fp8_ocp)
add_dependencies(test_fp8 test_bf8_ocp)
endif()
add_gtest_executable(test_bf8 test_bf8.cpp)
if(result EQUAL 0)
target_link_libraries(test_bf8 PRIVATE utility)
if (CK_USE_FNUZ_FP8)
add_gtest_executable(test_fp8_fnuz test_fp8_fnuz.cpp)
if(result EQUAL 0)
target_link_libraries(test_fp8_fnuz PRIVATE utility)
endif()
add_gtest_executable(test_bf8_fnuz test_bf8_fnuz.cpp)
if(result EQUAL 0)
target_link_libraries(test_bf8_fnuz PRIVATE utility)
endif()
add_dependencies(test_fp8 test_fp8_fnuz)
add_dependencies(test_fp8 test_bf8_fnuz)
endif()
add_gtest_executable(test_custom_type test_custom_type.cpp)
......
test/data_type/test_bf8.cpp → test/data_type/test_bf8_fnuz.cpp
View file @ c8c016dd
......@@ -5,158 +5,169 @@
#include "ck/utility/data_type.hpp"
#include "ck/utility/type_convert.hpp"
using ck::bf8_t;
using ck::bf8_fnuz_t;
using ck::f8_convert_rne;
using ck::f8_convert_sr;
using ck::half_t;
using ck::type_convert;
TEST(BF8, NumericLimits)
TEST(BF8FNUZ, NumericLimits)
{
// constants given for negative zero nan mode
EXPECT_EQ(ck::NumericLimits<bf8_t>::Min(), type_convert<bf8_t>(0x04));
EXPECT_EQ(ck::NumericLimits<bf8_t>::Max(), type_convert<bf8_t>(0x7F));
EXPECT_EQ(ck::NumericLimits<bf8_t>::Lowest(), type_convert<bf8_t>(0xFF));
EXPECT_EQ(ck::NumericLimits<bf8_t>::QuietNaN(), type_convert<bf8_t>(0x80));
EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::Min(), type_convert<bf8_fnuz_t>(0x04));
EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::Max(), type_convert<bf8_fnuz_t>(0x7F));
EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::Lowest(), type_convert<bf8_fnuz_t>(0xFF));
EXPECT_EQ(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(), type_convert<bf8_fnuz_t>(0x80));
}
TEST(BF8, ConvertFP32Nearest)
TEST(BF8FNUZ, ConvertFP32Nearest)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to bf8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_t>(0.0f)), abs_tol);
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(0.0f)), abs_tol);
// don't run the next test on gfx11 devices
#ifndef CK_SKIP_FLAKY_F8_TEST
// convert minimal float to bf8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_rne<bf8_t>(std::numeric_limits<float>::min())),
type_convert<float>(f8_convert_rne<bf8_fnuz_t>(std::numeric_limits<float>::min())),
abs_tol);
#endif
// convert maximal bf8_t to float and check if equal to 57344.0
ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_rne<bf8_t>(57344.0f)), abs_tol);
const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_fnuz_t>::Max());
// convert maximal bf8_fnuz_t to float and check if equal to 57344.0
ASSERT_NEAR(
max_bf8_t_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(max_bf8_t_float)), abs_tol);
// convert maximal float to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(57344.0f,
type_convert<float>(f8_convert_rne<bf8_t>(std::numeric_limits<float>::max())),
ASSERT_NEAR(max_bf8_t_float,
type_convert<float>(f8_convert_rne<bf8_fnuz_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to bf8_t and check if it is qNan
ASSERT_NEAR(type_convert<bf8_t>(0x80),
f8_convert_rne<bf8_t>(std::numeric_limits<float>::infinity()),
// convert inf float to bf8_fnuz_t and check if it is qNan
ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
f8_convert_rne<bf8_fnuz_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to bf8 and back, check if holds
float pos_float = 0.0000762939f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(pos_float)), abs_tol);
// negative norm float value to bf8 and back, check if holds
float neg_float = -0.0000610351f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(neg_float)), abs_tol);
// positive subnorm float value to bf8 and back, check if holds
pos_float = 0.0000305175f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(pos_float)), abs_tol);
// negative subnorm float value to bf8 and back, check if holds
neg_float = -0.0000152587f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<bf8_fnuz_t>(neg_float)), abs_tol);
}
TEST(BF8, ConvertFP32Stochastic)
TEST(BF8FNUZ, ConvertFP32Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to bf8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_t>(0.0f)), abs_tol);
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(0.0f)), abs_tol);
// convert minimal float to bf8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::min())),
type_convert<float>(f8_convert_sr<bf8_fnuz_t>(std::numeric_limits<float>::min())),
abs_tol);
// convert maximal bf8_t to float and check if equal to 57344.0
ASSERT_NEAR(57344.0f, type_convert<float>(f8_convert_sr<bf8_t>(57344.0f)), abs_tol);
const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_fnuz_t>::Max());
// convert maximal bf8_fnuz_t to float and check if equal to 57344.0
ASSERT_NEAR(
max_bf8_t_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(max_bf8_t_float)), abs_tol);
// convert maximal float to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(57344.0f,
type_convert<float>(f8_convert_sr<bf8_t>(std::numeric_limits<float>::max())),
ASSERT_NEAR(max_bf8_t_float,
type_convert<float>(f8_convert_sr<bf8_fnuz_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to bf8_t and check if it is qNan
ASSERT_NEAR(type_convert<bf8_t>(0x80),
f8_convert_sr<bf8_t>(std::numeric_limits<float>::infinity()),
// convert inf float to bf8_fnuz_t and check if it is qNan
ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
f8_convert_sr<bf8_fnuz_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to bf8 and back, check if holds
float pos_float = 0.0000762939f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(pos_float)), abs_tol);
// negative norm float value to bf8 and back, check if holds
float neg_float = -0.0000610351f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(neg_float)), abs_tol);
// positive subnorm float value to bf8 and back, check if holds
pos_float = 0.0000305175f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(pos_float)), abs_tol);
// negative subnorm float value to bf8 and back, check if holds
neg_float = -0.0000152587f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<bf8_fnuz_t>(neg_float)), abs_tol);
}
TEST(BF8, ConvertFP16Nearest)
TEST(BF8FNUZ, ConvertFP16Nearest)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to bf8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_rne<bf8_t>(half_t{0.0})), abs_tol);
ASSERT_NEAR(
half_t{0.0}, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to bf8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_rne<bf8_t>(ck::NumericLimits<half_t>::Min())),
type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal bf8_t to fp16 and check if equal to 57344.0
const auto max_bf8_t_half = type_convert<half_t>(ck::NumericLimits<bf8_fnuz_t>::Max());
// convert maximal bf8_fnuz_t to fp16 and check if equal to 57344.0
ASSERT_NEAR(
half_t{57344.0}, type_convert<half_t>(f8_convert_rne<bf8_t>(half_t{57344.0})), abs_tol);
max_bf8_t_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(max_bf8_t_half)), abs_tol);
// convert maximal fp16 to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(half_t{57344.0},
type_convert<half_t>(f8_convert_rne<bf8_t>(ck::NumericLimits<half_t>::Max())),
ASSERT_NEAR(max_bf8_t_half,
type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<bf8_t>(0x80),
f8_convert_rne<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
// convert QuietNaN fp16 to bf8_fnuz_t and check if it is QuietNaN
ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
f8_convert_rne<bf8_fnuz_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to bf8 and back, check if holds
half_t pos_half = half_t{0.0000762939};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(pos_half)), abs_tol);
// negative norm fp16 value to bf8 and back, check if holds
half_t neg_half = half_t{-0.0000610351};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to bf8 and back, check if holds
pos_half = half_t{0.0000305175};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(pos_half)), abs_tol);
// negative subnorm fp16 value to bf8 and back, check if holds
neg_half = half_t{-0.0000152587};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<bf8_fnuz_t>(neg_half)), abs_tol);
