AI/rocm-systems
2
0
انشعاب 0
کد مسائل تقاضاهای واکشی Actions Packages پروژه‌ها انتشارها دانشنامه فعالیت

SWDEV-468039 - Add Catch Test for fp8 OCP type conversions

فهرست منبع 
Change-Id: I2648909483b8dc32fcd720c18608c5ca32399045


[ROCm/hip-tests commit: e3e662b52b]
This commit is contained in:
Rahul Manocha
2024-06-26 16:24:23 -07:00
کامیت شده توسط Rahul Manocha
والد d28d65607a
کامیت f74c5c4972
6فایلهای تغییر یافته به همراه2357 افزوده شده و 300 حذف شده
دانلود پرونده وصله دانلود فایل تغییرات diff Expand all files Collapse all files
projects/hip-tests/catch/hipTestMain/config/config_amd_linux
-11
مشاهده پرونده
@@ -1081,17 +1081,6 @@
"=== SWDEV-457316 Below tests are disabled temporarily to avoid combined PSDB ===",
"Unit_hipGraphAddMemAllocNode_Positive_FreeSeparateGraph",
"Unit_hipGraphAddMemFreeNode_Negative_NotSupported",
"Unit_fp8_compare_host_device - float",
"Unit_fp8_compare_host_device - double",
"Unit_fp8x2_compare_host_device",
"Unit_fp8x2_split_compare",
"Unit_fp8x4_split_compare",
"Unit_fp8_bool",
"Unit_all_fp8_cvt",
"Unit_all_fp8_vector_cvt",
"Unit_fp8_correctness - float",
"Unit_fp8_correctness - double",
"Unit_fp8_vector_basic_conversions",
#endif
#if defined gfx908
"=== Below test soft hang in stress test on 29/08/23 ===",
projects/hip-tests/catch/hipTestMain/config/config_amd_windows
-11
مشاهده پرونده
@@ -1255,17 +1255,6 @@
"Unit_hipStreamBeginCaptureToGraph_CaptureDepGraph",
"=== SWDEV-465071 Below test is temporarily disabled due to compiler change ===",
"Unit_hipClassKernel_Virtual",
"Unit_fp8_compare_host_device - float",
"Unit_fp8_compare_host_device - double",
"Unit_fp8x2_compare_host_device",
"Unit_fp8x2_split_compare",
"Unit_fp8x4_split_compare",
"Unit_fp8_bool",
"Unit_all_fp8_cvt",
"Unit_all_fp8_vector_cvt",
"Unit_fp8_correctness - float",
"Unit_fp8_correctness - double",
"Unit_fp8_vector_basic_conversions",
"=== SWDEV-486448 - Following tests disabled due to taking too much time to execute, ~700s per test",
"Unit_Coalesced_Group_Tiled_Partition_Getters_Positive_Basic",
"Unit_Coalesced_Group_Tiled_Partition_Shfl_Up_Positive_Basic - int",
projects/hip-tests/catch/unit/deviceLib/CMakeLists.txt
+27
مشاهده پرونده
@@ -88,6 +88,7 @@ set(AMD_TEST_SRC
bfloat16.cc
AtomicsWithRandomActiveLanesInWavefront.cc
fp16_ops.cc
fp8_host.cc
)
set(AMD_ARCH_SPEC_TEST_SRC
@@ -109,6 +110,14 @@ set(AMD_ARCH_SPEC_TEST_SRC
unsafeAtomicAdd_NonCoherent_withunsafeflag.cc
)
set(AMD_GFX940_SPEC_TEST_SRC
fp8_fnuz.cc
)
set(AMD_GFX1200_SPEC_TEST_SRC
fp8_ocp.cc
)
# Note to pass arch use format like -DOFFLOAD_ARCH_STR="--offload-arch=gfx900 --offload-arch=gfx906"
# having space at the start/end of OFFLOAD_ARCH_STR can cause build failures
add_custom_target(kerDevAllocMultCO.code
@@ -138,8 +147,12 @@ add_custom_target(kerDevAllocSingleKer.code
# Accepted archs to compile this cmake file
set(ACCEPTED_OFFLOAD_ARCHS gfx90a gfx940 gfx941 gfx942)
set(ACCEPTED_GFX940_ARCH gfx940 gfx941 gfx942)
set(ACCEPTED_GFX1200_ARCH gfx1200 gfx1201)
function(CheckAcceptedArchs OFFLOAD_ARCH_STR_LOCAL)
set(ARCH_CHECK -1 PARENT_SCOPE)
set(ARCH_GFX940 -1 PARENT_SCOPE)
set(ARCH_GFX1200 -1 PARENT_SCOPE)
string(REGEX MATCHALL "--offload-arch=gfx[0-9a-z]+" OFFLOAD_ARCH_LIST ${OFFLOAD_ARCH_STR_LOCAL})
foreach(OFFLOAD_ARCH IN LISTS OFFLOAD_ARCH_LIST)
string(REGEX MATCHALL "--offload-arch=(gfx[0-9a-z]+)" matches ${OFFLOAD_ARCH})
@@ -147,6 +160,12 @@ function(CheckAcceptedArchs OFFLOAD_ARCH_STR_LOCAL)
if (CMAKE_MATCH_1 IN_LIST ACCEPTED_OFFLOAD_ARCHS)
set(ARCH_CHECK 1 PARENT_SCOPE)
endif() # CMAKE_MATCH_1
if (CMAKE_MATCH_1 IN_LIST ACCEPTED_GFX940_ARCH)
set(ARCH_GFX940 1 PARENT_SCOPE)
endif()
if (CMAKE_MATCH_1 IN_LIST ACCEPTED_GFX1200_ARCH)
set(ARCH_GFX1200 1 PARENT_SCOPE)
endif()
endif() # CMAKE_MATCH_COUNT
endforeach() # OFFLOAD_ARCH_LIST
endfunction() # CheckAcceptedArchs
@@ -158,6 +177,8 @@ if(HIP_PLATFORM MATCHES "amd")
CheckAcceptedArchs($ENV{HCC_AMDGPU_TARGET})
else()
set(ARCH_CHECK -1)
set(ARCH_GFX940 -1)
set(ARCH_GFX1200 -1)
endif()
set(TEST_SRC ${TEST_SRC} ${AMD_TEST_SRC})
set_source_files_properties(floatTM.cc PROPERTIES COMPILE_FLAGS -std=c++17)
@@ -177,6 +198,12 @@ if(${ARCH_CHECK} GREATER_EQUAL 0)
set_property(SOURCE ${AtomicAdd_files} PROPERTY COMPILE_FLAGS --save-temps)
file(GLOB unsafeAtomicAdd_files *unsafeAtomicAdd_*_*.cc)
set_property(SOURCE ${unsafeAtomicAdd_files} PROPERTY COMPILE_FLAGS --save-temps)
endif()
if(${ARCH_GFX940} GREATER_EQUAL 0)
set(TEST_SRC ${TEST_SRC} ${AMD_GFX940_SPEC_TEST_SRC})
endif()
if(${ARCH_GFX1200} GREATER_EQUAL 0)
set(TEST_SRC ${TEST_SRC} ${AMD_GFX1200_SPEC_TEST_SRC})
endif()
hip_add_exe_to_target(NAME UnitDeviceTests
TEST_SRC ${TEST_SRC}
projects/hip-tests/catch/unit/deviceLib/fp8.cpp → projects/hip-tests/catch/unit/deviceLib/fp8_fnuz.cc
+295 -278
مشاهده پرونده
@@ -27,24 +27,72 @@ THE SOFTWARE.
#include <vector>
#include <bitset>
/*
* This catch test is meant for FP8 FNUZ conversion checking
* tests only supported on gfx940,gfx941,gfx942 archs.