}
TEST(BF8, ConvertFP16Stochastic)
TEST(BF8FNUZ, ConvertFP16Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to bf8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{0.0})), abs_tol);
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to bf8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Min())),
type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal bf8_t to fp16 and check if equal to 57344.0
const auto max_bf8_t_half = type_convert<half_t>(ck::NumericLimits<bf8_fnuz_t>::Max());
// convert maximal bf8_fnuz_t to fp16 and check if equal to 57344.0
ASSERT_NEAR(
half_t{57344.0}, type_convert<half_t>(f8_convert_sr<bf8_t>(half_t{57344.0})), abs_tol);
max_bf8_t_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(max_bf8_t_half)), abs_tol);
// convert maximal fp16 to bf8 and back, check if clipped to 57344.0
ASSERT_NEAR(half_t{57344.0},
type_convert<half_t>(f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::Max())),
ASSERT_NEAR(max_bf8_t_half,
type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to bf8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<bf8_t>(0x80),
f8_convert_sr<bf8_t>(ck::NumericLimits<half_t>::QuietNaN()),
// convert QuietNaN fp16 to bf8_fnuz_t and check if it is QuietNaN
ASSERT_NEAR(ck::NumericLimits<bf8_fnuz_t>::QuietNaN(),
f8_convert_sr<bf8_fnuz_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to bf8 and back, check if holds
half_t pos_half = half_t{0.0000762939};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(pos_half)), abs_tol);
// negative norm fp16 value to bf8 and back, check if holds
half_t neg_half = half_t{-0.0000610351};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to bf8 and back, check if holds
pos_half = half_t{0.0000305175};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(pos_half)), abs_tol);
// negative subnorm fp16 value to bf8 and back, check if holds
neg_half = half_t{-0.0000152587};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<bf8_fnuz_t>(neg_half)), abs_tol);
}
test/data_type/test_bf8_ocp.cpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/utility/data_type.hpp"
#include "ck/utility/type_convert.hpp"
using ck::bf8_ocp_t;
using ck::f8_convert_rne;
using ck::f8_convert_sr;
using ck::half_t;
using ck::type_convert;
TEST(BF8OCP, NumericLimits)
{ // constants given for OCP FP8
EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::Min(),
type_convert<bf8_ocp_t>(0x04)); // 0b00000100 = 2^-14
EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::Max(),
type_convert<bf8_ocp_t>(0x7B)); // 0b01111011 = 57344
EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::Lowest(),
type_convert<bf8_ocp_t>(0xFB)); // 0b11111011 = -57344
EXPECT_EQ(ck::NumericLimits<bf8_ocp_t>::QuietNaN().data,
type_convert<bf8_ocp_t>(0x7D).data); // 0b01111101
EXPECT_FALSE(ck::NumericLimits<bf8_ocp_t>::QuietNaN() ==
ck::NumericLimits<bf8_ocp_t>::QuietNaN());
EXPECT_TRUE(ck::fp8_is_inf(type_convert<bf8_ocp_t>(0xFC)) &&
ck::fp8_is_inf(type_convert<bf8_ocp_t>(0x7C)));
}
TEST(BF8OCP, ConvertFP32Nearest)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to bfp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<bf8_ocp_t>(0.0f)), 0.0f);
// convert minimal float to bf8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::min())),
abs_tol);
const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_ocp_t>::Max());
// convert maximal bf8_ocp_t to float and check if equal to bf8 max
ASSERT_NEAR(
max_bf8_t_float, type_convert<float>(f8_convert_rne<bf8_ocp_t>(max_bf8_t_float)), 0.0f);
// convert maximal float to bf8 and back, check if clipped to bf8 max (saturation to finite)
ASSERT_NEAR(max_bf8_t_float,
type_convert<float>(f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::max())),
0.0f);
// convert float infinity to bf8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(ck::NumericLimits<bf8_ocp_t>::Max(),
f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::infinity()));
// positive normal float value to bf8 and back, check if holds
float pos_float = 0.0000762939f; // 10*2^-17
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<bf8_ocp_t>(pos_float)), abs_tol);
// negative smallest normal bf8 value to bf8 and back, check if holds
constexpr auto neg_min_bf8 = -0.00006103515625f; //-2^-14
ASSERT_NEAR(neg_min_bf8, type_convert<float>(f8_convert_rne<bf8_ocp_t>(neg_min_bf8)), 0.0f);
// positive subnorm float value to bf8 and back, check if holds
constexpr auto pos_subnorm_bf8 = 0.000030517578125f; // 2^-15
ASSERT_NEAR(
pos_subnorm_bf8, type_convert<float>(f8_convert_rne<bf8_ocp_t>(pos_subnorm_bf8)), 0.0f);
// min subnorm bf8 value to bf8 and back, check if holds
constexpr auto min_subnorm_bf8 = -0.0000152587890625f; //-2^-16
ASSERT_NEAR(
min_subnorm_bf8, type_convert<float>(f8_convert_rne<bf8_ocp_t>(min_subnorm_bf8)), 0.0f);
// smaller than min subnorm bf8 value to bf8 must be zero
constexpr auto less_than_min_subnorm = 0.00000762939453125f; // 2^-17
ASSERT_EQ(0.0f, type_convert<float>(f8_convert_rne<bf8_ocp_t>(less_than_min_subnorm)));
// convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
const auto bf8_nan = f8_convert_rne<bf8_ocp_t>(std::numeric_limits<float>::quiet_NaN());
ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
}
TEST(BF8OCP, ConvertFP32Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to bfp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<bf8_ocp_t>(0.0f)), 0.0f);
// convert minimal float to bf8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::min())),
abs_tol);
const auto max_bf8_t_float = type_convert<float>(ck::NumericLimits<bf8_ocp_t>::Max());
// convert maximal bf8_ocp_t to float and check if equal to bf8 max
ASSERT_NEAR(
max_bf8_t_float, type_convert<float>(f8_convert_sr<bf8_ocp_t>(max_bf8_t_float)), 0.0f);
// convert maximal float to bf8 and back, check if clipped to bf8 max (saturation to finite)
ASSERT_NEAR(max_bf8_t_float,
type_convert<float>(f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::max())),
0.0f);
// convert float infinity to bf8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(ck::NumericLimits<bf8_ocp_t>::Max(),
f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::infinity()));
// positive normal float value to bf8 and back, check if holds
float pos_float = 0.0000762939f; // 10*2^-17
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<bf8_ocp_t>(pos_float)), abs_tol);
// negative smallest normal bf8 value to bf8 and back, check if holds
constexpr auto neg_min_bf8 = -0.00006103515625f; //-2^-14
ASSERT_NEAR(neg_min_bf8, type_convert<float>(f8_convert_sr<bf8_ocp_t>(neg_min_bf8)), 0.0f);
// positive subnorm float value to bf8 and back, check if holds
constexpr auto pos_subnorm_bf8 = 0.000030517578125f; // 2^-15
ASSERT_NEAR(
pos_subnorm_bf8, type_convert<float>(f8_convert_sr<bf8_ocp_t>(pos_subnorm_bf8)), 0.0f);
// min subnorm bf8 value to bf8 and back, check if holds
constexpr auto min_subnorm_bf8 = -0.0000152587890625f; //-2^-16
ASSERT_NEAR(
min_subnorm_bf8, type_convert<float>(f8_convert_sr<bf8_ocp_t>(min_subnorm_bf8)), 0.0f);
// smaller than min subnorm bf8 value to bf8 alternates between 0 and 2^-16
constexpr auto less_than_min_subnorm = 0.00000762939453125f; // 2^-17
ASSERT_NEAR(0.0f,
type_convert<float>(f8_convert_sr<bf8_ocp_t>(less_than_min_subnorm)),
0.0000152587890625f);
// convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
const auto bf8_nan = f8_convert_sr<bf8_ocp_t>(std::numeric_limits<float>::quiet_NaN());
ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
}
TEST(BF8OCP, ConvertFP16Nearest)
{
// fix the tolerance value
constexpr half_t half_t_tol = 1e-3;
constexpr half_t half_t_zero = 0.0;
// convert 0 half_t to bfp8 and back, check if holds
ASSERT_NEAR(
half_t_zero, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(half_t_zero)), half_t_zero);
// convert minimal half_t to bf8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(ck::NumericLimits<half_t>::Min())),
half_t_tol);
const auto max_bf8_t_half_t = type_convert<half_t>(ck::NumericLimits<bf8_ocp_t>::Max());
// convert maximal bf8_ocp_t to half_t and check if equal to bf8 max
ASSERT_NEAR(max_bf8_t_half_t,
type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(max_bf8_t_half_t)),
half_t_zero);
// convert maximal half_t to bf8 and back, check if clipped to bf8 max (saturation to finite)
ASSERT_NEAR(max_bf8_t_half_t,
type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(ck::NumericLimits<half_t>::Max())),
half_t_zero);
// convert half_t infinity to bf8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(
ck::NumericLimits<bf8_ocp_t>::Max(),
f8_convert_rne<bf8_ocp_t>(type_convert<half_t>(std::numeric_limits<float>::infinity())));
// positive normal bf8 value to bf8 and back, check if holds
constexpr half_t pos_norm_bf8{0.0000762939f}; // 10*2^-17
ASSERT_NEAR(
pos_norm_bf8, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(pos_norm_bf8)), half_t_tol);
// negative smallest normal bf8 value to bf8 and back, check if holds
constexpr half_t neg_min_bf8{-0.00006103515625f}; //-2^-14
ASSERT_NEAR(
neg_min_bf8, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(neg_min_bf8)), half_t_zero);
// positive subnorm bf8 value to bf8 and back, check if holds
constexpr half_t pos_subnorm_bf8{0.000030517578125f}; // 2^-15
ASSERT_NEAR(pos_subnorm_bf8,
type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(pos_subnorm_bf8)),
half_t_zero);
// min subnorm bf8 value to bf8 and back, check if holds
constexpr half_t min_subnorm_bf8{-0.0000152587890625f}; //-2^-16
ASSERT_NEAR(min_subnorm_bf8,
type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(min_subnorm_bf8)),
half_t_zero);
// smaller than min subnorm bf8 value to bf8 must be zero
constexpr half_t less_than_min_subnorm{0.00000762939453125f}; // 2^-17
ASSERT_EQ(half_t_zero, type_convert<half_t>(f8_convert_rne<bf8_ocp_t>(less_than_min_subnorm)));
// convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
const auto bf8_nan = f8_convert_rne<bf8_ocp_t>(ck::NumericLimits<half_t>::QuietNaN());
ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
}
TEST(BF8OCP, ConvertFP16Stochastic)
{
// fix the tolerance value
constexpr half_t half_t_tol = 1e-3;
constexpr half_t half_t_zero = 0.0;
constexpr auto min_subnorm_bf8 = 0.0000152587890625f; // 2^-16
// convert 0 half_t to bfp8 and back, check if holds
ASSERT_NEAR(
half_t_zero, type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(half_t_zero)), half_t_zero);
// convert minimal half_t (6.103515625e-05) to fp8 and back
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(ck::NumericLimits<half_t>::Min())),
half_t_zero);
const auto max_bf8_t_half_t = type_convert<half_t>(ck::NumericLimits<bf8_ocp_t>::Max());
// convert maximal bf8_ocp_t to half_t and check if equal to bf8 max
ASSERT_NEAR(max_bf8_t_half_t,
type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(max_bf8_t_half_t)),
half_t_zero);
// convert maximal half_t to bf8 and back, check if clipped to bf8 max (saturation to finite)
ASSERT_NEAR(max_bf8_t_half_t,
type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(ck::NumericLimits<half_t>::Max())),
half_t_zero);
// convert half_t infinity to bf8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(
ck::NumericLimits<bf8_ocp_t>::Max(),
f8_convert_sr<bf8_ocp_t>(type_convert<half_t>(std::numeric_limits<float>::infinity())));
// positive normal bf8 value to bf8 and back, check if holds
constexpr half_t pos_norm_bf8{0.0000762939f}; // 10*2^-17
ASSERT_NEAR(
pos_norm_bf8, type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(pos_norm_bf8)), half_t_tol);
// negative smallest normal bf8 value to bf8 and back, check if holds
constexpr half_t neg_min_bf8{-0.00006103515625f}; //-2^-14
ASSERT_NEAR(
neg_min_bf8, type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(neg_min_bf8)), half_t_zero);
// positive subnorm bf8 value to bf8 and back, check if holds
constexpr half_t pos_subnorm_bf8{0.000030517578125f}; // 2^-15
ASSERT_NEAR(pos_subnorm_bf8,
type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(pos_subnorm_bf8)),
half_t_zero);
// min subnorm bf8 value to bf8 and back, check if holds
ASSERT_NEAR(half_t{-min_subnorm_bf8},
type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(half_t{-min_subnorm_bf8})),
half_t_zero);
// smaller than min subnorm bf8 value to bf8 alternates between 0 and 2^-16
constexpr half_t less_than_min_subnorm{0.00000762939453125f}; // 2^-17
ASSERT_NEAR(half_t_zero,
type_convert<half_t>(f8_convert_sr<bf8_ocp_t>(less_than_min_subnorm)),
half_t{min_subnorm_bf8});
// convert quiet NaN to bf8_ocp_t and check if it is quiet NaN
const auto bf8_nan = f8_convert_sr<bf8_ocp_t>(ck::NumericLimits<half_t>::QuietNaN());
ASSERT_TRUE(ck::fp8_impl::ocp_bf8_is_nan(bf8_nan.data));
}
test/data_type/test_custom_type.cpp
View file @ c8c016dd
......@@ -51,8 +51,11 @@ TEST(Custom_bool, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bool_t>()(Number<i>{}) = custom_bool_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bool_t, size> left_vec{right_vec};
vector_type<custom_bool_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_bool_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bool_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -129,8 +132,11 @@ TEST(Custom_int8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_int8_t>()(Number<i>{}) = custom_int8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_int8_t, size> left_vec{right_vec};
vector_type<custom_int8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_int8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_int8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -207,8 +213,11 @@ TEST(Custom_uint8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_uint8_t>()(Number<i>{}) = custom_uint8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_uint8_t, size> left_vec{right_vec};
vector_type<custom_uint8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_uint8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -287,8 +296,11 @@ TEST(Custom_f8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_f8_t>()(Number<i>{}) = custom_f8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_f8_t, size> left_vec{right_vec};
vector_type<custom_f8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_f8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_f8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -369,8 +381,11 @@ TEST(Custom_bf8, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bf8_t>()(Number<i>{}) = custom_bf8_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bf8_t, size> left_vec{right_vec};
vector_type<custom_bf8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_bf8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -450,8 +465,11 @@ TEST(Custom_half, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_half_t>()(Number<i>{}) = custom_half_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_half_t, size> left_vec{right_vec};
vector_type<custom_half_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_half_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_half_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -533,8 +551,11 @@ TEST(Custom_bhalf, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_bhalf_t>()(Number<i>{}) = custom_bhalf_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_bhalf_t, size> left_vec{right_vec};
vector_type<custom_bhalf_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_bhalf_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -615,8 +636,11 @@ TEST(Custom_float, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_float_t>()(Number<i>{}) = custom_float_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_float_t, size> left_vec{right_vec};
vector_type<custom_float_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_float_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_float_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -693,8 +717,11 @@ TEST(Custom_double, TestAsType)
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<custom_double_t>()(Number<i>{}) = custom_double_t{test_vec.at(i)};
});
// copy the vector
vector_type<custom_double_t, size> left_vec{right_vec};