*/
static_assert(sizeof(unsigned int) == sizeof(float));
template <typename T, bool is_e4m3_fnuz> __global__ void cvt_float_fp8_float(T* in, size_t len) {
std::string get_arch_type() {
hipDeviceProp_t prop;
int device;
HIP_CHECK(hipGetDevice(&device));
HIP_CHECK(hipGetDeviceProperties(&prop, device));
std::string gfxName(prop.gcnArchName);
return gfxName;
}
#define ARCH_TYPE_GFX940(name) \
(name.find("gfx940") != std::string::npos) || \
(name.find("gfx941") != std::string::npos) || \
(name.find("gfx942") != std::string::npos)
#define FP8_FNUZ_SKIP_TEST \
std::string gfxName = get_arch_type(); \
if (!(ARCH_TYPE_GFX940(gfxName))) { \
HipTest::HIP_SKIP_TEST("This test can only be run on gfx940,gfx941,gfx42 arch"); \
return; \
}
#define __FP8_DEVICE__ __device__ static inline
template<typename T> __FP8_DEVICE__ void e4m3_fnuz_device(T *val)
{
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3_fnuz tmp(*val);
*val = tmp;
#else
*val = 0;
#endif
}
template<typename T> __FP8_DEVICE__ void e5m2_fnuz_device(T *val)
{
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2_fnuz tmp(*val);
*val = tmp;
#else
*val = 0;
#endif
}
template <typename T, bool is_e4m3_fnuz> __global__ void cvt_float_fp8_float_fnuz(T* in, size_t len) {
int i = threadIdx.x;
if (i < len) {
float val = in[i];
T val = in[i];
if constexpr (is_e4m3_fnuz) {
__hip_fp8_e4m3_fnuz tmp(val);
in[i] = tmp;
e4m3_fnuz_device<T>(&val);
in[i] = val;
} else {
__hip_fp8_e5m2_fnuz tmp(val);
in[i] = tmp;
e5m2_fnuz_device<T>(&val);
in[i] = val;
}
}
}
template <typename T, bool is_e4m3_fnuz>
std::vector<T> cpu_cvt_float_fp8_float(const std::vector<T>& nums) {
std::vector<T> cpu_cvt_float_fp8_float_fnuz(const std::vector<T>& nums) {
std::vector<T> ret;
ret.reserve(nums.size());
for (const auto& num : nums) {
@@ -61,9 +109,9 @@ std::vector<T> cpu_cvt_float_fp8_float(const std::vector<T>& nums) {
return ret;
}
// This test only makes sense on gfx94x where device side convert will use the builtins to convert
// floats to fp8
TEMPLATE_TEST_CASE("Unit_fp8_compare_host_device", "", float, double) {
TEMPLATE_TEST_CASE("Unit_fp8_fnuz_compare_host_device", "", float, double) {
FP8_FNUZ_SKIP_TEST
std::vector<TestType> numbers = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
TestType* d_numbers;
@@ -75,16 +123,16 @@ TEMPLATE_TEST_CASE("Unit_fp8_compare_host_device", "", float, double) {
std::vector<TestType> cpu_result;
SECTION("e4m3_fnuz") {
cpu_result = cpu_cvt_float_fp8_float<TestType, true>(numbers);
auto kernel = cvt_float_fp8_float<TestType, true>;
cpu_result = cpu_cvt_float_fp8_float_fnuz<TestType, true>(numbers);
auto kernel = cvt_float_fp8_float_fnuz<TestType, true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(TestType) * numbers.size(),
hipMemcpyDeviceToHost));
}
SECTION("e5m2_fnuz") {
cpu_result = cpu_cvt_float_fp8_float<TestType, false>(numbers);
auto kernel = cvt_float_fp8_float<TestType, false>;
cpu_result = cpu_cvt_float_fp8_float_fnuz<TestType, false>(numbers);
auto kernel = cvt_float_fp8_float_fnuz<TestType, false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(TestType) * numbers.size(),
hipMemcpyDeviceToHost));
@@ -98,22 +146,42 @@ TEMPLATE_TEST_CASE("Unit_fp8_compare_host_device", "", float, double) {
}
}
template <bool is_e4m3_fnuz> __global__ void cvt_float2_fp8x2_float2(float2* in, size_t size) {
__FP8_DEVICE__ void e4m3_fp8x2_fnuz_device(float2 *val)
{
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x2_e4m3_fnuz tmp(*val);
*val = tmp;
#else
*val = float2(0.0,0.0);
#endif
}
__FP8_DEVICE__ void e5m2_fp8x2_fnuz_device(float2 *val)
{
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x2_e5m2_fnuz tmp(*val);
*val = tmp;
#else
*val = float2(0.0,0.0);
#endif
}
template <bool is_e4m3_fnuz> __global__ void cvt_float2_fp8x2_float2_fnuz(float2* in, size_t size) {
int i = threadIdx.x;
if (i < size) {
float2 val = in[i];
if constexpr (is_e4m3_fnuz) {
__hip_fp8x2_e4m3_fnuz tmp(val);
in[i] = tmp;
e4m3_fp8x2_fnuz_device(&val);
in[i] = val;
} else {
__hip_fp8x2_e5m2_fnuz tmp(val);
in[i] = tmp;
e5m2_fp8x2_fnuz_device(&val);
in[i] = val;
}
}
}
template <bool is_e4m3_fnuz>
std::vector<float2> cpu_cvt_float2_fp8x2_float2(const std::vector<float2>& nums) {
std::vector<float2> cpu_cvt_float2_fp8x2_float2_fnuz(const std::vector<float2>& nums) {
std::vector<float2> ret;
ret.reserve(nums.size());
for (const auto& num : nums) {
@@ -130,7 +198,8 @@ std::vector<float2> cpu_cvt_float2_fp8x2_float2(const std::vector<float2>& nums)
return ret;
}
TEST_CASE("Unit_fp8x2_compare_host_device") {
TEST_CASE("Unit_fp8x2_fnuz_compare_host_device") {
FP8_FNUZ_SKIP_TEST
std::vector<float> numbers_input = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
@@ -150,16 +219,16 @@ TEST_CASE("Unit_fp8x2_compare_host_device") {
std::vector<float2> cpu_result;
SECTION("e4m3_fnuz") {
cpu_result = cpu_cvt_float2_fp8x2_float2<true>(numbers);
auto kernel = cvt_float2_fp8x2_float2<true>;
cpu_result = cpu_cvt_float2_fp8x2_float2_fnuz<true>(numbers);
auto kernel = cvt_float2_fp8x2_float2_fnuz<true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
}
SECTION("e5m2_fnuz") {
cpu_result = cpu_cvt_float2_fp8x2_float2<false>(numbers);
auto kernel = cvt_float2_fp8x2_float2<false>;
cpu_result = cpu_cvt_float2_fp8x2_float2_fnuz<false>(numbers);
auto kernel = cvt_float2_fp8x2_float2_fnuz<false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
@@ -171,7 +240,8 @@ TEST_CASE("Unit_fp8x2_compare_host_device") {
}
}
TEST_CASE("Unit_fp8x2_split_compare") {
TEST_CASE("Unit_fp8x2_fnuz_split_compare") {
FP8_FNUZ_SKIP_TEST
std::vector<float> numbers_input = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
@@ -198,7 +268,7 @@ TEST_CASE("Unit_fp8x2_split_compare") {
float a = t_a, b = t_b;
cpu_result.push_back(float2(a, b));
}
auto kernel = cvt_float2_fp8x2_float2<true>;
auto kernel = cvt_float2_fp8x2_float2_fnuz<true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
@@ -211,7 +281,7 @@ TEST_CASE("Unit_fp8x2_split_compare") {
float a = t_a, b = t_b;
cpu_result.push_back(float2(a, b));
}
auto kernel = cvt_float2_fp8x2_float2<false>;
auto kernel = cvt_float2_fp8x2_float2_fnuz<false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
@@ -226,21 +296,42 @@ TEST_CASE("Unit_fp8x2_split_compare") {
}
}
template <bool is_e4m3_fnuz> __global__ void cvt_float4_fp8x4_float4(float4* in, size_t size) {
__FP8_DEVICE__ void e4m3_fp8x4_fnuz_device(float4 *val)
{
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x4_e4m3_fnuz tmp(*val);
*val = tmp;
#else
*val = float4(0.0,0.0,0.0,0.0);
#endif
}
__FP8_DEVICE__ void e5m2_fp8x4_fnuz_device(float4 *val)
{
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x4_e5m2_fnuz tmp(*val);
*val = tmp;
#else
*val = float4(0.0,0.0,0.0,0.0);
#endif
}
template <bool is_e4m3_fnuz> __global__ void cvt_float4_fp8x4_float4_fnuz(float4* in, size_t size) {
int i = threadIdx.x;
if (i < size) {
float4 val = in[i];
if constexpr (is_e4m3_fnuz) {
__hip_fp8x4_e4m3_fnuz tmp(val);
in[i] = tmp;
e4m3_fp8x4_fnuz_device(&val);
in[i] = val;
} else {
__hip_fp8x4_e5m2_fnuz tmp(val);
in[i] = tmp;
e5m2_fp8x4_fnuz_device(&val);
in[i] = val;
}
}
}
TEST_CASE("Unit_fp8x4_split_compare") {
TEST_CASE("Unit_fp8x4_fnuz_split_compare") {
FP8_FNUZ_SKIP_TEST
std::vector<float> numbers_input = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
std::vector<float> numbers_input2 = {1.3f, 1.6f, 1.8f, 2.5f, 2.9f, 3.8f, 3.9f,
@@ -271,7 +362,7 @@ TEST_CASE("Unit_fp8x4_split_compare") {
float a = t_a, b = t_b, c = t_c, d = t_d;
cpu_result.push_back(float4(a, b, c, d));
}
auto kernel = cvt_float4_fp8x4_float4<true>;
auto kernel = cvt_float4_fp8x4_float4_fnuz<true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float4) * numbers.size(),
hipMemcpyDeviceToHost));
@@ -286,7 +377,7 @@ TEST_CASE("Unit_fp8x4_split_compare") {
float a = t_a, b = t_b, c = t_c, d = t_d;
cpu_result.push_back(float4(a, b, c, d));
}
auto kernel = cvt_float4_fp8x4_float4<false>;
auto kernel = cvt_float4_fp8x4_float4_fnuz<false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float4) * numbers.size(),
hipMemcpyDeviceToHost));
@@ -304,23 +395,47 @@ TEST_CASE("Unit_fp8x4_split_compare") {
CHECK(cpu_result[i] == result[i]);
}
}
__FP8_DEVICE__ bool e4m3_bool_fnuz_device(float val)
{
bool x = false;
float y = val;
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3_fnuz tmp(y);
x = tmp;
#else
x = (y == 0);
#endif
return x;
}
template <bool is_e4m3_fnuz> __global__ void fp8_2_bool(float* f, bool* ret, size_t size) {
__FP8_DEVICE__ bool e5m2_bool_fnuz_device(float val)
{
bool x = false;
float y = val;
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2_fnuz tmp(y);
x = tmp;
#else
x = (y == 0);
#endif
return x;
}
template <bool is_e4m3_fnuz> __global__ void fp8_2_bool_fnuz(float* f, bool* ret, size_t size) {
int i = threadIdx.x;
bool r = false;
if (i < size) {
if constexpr (is_e4m3_fnuz) {
__hip_fp8_e4m3_fnuz fp8(f[i]);
r = fp8;
r = e4m3_bool_fnuz_device(f[i]);
} else {
__hip_fp8_e5m2_fnuz fp8(f[i]);
r = fp8;
r = e5m2_bool_fnuz_device(f[i]);
}
ret[i] = r;
}
}
TEST_CASE("Unit_fp8_bool") {
TEST_CASE("Unit_fp8_fnuz_bool_device") {
FP8_FNUZ_SKIP_TEST
// clang-format off
std::vector<float> fvals{-10.0f, -1.0f, -0.0f, 0.0f, 1.0f, 10.0f};
std::vector<bool> tvals {true, true, false, false, true, true};
@@ -329,27 +444,13 @@ TEST_CASE("Unit_fp8_bool") {
bool result[] = {false, false, false,
false, false, false}; // cant use std::vector coz data() = delete
SECTION("e4m3_fnuz-cpu") {
for (size_t i = 0; i < tvals.size(); i++) {
__hip_fp8_e4m3_fnuz fp8(fvals[i]);
result[i] = fp8;
}
}
SECTION("e5m2_fnuz-cpu") {
for (size_t i = 0; i < tvals.size(); i++) {
__hip_fp8_e5m2_fnuz fp8(fvals[i]);
result[i] = fp8;
}
}
SECTION("e4m3_fnuz-gpu") {
float* d_in{nullptr};
bool* d_res{nullptr};
HIP_CHECK(hipMalloc(&d_in, sizeof(float) * tvals.size()));
HIP_CHECK(hipMalloc(&d_res, sizeof(bool) * tvals.size()));
auto kernel = fp8_2_bool<true>;
auto kernel = fp8_2_bool_fnuz<true>;
HIP_CHECK(hipMemcpy(d_in, fvals.data(), sizeof(float) * fvals.size(), hipMemcpyHostToDevice));
kernel<<<1, tvals.size()>>>(d_in, d_res, tvals.size());
@@ -366,7 +467,7 @@ TEST_CASE("Unit_fp8_bool") {
HIP_CHECK(hipMalloc(&d_res, sizeof(bool) * tvals.size()));
HIP_CHECK(hipMemcpy(d_in, fvals.data(), sizeof(float) * fvals.size(), hipMemcpyHostToDevice));
auto kernel = fp8_2_bool<false>;
auto kernel = fp8_2_bool_fnuz<false>;
kernel<<<1, tvals.size()>>>(d_in, d_res, tvals.size());
HIP_CHECK(hipMemcpy(result, d_res, sizeof(bool) * tvals.size(), hipMemcpyDeviceToHost));
@@ -381,7 +482,7 @@ TEST_CASE("Unit_fp8_bool") {
}
}
std::vector<__hip_fp8_storage_t> get_all_fp8_nums() {
std::vector<__hip_fp8_storage_t> get_all_fp8_fnuz_nums() {
std::vector<__hip_fp8_storage_t> ret;
constexpr unsigned short max_fp8_num = 0b1111'1111;
ret.reserve(max_fp8_num + 1 /* 0 */);
@@ -394,27 +495,52 @@ std::vector<__hip_fp8_storage_t> get_all_fp8_nums() {
return ret;
}
__FP8_DEVICE__ __hip_fp8_storage_t e4m3_fnuz_fp8_device(float val)
{
__hip_fp8_storage_t x = 0;
float y = val;
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3_fnuz tmp(y);
x = tmp.__x;
#else
x = (y == 0) ? 0x0 : 0x80;
#endif
return x;
}
__FP8_DEVICE__ __hip_fp8_storage_t e5m2_fnuz_fp8_device(float val)
{
__hip_fp8_storage_t x = 0;
float y = val;
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2_fnuz tmp(y);
x = tmp.__x;
#else
x = (y == 0) ? 0x0 : 0x80;
#endif
return x;
}
template <bool is_e4m3_fnuz>
__global__ void Type_to_fp8(float* f, __hip_fp8_storage_t* res, size_t size) {
__global__ void Type_to_fp8_fnuz(float* f, __hip_fp8_storage_t* res, size_t size) {
auto i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < size) {
if constexpr (is_e4m3_fnuz) {
__hip_fp8_e4m3_fnuz a(f[i]);
res[i] = a.__x;
res[i] = e4m3_fnuz_fp8_device(f[i]);
} else {
__hip_fp8_e5m2_fnuz a(f[i]);
res[i] = a.__x;
res[i] = e5m2_fnuz_fp8_device(f[i]);
}
}
}
TEST_CASE("Unit_all_fp8_cvt") {
TEST_CASE("Unit_all_fp8_fnuz_cvt") {
FP8_FNUZ_SKIP_TEST
bool is_e4m3_fnuz = GENERATE(true, false);
std::vector<float> f_vals;
std::vector<__hip_fp8_storage_t> all_vals;
SECTION("all representable number") {
all_vals = get_all_fp8_nums();
all_vals = get_all_fp8_fnuz_nums();
f_vals.reserve(all_vals.size());
for (const auto& fp8 : all_vals) {
@@ -461,7 +587,7 @@ TEST_CASE("Unit_all_fp8_cvt") {
HIP_CHECK(
hipMemcpy(d_f_vals, f_vals.data(), sizeof(float) * f_vals.size(), hipMemcpyHostToDevice));
auto fp8_kernel = is_e4m3_fnuz ? Type_to_fp8<true> : Type_to_fp8<false>;
auto fp8_kernel = is_e4m3_fnuz ? Type_to_fp8_fnuz<true> : Type_to_fp8_fnuz<false>;
fp8_kernel<<<(f_vals.size() / 256) + 1, 256>>>(d_f_vals, d_res, f_vals.size());
std::vector<__hip_fp8_storage_t> final_res(f_vals.size(), static_cast<__hip_fp8_storage_t>(0));
@@ -498,66 +624,60 @@ TEST_CASE("Unit_all_fp8_cvt") {
HIP_CHECK(hipFree(d_res));
}
// test to check we are putting in data correctly in vector types
TEST_CASE("Unit_all_fp8_vector_cvt") {
float2 in2{1.0f, 2.0f};
float4 in4{3.0f, 4.0f, 5.0f, 6.0f};
template<typename T> __FP8_DEVICE__ void e4m3_fnuz_fp8_cvt(T val, float *cvt1, float *cvt2)
{
T y = val;
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3_fnuz tmp(y);
*cvt1 = tmp;
SECTION("e4m3_fnuz x2") {
__hip_fp8x2_e4m3_fnuz in(in2);
float2 out = in;
INFO("In: " << in2.x << " - " << in2.y);
INFO("Out: " << out.x << " - " << out.y);
REQUIRE(out == in2);
}
SECTION("e4m3_fnuz x4") {
__hip_fp8x4_e4m3_fnuz in(in4);
float4 out = in;
INFO("In: " << in4.x << " - " << in4.y << " - " << in4.z << " - " << in4.w);
INFO("Out: " << out.x << " - " << out.y << " - " << out.z << " - " << out.w);
REQUIRE(out == in4);
}
__hip_fp8_e4m3_fnuz tmp1;
tmp1.__x = std::is_same<T, float>::value
? __hip_cvt_float_to_fp8(val, __HIP_SATFINITE, __HIP_E4M3_FNUZ)
: __hip_cvt_double_to_fp8(val, __HIP_SATFINITE, __HIP_E4M3_FNUZ);
;
*cvt2 = tmp1;
#else
*cvt1 = (y == 0) ? 0 : y;
*cvt2 = (y == 0) ? 0 : y;
#endif
}
SECTION("e5m2_fnuz x2") {
__hip_fp8x2_e5m2_fnuz in(in2);
float2 out = in;
INFO("In: " << in2.x << " - " << in2.y);
INFO("Out: " << out.x << " - " << out.y);
REQUIRE(out == in2);
}
template<typename T> __FP8_DEVICE__ void e5m2_fnuz_fp8_cvt(T val, float *cvt1, float *cvt2)
{
T y = val;
#if (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2_fnuz tmp(y);
*cvt1 = tmp;
SECTION("e5m2_fnuz x4") {
__hip_fp8x4_e5m2_fnuz in(in4);
float4 out = in;
INFO("In: " << in4.x << " - " << in4.y << " - " << in4.z << " - " << in4.w);
INFO("Out: " << out.x << " - " << out.y << " - " << out.z << " - " << out.w);
REQUIRE(out == in4);
}
__hip_fp8_e5m2_fnuz tmp1;
tmp1.__x = std::is_same<T, float>::value
? __hip_cvt_float_to_fp8(val, __HIP_SATFINITE, __HIP_E5M2_FNUZ)
: __hip_cvt_double_to_fp8(val, __HIP_SATFINITE, __HIP_E5M2_FNUZ);
;
*cvt2 = tmp1;
#else
*cvt1 = (y == 0) ? 0 : y;
*cvt2 = (y == 0) ? 0 : y;
#endif
}
SECTION("half x2 e4m3_fnuz") {
__hip_fp8x2_e4m3_fnuz in(in2);
auto hr2 = __hip_cvt_fp8x2_to_halfraw2(in.__x, __HIP_E4M3_FNUZ);
float2 fout1 = in;
float2 fout2 = __half22float2(__half2(hr2));
INFO("In: " << in2.x << " - " << in2.y);
INFO("Out from f8 : " << fout1.x << " - " << fout1.y);
INFO("Out from half: " << fout2.x << " - " << fout2.y);
REQUIRE(fout1 == fout2);
}
SECTION("half x2 e5m2_fnuz") {
__hip_fp8x2_e5m2_fnuz in(in2);
auto hr2 = __hip_cvt_fp8x2_to_halfraw2(in.__x, __HIP_E5M2_FNUZ);
float2 fout1 = in;
float2 fout2 = __half22float2(__half2(hr2));
INFO("In: " << in2.x << " - " << in2.y);
INFO("Out from f8 : " << fout1.x << " - " << fout1.y);
INFO("Out from half: " << fout2.x << " - " << fout2.y);
REQUIRE(fout1 == fout2);
template <typename T, bool is_e4m3_fnuz>
__global__ void Type_to_fp8_fnuz_cvt(T* f, float *cvt1, float *cvt2, size_t size) {
auto i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < size) {
if constexpr (is_e4m3_fnuz) {
e4m3_fnuz_fp8_cvt(f[i], &cvt1[i], &cvt2[i]);
} else {
e5m2_fnuz_fp8_cvt(f[i], &cvt1[i], &cvt2[i]);
}
}
}
TEMPLATE_TEST_CASE("Unit_fp8_correctness", "", float, double) {
TEMPLATE_TEST_CASE("Unit_fp8_fnuz_correctness_device", "", float, double) {
FP8_FNUZ_SKIP_TEST
SECTION("e4m3_fnuz") {
/* These are basically all the fp8 - e4m3_fnuz type numbers.