vector_type<custom_double_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<custom_double_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<custom_double_t>()(Number<i>{}).data, test_vec.at(i));
......@@ -813,8 +840,11 @@ TEST(Complex_half, TestAsType)
right_vec.template AsType<complex_half_t>()(Number<i>{}) =
complex_half_t{test_vec.at(num_elem * i), test_vec.at(num_elem * i + 1)};
});
// copy the vector
vector_type<complex_half_t, size> left_vec{right_vec};
vector_type<complex_half_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<complex_half_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).real,
......@@ -872,3 +902,167 @@ TEST(Complex_half, TestAsTypeReshape)
test_vec.at(num_elem * i + 1));
});
}
#if CK_USE_OCP_FP8
TEST(FP8OCP, TestSize)
{
static_assert(std::is_same_v<f8_t, ck::f8_ocp_t>, "OCP FP8 is not enabled");
ASSERT_EQ(sizeof(f8_t), sizeof(ck::fp8_storage_t));
ASSERT_EQ(sizeof(vector_type<f8_t, 2>), sizeof(vector_type<ck::fp8_storage_t, 2>));
ASSERT_EQ(sizeof(vector_type<f8_t, 4>), sizeof(vector_type<ck::fp8_storage_t, 4>));
ASSERT_EQ(sizeof(vector_type<f8_t, 8>), sizeof(vector_type<ck::fp8_storage_t, 8>));
ASSERT_EQ(sizeof(vector_type<f8_t, 16>), sizeof(vector_type<ck::fp8_storage_t, 16>));
ASSERT_EQ(sizeof(vector_type<f8_t, 32>), sizeof(vector_type<ck::fp8_storage_t, 32>));
ASSERT_EQ(sizeof(vector_type<f8_t, 64>), sizeof(vector_type<ck::fp8_storage_t, 64>));
}
TEST(FP8OCP, TestAsType)
{
static_assert(std::is_same_v<f8_t, ck::f8_ocp_t>, "OCP FP8 is not enabled");
// test size
std::array<float, 8> test_vec = {-4, -2, -0.5, -0.25, 1.0 / 8.0, 1, 1.5, 16};
constexpr int size = test_vec.size();
// reference vector
vector_type<f8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<f8_t>()(Number<i>{}), f8_t{0}); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<f8_t>()(Number<i>{}) = ck::type_convert<f8_t>(test_vec.at(i));
});
vector_type<f8_t, size> left_vec;
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<f8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<f8_t>()(Number<i>{}),
ck::type_convert<f8_t>(test_vec.at(i)));
});
ck::non_native_vector_base<ck::f8_ocp_t, 2> nnvb_f8x2(ck::type_convert<f8_t>(-10.0f));
ASSERT_EQ(nnvb_f8x2.template AsType<f8_t>()(Number<0>{}), ck::type_convert<f8_t>(-10.0f));
ASSERT_EQ(nnvb_f8x2.template AsType<f8_t>()(Number<1>{}), ck::type_convert<f8_t>(-10.0f));
}
TEST(FP8OCP, TestAsTypeReshape)
{
static_assert(std::is_same_v<f8_t, ck::f8_ocp_t>, "OCP FP8 is not enabled");
// test size
std::array<float, 8> test_vec = {-8, -0.5, -0.25, 1.0 / 8.0, 1 / 256, 1, 1.5, 16};
constexpr int size = test_vec.size();
// reference vector
vector_type<f8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<f8_t>()(Number<i>{}), f8_t{0}); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<f8_t>()(Number<i>{}) = ck::type_convert<f8_t>(test_vec.at(i));
});
// copy the first half of a vector
vector_type<f8_t, size / 2> left_vec{
right_vec.template AsType<vector_type<f8_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<f8_t>()(Number<i>{}),
ck::type_convert<f8_t>(test_vec.at(i)));
});
}
TEST(BF8OCP, TestSize)
{
static_assert(std::is_same_v<bf8_t, ck::bf8_ocp_t>, "OCP BF8 is not enabled");
ASSERT_EQ(sizeof(bf8_t), sizeof(ck::fp8_storage_t));
ASSERT_EQ(sizeof(vector_type<bf8_t, 2>), sizeof(vector_type<ck::fp8_storage_t, 2>));
ASSERT_EQ(sizeof(vector_type<bf8_t, 4>), sizeof(vector_type<ck::fp8_storage_t, 4>));
ASSERT_EQ(sizeof(vector_type<bf8_t, 8>), sizeof(vector_type<ck::fp8_storage_t, 8>));
ASSERT_EQ(sizeof(vector_type<bf8_t, 16>), sizeof(vector_type<ck::fp8_storage_t, 16>));
ASSERT_EQ(sizeof(vector_type<bf8_t, 32>), sizeof(vector_type<ck::fp8_storage_t, 32>));
ASSERT_EQ(sizeof(vector_type<bf8_t, 64>), sizeof(vector_type<ck::fp8_storage_t, 64>));
}
TEST(BF8OCP, TestAsType)
{
static_assert(std::is_same_v<bf8_t, ck::bf8_ocp_t>, "OCP BF8 is not enabled");
// test size
std::array<float, 8> test_vec = {-4, -2, -0.5, -0.25, 1.0 / 8.0, 1, 1.5, 16};
constexpr int size = test_vec.size();
// reference vector
vector_type<bf8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<bf8_t>()(Number<i>{}), bf8_t{0}); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<bf8_t>()(Number<i>{}) = ck::type_convert<bf8_t>(test_vec.at(i));
});
vector_type<bf8_t, size> left_vec{right_vec};
// check copy assignment op
left_vec = right_vec;
// overwrite right_vec with 0s
right_vec = vector_type<bf8_t, size>{};
// check if values were copied correctly
ck::static_for<0, size, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<bf8_t>()(Number<i>{}),
ck::type_convert<bf8_t>(test_vec.at(i)));
});
ck::non_native_vector_base<bf8_t, 2> nnvb_bf8x2(ck::type_convert<bf8_t>(-10.0f));
ASSERT_EQ(nnvb_bf8x2.template AsType<bf8_t>()(Number<0>{}), ck::type_convert<bf8_t>(-10.0f));
ASSERT_EQ(nnvb_bf8x2.template AsType<bf8_t>()(Number<1>{}), ck::type_convert<bf8_t>(-10.0f));
}
TEST(BF8OCP, TestAsTypeReshape)
{
static_assert(std::is_same_v<bf8_t, ck::bf8_ocp_t>, "OCP BF8 is not enabled");
// test size
std::array<float, 8> test_vec = {-8, -0.5, -0.25, 1.0 / 8.0, 1 / 256, 1, 1.5, 16};
constexpr int size = test_vec.size();
// reference vector
vector_type<bf8_t, size> right_vec;
// check default CTOR
ck::static_for<0, size, 1>{}(
[&](auto i) { ASSERT_EQ(right_vec.template AsType<bf8_t>()(Number<i>{}), bf8_t{0}); });
// assign test values to the vector
ck::static_for<0, size, 1>{}([&](auto i) {
right_vec.template AsType<bf8_t>()(Number<i>{}) = ck::type_convert<bf8_t>(test_vec.at(i));
});
// copy the first half of a vector
vector_type<bf8_t, size / 2> left_vec{
right_vec.template AsType<vector_type<bf8_t, size / 2>::type>()(Number<0>{})};
// check if values were copied correctly
ck::static_for<0, size / 2, 1>{}([&](auto i) {
ASSERT_EQ(left_vec.template AsType<bf8_t>()(Number<i>{}),
ck::type_convert<bf8_t>(test_vec.at(i)));
});
}
#endif
test/data_type/test_fp8.cpp → test/data_type/test_fp8_fnuz.cpp
View file @ c8c016dd
......@@ -7,154 +7,171 @@
using ck::f8_convert_rne;
using ck::f8_convert_sr;
using ck::f8_t;
using ck::f8_fnuz_t;
using ck::half_t;
using ck::type_convert;
TEST(FP8, NumericLimits)
TEST(FP8FNUZ, NumericLimits)
{
// constants given for negative zero nan mode
EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), type_convert<f8_t>(0x08));
EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), type_convert<f8_t>(0x7F));
EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), type_convert<f8_t>(0xFF));
EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), type_convert<f8_t>(0x80));
EXPECT_EQ(ck::NumericLimits<f8_fnuz_t>::Min(), type_convert<f8_fnuz_t>(0x08));
EXPECT_EQ(ck::NumericLimits<f8_fnuz_t>::Max(), type_convert<f8_fnuz_t>(0x7F));
EXPECT_EQ(ck::NumericLimits<f8_fnuz_t>::Lowest(), type_convert<f8_fnuz_t>(0xFF));
EXPECT_EQ(ck::NumericLimits<f8_fnuz_t>::QuietNaN(), type_convert<f8_fnuz_t>(0x80));
}
TEST(FP8, ConvertFP32Nearest)
TEST(FP8FNUZ, ConvertFP32Nearest)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<f8_t>(0.0f)), abs_tol);
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<f8_fnuz_t>(0.0f)), abs_tol);
// don't run the next test on gfx11 devices
#ifndef CK_SKIP_FLAKY_F8_TEST
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_rne<f8_t>(std::numeric_limits<float>::min())),
type_convert<float>(f8_convert_rne<f8_fnuz_t>(std::numeric_limits<float>::min())),
abs_tol);
#endif
// convert maximal f8_t to float and check if equal to 240.0
ASSERT_NEAR(240.0f, type_convert<float>(f8_convert_rne<f8_t>(240.0f)), abs_tol);
// convert maximal float to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(240.0f,
type_convert<float>(f8_convert_rne<f8_t>(std::numeric_limits<float>::max())),
const auto max_f8_t_float = type_convert<float>(ck::NumericLimits<f8_fnuz_t>::Max());
// convert maximal f8_fnuz_t to float and check if equal to fp8 max
ASSERT_NEAR(
max_f8_t_float, type_convert<float>(f8_convert_rne<f8_fnuz_t>(max_f8_t_float)), abs_tol);
// XXX: FNUZ f8_convert_rne behavior is inconsistent.
// Clipping large values to fp8 max (saturation to finite) contradicts converting inf float to
// fp8 qNAN (no saturation).