* They can be generated by iterating over 0'0000'000 and converting them to fp32 number
@@ -648,22 +768,41 @@ TEMPLATE_TEST_CASE("Unit_fp8_correctness", "", float, double) {
-160, -176, -192,
-208, -224, -240};
for (const auto& orig : e4m3_fnuz_nums) {
__hip_fp8_e4m3_fnuz fp8(orig);
float cvt1 = fp8;
size_t totalnums = e4m3_fnuz_nums.size();
TestType *fnums; HIP_CHECK(hipMalloc((void **)&fnums, totalnums * sizeof(TestType)));
float *cvt1_dev; HIP_CHECK(hipMalloc((void **)&cvt1_dev, totalnums * sizeof(TestType)));
float *cvt2_dev; HIP_CHECK(hipMalloc((void **)&cvt2_dev, totalnums * sizeof(TestType)));
__hip_fp8_e4m3_fnuz tmp;
tmp.__x = std::is_same<TestType, float>::value
? __hip_cvt_float_to_fp8(orig, __HIP_SATFINITE, __HIP_E4M3_FNUZ)
: __hip_cvt_double_to_fp8(orig, __HIP_SATFINITE, __HIP_E4M3_FNUZ);
;
float cvt2 = tmp;
HIP_CHECK(hipMemcpy(fnums, e4m3_fnuz_nums.data(), totalnums * sizeof(TestType),
hipMemcpyHostToDevice));
auto fp8_kernel = Type_to_fp8_fnuz_cvt<TestType, true>;
fp8_kernel<<<totalnums / 256 + 1, 256>>>(fnums, cvt1_dev, cvt2_dev, totalnums);
float *cvt1_host = (float *)malloc(sizeof(float) * totalnums);
float *cvt2_host = (float *)malloc(sizeof(float) * totalnums);
HIP_CHECK(hipMemcpy(cvt1_host, cvt1_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipMemcpy(cvt2_host, cvt2_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipDeviceSynchronize());
for (size_t idx = 0; idx < totalnums; idx++) {
TestType orig = e4m3_fnuz_nums[idx];
float cvt1 = cvt1_host[idx];
float cvt2 = cvt2_host[idx];
INFO("Original: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&orig)));
INFO("Cvt back: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&cvt1)));
REQUIRE(cvt1 == Approx(orig));
REQUIRE(cvt2 == cvt1);
}
HIP_CHECK(hipFree(fnums));
HIP_CHECK(hipFree(cvt1_dev));
HIP_CHECK(hipFree(cvt2_dev));
free(cvt1_host);
free(cvt2_host);
}
SECTION("e5m2_fnuz") {
@@ -926,162 +1065,40 @@ TEMPLATE_TEST_CASE("Unit_fp8_correctness", "", float, double) {
-49152,
-57344};
for (const auto& orig : e5m2_fnuz_nums) {
__hip_fp8_e5m2_fnuz fp8(orig);
float cvt1 = fp8;
size_t totalnums = e5m2_fnuz_nums.size();
TestType *fnums; HIP_CHECK(hipMalloc((void **)&fnums, totalnums * sizeof(TestType)));
float *cvt1_dev; HIP_CHECK(hipMalloc((void **)&cvt1_dev, totalnums * sizeof(TestType)));
float *cvt2_dev; HIP_CHECK(hipMalloc((void **)&cvt2_dev, totalnums * sizeof(TestType)));
__hip_fp8_e5m2_fnuz tmp;
tmp.__x = std::is_same<TestType, float>::value
? __hip_cvt_float_to_fp8(orig, __HIP_SATFINITE, __HIP_E5M2_FNUZ)
: __hip_cvt_double_to_fp8(orig, __HIP_SATFINITE, __HIP_E5M2_FNUZ);
;
float cvt2 = tmp;
HIP_CHECK(hipMemcpy(fnums, e5m2_fnuz_nums.data(), totalnums * sizeof(TestType),
hipMemcpyHostToDevice));
auto fp8_kernel = Type_to_fp8_fnuz_cvt<TestType, false>;
fp8_kernel<<<totalnums / 256 + 1, 256>>>(fnums, cvt1_dev, cvt2_dev, totalnums);
float *cvt1_host = (float *)malloc(sizeof(float) * totalnums);
float *cvt2_host = (float *)malloc(sizeof(float) * totalnums);
HIP_CHECK(hipMemcpy(cvt1_host, cvt1_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipMemcpy(cvt2_host, cvt2_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipDeviceSynchronize());
for (size_t idx = 0; idx < totalnums; idx++) {
TestType orig = e5m2_fnuz_nums[idx];
float cvt1 = cvt1_host[idx];
float cvt2 = cvt2_host[idx];
INFO("Original: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&orig)));
INFO("Cvt back: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&cvt1)));
REQUIRE(cvt1 == Approx(orig));
REQUIRE(cvt1 == cvt2);
REQUIRE(cvt2 == cvt1);
}
}
}
// Check the orientation encoded is correct
TEST_CASE("Unit_fp8_vector_basic_conversions") {
float f1 = 1.0f;
float2 f2 = {1.0f, 2.0f};
float4 f4 = {1.0f, 2.0f, 3.0f, 4.0f};
SECTION("e4m3-fnuz cvt float") {
__hip_fp8_e4m3_fnuz f8_1 = f1;
__hip_fp8x2_e4m3_fnuz f8_2 = f2;
__hip_fp8x4_e4m3_fnuz f8_4 = f4;
float cf1 = f8_1;
float2 cf2 = f8_2;
float4 cf4 = f8_4;
__hip_fp8x2_e4m3_fnuz tmp;
tmp.__x = __hip_cvt_float2_to_fp8x2(cf2, __HIP_SATFINITE, __HIP_E4M3_FNUZ);
float2 xtmp = tmp;
REQUIRE(f1 == cf1);
REQUIRE(f2 == cf2);
REQUIRE(f4 == cf4);
REQUIRE(xtmp == f2);
}
SECTION("e5m2-fnuz cvt float") {
__hip_fp8_e5m2_fnuz f8_1 = f1;
__hip_fp8x2_e5m2_fnuz f8_2 = f2;
__hip_fp8x4_e5m2_fnuz f8_4 = f4;
float cf1 = f8_1;
float2 cf2 = f8_2;
float4 cf4 = f8_4;
__hip_fp8x2_e5m2_fnuz tmp;
tmp.__x = __hip_cvt_float2_to_fp8x2(cf2, __HIP_SATFINITE, __HIP_E5M2_FNUZ);
float2 xtmp = tmp;
REQUIRE(f1 == cf1);
REQUIRE(f2 == cf2);
REQUIRE(f4 == cf4);
REQUIRE(xtmp == f2);
}
SECTION("e4m3-fnuz cvt double") {
double d1 = f1;
double2 d2 = {f2.x, f2.y};
double4 d4 = {f4.x, f4.y, f4.z, f4.w};
__hip_fp8_e4m3_fnuz f8_1 = d1;
__hip_fp8x2_e4m3_fnuz f8_2 = d2;
__hip_fp8x4_e4m3_fnuz f8_4 = d4;
double cf1 = f8_1;
float2 cf2 = f8_2;
float4 cf4 = f8_4;
__hip_fp8x2_e4m3_fnuz tmp;
tmp.__x = __hip_cvt_double2_to_fp8x2(d2, __HIP_SATFINITE, __HIP_E4M3_FNUZ);
float2 xtmp = tmp;
REQUIRE(d1 == cf1);
REQUIRE(d2 == double2{cf2.x, cf2.y});
REQUIRE(d4 == double4{cf4.x, cf4.y, cf4.z, cf4.w});
REQUIRE(double2{xtmp.x, xtmp.y} == d2);
}
SECTION("e5m2-fnuz cvt double") {
double d1 = f1;
double2 d2 = {f2.x, f2.y};
double4 d4 = {f4.x, f4.y, f4.z, f4.w};
__hip_fp8_e5m2_fnuz f8_1 = d1;
__hip_fp8x2_e5m2_fnuz f8_2 = d2;
__hip_fp8x4_e5m2_fnuz f8_4 = d4;
double cf1 = f8_1;
float2 cf2 = f8_2;
float4 cf4 = f8_4;
__hip_fp8x2_e5m2_fnuz tmp;
tmp.__x = __hip_cvt_double2_to_fp8x2(d2, __HIP_SATFINITE, __HIP_E5M2_FNUZ);
float2 xtmp = tmp;
REQUIRE(d1 == cf1);
REQUIRE(d2 == double2{cf2.x, cf2.y});
REQUIRE(d4 == double4{cf4.x, cf4.y, cf4.z, cf4.w});
REQUIRE(double2{xtmp.x, xtmp.y} == d2);
}
SECTION("e4m3-fnuz half2/bfloat162") {
auto bf16_val = __float22bfloat162_rn(f2);
auto half2_val = __float22half2_rn(f2);
__hip_fp8x2_e4m3_fnuz x1(bf16_val);
__hip_fp8x2_e4m3_fnuz x2(half2_val);
__hip_fp8x2_e4m3_fnuz tmp1;
tmp1.__x = __hip_cvt_bfloat16raw2_to_fp8x2(bf16_val, __HIP_SATFINITE, __HIP_E4M3_FNUZ);
float2 bf2_1 = tmp1;
tmp1.__x = __hip_cvt_halfraw2_to_fp8x2(half2_val, __HIP_SATFINITE, __HIP_E4M3_FNUZ);