// convert maximal float to fp8 and back, check if clipped to fp8 max
ASSERT_NEAR(max_f8_t_float,
type_convert<float>(f8_convert_rne<f8_fnuz_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(type_convert<f8_t>(0x80),
f8_convert_rne<f8_t>(std::numeric_limits<float>::infinity()),
// convert inf float to f8_fnuz_t and check if it is qNan
ASSERT_NEAR(ck::NumericLimits<f8_fnuz_t>::QuietNaN(),
f8_convert_rne<f8_fnuz_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to fp8 and back, check if holds
float pos_float = 0.017578125f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_fnuz_t>(pos_float)), abs_tol);
// negative norm float value to fp8 and back, check if holds
float neg_float = -0.015625f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_fnuz_t>(neg_float)), abs_tol);
// positive subnorm float value to fp8 and back, check if holds
pos_float = 0.00390625f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_fnuz_t>(pos_float)), abs_tol);
// negative subnorm float value to fp8 and back, check if holds
neg_float = -0.001953125f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_fnuz_t>(neg_float)), abs_tol);
}
TEST(FP8, ConvertFP32Stochastic)
TEST(FP8FNUZ, ConvertFP32Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<f8_t>(0.0f)), abs_tol);
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<f8_fnuz_t>(0.0f)), abs_tol);
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::min())),
type_convert<float>(f8_convert_sr<f8_fnuz_t>(std::numeric_limits<float>::min())),
abs_tol);
// convert maximal f8_t to float and check if equal to 240.0
ASSERT_NEAR(240.0f, type_convert<float>(f8_convert_sr<f8_t>(240.0f)), abs_tol);
// convert maximal float to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(240.0f,
type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::max())),
const auto max_f8_t_float = type_convert<float>(ck::NumericLimits<f8_fnuz_t>::Max());
// convert maximal f8_fnuz_t to float and check if equal to fp8 max
ASSERT_NEAR(
max_f8_t_float, type_convert<float>(f8_convert_sr<f8_fnuz_t>(max_f8_t_float)), abs_tol);
// convert maximal float to fp8 and back, check if clipped to fp8 max
ASSERT_NEAR(max_f8_t_float,
type_convert<float>(f8_convert_sr<f8_fnuz_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(type_convert<f8_t>(0x80),
f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()),
// convert inf float to f8_fnuz_t and check if it is qNan
ASSERT_NEAR(ck::NumericLimits<f8_fnuz_t>::QuietNaN(),
f8_convert_sr<f8_fnuz_t>(std::numeric_limits<float>::infinity()),
abs_tol);
// positive norm float value to fp8 and back, check if holds
float pos_float = 0.017578125f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_fnuz_t>(pos_float)), abs_tol);
// negative norm float value to fp8 and back, check if holds
float neg_float = -0.015625f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_fnuz_t>(neg_float)), abs_tol);
// positive subnorm float value to fp8 and back, check if holds
pos_float = 0.00390625f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_fnuz_t>(pos_float)), abs_tol);
// negative subnorm float value to fp8 and back, check if holds
neg_float = -0.001953125f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_fnuz_t>(neg_float)), abs_tol);
}
TEST(FP8, ConvertFP16Nearest)
TEST(FP8FNUZ, ConvertFP16Nearest)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to fp8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_rne<f8_t>(half_t{0.0})), abs_tol);
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to fp8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_rne<f8_t>(ck::NumericLimits<half_t>::Min())),
type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal f8_t to fp16 and check if equal to 240.0
ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(f8_convert_rne<f8_t>(half_t{240.0})), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(half_t{240.0},
type_convert<half_t>(f8_convert_rne<f8_t>(ck::NumericLimits<half_t>::Max())),
const auto max_f8_t_half = type_convert<half_t>(ck::NumericLimits<f8_fnuz_t>::Max());
// convert maximal f8_fnuz_t to fp16 and check if equal to fp8 max
ASSERT_NEAR(
max_f8_t_half, type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(max_f8_t_half)), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to fp8 max
ASSERT_NEAR(max_f8_t_half,
type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<f8_t>(0x80),
f8_convert_rne<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
// convert QuietNaN fp16 to f8_fnuz_t and check if it is QuietNaN
ASSERT_NEAR(ck::NumericLimits<f8_fnuz_t>::QuietNaN(),
f8_convert_rne<f8_fnuz_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.017578125};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<f8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(pos_half)), abs_tol);
// negative norm fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.015625};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<f8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to fp8 and back, check if holds
pos_half = half_t{0.00390625};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<f8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(pos_half)), abs_tol);
// negative subnorm fp16 value to fp8 and back, check if holds
neg_half = half_t{-0.001953125};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<f8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_rne<f8_fnuz_t>(neg_half)), abs_tol);
}
TEST(FP8, ConvertFP16Stochastic)
TEST(FP8FNUZ, ConvertFP16Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to fp8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<f8_t>(half_t{0.0})), abs_tol);
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to fp8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Min())),
type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal f8_t to fp16 and check if equal to 240.0
ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(f8_convert_sr<f8_t>(half_t{240.0})), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(half_t{240.0},
type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Max())),
const auto max_f8_t_half = type_convert<half_t>(ck::NumericLimits<f8_fnuz_t>::Max());
// convert maximal f8_fnuz_t to fp16 and check if equal to fp8 max
ASSERT_NEAR(
max_f8_t_half, type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(max_f8_t_half)), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to fp8 max
ASSERT_NEAR(max_f8_t_half,
type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(type_convert<f8_t>(0x80),
f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()),
// convert QuietNaN fp16 to f8_fnuz_t and check if it is QuietNaN
ASSERT_NEAR(ck::NumericLimits<f8_fnuz_t>::QuietNaN(),
f8_convert_sr<f8_fnuz_t>(ck::NumericLimits<half_t>::QuietNaN()),
abs_tol);
// positive norm fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.017578125};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(pos_half)), abs_tol);
// negative norm fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.015625};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(neg_half)), abs_tol);
// positive subnorm fp16 value to fp8 and back, check if holds
pos_half = half_t{0.00390625};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(pos_half)), abs_tol);
// negative subnorm fp16 value to fp8 and back, check if holds
neg_half = half_t{-0.001953125};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_fnuz_t>(neg_half)), abs_tol);
}
test/data_type/test_fp8_ocp.cpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/utility/data_type.hpp"
#include "ck/utility/type_convert.hpp"
using ck::f8_convert_rne;
using ck::f8_convert_sr;
using ck::f8_ocp_t;
using ck::half_t;
using ck::type_convert;
TEST(FP8OCP, NumericLimits)
{
// constants given for OCP FP8
EXPECT_EQ(ck::NumericLimits<f8_ocp_t>::Min(),
type_convert<f8_ocp_t>(0x08)); // 0b00001000 = 2^-6
EXPECT_EQ(ck::NumericLimits<f8_ocp_t>::Max(), type_convert<f8_ocp_t>(0x7E)); // 0b01111110 = 448
EXPECT_EQ(ck::NumericLimits<f8_ocp_t>::Lowest(),
type_convert<f8_ocp_t>(0xFE)); // 0b11111110 = -448
EXPECT_EQ(ck::NumericLimits<f8_ocp_t>::QuietNaN().data,
type_convert<f8_ocp_t>(0x7F).data); // 0b01111111
EXPECT_FALSE(ck::NumericLimits<f8_ocp_t>::QuietNaN() ==
ck::NumericLimits<f8_ocp_t>::QuietNaN());
}
TEST(FP8OCP, ConvertFP32Nearest)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_rne<f8_ocp_t>(0.0f)), 0.0f);
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_rne<f8_ocp_t>(std::numeric_limits<float>::min())),
abs_tol);
const auto max_f8_t_float = type_convert<float>(ck::NumericLimits<f8_ocp_t>::Max());
// convert maximal f8_ocp_t to float and check if equal to fp8 max
ASSERT_NEAR(
max_f8_t_float, type_convert<float>(f8_convert_rne<f8_ocp_t>(max_f8_t_float)), 0.0f);
// convert maximal float to fp8 and back, check if clipped to fp8 max (saturation to finite)
ASSERT_NEAR(max_f8_t_float,
type_convert<float>(f8_convert_rne<f8_ocp_t>(std::numeric_limits<float>::max())),
0.0f);
// convert float infinity to f8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(ck::NumericLimits<f8_ocp_t>::Max(),