float2 h2_1 = tmp1;
float2 f2_1 = x1;
float2 f2_2 = x2;
REQUIRE(f2_1 == f2);
REQUIRE(f2_2 == f2);
REQUIRE(f2 == bf2_1);
REQUIRE(f2 == h2_1);
}
SECTION("e5m2-fnuz half2/bfloat162") {
auto bf16_val = __float22bfloat162_rn(f2);
auto half2_val = __float22half2_rn(f2);
__hip_fp8x2_e5m2_fnuz x1(bf16_val);
__hip_fp8x2_e5m2_fnuz x2(half2_val);
__hip_fp8x2_e5m2_fnuz tmp1;
tmp1.__x = __hip_cvt_bfloat16raw2_to_fp8x2(bf16_val, __HIP_SATFINITE, __HIP_E5M2_FNUZ);
float2 bf2_1 = tmp1;
tmp1.__x = __hip_cvt_halfraw2_to_fp8x2(half2_val, __HIP_SATFINITE, __HIP_E5M2_FNUZ);
float2 h2_1 = tmp1;
float2 f2_1 = x1;
float2 f2_2 = x2;
REQUIRE(f2_1 == f2);
REQUIRE(f2_2 == f2);
REQUIRE(f2 == bf2_1);
REQUIRE(f2 == h2_1);
HIP_CHECK(hipFree(fnums));
HIP_CHECK(hipFree(cvt1_dev));
HIP_CHECK(hipFree(cvt2_dev));
free(cvt1_host);
free(cvt2_host);
}
}
projects/hip-tests/catch/unit/deviceLib/fp8_host.cc
+1096
مشاهده پرونده
تفاوت فایلی نمایش داده نمی شود زیرا این فایل بسیار بزرگ است Diff را بارگزاری کن
projects/hip-tests/catch/unit/deviceLib/fp8_ocp.cc
+939
مشاهده پرونده
@@ -0,0 +1,939 @@
/*
Copyright (c) 2024 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
#include <hip_test_common.hh>
#include <hip/hip_fp8.h>
#include <type_traits>
#include <vector>
#include <bitset>
/*
* This catch test is meant for FP8 OCP conversions
* tests only supported on gfx1200 and gfx1201 archs
*/
static_assert(sizeof(unsigned int) == sizeof(float));
std::string arch_type() {
hipDeviceProp_t prop;
int device;
HIP_CHECK(hipGetDevice(&device));
HIP_CHECK(hipGetDeviceProperties(&prop, device));
std::string gfxName(prop.gcnArchName);
return gfxName;
}
#define ARCH_TYPE_GFX1200(name) \
(name.find("gfx1200") != std::string::npos) || \
(name.find("gfx1201") != std::string::npos)
#define FP8_OCP_SKIP_TEST \
std::string gfxName = arch_type(); \
if (!(ARCH_TYPE_GFX1200(gfxName))) { \
HipTest::HIP_SKIP_TEST("This test can only be run on GFX1200"); \
return; \
}
#define __FP8_DEVICE__ __device__ static inline
template<typename T> __FP8_DEVICE__ void e4m3_ocp_device(T *val)
{
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3 tmp(*val);
*val = tmp;
#else
*val = 0;
#endif
}
template<typename T> __FP8_DEVICE__ void e5m2_ocp_device(T *val)
{
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2 tmp(*val);
*val = tmp;
#else
*val = 0;
#endif
}
template <typename T, bool is_e4m3_ocp> __global__ void cvt_float_fp8_float(T* in, size_t len) {
int i = threadIdx.x;
if (i < len) {
T val = in[i];
if constexpr (is_e4m3_ocp) {
e4m3_ocp_device<T>(&val);
in[i] = val;
} else {
e5m2_ocp_device<T>(&val);
in[i] = val;
}
}
}
template <typename T, bool is_e4m3_ocp>
std::vector<T> cpu_cvt_float_fp8_float(const std::vector<T>& nums) {
std::vector<T> ret;
ret.reserve(nums.size());
for (const auto& num : nums) {
T out = 0.0;
if constexpr (is_e4m3_ocp) {
__hip_fp8_e4m3 tmp(num);
out = tmp;
} else {
__hip_fp8_e5m2 tmp(num);
out = tmp;
}
ret.push_back(out);
}
return ret;
}
// This test only makes sense on gfx94x where device side convert will use the builtins to convert
// floats to fp8
TEMPLATE_TEST_CASE("Unit_fp8_ocp_compare_host_device", "", float, double) {
FP8_OCP_SKIP_TEST
std::vector<TestType> numbers = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
TestType* d_numbers;
HIP_CHECK(hipMalloc(&d_numbers, sizeof(TestType) * numbers.size()));
HIP_CHECK(hipMemcpy(d_numbers, numbers.data(), sizeof(TestType) * numbers.size(),
hipMemcpyHostToDevice));
std::vector<TestType> result(numbers.size(), 0.0f);
std::vector<TestType> cpu_result;
SECTION("e4m3_ocp") {
cpu_result = cpu_cvt_float_fp8_float<TestType, true>(numbers);
auto kernel = cvt_float_fp8_float<TestType, true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(TestType) * numbers.size(),
hipMemcpyDeviceToHost));
}
SECTION("e5m2_ocp") {
cpu_result = cpu_cvt_float_fp8_float<TestType, false>(numbers);
auto kernel = cvt_float_fp8_float<TestType, false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(TestType) * numbers.size(),
hipMemcpyDeviceToHost));
}
REQUIRE(cpu_result.size() == result.size());
for (size_t i = 0; i < result.size(); i++) {
INFO("Checking: " << numbers[i] << " cpu convert: " << cpu_result[i]
<< " - gpu_result: " << result[i]);
CHECK(cpu_result[i] == result[i]);
}
}
__FP8_DEVICE__ void e4m3_fp8x2_ocp_device(float2 *val)
{
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x2_e4m3 tmp(*val);
*val = tmp;
#else
*val = float2(0.0,0.0);
#endif
}
__FP8_DEVICE__ void e5m2_fp8x2_ocp_device(float2 *val)
{
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x2_e5m2 tmp(*val);
*val = tmp;
#else
*val = float2(0.0,0.0);
#endif
}
template <bool is_e4m3_ocp> __global__ void cvt_float2_fp8x2_float2(float2* in, size_t size) {
int i = threadIdx.x;
if (i < size) {
float2 val = in[i];
if constexpr (is_e4m3_ocp) {
e4m3_fp8x2_ocp_device(&val);
in[i] = val;
} else {
e5m2_fp8x2_ocp_device(&val);
in[i] = val;
}
}
}
template <bool is_e4m3_ocp>
std::vector<float2> cpu_cvt_float2_fp8x2_float2(const std::vector<float2>& nums) {
std::vector<float2> ret;
ret.reserve(nums.size());
for (const auto& num : nums) {
float2 out = {0.0f, 0.0f};
if constexpr (is_e4m3_ocp) {
__hip_fp8x2_e4m3 tmp(num);
out = tmp;
} else {
__hip_fp8x2_e5m2 tmp(num);
out = tmp;
}
ret.push_back(out);
}
return ret;
}
TEST_CASE("Unit_fp8x2_ocp_compare_host_device") {
FP8_OCP_SKIP_TEST
std::vector<float> numbers_input = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
std::vector<float2> numbers;
numbers.reserve(numbers_input.size());
for (size_t i = 0, end = numbers_input.size() - 1; i < numbers_input.size(); i++, end--) {
float2 ret(numbers_input[i], numbers_input[end]);
numbers.push_back(ret);
}
float2* d_numbers;
HIP_CHECK(hipMalloc(&d_numbers, sizeof(float2) * numbers.size()));
HIP_CHECK(
hipMemcpy(d_numbers, numbers.data(), sizeof(float2) * numbers.size(), hipMemcpyHostToDevice));
std::vector<float2> result(numbers.size(), float2{0.0f, 0.0f});
std::vector<float2> cpu_result;
SECTION("e4m3_ocp") {
cpu_result = cpu_cvt_float2_fp8x2_float2<true>(numbers);