f8_convert_rne<f8_ocp_t>(std::numeric_limits<float>::infinity()));
// positive norm float value to fp8 and back, check if holds
float pos_float = 0.017578125f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_ocp_t>(pos_float)), abs_tol);
// smallest normal fp8 value to fp8 and back, check if holds
float neg_float = -0.015625f; //-2^-6
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_ocp_t>(neg_float)), 0.0f);
// positive subnorm float value to fp8 and back, check if holds
pos_float = 0.00390625f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_rne<f8_ocp_t>(pos_float)), abs_tol);
// min subnorm fp8 value to fp8 and back, check if holds
neg_float = -0.001953125f; //-2^-9
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_rne<f8_ocp_t>(neg_float)), 0.0f);
// smaller than min subnorm fp8 value to fp8 must be zero
auto less_than_min_subnorm = 0.0009765625f; // 2^-10
ASSERT_EQ(0.0f, type_convert<float>(f8_convert_rne<f8_ocp_t>(less_than_min_subnorm)));
// convert quiet NaN to f8_ocp_t and check if it is quiet NaN
auto f8_nan = f8_convert_rne<f8_ocp_t>(std::numeric_limits<float>::quiet_NaN());
ASSERT_TRUE((f8_nan.data & 0x7f) == 0x7f);
}
TEST(FP8OCP, ConvertFP32Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<f8_ocp_t>(0.0f)), 0.0f);
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_sr<f8_ocp_t>(std::numeric_limits<float>::min())),
abs_tol);
const auto max_f8_t_float = type_convert<float>(ck::NumericLimits<f8_ocp_t>::Max());
// convert maximal f8_ocp_t to float and check if equal to fp8 max
ASSERT_NEAR(max_f8_t_float, type_convert<float>(f8_convert_sr<f8_ocp_t>(max_f8_t_float)), 0.0f);
// convert maximal float to fp8 and back, check if clipped to fp8 max (saturation to finite)
ASSERT_NEAR(max_f8_t_float,
type_convert<float>(f8_convert_sr<f8_ocp_t>(std::numeric_limits<float>::max())),
0.0f);
// convert float infinity to f8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(ck::NumericLimits<f8_ocp_t>::Max(),
f8_convert_sr<f8_ocp_t>(std::numeric_limits<float>::infinity()));
// positive norm float value to fp8 and back, check if holds
float pos_float = 0.017578125f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_ocp_t>(pos_float)), abs_tol);
// smallest normal fp8 value to fp8 and back, check if holds
float neg_float = -0.015625f; //-2^-6
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_ocp_t>(neg_float)), 0.0f);
// positive subnorm float value to fp8 and back, check if holds
pos_float = 0.00390625f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_ocp_t>(pos_float)), abs_tol);
// min subnorm fp8 value to fp8 and back, check if holds
constexpr auto min_subnorm_fp8 = -0.001953125f; //-2^-9
ASSERT_NEAR(
min_subnorm_fp8, type_convert<float>(f8_convert_sr<f8_ocp_t>(min_subnorm_fp8)), 0.0f);
// smaller than min subnorm fp8 value to fp8 alternates between 0 and 2^-9
auto less_than_min_subnorm = 0.0009765625f; // 2^-10
ASSERT_NEAR(
0.0f, type_convert<float>(f8_convert_sr<f8_ocp_t>(less_than_min_subnorm)), 0.001953125f);
// convert quiet NaN to f8_ocp_t and check if it is quiet NaN
auto f8_nan = f8_convert_sr<f8_ocp_t>(std::numeric_limits<float>::quiet_NaN());
ASSERT_TRUE((f8_nan.data & 0x7f) == 0x7f);
}
TEST(FP8OCP, ConvertFP16Nearest)
{
// fix the tolerance value
constexpr half_t half_t_tol = 1e-3;
constexpr half_t half_t_zero = 0.0;
// convert 0 half_t to fp8 and back, check if holds
ASSERT_NEAR(
half_t_zero, type_convert<half_t>(f8_convert_rne<f8_ocp_t>(half_t_zero)), half_t_zero);
// convert minimal half_t to fp8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_rne<f8_ocp_t>(ck::NumericLimits<half_t>::Min())),
half_t_tol);
const auto max_f8_t_half_t = type_convert<half_t>(ck::NumericLimits<f8_ocp_t>::Max());
// convert maximal f8_ocp_t to half_t and check if equal to fp8 max
ASSERT_NEAR(max_f8_t_half_t,
type_convert<half_t>(f8_convert_rne<f8_ocp_t>(max_f8_t_half_t)),
half_t_zero);
// convert maximal half_t to fp8 and back, check if clipped to fp8 max (saturation to finite)
ASSERT_NEAR(max_f8_t_half_t,
type_convert<half_t>(f8_convert_rne<f8_ocp_t>(ck::NumericLimits<half_t>::Max())),
half_t_zero);
// convert half_t infinity to f8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(
ck::NumericLimits<f8_ocp_t>::Max(),
f8_convert_rne<f8_ocp_t>(type_convert<half_t>(std::numeric_limits<float>::infinity())));
// positive norm half_t value to fp8 and back, check if holds
half_t pos_half_t{0.017578125f};
ASSERT_NEAR(pos_half_t, type_convert<half_t>(f8_convert_rne<f8_ocp_t>(pos_half_t)), half_t_tol);
// smallest normal fp8 value to fp8 and back, check if holds
half_t neg_half_t{-0.015625f}; //-2^-6
ASSERT_NEAR(
neg_half_t, type_convert<half_t>(f8_convert_rne<f8_ocp_t>(neg_half_t)), half_t_zero);
// positive subnorm half_t value to fp8 and back, check if holds
pos_half_t = half_t{0.00390625f};
ASSERT_NEAR(pos_half_t, type_convert<half_t>(f8_convert_rne<f8_ocp_t>(pos_half_t)), half_t_tol);
// min subnorm fp8 value to fp8 and back, check if holds
neg_half_t = half_t{-0.001953125f}; //-2^-9
ASSERT_NEAR(
neg_half_t, type_convert<half_t>(f8_convert_rne<f8_ocp_t>(neg_half_t)), half_t_zero);
// smaller than min subnorm fp8 value to fp8 must be zero
auto less_than_min_subnorm = half_t{0.0009765625f}; // 2^-10
ASSERT_EQ(half_t_zero, type_convert<half_t>(f8_convert_rne<f8_ocp_t>(less_than_min_subnorm)));
// convert quiet NaN to f8_ocp_t and check if it is quiet NaN
auto f8_nan = f8_convert_rne<f8_ocp_t>(ck::NumericLimits<half_t>::QuietNaN());
ASSERT_TRUE(ck::fp8_impl::ocp_f8_is_nan(f8_nan.data));
}
TEST(FP8OCP, ConvertFP16Stochastic)
{
// fix the tolerance value
constexpr half_t half_t_tol = 1e-3;
constexpr half_t half_t_zero = 0.0;
constexpr auto min_subnorm_fp8 = 0.001953125f; // 2^-9
// convert 0 half_t to fp8 and back, check if holds
ASSERT_NEAR(
half_t_zero, type_convert<half_t>(f8_convert_sr<f8_ocp_t>(half_t_zero)), half_t_zero);
// convert minimal half_t (6.103515625e-05) to fp8 and back
// alternates between 0 and 2^-9 (0.001953125)
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_sr<f8_ocp_t>(ck::NumericLimits<half_t>::Min())),
type_convert<half_t>(min_subnorm_fp8));
const auto max_f8_t_half_t = type_convert<half_t>(ck::NumericLimits<f8_ocp_t>::Max());
// convert maximal f8_ocp_t to half_t and check if equal to fp8 max
ASSERT_NEAR(max_f8_t_half_t,
type_convert<half_t>(f8_convert_sr<f8_ocp_t>(max_f8_t_half_t)),
half_t_zero);
// convert maximal half_t to fp8 and back, check if clipped to fp8 max (saturation to finite)
ASSERT_NEAR(max_f8_t_half_t,
type_convert<half_t>(f8_convert_sr<f8_ocp_t>(ck::NumericLimits<half_t>::Max())),
half_t_zero);
// convert half_t infinity to f8_ocp_t and check if it is max value (saturation to finite)
ASSERT_EQ(
ck::NumericLimits<f8_ocp_t>::Max(),
f8_convert_sr<f8_ocp_t>(type_convert<half_t>(std::numeric_limits<float>::infinity())));
// positive norm half_t value to fp8 and back, check if holds
half_t pos_half_t{0.017578125f};
ASSERT_NEAR(pos_half_t, type_convert<half_t>(f8_convert_sr<f8_ocp_t>(pos_half_t)), half_t_tol);
// smallest normal fp8 value to fp8 and back, check if holds
half_t neg_half_t{-0.015625f}; //-2^-6
ASSERT_NEAR(neg_half_t, type_convert<half_t>(f8_convert_sr<f8_ocp_t>(neg_half_t)), half_t_zero);
// positive subnorm half_t value to fp8 and back, check if holds
pos_half_t = half_t{0.00390625f};
ASSERT_NEAR(pos_half_t, type_convert<half_t>(f8_convert_sr<f8_ocp_t>(pos_half_t)), half_t_tol);
// min subnorm fp8 value to fp8 and back, check if holds
neg_half_t = half_t{-min_subnorm_fp8}; //-2^-9
ASSERT_NEAR(neg_half_t, type_convert<half_t>(f8_convert_sr<f8_ocp_t>(neg_half_t)), half_t_zero);
// smaller than min subnorm fp8 value to fp8 alternates between 0 and 2^-9
auto less_than_min_subnorm = half_t{0.0009765625f}; // 2^-10
ASSERT_NEAR(
type_convert<float>(half_t_zero),
type_convert<float>(type_convert<half_t>(f8_convert_sr<f8_ocp_t>(less_than_min_subnorm))),
min_subnorm_fp8);
// convert quiet NaN to f8_ocp_t and check if it is quiet NaN
auto f8_nan = f8_convert_sr<f8_ocp_t>(ck::NumericLimits<half_t>::QuietNaN());
ASSERT_TRUE(ck::fp8_impl::ocp_f8_is_nan(f8_nan.data));
}
test/grouped_convnd_bwd_data/CMakeLists.txt
View file @ c8c016dd
add_gtest_executable(test_grouped_convnd_bwd_data test_grouped_convnd_bwd_data_xdl_wmma.cpp)
add_gtest_executable(test_grouped_convnd_bwd_data_xdl test_grouped_convnd_bwd_data_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_data PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
target_link_libraries(test_grouped_convnd_bwd_data_xdl PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
endif()
add_gtest_executable(test_grouped_convnd_bwd_data_wmma test_grouped_convnd_bwd_data_wmma.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_data_wmma PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
endif()
add_gtest_executable(test_grouped_convnd_bwd_data_interface_xdl test_grouped_convnd_bwd_data_interface_xdl.cpp)
if(result EQUAL 0)
......