auto kernel = cvt_float2_fp8x2_float2<true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
}
SECTION("e5m2_ocp") {
cpu_result = cpu_cvt_float2_fp8x2_float2<false>(numbers);
auto kernel = cvt_float2_fp8x2_float2<false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
}
REQUIRE(cpu_result.size() == result.size());
for (size_t i = 0; i < result.size(); i++) {
CHECK(cpu_result[i] == result[i]);
}
}
TEST_CASE("Unit_fp8x2_ocp_split_compare") {
FP8_OCP_SKIP_TEST
std::vector<float> numbers_input = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
std::vector<float2> numbers;
numbers.reserve(numbers_input.size());
for (size_t i = 0, end = numbers_input.size() - 1; i < numbers_input.size(); i++, end--) {
float2 ret(numbers_input[i], numbers_input[end]);
numbers.push_back(ret);
}
float2* d_numbers;
HIP_CHECK(hipMalloc(&d_numbers, sizeof(float2) * numbers.size()));
HIP_CHECK(
hipMemcpy(d_numbers, numbers.data(), sizeof(float2) * numbers.size(), hipMemcpyHostToDevice));
std::vector<float2> result(numbers.size(), float2{0.0f, 0.0f});
std::vector<float2> cpu_result;
cpu_result.reserve(result.size());
SECTION("e4m3_ocp") {
for (const auto& num : numbers) {
__hip_fp8_e4m3 t_a(num.x);
__hip_fp8_e4m3 t_b(num.y);
float a = t_a, b = t_b;
cpu_result.push_back(float2(a, b));
}
auto kernel = cvt_float2_fp8x2_float2<true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
}
SECTION("e5m2_ocp") {
for (const auto& num : numbers) {
__hip_fp8_e5m2 t_a(num.x);
__hip_fp8_e5m2 t_b(num.y);
float a = t_a, b = t_b;
cpu_result.push_back(float2(a, b));
}
auto kernel = cvt_float2_fp8x2_float2<false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float2) * numbers.size(),
hipMemcpyDeviceToHost));
}
HIP_CHECK(hipDeviceSynchronize());
REQUIRE(cpu_result.size() == result.size());
for (size_t i = 0; i < result.size(); i++) {
INFO("cpu x: " << cpu_result[i].x << " y: " << cpu_result[i].y << " gpu x: " << result[i].x
<< " y: " << result[i].y);
CHECK(cpu_result[i] == result[i]);
}
}
__FP8_DEVICE__ void e4m3_fp8x4_ocp_device(float4 *val)
{
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x4_e4m3 tmp(*val);
*val = tmp;
#else
*val = float4(0.0,0.0,0.0,0.0);
#endif
}
__FP8_DEVICE__ void e5m2_fp8x4_ocp_device(float4 *val)
{
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8x4_e5m2 tmp(*val);
*val = tmp;
#else
*val = float4(0.0,0.0,0.0,0.0);
#endif
}
template <bool is_e4m3_ocp> __global__ void cvt_float4_fp8x4_float4(float4* in, size_t size) {
int i = threadIdx.x;
if (i < size) {
float4 val = in[i];
if constexpr (is_e4m3_ocp) {
e4m3_fp8x4_ocp_device(&val);
in[i] = val;
} else {
e5m2_fp8x4_ocp_device(&val);
in[i] = val;
}
}
}
TEST_CASE("Unit_fp8x4_ocp_split_compare") {
FP8_OCP_SKIP_TEST
std::vector<float> numbers_input = {0.0f, 1.0f, 1.1f, 2.0f, 2.1f, 3.0f, 3.2f,
3.3f, 4.0f, 4.5f, 10.0f, 11.0f, 12.2f, 14.1f};
std::vector<float> numbers_input2 = {1.3f, 1.6f, 1.8f, 2.5f, 2.9f, 3.8f, 3.9f,
5.5f, 7.1f, 8.5f, 11.2f, 13.5f, 16.1f, 19.4f};
std::vector<float4> numbers;
numbers.reserve(numbers_input.size());
for (size_t i = 0, end = numbers_input.size() - 1; i < numbers_input.size(); i++, end--) {
float4 ret(numbers_input[i], numbers_input[end], numbers_input2[i], numbers_input2[end]);
numbers.push_back(ret);
}
float4* d_numbers;
HIP_CHECK(hipMalloc(&d_numbers, sizeof(float4) * numbers.size()));
HIP_CHECK(
hipMemcpy(d_numbers, numbers.data(), sizeof(float4) * numbers.size(), hipMemcpyHostToDevice));
std::vector<float4> result(numbers.size(), float4{0.0f, 0.0f, 0.0f, 0.0f});
std::vector<float4> cpu_result;
cpu_result.reserve(result.size());
SECTION("e4m3_ocp") {
for (const auto& num : numbers) {
__hip_fp8_e4m3 t_a(num.x);
__hip_fp8_e4m3 t_b(num.y);
__hip_fp8_e4m3 t_c(num.z);
__hip_fp8_e4m3 t_d(num.w);
float a = t_a, b = t_b, c = t_c, d = t_d;
cpu_result.push_back(float4(a, b, c, d));
}
auto kernel = cvt_float4_fp8x4_float4<true>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float4) * numbers.size(),
hipMemcpyDeviceToHost));
}
SECTION("e5m2_ocp") {
for (const auto& num : numbers) {
__hip_fp8_e5m2 t_a(num.x);
__hip_fp8_e5m2 t_b(num.y);
__hip_fp8_e5m2 t_c(num.z);
__hip_fp8_e5m2 t_d(num.w);
float a = t_a, b = t_b, c = t_c, d = t_d;
cpu_result.push_back(float4(a, b, c, d));
}
auto kernel = cvt_float4_fp8x4_float4<false>;
kernel<<<1, numbers.size()>>>(d_numbers, numbers.size());
HIP_CHECK(hipMemcpy(result.data(), d_numbers, sizeof(float4) * numbers.size(),
hipMemcpyDeviceToHost));
}
HIP_CHECK(hipDeviceSynchronize());
REQUIRE(cpu_result.size() == result.size());
for (size_t i = 0; i < result.size(); i++) {
INFO("original: x: " << numbers[i].x << " y: " << numbers[i].y << " z: " << numbers[i].z
<< " w: " << numbers[i].w);
INFO("cpu x: " << cpu_result[i].x << " y: " << cpu_result[i].y << " z: " << cpu_result[i].z
<< " w: " << cpu_result[i].w);
INFO("gpu x: " << result[i].x << " y: " << result[i].y << " z: " << result[i].z
<< " w: " << result[i].w);
CHECK(cpu_result[i] == result[i]);
}
}
__FP8_DEVICE__ bool e4m3_bool_ocp_device(float val)
{
bool x = false;
float y = val;
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3 tmp(y);
x = tmp;
#else
x = (y == 0);
#endif
return x;
}
__FP8_DEVICE__ bool e5m2_bool_ocp_device(float val)
{
bool x = false;
float y = val;
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2 tmp(y);
x = tmp;
#else
x = (y == 0);
#endif
return x;
}
template <bool is_e4m3_ocp> __global__ void fp8_2_bool(float* f, bool* ret, size_t size) {
int i = threadIdx.x;
bool r = false;
if (i < size) {
if constexpr (is_e4m3_ocp) {
r = e4m3_bool_ocp_device(f[i]);
} else {
r = e5m2_bool_ocp_device(f[i]);
}
ret[i] = r;
}
}
TEST_CASE("Unit_fp8_ocp_bool_device") {
FP8_OCP_SKIP_TEST
// clang-format off
std::vector<float> fvals{-10.0f, -1.0f, -0.0f, 0.0f, 1.0f, 10.0f};
std::vector<bool> tvals {true, true, false, false, true, true};
// clang-format on
bool result[] = {false, false, false,
false, false, false}; // cant use std::vector coz data() = delete
SECTION("e4m3_ocp-gpu") {
float* d_in{nullptr};
bool* d_res{nullptr};
HIP_CHECK(hipMalloc(&d_in, sizeof(float) * tvals.size()));