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_wmma.cpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdDataWmma : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
using OutLayout = std::tuple_element_t<1, Tuple>;
using WeiLayout = std::tuple_element_t<2, Tuple>;
using InLayout = std::tuple_element_t<3, Tuple>;
std::vector<ck::utils::conv::ConvParam> conv_params;
template <ck::index_t NDimSpatial>
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto& param : conv_params)
{
pass = pass && ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
OutLayout,
WeiLayout,
InLayout,
DataType,
DataType,
DataType>(
true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param);
}
EXPECT_TRUE(pass);
}
};
using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<int8_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<int8_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<int8_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<int8_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdDataWmma2d : public TestGroupedConvndBwdDataWmma<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdDataWmma3d : public TestGroupedConvndBwdDataWmma<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdDataWmma2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataWmma3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdDataWmma2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdDataWmma3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->template Run<3>();
}
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_xdl_wmma.cpp → test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_xdl.cpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
......@@ -12,7 +12,7 @@
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdData : public ::testing::Test
class TestGroupedConvndBwdDataXdl : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
......@@ -51,35 +51,31 @@ using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>,
std::tuple<int8_t, GNHWK, GKYXC, GNHWC>,
std::tuple<float, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>,
std::tuple<int8_t, NHWGK, GKYXC, NHWGC>>;
std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::bhalf_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<int8_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<float, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<int8_t, NDHWGK, GKZYXC, NDHWGC>>;
std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdData2d : public TestGroupedConvndBwdData<Tuple>
class TestGroupedConvndBwdDataXdl2d : public TestGroupedConvndBwdDataXdl<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdData3d : public TestGroupedConvndBwdData<Tuple>
class TestGroupedConvndBwdDataXdl3d : public TestGroupedConvndBwdDataXdl<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdData2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdData3d, KernelTypes3d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataXdl2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataXdl3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
TYPED_TEST(TestGroupedConvndBwdDataXdl2d, Test2D)
{
this->conv_params.clear();
......@@ -94,10 +90,13 @@ TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
// SplitN case
this->conv_params.push_back(
{2, 1, 128, 4, 192, {2, 2}, {224, 224}, {224, 224}, {1, 1}, {0, 0}, {0, 0}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
TYPED_TEST(TestGroupedConvndBwdDataXdl3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
......@@ -112,5 +111,17 @@ TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
// SplitN case
this->conv_params.push_back({3,
1,
128,
4,
192,
{2, 2, 2},
{2, 224, 224},
{1, 224, 224},
{1, 1, 1},
{0, 0, 0},
{0, 0, 0}});
this->template Run<3>();
}
test/grouped_gemm/CMakeLists.txt
View file @ c8c016dd
......@@ -6,12 +6,6 @@ if(result EQUAL 0)
add_dependencies(test_grouped_gemm test_grouped_gemm_splitk)
endif()
add_gtest_executable(test_grouped_gemm_two_stage_splitk test_grouped_gemm_two_stage_multiple_d_splitk_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_gemm_two_stage_splitk PRIVATE utility device_grouped_gemm_instance)
add_dependencies(test_grouped_gemm test_grouped_gemm_two_stage_splitk)
endif()
add_gtest_executable(test_grouped_gemm_interface test_grouped_gemm_interface_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_gemm_interface PRIVATE utility device_grouped_gemm_instance)
......
test/grouped_gemm/test_grouped_gemm_splitk_xdl.cpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <vector>
......@@ -11,24 +11,34 @@
#include "test_grouped_gemm_util.hpp"
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F8 = ck::f8_t;
using I8 = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using RRR_F16_F16_F16 = ck::test::TestGroupedGemm<std::tuple<Row, Row, Row, F16, F16, F16>>;
using RCR_F16_F16_F16 = ck::test::TestGroupedGemm<std::tuple<Row, Col, Row, F16, F16, F16>>;
using RRR_F16_F16_F16_LargeK = ck::test::TestGroupedGemm<std::tuple<Row, Row, Row, F16, F16, F16>>;
using RCR_F16_F16_F16_LargeK = ck::test::TestGroupedGemm<std::tuple<Row, Col, Row, F16, F16, F16>>;
const std::vector<int> KBATCH{1, 2, 3, 5, 8};
INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_MK_KN, RRR_F16_F16_F16, testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_MK_NK, RCR_F16_F16_F16, testing::ValuesIn(KBATCH));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_LargeK_MK_KN,
RRR_F16_F16_F16_LargeK,
testing::Values(32, 64));
INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_LargeK_MK_NK,
RCR_F16_F16_F16_LargeK,
testing::Values(32, 64));
template <typename Tuple>
class TestGroupedGemm : public ck::test::TestGroupedGemm<Tuple>
{
};
// clang-format off
using KernelTypes = ::testing::Types<
std::tuple< Row, Row, Row, F16, F16, F16>,
std::tuple< Row, Col, Row, F16, F16, F16>,
std::tuple< Col, Row, Row, F16, F16, F16>,
std::tuple< Col, Col, Row, F16, F16, F16>,
std::tuple< Row, Row, Row, BF16, BF16, BF16>,
std::tuple< Row, Col, Row, BF16, BF16, BF16>,
std::tuple< Col, Row, Row, BF16, BF16, BF16>,
std::tuple< Row, Row, Row, BF16, I8, BF16>,
std::tuple< Row, Col, Row, BF16, I8, BF16>,
std::tuple< Row, Row, Row, F16, F8, F16>,
std::tuple< Row, Row, Row, F8, F16, F16>
>;
// clang-format on
TYPED_TEST_SUITE(TestGroupedGemm, KernelTypes);
#include "test_grouped_gemm_ut_cases.inc"
test/grouped_gemm/test_grouped_gemm_ut_cases.inc
View file @ c8c016dd
#pragma once
TEST_P(RRR_F16_F16_F16, TinyCases)
TYPED_TEST(TestGroupedGemm, TinyCases)
{
const std::vector<int> Ms{0, 1};
constexpr int N = 768;
......@@ -8,14 +8,11 @@ TEST_P(RRR_F16_F16_F16, TinyCases)
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
this->Run(Ms, Ns, Ks);
}
TEST_P(RRR_F16_F16_F16, SmallCases)
TYPED_TEST(TestGroupedGemm, SmallCases)
{
const std::vector<int> Ms{2, 1, 3, 4, 5, 0};
constexpr int N = 768;
......@@ -23,14 +20,11 @@ TEST_P(RRR_F16_F16_F16, SmallCases)
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
this->Run(Ms, Ns, Ks);
}
TEST_P(RRR_F16_F16_F16, MidCases)
TYPED_TEST(TestGroupedGemm, MidCases)
{
const std::vector<int> Ms{167, 183, 177, 153, 139, 204};
constexpr int N = 768;
......@@ -38,14 +32,11 @@ TEST_P(RRR_F16_F16_F16, MidCases)
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
this->Run(Ms, Ns, Ks);
}
TEST_P(RRR_F16_F16_F16, Regular)
TYPED_TEST(TestGroupedGemm, Regular)
{
const std::vector<int> Ms{64, 128, 256};
constexpr int N = 768;
......@@ -53,14 +44,11 @@ TEST_P(RRR_F16_F16_F16, Regular)
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
this->Run(Ms, Ns, Ks);
}
TEST_P(RRR_F16_F16_F16, MNKPadded)
TYPED_TEST(TestGroupedGemm, MNKPadded)
{
const std::vector<int> Ms{127, 150, 188, 210};
constexpr int N = 136;