HIP_CHECK(hipMalloc(&d_res, sizeof(bool) * tvals.size()));
auto kernel = fp8_2_bool<true>;
HIP_CHECK(hipMemcpy(d_in, fvals.data(), sizeof(float) * fvals.size(), hipMemcpyHostToDevice));
kernel<<<1, tvals.size()>>>(d_in, d_res, tvals.size());
HIP_CHECK(hipMemcpy(result, d_res, sizeof(bool) * tvals.size(), hipMemcpyDeviceToHost));
HIP_CHECK(hipFree(d_in));
HIP_CHECK(hipFree(d_res));
}
SECTION("e5m2_ocp-gpu") {
float* d_in{nullptr};
bool* d_res{nullptr};
HIP_CHECK(hipMalloc(&d_in, sizeof(float) * tvals.size()));
HIP_CHECK(hipMalloc(&d_res, sizeof(bool) * tvals.size()));
HIP_CHECK(hipMemcpy(d_in, fvals.data(), sizeof(float) * fvals.size(), hipMemcpyHostToDevice));
auto kernel = fp8_2_bool<false>;
kernel<<<1, tvals.size()>>>(d_in, d_res, tvals.size());
HIP_CHECK(hipMemcpy(result, d_res, sizeof(bool) * tvals.size(), hipMemcpyDeviceToHost));
HIP_CHECK(hipFree(d_in));
HIP_CHECK(hipFree(d_res));
}
for (size_t i = 0; i < tvals.size(); i++) {
INFO("Check for: " << fvals[i] << " expected: " << tvals[i] << " result: " << result[i]);
REQUIRE(result[i] == tvals[i]);
}
}
std::vector<__hip_fp8_storage_t> get_all_fp8_nums(bool is_e4m3_ocp) {
std::vector<__hip_fp8_storage_t> ret;
constexpr unsigned short max_fp8_num = 0b1111'1111;
ret.reserve(max_fp8_num + 1 );
for (unsigned short i = 0; i <= max_fp8_num; i++) {
if(is_e4m3_ocp) {
if ((i & 0x7f) != 0x7f) { // 0xff and 0x7f are nan
ret.push_back(static_cast<__hip_fp8_storage_t>(i));
}
}
else {
if ((i & 0x7f) < 0x7c) { // 0x7c 0x7d 0x7e are nan and 0x7f is inf
ret.push_back(static_cast<__hip_fp8_storage_t>(i));
}
}
}
return ret;
}
__FP8_DEVICE__ __hip_fp8_storage_t e4m3_ocp_fp8_device(float val)
{
__hip_fp8_storage_t x = 0;
float y = val;
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3 tmp(y);
x = tmp.__x;
#else
x = (y == 0) ? 0x0 : 0x7f;
#endif
return x;
}
__FP8_DEVICE__ __hip_fp8_storage_t e5m2_ocp_fp8_device(float val)
{
__hip_fp8_storage_t x = 0;
float y = val;
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2 tmp(y);
x = tmp.__x;
#else
x = (y == 0) ? 0x0 : 0x7f;
#endif
return x;
}
template <bool is_e4m3_ocp>
__global__ void Type_to_fp8(float* f, __hip_fp8_storage_t* res, size_t size) {
auto i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < size) {
if constexpr (is_e4m3_ocp) {
res[i] = e4m3_ocp_fp8_device(f[i]);
} else {
res[i] = e5m2_ocp_fp8_device(f[i]);
}
}
}
TEST_CASE("Unit_all_fp8_ocp_cvt") {
FP8_OCP_SKIP_TEST
bool is_e4m3_ocp = GENERATE(true, false);
std::vector<float> f_vals;
std::vector<__hip_fp8_storage_t> all_vals;
SECTION("all representable number") {
all_vals = get_all_fp8_nums(is_e4m3_ocp);
f_vals.reserve(all_vals.size());
for (const auto& fp8 : all_vals) {
float f = 0.0f;
if (is_e4m3_ocp) {
__hip_fp8_e4m3 tmp;
tmp.__x = fp8;
f = tmp;
} else {
__hip_fp8_e5m2 tmp;
tmp.__x = fp8;
f = tmp;
}
f_vals.push_back(f);
}
}
SECTION("Range stepped numbers") {
constexpr float lhs = -200.0f;
constexpr float rhs = 200.0f;
constexpr float step = 0.1234f;
f_vals.reserve(4000);
all_vals.reserve(4000);
for (float fval = lhs; fval <= rhs; fval += step) {
if (is_e4m3_ocp) {
__hip_fp8_e4m3 tmp = fval;
all_vals.push_back(tmp.__x);
} else {
__hip_fp8_e5m2 tmp = fval;
all_vals.push_back(tmp.__x);
}
f_vals.push_back(fval);
}
}
float* d_f_vals;
__hip_fp8_storage_t* d_res;
HIP_CHECK(hipMalloc(&d_f_vals, sizeof(float) * f_vals.size()));
HIP_CHECK(hipMalloc(&d_res, sizeof(__hip_fp8_storage_t) * f_vals.size()));
HIP_CHECK(
hipMemcpy(d_f_vals, f_vals.data(), sizeof(float) * f_vals.size(), hipMemcpyHostToDevice));
auto fp8_kernel = is_e4m3_ocp ? Type_to_fp8<true> : Type_to_fp8<false>;
fp8_kernel<<<(f_vals.size() / 256) + 1, 256>>>(d_f_vals, d_res, f_vals.size());
std::vector<__hip_fp8_storage_t> final_res(f_vals.size(), static_cast<__hip_fp8_storage_t>(0));
HIP_CHECK(hipMemcpy(final_res.data(), d_res, sizeof(__hip_fp8_storage_t) * final_res.size(),
hipMemcpyDeviceToHost));
for (size_t i = 0; i < final_res.size(); i++) {
INFO("Checking: " << f_vals[i] << " for: " << (is_e4m3_ocp ? "e4m3_ocp" : "e5m2_ocp")
<< " original: " << (int)all_vals[i]
<< " convert back: " << (int)final_res[i]);
float gpu_cvt_res = 0.0f, cpu_cvt_res = 0.0f;
if (is_e4m3_ocp) {
__hip_fp8_e4m3 gtmp;
gtmp.__x = final_res[i];
gpu_cvt_res = gtmp;
__hip_fp8_e4m3 ctmp;
ctmp.__x = all_vals[i];
cpu_cvt_res = ctmp;
} else {
__hip_fp8_e5m2 gtmp;
gtmp.__x = final_res[i];
gpu_cvt_res = gtmp;
__hip_fp8_e5m2 ctmp;
ctmp.__x = all_vals[i];
cpu_cvt_res = ctmp;
}
INFO("cpu cvt val: " << cpu_cvt_res << " gpu cvt val: " << gpu_cvt_res);
REQUIRE(cpu_cvt_res == gpu_cvt_res);
}
HIP_CHECK(hipFree(d_f_vals));
HIP_CHECK(hipFree(d_res));
}
template<typename T> __FP8_DEVICE__ void e4m3_ocp_fp8_cvt(T val, float *cvt1, float *cvt2)
{
T y = val;
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e4m3 tmp(y);
*cvt1 = tmp;
__hip_fp8_e4m3 tmp1;
tmp1.__x = std::is_same<T, float>::value
? __hip_cvt_float_to_fp8(val, __HIP_SATFINITE, __HIP_E4M3)
: __hip_cvt_double_to_fp8(val, __HIP_SATFINITE, __HIP_E4M3);
;
*cvt2 = tmp1;
#else
*cvt1 = (y == 0) ? 0 : y;
*cvt2 = (y == 0) ? 0 : y;
#endif
}
template<typename T> __FP8_DEVICE__ void e5m2_ocp_fp8_cvt(T val, float *cvt1, float *cvt2)
{
T y = val;
#if (defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
__hip_fp8_e5m2 tmp(y);
*cvt1 = tmp;
__hip_fp8_e5m2 tmp1;
tmp1.__x = std::is_same<T, float>::value
? __hip_cvt_float_to_fp8(val, __HIP_SATFINITE, __HIP_E5M2)
: __hip_cvt_double_to_fp8(val, __HIP_SATFINITE, __HIP_E5M2);
;
*cvt2 = tmp1;
#else
*cvt1 = (y == 0) ? 0 : y;
*cvt2 = (y == 0) ? 0 : y;
#endif
}
template <typename T, bool is_e4m3_ocp>
__global__ void Type_to_fp8_cvt(T* f, float *cvt1, float *cvt2, size_t size) {
auto i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < size) {
if constexpr (is_e4m3_ocp) {
e4m3_ocp_fp8_cvt(f[i], &cvt1[i], &cvt2[i]);
} else {
e5m2_ocp_fp8_cvt(f[i], &cvt1[i], &cvt2[i]);
}
}
}
TEMPLATE_TEST_CASE("Unit_fp8_ocp_correctness_device", "", float, double) {
FP8_OCP_SKIP_TEST
SECTION("e4m3_ocp") {
/* These are basically all the fp8 - e4m3_ocp type numbers.
* They can be generated by iterating over 0'0000'000 and converting them to fp32 number
* skipping the nan/inf */
std::vector<TestType> e4m3_ocp_nums = { 0, 0.00195312, 0.00390625,
0.00585938, 0.0078125, 0.00976562,
0.0117188, 0.0136719, 0.015625,
0.0175781, 0.0195312, 0.0214844,
0.0234375, 0.0253906, 0.0273438,
0.0292969, 0.03125, 0.0351562,
0.0390625, 0.0429688, 0.046875,
0.0507812, 0.0546875, 0.0585938,
0.0625, 0.0703125, 0.078125,
0.0859375, 0.09375, 0.101562,
0.109375, 0.117188, 0.125,
0.140625, 0.15625, 0.171875,
0.1875, 0.203125, 0.21875,
0.234375, 0.25, 0.28125,
0.3125, 0.34375, 0.375,
0.40625, 0.4375, 0.46875,
0.5, 0.5625, 0.625,
0.6875, 0.75, 0.8125,
0.875, 0.9375, 1,
1.125, 1.25, 1.375,
1.5, 1.625, 1.75,
1.875, 2, 2.25,
2.5, 2.75, 3,
3.25, 3.5, 3.75,
4, 4.5, 5,
5.5, 6, 6.5,
7, 7.5, 8,
9, 10, 11,
12, 13, 14,
15, 16, 18,
20, 22, 24,
26, 28, 30,
32, 36, 40,
44, 48, 52,
56, 60, 64,
72, 80, 88,
96, 104, 112,
120, 128, 144,
160, 176, 192,
208, 224, 240,
256, 288, 320,
352, 384, 416,
448, -0, -0.00195312,
-0.00390625, -0.00585938, -0.0078125,
-0.00976562, -0.0117188, -0.0136719,
-0.015625, -0.0175781, -0.0195312,
-0.0214844, -0.0234375, -0.0253906,
-0.0273438, -0.0292969, -0.03125,
-0.0351562, -0.0390625, -0.0429688,
-0.046875, -0.0507812, -0.0546875,
-0.0585938, -0.0625, -0.0703125,
-0.078125, -0.0859375, -0.09375,
-0.101562, -0.109375, -0.117188,
-0.125, -0.140625, -0.15625,
-0.171875, -0.1875, -0.203125,
-0.21875, -0.234375, -0.25,
-0.28125, -0.3125, -0.34375,
-0.375, -0.40625, -0.4375,
-0.46875, -0.5, -0.5625,
-0.625, -0.6875, -0.75,
-0.8125, -0.875, -0.9375,
-1, -1.125, -1.25,
-1.375, -1.5, -1.625,
-1.75, -1.875, -2,
-2.25, -2.5, -2.75,
-3, -3.25, -3.5,
-3.75, -4, -4.5,
-5, -5.5, -6,
-6.5, -7, -7.5,
-8, -9, -10,
-11, -12, -13,
-14, -15, -16,
-18, -20, -22,
-24, -26, -28,
-30, -32, -36,
-40, -44, -48,
-52, -56, -60,
-64, -72, -80,
-88, -96, -104,
-112, -120, -128,
-144, -160, -176,
-192, -208, -224,
-240, -256, -288,
-320, -352, -384,
-416, -448};
size_t totalnums = e4m3_ocp_nums.size();
TestType *fnums; HIP_CHECK(hipMalloc((void **)&fnums, totalnums * sizeof(TestType)));
float *cvt1_dev; HIP_CHECK(hipMalloc((void **)&cvt1_dev, totalnums * sizeof(TestType)));
float *cvt2_dev; HIP_CHECK(hipMalloc((void **)&cvt2_dev, totalnums * sizeof(TestType)));
HIP_CHECK(hipMemcpy(fnums, e4m3_ocp_nums.data(), totalnums * sizeof(TestType),
hipMemcpyHostToDevice));
auto fp8_kernel = Type_to_fp8_cvt<TestType, true>;
fp8_kernel<<<totalnums / 256 + 1, 256>>>(fnums, cvt1_dev, cvt2_dev, totalnums);
float *cvt1_host = (float *)malloc(sizeof(float) * totalnums);
float *cvt2_host = (float *)malloc(sizeof(float) * totalnums);
HIP_CHECK(hipMemcpy(cvt1_host, cvt1_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipMemcpy(cvt2_host, cvt2_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipDeviceSynchronize());
for (size_t idx = 0; idx < totalnums; idx++) {
TestType orig = e4m3_ocp_nums[idx];
float cvt1 = cvt1_host[idx];
float cvt2 = cvt2_host[idx];
INFO("Original: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&orig)));
INFO("Cvt back: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&cvt1)));
REQUIRE(cvt1 == Approx(orig));
REQUIRE(cvt2 == cvt1);
}
HIP_CHECK(hipFree(fnums));
HIP_CHECK(hipFree(cvt1_dev));
HIP_CHECK(hipFree(cvt2_dev));
free(cvt1_host);
free(cvt2_host);
}
SECTION("e5m2_ocp") {
/* These are basically all the fp8 - e5m2_ocp type numbers.
* They can be generated by iterating over 0'00000'00 converting them to fp32 number skipping
* the nan/inf */
std::vector<TestType> e5m2_ocp_nums = { 0, 1.52588e-05, 3.05176e-05,
4.57764e-05, 6.10352e-05, 7.62939e-05,
9.15527e-05, 0.000106812, 0.00012207,
0.000152588, 0.000183105, 0.000213623,
0.000244141, 0.000305176, 0.000366211,
0.000427246, 0.000488281, 0.000610352,
0.000732422, 0.000854492, 0.000976562,
0.0012207, 0.00146484, 0.00170898,
0.00195312, 0.00244141, 0.00292969,
0.00341797, 0.00390625, 0.00488281,
0.00585938, 0.00683594, 0.0078125,
0.00976562, 0.0117188, 0.0136719,
0.015625, 0.0195312, 0.0234375,
0.0273438, 0.03125, 0.0390625,
0.046875, 0.0546875, 0.0625,
0.078125, 0.09375, 0.109375,
0.125, 0.15625, 0.1875,
0.21875, 0.25, 0.3125,
0.375, 0.4375, 0.5,
0.625, 0.75, 0.875,
1, 1.25, 1.5,
1.75, 2, 2.5,
3, 3.5, 4,
5, 6, 7,
8, 10, 12,
14, 16, 20,
24, 28, 32,
40, 48, 56,
64, 80, 96,
112, 128, 160,
192, 224, 256,
320, 384, 448,
512, 640, 768,
896, 1024, 1280,
1536, 1792, 2048,
2560, 3072, 3584,
4096, 5120, 6144,
7168, 8192, 10240,
12288, 14336, 16384,
20480, 24576, 28672,
32768, 40960, 49152,
57344, -0, -1.52588e-05,
-3.05176e-05, -4.57764e-05, -6.10352e-05,
-7.62939e-05, -9.15527e-05, -0.000106812,
-0.00012207, -0.000152588, -0.000183105,
-0.000213623, -0.000244141, -0.000305176,
-0.000366211, -0.000427246, -0.000488281,
-0.000610352, -0.000732422, -0.000854492,
-0.000976562, -0.0012207, -0.00146484,
-0.00170898, -0.00195312, -0.00244141,
-0.00292969, -0.00341797, -0.00390625,
-0.00488281, -0.00585938, -0.00683594,
-0.0078125, -0.00976562, -0.0117188,
-0.0136719, -0.015625, -0.0195312,
-0.0234375, -0.0273438, -0.03125,
-0.0390625, -0.046875, -0.0546875,
-0.0625, -0.078125, -0.09375,
-0.109375, -0.125, -0.15625,
-0.1875, -0.21875, -0.25,
-0.3125, -0.375, -0.4375,
-0.5, -0.625, -0.75,
-0.875, -1, -1.25,
-1.5, -1.75, -2,
-2.5, -3, -3.5,
-4, -5, -6,
-7, -8, -10,
-12, -14, -16,
-20, -24, -28,
-32, -40, -48,
-56, -64, -80,
-96, -112, -128,
-160, -192, -224,
-256, -320, -384,
-448, -512, -640,
-768, -896, -1024,
-1280, -1536, -1792,
-2048, -2560, -3072,
-3584, -4096, -5120,
-6144, -7168, -8192,
-10240, -12288, -14336,
-16384, -20480, -24576,
-28672, -32768, -40960,
-49152, -57344};
size_t totalnums = e5m2_ocp_nums.size();
TestType *fnums; HIP_CHECK(hipMalloc((void **)&fnums, totalnums * sizeof(TestType)));
float *cvt1_dev; HIP_CHECK(hipMalloc((void **)&cvt1_dev, totalnums * sizeof(TestType)));
float *cvt2_dev; HIP_CHECK(hipMalloc((void **)&cvt2_dev, totalnums * sizeof(TestType)));
HIP_CHECK(hipMemcpy(fnums, e5m2_ocp_nums.data(), totalnums * sizeof(TestType),
hipMemcpyHostToDevice));
auto fp8_kernel = Type_to_fp8_cvt<TestType, false>;
fp8_kernel<<<totalnums / 256 + 1, 256>>>(fnums, cvt1_dev, cvt2_dev, totalnums);
float *cvt1_host = (float *)malloc(sizeof(float) * totalnums);
float *cvt2_host = (float *)malloc(sizeof(float) * totalnums);
HIP_CHECK(hipMemcpy(cvt1_host, cvt1_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipMemcpy(cvt2_host, cvt2_dev, totalnums * sizeof(float) , hipMemcpyDeviceToHost));
HIP_CHECK(hipDeviceSynchronize());
for (size_t idx = 0; idx < totalnums; idx++) {
TestType orig = e5m2_ocp_nums[idx];
float cvt1 = cvt1_host[idx];
float cvt2 = cvt2_host[idx];
INFO("Original: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&orig)));
INFO("Cvt back: " << std::bitset<32>(*reinterpret_cast<const unsigned int*>(&cvt1)));
REQUIRE(cvt1 == Approx(orig));
REQUIRE(cvt2 == cvt1);
}
HIP_CHECK(hipFree(fnums));
HIP_CHECK(hipFree(cvt1_dev));
HIP_CHECK(hipFree(cvt2_dev));
free(cvt1_host);
free(cvt2_host);
}
}