......@@ -68,88 +56,11 @@ TEST_P(RRR_F16_F16_F16, MNKPadded)
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
this->Run(Ms, Ns, Ks);
}
TEST_P(RCR_F16_F16_F16, TinyCases)
{
const std::vector<int> Ms{0, 1};
constexpr int N = 768;
constexpr int K = 544;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_F16_F16_F16, SmallCases)
{
const std::vector<int> Ms{2, 1, 3, 4, 5, 0};
constexpr int N = 768;
constexpr int K = 544;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_F16_F16_F16, MidCases)
{
const std::vector<int> Ms{167, 183, 177, 153, 139, 204};
constexpr int N = 768;
constexpr int K = 544;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_F16_F16_F16, Regular)
{
const std::vector<int> Ms{32, 64, 128, 256};
constexpr int N = 768;
constexpr int K = 320;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_F16_F16_F16, MNKPadded)
{
const std::vector<int> Ms{127, 150, 188, 210};
constexpr int N = 136;
constexpr int K = 280;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RRR_F16_F16_F16_LargeK, TestLargeKBatch)
TYPED_TEST(TestGroupedGemm, TestLargeKBatch)
{
const std::vector<int> Ms{188, 210};
constexpr int N = 768;
......@@ -157,24 +68,8 @@ TEST_P(RRR_F16_F16_F16_LargeK, TestLargeKBatch)
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), N);
const std::vector<int> StrideCs(Ms.size(), N);
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
}
TEST_P(RCR_F16_F16_F16_LargeK, TestLargeKBatch)
{
const std::vector<int> Ms{188, 210};
constexpr int N = 768;
constexpr int K = 4096;
const std::vector<int> Ns(Ms.size(), N);
const std::vector<int> Ks(Ms.size(), K);
const std::vector<int> StrideAs(Ms.size(), K);
const std::vector<int> StrideBs(Ms.size(), K);
const std::vector<int> StrideCs(Ms.size(), N);
this->k_batches_ = {32, 64};
this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
this->Run(Ms, Ns, Ks);
}
test/grouped_gemm/test_grouped_gemm_util.hpp
View file @ c8c016dd
......@@ -22,7 +22,6 @@
#include "ck/utility/tuple.hpp"
#include "ck/utility/number.hpp"
#include "profiler/profile_grouped_gemm_impl.hpp"
#include "profiler/profile_grouped_gemm_two_stage_impl.hpp"
namespace ck {
namespace test {
......@@ -40,7 +39,7 @@ std::string serialize_range(const Range& range)
}
template <typename Tuple>
class TestGroupedGemm : public testing::TestWithParam<int>
class TestGroupedGemm : public testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
......@@ -50,77 +49,79 @@ class TestGroupedGemm : public testing::TestWithParam<int>
using BDataType = std::tuple_element_t<4, Tuple>;
using EDataType = std::tuple_element_t<5, Tuple>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
public:
static constexpr bool verify_ = true;
static constexpr int init_method_ = 1; // decimal value initialization
static constexpr int init_method_ = 1; // integer value initialization
static constexpr bool log_ = false;
static constexpr bool bench_ = false; // measure kernel performance
static constexpr int n_warmup_ = 0;
static constexpr int n_iter_ = 1;
std::vector<int> k_batches_;
void SetUp() override {}
void SetUp() override { k_batches_ = {1, 2, 3, 5, 8}; }
private:
template <typename Layout>
void SetStrides(std::vector<int>& strides,
const std::vector<int>& rows,
const std::vector<int>& cols) const
{
if(std::is_same_v<Layout, Row>)
{
for(const auto c : cols)
{
strides.emplace_back(c);
}
}
else if(std::is_same_v<Layout, Col>)
{
for(const auto r : rows)
{
strides.emplace_back(r);
}
}
}
public:
void Run(const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
const std::vector<int>& StrideAs = {},
const std::vector<int>& StrideBs = {},
const std::vector<int>& StrideCs = {})
{
bool pass = ck::profiler::profile_grouped_gemm_impl<ADataType,
BDataType,
EDataType,
float,
ALayout,
BLayout,
ELayout>(verify_,
init_method_,
log_,
bench_,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatch,
n_warmup,
n_iter);
EXPECT_TRUE(pass);
}
};
template <typename Tuple>
class TestGroupedGemmTwoStage : public testing::TestWithParam<int>
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using ELayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using EDataType = std::tuple_element_t<5, Tuple>;
std::vector<int> stride_as = StrideAs;
std::vector<int> stride_bs = StrideBs;
std::vector<int> stride_cs = StrideCs;
public:
static constexpr bool verify_ = true;
static constexpr int init_method_ = 1; // decimal value initialization
static constexpr bool log_ = false;
static constexpr bool bench_ = false; // measure kernel performance
if(stride_as.empty())
{
SetStrides<ALayout>(stride_as, Ms, Ks);
}
if(stride_bs.empty())
{
SetStrides<BLayout>(stride_bs, Ks, Ns);
}
if(stride_cs.empty())
{
SetStrides<ELayout>(stride_cs, Ms, Ns);
}
void SetUp() override {}
RunSingle(Ms, Ns, Ks, stride_as, stride_bs, stride_cs, k_batches_);
}
void Run(const std::vector<int>& Ms,
void RunSingle(const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
const std::vector<int>& kbatches)
{
bool pass = ck::profiler::profile_grouped_gemm_two_stage_impl<ADataType,
bool pass = ck::profiler::profile_grouped_gemm_impl<ADataType,
BDataType,
EDataType,
float,
......@@ -136,9 +137,9 @@ class TestGroupedGemmTwoStage : public testing::TestWithParam<int>
StrideAs,
StrideBs,
StrideCs,
kbatch,
n_warmup,
n_iter);
kbatches,
n_warmup_,
n_iter_);
EXPECT_TRUE(pass);
}
};
......@@ -263,7 +264,7 @@ struct DeviceGroupedGemmSplitkInstanceWrapper
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, PassThrough{}, PassThrough{}, PassThrough{});
if(kbatch > 1)
{
ggemm_instance.SetKBatchSize(argument, kbatch);
ggemm_instance.SetKBatchSize(&argument, kbatch);
}
return ggemm_instance.IsSupportedArgument(argument);
......@@ -300,13 +301,13 @@ struct DeviceGroupedGemmSplitkInstanceWrapper
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, PassThrough{}, PassThrough{}, PassThrough{});
if(kbatch > 1)
{
ggemm_instance.SetKBatchSize(argument, kbatch);
ggemm_instance.SetKBatchSize(&argument, kbatch);
}
EXPECT_TRUE(ggemm_instance.IsSupportedArgument(argument));
auto invoker = ggemm_instance.MakeInvoker();
DeviceMem gemm_desc_workspace(ggemm_instance.GetWorkSpaceSize(&argument));
ggemm_instance.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
DeviceMem dev_gemm_kargs(ggemm_instance.GetDeviceKernelArgSize(&argument));
ggemm_instance.SetDeviceKernelArgs(&argument, dev_gemm_kargs.GetDeviceBuffer());
return invoker.Run(argument, StreamConfig{nullptr, false});
}
};
......
test/pool/test_avg_pool2d_fwd.cpp
View file @ c8c016dd
......@@ -138,7 +138,7 @@ TYPED_TEST_SUITE(AvgPool2D_BF16, AvgPool2D_BF16_Types);
TYPED_TEST_SUITE(AvgPool2D_I8, AvgPool2D_I8_Types);
TYPED_TEST_SUITE(AvgPool2D_F8, AvgPool2D_F8_Types);
TYPED_TEST(AvgPool2D_F32, AvgPool2D_I8_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_F32, AvgPool2D_F32_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_F16, AvgPool2D_F16_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_BF16, AvgPool2D_BF16_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_I8, AvgPool2D_I8_Test) { this->Run(); }
......
test/pool/test_max_pool2d_fwd.cpp
View file @ c8c016dd
......@@ -143,7 +143,7 @@ TYPED_TEST_SUITE(MaxPool2D_BF16, MaxPool2D_BF16_Types);
TYPED_TEST_SUITE(MaxPool2D_I8, MaxPool2D_I8_Types);
TYPED_TEST_SUITE(MaxPool2D_F8, MaxPool2D_F8_Types);
TYPED_TEST(MaxPool2D_F32, MaxPool2D_I8_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_F32, MaxPool2D_F32_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_F16, MaxPool2D_F16_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_BF16, MaxPool2D_BF16_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_I8, MaxPool2D_I8_Test) { this->Run(); }
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment