ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 02:02:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Rebase the PR #1520 to ROCm repo. (#1574)

Browse Source 
* Implement hiprtc for codegen tests

* Introduce gemm_softmax_gemm to codegen.

* Fix codegen build issues.

* Address PR comments.

* Separate ck_host lib and gemm_softmax_gemm into different PR.

* Fix cmake.

* Replace ENV variable with CMake option for toggling hipRTC in codegen
tests.

* Address PR comments.

* fix clang format

* Add missing header in magic_division.hpp

* - Workaround for hipRTC content wrapper
- Move descriptor for gemm_softmax_gemm to different branch

* Fix formatting.

* Revert "Fix formatting."

This reverts commit b5209eaef4.

* formatting fix

* fixed header guard issues

* updated header guards

* updated data_type for new types

* fixed redefinition error

* Add codegen test for batched_gemm_softmax_gemm.

Signed-off-by: Mirza Halilcevic <mirza.halilcevic@amd.com>

* formatting fix

---------

Signed-off-by: Mirza Halilcevic <mirza.halilcevic@amd.com>
Co-authored-by: Dino Musić <dino.music@htecgroup.com>
Co-authored-by: Mirza Halilcevic <mirza.halilcevic@htecgroup.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: arai713 <67439843+arai713@users.noreply.github.com>
Co-authored-by: Astha Rai <astha.rai713@gmail.com>
Co-authored-by: Mirza Halilcevic <mirza.halilcevic@amd.com>

[ROCm/composable_kernel commit: 68a08c872e]
This commit is contained in:
Illia Silin
2025-02-20 18:58:14 -08:00
committed by GitHub
parent bd738ccf57
commit ff04241799
32 changed files with 880 additions and 517 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
codegen/test/batched_gemm_softmax_gemm.cpp Normal file
View File

@@ -0,0 +1,87 @@
#include "ck/host/device_batched_gemm_softmax_gemm/problem.hpp"
#include "ck/host/stringutils.hpp"
#include "ck/host/utils.hpp"
#include "common.hpp"
#include <rtc/compile_kernel.hpp>
#include <rtc/hip.hpp>
#include <test.hpp>
#include <cmath>
using half = _Float16;
const std::string gemm_compile_check = R"__ck__(
#include <${include}>
extern "C" __global__ void f(const ck::half_t* a, const ck::half_t* b, const ck::half_t* b1, ck::half_t* c) {
using G = ${template};
constexpr auto desc = G::make_descriptor(ck::make_naive_tensor_descriptor(ck::make_tuple(${m}, ${k}), ck::make_tuple(${m}, 1)),
ck::make_naive_tensor_descriptor(ck::make_tuple(${n}, ${k}), ck::make_tuple(${n}, 1)),
ck::make_naive_tensor_descriptor(ck::make_tuple(${n}, ${o}), ck::make_tuple(1, ${n})),
ck::make_naive_tensor_descriptor(ck::make_tuple(${m}, ${o}), ck::make_tuple(${m}, 1)));
static_assert(desc.IsValid(), "Invalid ck gemm.");
if constexpr(desc.IsValid())
{
${template}::Run(desc,
1.0,
a,
b,
b1,
c);
}
}
)__ck__";
TEST_CASE(test_problem_kernel)
{
ck::host::device_batched_gemm_softmax_gemm::Problem prob;
prob.M = 1024;
prob.N = 1024;
prob.K = 1024;
prob.O = 1024;
prob.TransB = true;
check_all<half> check1, check2;
auto a = to_gpu(generate_buffer<half>(1024 * 1024, 0));
auto b = to_gpu(generate_buffer<half>(1024 * 1024, 1));
auto b1 = to_gpu(generate_buffer<half>(1024 * 1024, 2));
auto c = to_gpu(generate_buffer<half>(1024 * 1024, 3));
std::string epilogue = "";
std::string prologue = "";
auto solutions = prob.GetSolutions("gfx90a", prologue, epilogue);
std::cout << "Num solutions: " << solutions.size() << std::endl;
for(auto i = 0; i < solutions.size(); ++i)
{
std::cout << "Testing solution " << std::to_string(i + 1) << std::endl;
auto&& solution = solutions[i];
auto src = ck::host::InterpolateString(gemm_compile_check,
{{"include", prob.GetIncludeHeader()},
{"template", solution.ToTemplateString()},
{"m", std::to_string(prob.M)},
{"n", std::to_string(prob.N)},
{"k", std::to_string(prob.K)},
{"o", std::to_string(prob.O)}});
auto srcs = get_headers_for_test();
srcs.push_back({"main.cpp", src});
rtc::compile_options options;
options.kernel_name = "f";
auto k = rtc::compile_kernel(srcs, options);
auto block_size = solution.GetTemplateParameter<std::size_t>("BlockSize");
auto m_per_block = solution.GetTemplateParameter<std::size_t>("Gemm01MPerBlock");
auto n_per_block = solution.GetTemplateParameter<std::size_t>("Gemm1NPerBlock");
auto grid_size = ck::host::integer_divide_ceil(prob.M, m_per_block) *
ck::host::integer_divide_ceil(prob.N, n_per_block);
k.launch(nullptr, grid_size * block_size, block_size)(
a.data(), b.data(), b1.data(), c.data());
if(solution.GetTemplateParameter<bool>("MaskOutUpperTriangle"))
CHECK(report(solution, check1(rtc::from_gpu(c))));
else
CHECK(report(solution, check2(rtc::from_gpu(c))));
}
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }

134
codegen/test/gemm_multiple_d.cpp
View File

@@ -6,134 +6,24 @@
#include "ck/host/headers.hpp"
#include "ck/host/stringutils.hpp"
#include "ck/host/utils.hpp"
#include <algorithm>
#include <cmath>
#include <iterator>
#include <random>
#include <test.hpp>
#include "common.hpp"
#include <rtc/compile_kernel.hpp>
#include <rtc/hip.hpp>
#include <test.hpp>
#include <algorithm>
#include <cmath>
#include <fstream>
#include <iterator>
#include <random>
using half = _Float16;
// using half = __fp16;
std::vector<rtc::src_file> get_headers_for_test()
{
std::vector<rtc::src_file> result;
auto hs = ck::host::GetHeaders();
std::transform(
hs.begin(), hs.end(), std::back_inserter(result), [&](const auto& p) -> rtc::src_file {
return {p.first, p.second};
});
return result;
}
template <class T>
rtc::buffer<T> generate_buffer(std::size_t n, std::size_t seed = 0)
{
rtc::buffer<T> result(n);
std::mt19937 gen(seed);
std::uniform_real_distribution<double> dis(-1.0);
std::generate(result.begin(), result.end(), [&] { return dis(gen); });
return result;
}
template <class T, class U>
bool allclose(const T& a, const U& b, double atol = 0.01, double rtol = 0.01)
{
return std::equal(a.begin(), a.end(), b.begin(), b.end(), [&](double x, double y) {
return fabs(x - y) < atol + rtol * fabs(y);
});
}
std::string classify(double x)
{
switch(std::fpclassify(x))
{
case FP_INFINITE: return "inf";
case FP_NAN: return "nan";
case FP_NORMAL: return "normal";
case FP_SUBNORMAL: return "subnormal";
case FP_ZERO: return "zero";
default: return "unknown";
}
}
template <class Buffer>
void print_classification(const Buffer& x)
{
std::unordered_set<std::string> result;
for(const auto& i : x)
result.insert(classify(i));
for(const auto& c : result)
std::cout << c << ", ";
std::cout << std::endl;
}
template <class Buffer>
void print_statistics(const Buffer& x)
{
std::cout << "Min value: " << *std::min_element(x.begin(), x.end()) << ", ";
std::cout << "Max value: " << *std::max_element(x.begin(), x.end()) << ", ";
double num_elements = x.size();
auto mean =
std::accumulate(x.begin(), x.end(), double{0.0}, std::plus<double>{}) / num_elements;
auto stddev = std::sqrt(
std::accumulate(x.begin(),
x.end(),
double{0.0},
[&](double r, double v) { return r + std::pow((v - mean), 2.0); }) /
num_elements);
std::cout << "Mean: " << mean << ", ";
std::cout << "StdDev: " << stddev << "\n";
}
template <class Buffer>
void print_preview(const Buffer& x)
{
if(x.size() <= 10)
{
std::for_each(x.begin(), x.end(), [&](double i) { std::cout << i << ", "; });
}
else
{
std::for_each(x.begin(), x.begin() + 5, [&](double i) { std::cout << i << ", "; });
std::cout << "..., ";
std::for_each(x.end() - 5, x.end(), [&](double i) { std::cout << i << ", "; });
}
std::cout << std::endl;
}
template <class T>
struct check_all
{
rtc::buffer<T> data{};
bool operator()(const rtc::buffer<T>& x)
{
if(data.empty())
{
data = x;
return true;
}
if(std::any_of(x.begin(), x.end(), [](double y) { return std::isnan(y); }))
return false;
return allclose(data, x);
}
};
template <class Solution>
auto report(const Solution& solution, bool pass)
{
return test::make_predicate(solution.ToTemplateString(), [=] { return pass; });
}
const std::string gemm_compile_check = R"__ck__(
#include <${include}>
extern "C" __global__ void f(const ck::half_t* a, const ck::half_t* b, ck::half_t* c) {
using G = ${template};
constexpr auto desc = ${template}::make_descriptor(ck::make_naive_tensor_descriptor_packed(ck::make_tuple(${m}, ${k})),
constexpr auto desc = G::make_descriptor(ck::make_naive_tensor_descriptor_packed(ck::make_tuple(${m}, ${k})),
ck::make_naive_tensor_descriptor(ck::make_tuple(${n}, ${k}), ck::make_tuple(1, ${n})),
ck::make_tuple(),
ck::make_naive_tensor_descriptor_packed(ck::make_tuple(${m}, ${n})));
@@ -166,15 +56,19 @@ TEST_CASE(test_problem_kernel)
std::string epilogue = "";
std::string prologue = "";
for(auto solution : prob.GetSolutions("gfx90a", prologue, epilogue))
auto solutions = prob.GetSolutions("gfx90a", prologue, epilogue);
std::cout << "Num solutions: " << solutions.size() << std::endl;
for(auto i = 0; i < solutions.size(); ++i)
{
auto src = ck::host::InterpolateString(gemm_compile_check,
std::cout << "Testing solution " << std::to_string(i + 1) << std::endl;
auto&& solution = solutions[i];
auto src = ck::host::InterpolateString(gemm_compile_check,
{{"include", prob.GetIncludeHeader()},
{"template", solution.ToTemplateString()},
{"m", std::to_string(prob.M)},
{"n", std::to_string(prob.N)},
{"k", std::to_string(prob.K)}});
auto srcs = get_headers_for_test();
auto srcs = get_headers_for_test();
srcs.push_back({"main.cpp", src});
rtc::compile_options options;
options.kernel_name = "f";

49
codegen/test/include/common.hpp
View File

@@ -2,27 +2,38 @@
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/host/headers.hpp"
#include <rtc/compile_kernel.hpp>
#include <rtc/hip.hpp>
#include <test.hpp>
#include <algorithm>
#include <cmath>
#include <iterator>
#include <numeric>
#include <random>
#include <test.hpp>
#include <rtc/compile_kernel.hpp>
#include <rtc/hip.hpp>
#include <fstream>
#include <unordered_set>
std::vector<rtc::src_file> get_headers_for_test()
inline std::vector<rtc::src_file> create_headers_for_test()
{
auto ck_headers = ck::host::GetHeaders();
std::vector<rtc::src_file> result;
auto hs = ck::host::GetHeaders();
std::transform(
hs.begin(), hs.end(), std::back_inserter(result), [&](const auto& p) -> rtc::src_file {
return {p.first, p.second};
});
std::transform(ck_headers.begin(), ck_headers.end(), std::back_inserter(result), [](auto& p) {
std::string content;
content.reserve(p.second.size() + 1);
content.push_back(' '); // We need a whitespace before the content for hipRTC to work
content.append(p.second.data(), p.second.size());
return rtc::src_file{p.first, std::move(content)};
});
return result;
}
inline const std::vector<rtc::src_file>& get_headers_for_test()
{
static const std::vector<rtc::src_file> headers = create_headers_for_test();
return headers;
}
template <typename V>
std::size_t GetSize(V mLens, V mStrides)
{
@@ -37,18 +48,24 @@ std::size_t GetSize(V mLens, V mStrides)
return space;
}
template <class T, typename V>
rtc::buffer<T> generate_buffer(V mLens, V mStrides, std::size_t seed = 0)
template <class T>
rtc::buffer<T> generate_buffer(std::size_t n, std::size_t seed = 0)
{
std::size_t space = GetSize(mLens, mStrides);
rtc::buffer<T> result(space);
rtc::buffer<T> result(n);
std::mt19937 gen(seed);
std::uniform_real_distribution<double> dis(-1.0);
std::generate(result.begin(), result.end(), [&] { return dis(gen); });
// std::fill(result.begin(), result.end(), 1);
return result;
}
template <class T, typename V>
std::enable_if_t<!std::is_integral_v<V>, rtc::buffer<T>>
generate_buffer(V mLens, V mStrides, std::size_t seed = 0)
{
std::size_t space = GetSize(mLens, mStrides);
return generate_buffer<T>(space, seed);
}
template <class T, class U>
bool allclose(const T& a, const U& b, double atol = 0.01, double rtol = 0.01)
{
@@ -57,7 +74,7 @@ bool allclose(const T& a, const U& b, double atol = 0.01, double rtol = 0.01)
});
}
std::string classify(double x)
inline std::string classify(double x)
{
switch(std::fpclassify(x))
{

6
codegen/test/rtc/CMakeLists.txt
View File

@@ -4,3 +4,9 @@ add_library(ck_rtc ${RTC_SOURCES})
target_include_directories(ck_rtc PUBLIC include)
target_link_libraries(ck_rtc PUBLIC hip::host)
target_link_libraries(ck_rtc PUBLIC -lstdc++fs)
option(USE_HIPRTC_FOR_CODEGEN_TESTS "Whether to enable hipRTC for codegen tests." ON)
if(USE_HIPRTC_FOR_CODEGEN_TESTS)
target_compile_definitions(ck_rtc PUBLIC HIPRTC_FOR_CODEGEN_TESTS)
message("CK compiled with USE_HIPRTC_FOR_CODEGEN_TESTS set to ${USE_HIPRTC_FOR_CODEGEN_TESTS}")
endif()

5
codegen/test/rtc/include/rtc/compile_kernel.hpp
View File

@@ -12,8 +12,9 @@ namespace rtc {
struct src_file
{
src_file(std::filesystem::path p, std::string c) : path{std::move(p)}, content{std::move(c)} {}
fs::path path;
std::string_view content;
std::string content;
};
struct compile_options
@@ -22,7 +23,7 @@ struct compile_options
std::string kernel_name = "main";
};
kernel compile_kernel(const std::vector<src_file>& src,
kernel compile_kernel(const std::vector<src_file>& srcs,
compile_options options = compile_options{});
} // namespace rtc

203
codegen/test/rtc/src/compile_kernel.cpp
View File

@@ -3,14 +3,41 @@
#include <rtc/hip.hpp>
#include <rtc/compile_kernel.hpp>
#ifdef HIPRTC_FOR_CODEGEN_TESTS
#include <hip/hiprtc.h>
#include <rtc/manage_ptr.hpp>
#endif
#include <rtc/tmp_dir.hpp>
#include <stdexcept>
#include <iostream>
#include <fstream>
#include <algorithm>
#include <cassert>
#include <deque>
#include <fstream>
#include <iostream>
#include <numeric>
#include <stdexcept>
namespace rtc {
bool EndsWith(const std::string& value, const std::string& suffix)
{
if(suffix.size() > value.size())
return false;
else
return std::equal(suffix.rbegin(), suffix.rend(), value.rbegin());
}
std::vector<std::string> SplitString(const std::string& s, char delim)
{
std::vector<std::string> elems;
std::stringstream ss(s + delim);
std::string item;
while(std::getline(ss, item, delim))
{
elems.push_back(item);
}
return elems;
}
template <class T>
T generic_read_file(const std::string& filename, size_t offset = 0, size_t nbytes = 0)
{
@@ -62,7 +89,7 @@ std::string compiler() { return "/opt/rocm/llvm/bin/clang++ -x hip --cuda-device
// TODO: undo after extracting the codeobj
// std::string compiler() { return "/opt/rocm/llvm/bin/clang++ -x hip"; }
kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options)
kernel clang_compile_kernel(const std::vector<src_file>& srcs, compile_options options)
{
assert(not srcs.empty());
tmp_dir td{"compile"};
@@ -103,4 +130,172 @@ kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options
return kernel{obj.data(), options.kernel_name};
}
#ifdef HIPRTC_FOR_CODEGEN_TESTS
std::string hiprtc_error(hiprtcResult err, const std::string& msg)
{
return "hiprtc: " + (hiprtcGetErrorString(err) + (": " + msg));
}
void hiprtc_check_error(hiprtcResult err, const std::string& msg = "")
{
if(err != HIPRTC_SUCCESS)
throw std::runtime_error(hiprtc_error(err, msg));
}
struct hiprtc_src_file
{
hiprtc_src_file() = default;
hiprtc_src_file(const src_file& s) : path(s.path.string()), content(s.content) {}
std::string path;
std::string content;
};
void hiprtc_program_destroy(hiprtcProgram prog) { hiprtcDestroyProgram(&prog); }
using hiprtc_program_ptr = RTC_MANAGE_PTR(hiprtcProgram, hiprtc_program_destroy);
template <class... Ts>
hiprtc_program_ptr hiprtc_program_create(Ts... xs)
{
hiprtcProgram prog = nullptr;
auto result = hiprtcCreateProgram(&prog, xs...);
hiprtc_program_ptr p{prog};
hiprtc_check_error(result, "Create program failed.");
return p;
}
struct hiprtc_program
{
struct string_array
{
std::deque<std::string> strings{};
std::vector<const char*> c_strs{};
string_array() {}
string_array(const string_array&) = delete;
std::size_t size() const { return strings.size(); }
const char** data() { return c_strs.data(); }
void push_back(std::string s)
{
strings.push_back(std::move(s));
c_strs.push_back(strings.back().c_str());
}
};
hiprtc_program_ptr prog = nullptr;
string_array headers{};
string_array include_names{};
std::string cpp_src = "";
std::string cpp_name = "";
hiprtc_program(const std::string& src, const std::string& name = "main.cpp")
: cpp_src(src), cpp_name(name)
{
create_program();
}
hiprtc_program(std::vector<src_file> srcs)
{
for(auto&& src : srcs)
{
if(EndsWith(src.path, ".cpp"))
{
cpp_src = std::move(src.content);
cpp_name = std::move(src.path);
}
else
{
headers.push_back(std::move(src.content));
include_names.push_back(std::move(src.path));
}
}
create_program();
}
void create_program()
{
assert(not cpp_src.empty());
assert(not cpp_name.empty());
assert(headers.size() == include_names.size());
prog = hiprtc_program_create(cpp_src.c_str(),
cpp_name.c_str(),
headers.size(),
headers.data(),
include_names.data());
}
void compile(const std::vector<std::string>& options, bool quiet = false) const
{
std::vector<const char*> c_options;
std::transform(options.begin(),
options.end(),
std::back_inserter(c_options),
[](const std::string& s) { return s.c_str(); });
auto result = hiprtcCompileProgram(prog.get(), c_options.size(), c_options.data());
auto prog_log = log();
if(not prog_log.empty() and not quiet)
{
std::cerr << prog_log << std::endl;
}
if(result != HIPRTC_SUCCESS)
throw std::runtime_error("Compilation failed.");
}
std::string log() const
{
std::size_t n = 0;
hiprtc_check_error(hiprtcGetProgramLogSize(prog.get(), &n));
if(n == 0)
return {};
std::string buffer(n, '\0');
hiprtc_check_error(hiprtcGetProgramLog(prog.get(), buffer.data()));
assert(buffer.back() != 0);
return buffer;
}
std::vector<char> get_code_obj() const
{
std::size_t n = 0;
hiprtc_check_error(hiprtcGetCodeSize(prog.get(), &n));
std::vector<char> buffer(n);
hiprtc_check_error(hiprtcGetCode(prog.get(), buffer.data()));
return buffer;
}
};
std::vector<std::vector<char>> compile_hip_src_with_hiprtc(const std::vector<src_file>& srcs,
const compile_options& options)
{
hiprtc_program prog(srcs);
auto flags = SplitString(options.flags, ' ');
prog.compile(flags);
return {prog.get_code_obj()};
}
static kernel hiprtc_compile_kernel(const std::vector<src_file>& srcs, compile_options options)
{
options.flags += " -I. -O3";
options.flags += " -std=c++17";
options.flags += " --offload-arch=" + get_device_name();
auto cos = compile_hip_src_with_hiprtc(srcs, options);
if(cos.size() != 1)
std::runtime_error("No code object");
auto& obj = cos.front();
return kernel{obj.data(), options.kernel_name};
}
#endif
kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options)
{
#ifdef HIPRTC_FOR_CODEGEN_TESTS
return hiprtc_compile_kernel(srcs, options);
#else
return clang_compile_kernel(srcs, options);
#endif
}
} // namespace rtc

3
include/ck/ck.hpp
View File

@@ -4,8 +4,9 @@
#pragma once
#include "ck/config.h"
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include "ck/utility/env.hpp"
#ifndef CK_CODE_GEN_RTC
#ifndef CK_DONT_USE_HIP_RUNTIME_HEADERS
#include "hip/hip_runtime.h"
#include "hip/hip_fp16.h"

2
include/ck/host_utility/device_prop.hpp
View File

@@ -3,6 +3,7 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <string>
#include <map>
#include <hip/hip_runtime.h>
@@ -97,3 +98,4 @@ inline bool is_gfx12_supported()
}
} // namespace ck
#endif

3
include/ck/host_utility/kernel_launch.hpp
View File

@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef __HIPCC_RTC__
#include <hip/hip_runtime.h>
#include "ck/ck.hpp"
@@ -166,3 +166,4 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
return 0;
#endif
}
#endif

7
include/ck/tensor_operation/gpu/device/device_base.hpp
View File

@@ -3,11 +3,12 @@
#pragma once
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include <string>
#include <sstream>
#include <regex>
#include <optional>
#include "ck/stream_config.hpp"
#endif
@@ -15,7 +16,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#define GET_OBJECT_NAME_IMLP \
std::optional<std::string> GetObjectName() const override \
{ \
@@ -77,7 +78,7 @@ struct BaseOperator
BaseOperator() = default;
BaseOperator(const BaseOperator&) = default;
BaseOperator& operator=(const BaseOperator&) = default;
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
virtual bool IsSupportedArgument(const BaseArgument*) { return false; }
virtual std::string GetTypeString() const { return ""; }

5
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
View File

@@ -2,9 +2,10 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <vector>
#endif
#include "device_base.hpp"
@@ -28,6 +29,7 @@ template <typename ALayout,
bool MaskOutUpperTriangle> // TODO: enum for mask type
struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
{
#ifndef __HIPCC_RTC__
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b0,
@@ -53,6 +55,7 @@ struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
CElementwiseOperation c_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
#endif
};
} // namespace device

8
include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
View File

@@ -2,9 +2,11 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef __HIPCC_RTC__
#include <array>
#endif
#include "ck/utility/array.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
@@ -34,6 +36,7 @@ struct DeviceGemmMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
#ifndef __HIPCC_RTC__
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
@@ -51,6 +54,7 @@ struct DeviceGemmMultipleD : public BaseOperator
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
#endif
};
// GEMM:
@@ -76,6 +80,7 @@ struct DeviceGemmMultipleDSplitK : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
#ifndef __HIPCC_RTC__
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
@@ -94,6 +99,7 @@ struct DeviceGemmMultipleDSplitK : public BaseOperator
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
#endif
};
// GEMM:

3
include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
View File

@@ -28,8 +28,7 @@ enum struct GemmSpecialization
NKOPadding,
MNKOPadding,
};
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
inline std::string getGemmSpecializationString(const GemmSpecialization& s)
{
switch(s)

10
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
View File

@@ -3,8 +3,12 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <sstream>
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#endif
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
@@ -15,8 +19,6 @@
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
@@ -429,6 +431,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
matrix_padder.PadN,
MaskOutUpperTriangle>;
#ifndef __HIPCC_RTC__
// Argument
struct Argument : public BaseArgument
{
@@ -603,6 +606,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
#endif
static constexpr bool IsValidCompilationParameter()
{
@@ -610,6 +614,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return true;
}
#ifndef __HIPCC_RTC__
static constexpr bool
IsSupported(index_t MRaw_, index_t NRaw_, index_t KRaw_, index_t Gemm1NRaw_)
{
@@ -837,6 +842,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return str.str();
}
#endif
template <class ADesc, class BDesc, class B1Desc, class CDesc>
struct Descriptor

27
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -3,8 +3,12 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <sstream>
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#endif
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
@@ -14,8 +18,6 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
@@ -224,9 +226,9 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
const std::array<index_t, NumDTensor>& NRaws,
const std::array<index_t, NumDTensor>& DsStride)
static auto MakeDsGridDescriptor_M_N(const Array<index_t, NumDTensor>& MRaws,
const Array<index_t, NumDTensor>& NRaws,
const Array<index_t, NumDTensor>& DsStride)
{
return generate_tuple(
[&](auto i) {
@@ -308,6 +310,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
using Block2ETileMap =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
#ifndef __HIPCC_RTC__
// Argument
struct Argument : public BaseArgument
{
@@ -497,6 +500,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
}
};
#endif
static constexpr bool IsSupported(index_t MRaw_, index_t NRaw_, index_t KRaw_)
{
// check vector load/store
@@ -577,6 +582,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
return true;
}
#ifndef __HIPCC_RTC__
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
@@ -675,11 +681,13 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
{
auto str = std::stringstream();
std::map<LoopScheduler, std::string> LoopSchedToString{
{LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
std::map<LoopScheduler, std::string> LoopSchedToString{{LoopScheduler::Default, "Default"},
{ LoopScheduler::Interwave,
"Interwave" }};
std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
{PipelineVersion::v2, "v2"}};
{ PipelineVersion::v2,
"v2" }};
// clang-format off
str << "DeviceGemmMultipleD_Xdl_CShuffle"
@@ -708,6 +716,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
return str.str();
}
#endif
template <class ADesc, class BDesc, class DsDesc, class EDesc>
struct Descriptor
@@ -846,7 +855,9 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
EDataType* __restrict__ p_e_grid)
{
__shared__ char p_shared_block[GridwiseGemm::GetSharedMemoryNumberOfByte()];
#ifndef __HIPCC_RTC__
assert(desc.IsValid());
#endif
if(desc.has_main_k_block_loop)
{
GridwiseGemm::template Run<true>(p_a_grid,

4
include/ck/tensor_operation/gpu/device/masking_specialization.hpp
View File

@@ -13,6 +13,7 @@ enum struct MaskingSpecialization
MaskOutUpperTriangle
};
#ifndef __HIPCC_RTC__
inline std::string getMaskingSpecializationString(const MaskingSpecialization& s)
{
switch(s)
@@ -22,6 +23,7 @@ inline std::string getMaskingSpecializationString(const MaskingSpecialization& s
default: return "Unrecognized specialization!";
}
}
#endif
struct MaskDisabledPredicate
{
@@ -53,7 +55,7 @@ struct MaskOutUpperTrianglePredicate
template <typename MaskOutPredicate>
struct C0MatrixMask_impl
{
__host__ __device__ C0MatrixMask_impl(index_t NRaw)
__host__ __device__ constexpr C0MatrixMask_impl(index_t NRaw)
: NRaw_(NRaw), predicate_(MaskOutPredicate{})
{
}

2
include/ck/tensor_operation/gpu/device/tensor_layout.hpp
View File

@@ -436,7 +436,7 @@ struct G_NDHW : public BaseTensorLayout
} // namespace convolution
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
template <
typename Layout,
typename std::enable_if<std::is_base_of<BaseTensorLayout, Layout>::value, bool>::type = false>

3
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View File

@@ -697,7 +697,7 @@ struct FastGelu
template <typename Y, typename X>
__device__ void operator()(Y& y, const X& x) const;
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
template <>
__host__ void operator()<float, float>(float& y, const float& x) const
{
@@ -709,7 +709,6 @@ struct FastGelu
y = x / (1.f + emu);
}
#endif
// device code, use lower precision "__ocml_exp_f32" and "rcp"
template <>
__device__ void operator()<float, float>(float& y, const float& x) const

2
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View File

@@ -8,7 +8,7 @@
#include "ck/utility/tuple.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include <limits>
#include <stdlib.h>
#endif

5
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -473,7 +473,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
#ifdef CK_CODE_GEN_RTC
#if defined(__HIPCC_RTC__) || defined(CK_CODE_GEN_RTC)
template <typename DsLayout, GemmSpecialization GemmSpec>
__host__ __device__ static auto
MakeDsGridDescriptor_M_N(const ck::Array<index_t, NumDTensor>& MRaws,
@@ -486,6 +486,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
const std::array<index_t, NumDTensor>& NRaws,
const std::array<index_t, NumDTensor>& DsStride)
#endif
{
return generate_tuple(
[&](auto i) {
@@ -949,7 +950,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
const index_t K,
const index_t StrideA,
const index_t StrideB,
#ifdef CK_CODE_GEN_RTC
#if defined(__HIPCC_RTC__) || defined(CK_CODE_GEN_RTC)
const ck::Array<index_t, NumDTensor> StrideDs,
#else
const std::array<index_t, NumDTensor> StrideDs,

7
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
View File

@@ -2,7 +2,8 @@
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include <iostream>
#include <ostream>
#endif
@@ -54,7 +55,7 @@ constexpr auto GridwiseGemmPipeline_Selector()
}
else
{
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
std::cerr << "GridwiseGemmPipeline configuration is not available" << std::endl;
#endif
}
@@ -62,7 +63,7 @@ constexpr auto GridwiseGemmPipeline_Selector()
} // namespace ck
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
inline std::ostream& operator<<(std::ostream& os, const ck::PipelineVersion& p)
{
switch(p)

2
include/ck/utility/amd_buffer_addressing.hpp
View File

@@ -1008,6 +1008,7 @@ llvm_amdgcn_raw_buffer_load_lds(int32x4_t rsrc,
index_t offset,
index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds");
#ifndef __HIPCC_RTC__
template <typename T, index_t NumElemsPerThread>
__device__ void amd_direct_load_global_to_lds(const T* global_base_ptr,
const index_t global_offset,
@@ -1059,5 +1060,6 @@ __device__ void amd_direct_load_global_to_lds(const T* global_base_ptr,
src_resource, lds_ptr, sizeof(uint32_t), global_offset_bytes, 0, 0, 0);
#endif
}
#endif
} // namespace ck

2
include/ck/utility/amd_wave_read_first_lane.hpp
View File

@@ -7,7 +7,7 @@
#include "ck/utility/functional2.hpp"
#include "ck/utility/math.hpp"
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include <array>
#include <cstddef>
#include <cstdint>

128
include/ck/utility/data_type.hpp
View File

@@ -6,16 +6,20 @@
#include "ck/utility/amd_ck_fp8.hpp"
#include "ck/utility/e8m0.hpp"
#include "ck/utility/statically_indexed_array.hpp"
#ifdef CK_CODE_GEN_RTC
/// Definitions from <cstdint>, <cmath> conflict with
/// /opt/rocm/include/hip/amd_detail/amd_hip_vector_types.h.
#if defined(__HIPCC_RTC__) || defined(CK_CODE_GEN_RTC)
using int8_t = signed char;
using uint8_t = unsigned char;
using int16_t = signed short;
using uint16_t = unsigned short;
using float_t = float;
#endif
namespace ck {
#endif // __HIPCC_RTC__
#ifdef CK_CODE_GEN_RTC
namespace ck {
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
using byte = unsigned char;
#else
using std::byte;
@@ -2612,7 +2616,7 @@ using pk_i4x2_t = typename vector_type<pk_i4_t, 2>::type;
using pk_i4x4_t = typename vector_type<pk_i4_t, 4>::type;
using pk_i4x8_t = typename vector_type<pk_i4_t, 8>::type;
#ifdef CK_CODE_GEN_RTC
#if defined(__HIPCC_RTC__) || defined(CK_CODE_GEN_RTC)
template <typename T>
struct NumericLimits;
@@ -2825,6 +2829,118 @@ struct NumericLimits<bf8_ocp_t>
return bit_cast<bf8_ocp_t>(binary_qnan);
}
};
template <>
struct NumericLimits<f4_t>
{
static constexpr uint8_t binary_min_normal = 0x2; // 0b0010
static constexpr uint8_t binary_max_normal = 0x7; // 0b0111
static constexpr uint8_t binary_lowest_normal = 0xF; // 0b1111
static constexpr uint8_t binary_min_subnorm = 0x1; // 0b0001
static constexpr uint8_t binary_max_subnorm = 0x1; // 0b0001
static constexpr float data_max_normal_number = 6;
static constexpr float data_min_subnormal_number = 0.5;
__host__ __device__ static constexpr f4_t Min() { return f4_t(binary_min_normal); }
__host__ __device__ static constexpr f4_t Max() { return f4_t(binary_max_normal); }
__host__ __device__ static constexpr f4_t Lowest() { return f4_t(binary_lowest_normal); }
__host__ __device__ static constexpr f4_t MinSubnorm() { return f4_t(binary_min_subnorm); }
__host__ __device__ static constexpr f4_t MaxSubnorm() { return f4_t(binary_max_subnorm); }
__host__ __device__ static constexpr float DataMaxNorm() { return data_max_normal_number; }
__host__ __device__ static constexpr float DataMinSubnorm()
{
return data_min_subnormal_number;
}
};
template <>
struct NumericLimits<f6_t>
{
static constexpr uint8_t binary_min_normal = 0x08; // 0b001000
static constexpr uint8_t binary_max_normal = 0x1F; // 0b011111
static constexpr uint8_t binary_lowest_normal = 0x3F; // 0b111111
static constexpr uint8_t binary_min_subnorm = 0x01; // 0b000001
static constexpr uint8_t binary_max_subnorm = 0x07; // 0b000111
static constexpr float data_max_normal_number = 7.5;
static constexpr float data_min_subnormal_number = 0.125;
__host__ __device__ static constexpr f6_t Min() { return f6_t(binary_min_normal & 0b111111); }
__host__ __device__ static constexpr f6_t Max() { return f6_t(binary_max_normal & 0b111111); }
__host__ __device__ static constexpr f6_t Lowest()
{
return f6_t(binary_lowest_normal & 0b111111);
}
__host__ __device__ static constexpr f6_t MinSubnorm()
{
return f6_t(binary_min_subnorm & 0b111111);
}
__host__ __device__ static constexpr f6_t MaxSubnorm()
{
return f6_t(binary_max_subnorm & 0b111111);
}
__host__ __device__ static constexpr float DataMaxNorm() { return data_max_normal_number; }
__host__ __device__ static constexpr float DataMinSubnorm()
{
return data_min_subnormal_number;
}
};
template <>
struct NumericLimits<bf6_t>
{
static constexpr uint8_t binary_min_normal = 0x08; // 0b001000
static constexpr uint8_t binary_max_normal = 0x1F; // 0b011111
static constexpr uint8_t binary_lowest_normal = 0x3F; // 0b111111
static constexpr uint8_t binary_min_subnorm = 0x01; // 0b000001
static constexpr uint8_t binary_max_subnorm = 0x03; // 0b000011
static constexpr float data_max_normal_number = 28;
static constexpr float data_min_subnormal_number = 0.0625;
__host__ __device__ static constexpr bf6_t Min() { return bf6_t(binary_min_normal); }
__host__ __device__ static constexpr bf6_t Max() { return bf6_t(binary_max_normal); }
__host__ __device__ static constexpr bf6_t Lowest() { return bf6_t(binary_lowest_normal); }
__host__ __device__ static constexpr bf6_t MinSubnorm() { return bf6_t(binary_min_subnorm); }
__host__ __device__ static constexpr bf6_t MaxSubnorm() { return bf6_t(binary_max_subnorm); }
__host__ __device__ static constexpr float DataMaxNorm() { return data_max_normal_number; }
__host__ __device__ static constexpr float DataMinSubnorm()
{
return data_min_subnormal_number;
}
};
template <>
struct NumericLimits<e8m0_bexp_t>
{
static constexpr e8m0_bexp_t binary_min = 0x00; // 0b00000000
static constexpr e8m0_bexp_t binary_max = 0xFE; // 0b11111110
static constexpr e8m0_bexp_t binary_qnan = 0xFF; // 0b11111111
static constexpr e8m0_bexp_t binary_1 = 0x7F; // 0b01111111
static constexpr e8m0_bexp_t binary_2 = 0x80; // 0b10000000
static constexpr e8m0_bexp_t binary_3 = 0x82; // 0b10000010
static constexpr e8m0_bexp_t binary_135 = 0x87; // 0b10000111
static constexpr e8m0_bexp_t binary_142 = 0x8E; // 0b10001110
__host__ __device__ static constexpr e8m0_bexp_t Min() { return e8m0_bexp_t(binary_min); }
__host__ __device__ static constexpr e8m0_bexp_t Max() { return e8m0_bexp_t(binary_max); }
__host__ __device__ static constexpr e8m0_bexp_t QuietNaN() { return e8m0_bexp_t(binary_qnan); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_1() { return e8m0_bexp_t(binary_1); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_2() { return e8m0_bexp_t(binary_2); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_3() { return e8m0_bexp_t(binary_3); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_135()
{
return e8m0_bexp_t(binary_135);
}
__host__ __device__ static constexpr e8m0_bexp_t Binary_142()
{
return e8m0_bexp_t(binary_142);
}
};
#else
template <typename T>
struct NumericLimits
@@ -2959,7 +3075,6 @@ struct NumericLimits<bf8_ocp_t>
return bit_cast<bf8_ocp_t>(binary_qnan);
}
};
#endif
template <>
struct NumericLimits<f4_t>
@@ -3072,6 +3187,7 @@ struct NumericLimits<e8m0_bexp_t>
return e8m0_bexp_t(binary_142);
}
};
#endif
template <typename T>
struct NumericUtils

18
include/ck/utility/enable_if.hpp
View File

@@ -4,15 +4,7 @@
#pragma once
namespace ck {
#ifndef CK_CODE_GEN_RTC
template <bool B, typename T = void>
using enable_if = std::enable_if<B, T>;
template <bool B, typename T = void>
using enable_if_t = typename std::enable_if<B, T>::type;
#else
#if defined(__HIPCC_RTC__) || defined(CK_CODE_GEN_RTC)
template <bool B, class T = void>
struct enable_if
{
@@ -26,6 +18,12 @@ struct enable_if<true, T>
template <bool B, class T = void>
using enable_if_t = typename enable_if<B, T>::type;
#endif
#else
template <bool B, typename T = void>
using enable_if = std::enable_if<B, T>;
template <bool B, typename T = void>
using enable_if_t = typename std::enable_if<B, T>::type;
#endif
} // namespace ck

8
include/ck/utility/loop_scheduler.hpp
View File

@@ -1,12 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_CODE_GEN_RTC
#pragma once
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include <ostream>
#endif
#pragma once
#include "ck/utility/common_header.hpp"
namespace ck {
@@ -28,7 +28,7 @@ constexpr LoopScheduler make_default_loop_scheduler()
} // namespace ck
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
{
switch(s)

3
include/ck/utility/magic_division.hpp
View File

@@ -4,6 +4,7 @@
#pragma once
#include "ck/ck.hpp"
#include "data_type.hpp"
#include "integral_constant.hpp"
#include "number.hpp"
#include "type.hpp"
@@ -34,7 +35,7 @@ struct MagicDivision
// WARNING: magic division is only applicable for division inside this range.
// You should use the return value of CalculateMagicNumbers, if division is not inside this
// range. The "else" logic below is to quiet down run-time error.
if(divisor >= 1 && divisor <= INT32_MAX)
if(divisor >= 1 && divisor <= ck::NumericLimits<int32_t>::Max())
{
uint32_t shift = 0;
for(shift = 0; shift < 32; ++shift)

20
include/ck/utility/math_v2.hpp
View File

@@ -19,7 +19,7 @@ extern "C" __device__ float __ocml_native_recip_f32(float);
#endif
// math functions for the host, some are implemented by calling C++ std functions
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
static inline __host__ float abs(float x) { return std::abs(x); };
static inline __host__ double abs(double x) { return std::abs(x); };
@@ -924,5 +924,23 @@ inline __device__ double expm1<double>(double x)
return expm1(x);
};
template <typename T>
inline __device__ T cos(T x)
{
return ck::type_convert<T>(cosf(ck::type_convert<float>(x)));
};
template <>
inline __device__ float cos<float>(float x)
{
return cosf(x);
};
template <>
inline __device__ double cos<double>(double x)
{
return cos(x);
};
} // namespace math
} // namespace ck

4
include/ck/utility/sequence.hpp
View File

@@ -3,7 +3,7 @@
#pragma once
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
#include <ostream>
#endif
@@ -902,7 +902,7 @@ using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
} // namespace ck
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
template <ck::index_t... Is>
std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
{

4
include/ck/utility/tuple_helper.hpp
View File

@@ -159,7 +159,7 @@ __host__ __device__ constexpr auto TupleReduce(F&& f, const Tuple<Ts...>& tuple)
}
}
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
template <typename T>
using is_tuple = decltype(ck::declval<T&>().IsTuple());
#endif
@@ -167,7 +167,7 @@ using is_tuple = decltype(ck::declval<T&>().IsTuple());
template <typename... Ts>
__host__ __device__ constexpr auto IsNestedTuple(const Tuple<Ts...>&)
{
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
return (is_detected<is_tuple, Ts>::value || ...);
#endif
}

629
include/ck/utility/type.hpp
View File

@@ -1,316 +1,313 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/utility/enable_if.hpp"
#include "ck/utility/integral_constant.hpp"
namespace ck {
#ifdef CK_CODE_GEN_RTC
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAIT1(name) \
template <class T> \
struct name : bool_constant<__##name(T)> \
{ \
}
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAIT2(name) \
template <class T, class U> \
struct name : bool_constant<__##name(T, U)> \
{ \
}
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAITN(name) \
template <class... Ts> \
struct name : bool_constant<__##name(Ts...)> \
{ \
}
CK_BUILTIN_TYPE_TRAIT1(is_class);
CK_BUILTIN_TYPE_TRAIT1(is_pointer);
CK_BUILTIN_TYPE_TRAIT1(is_reference);
CK_BUILTIN_TYPE_TRAIT1(is_trivially_copyable);
CK_BUILTIN_TYPE_TRAIT1(is_unsigned);
CK_BUILTIN_TYPE_TRAIT2(is_base_of);
template <class T>
struct remove_cv
{
using type = T;
};
template <class T>
struct remove_cv<const T> : remove_cv<T>
{
};
template <class T>
struct remove_cv<volatile T> : remove_cv<T>
{
};
template <class T>
struct remove_reference
{
typedef T type;
};
template <class T>
struct remove_reference<T&>
{
typedef T type;
};
template <class T>
struct remove_reference<T&&>
{
typedef T type;
};
template <class T>
struct remove_pointer
{
typedef T type;
};
template <class T>
struct remove_pointer<T*>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* const>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* volatile>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* const volatile>
{
typedef T type;
};
template <typename T>
constexpr T&& forward(typename remove_reference<T>::type& t_) noexcept
{
return static_cast<T&&>(t_);
}
template <typename T>
constexpr T&& forward(typename remove_reference<T>::type&& t_) noexcept
{
return static_cast<T&&>(t_);
}
template <class T>
struct is_const : public integral_constant<bool, false>
{
};
template <class T>
struct is_const<const T> : public integral_constant<bool, true>
{
};
template <class T>
inline constexpr bool is_const_v = is_const<T>::value;
template <typename T>
inline constexpr bool is_reference_v = is_reference<T>::value;
template <class T>
struct remove_const
{
typedef T type;
};
template <class T>
struct remove_const<const T>
{
typedef T type;
};
template <class T>
using remove_const_t = typename remove_const<T>::type;
template <class T>
inline constexpr bool is_class_v = is_class<T>::value;
template <class T>
inline constexpr bool is_trivially_copyable_v = is_trivially_copyable<T>::value;
// template <typename T>
// T&& declval() noexcept;
template <class T, class U = T&&>
U private_declval(int);
template <class T>
T private_declval(long);
template <class T>
auto declval() noexcept -> decltype(private_declval<T>(0));
template <class...>
using void_t = void;
#else
#include <utility>
#include <type_traits>
using std::declval;
using std::forward;
using std::is_base_of;
using std::is_class;
using std::is_class_v;
using std::is_const_v;
using std::is_pointer;
using std::is_reference;
using std::is_reference_v;
using std::is_trivially_copyable;
using std::is_trivially_copyable_v;
using std::is_unsigned;
using std::remove_const_t;
using std::remove_cv;
using std::remove_pointer;
using std::remove_reference;
using std::void_t;
#endif
template <typename X, typename Y>
struct is_same : public integral_constant<bool, false>
{
};
template <typename X>
struct is_same<X, X> : public integral_constant<bool, true>
{
};
template <typename X>
struct is_floating_point : public integral_constant<bool, false>
{
};
template <>
struct is_floating_point<float> : public integral_constant<bool, true>
{
};
template <>
struct is_floating_point<double> : public integral_constant<bool, true>
{
};
template <>
struct is_floating_point<long double> : public integral_constant<bool, true>
{
};
template <typename X>
struct is_integral : public integral_constant<bool, false>
{
};
template <>
struct is_integral<int> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned int> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<short> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned short> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<long long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned long long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<char> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<signed char> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned char> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<wchar_t> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<char16_t> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<char32_t> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<bool> : public integral_constant<bool, true>
{
};
template <typename X, typename Y>
inline constexpr bool is_same_v = is_same<X, Y>::value;
template <typename X, typename Y>
inline constexpr bool is_base_of_v = is_base_of<X, Y>::value;
template <typename T>
inline constexpr bool is_unsigned_v = is_unsigned<T>::value;
template <typename T>
using remove_reference_t = typename remove_reference<T>::type;
template <typename T>
using remove_reference_t = typename remove_reference<T>::type;
template <typename T>
using remove_cv_t = typename remove_cv<T>::type;
template <typename T>
using remove_cvref_t = remove_cv_t<remove_reference_t<T>>;
template <typename T>
using remove_pointer_t = typename remove_pointer<T>::type;
template <typename T>
inline constexpr bool is_pointer_v = is_pointer<T>::value;
template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
__host__ __device__ constexpr Y bit_cast(const X& x)
{
static_assert(__has_builtin(__builtin_bit_cast), "");
static_assert(sizeof(X) == sizeof(Y), "Do not support cast between different size of type");
return __builtin_bit_cast(Y, x);
}
} // namespace ck
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/utility/enable_if.hpp"
#include "ck/utility/integral_constant.hpp"
namespace ck {
#if defined(__HIPCC_RTC__) || defined(CK_CODE_GEN_RTC)
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAIT1(name) \
template <class T> \
struct name : bool_constant<__##name(T)> \
{ \
}
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAIT2(name) \
template <class T, class U> \
struct name : bool_constant<__##name(T, U)> \
{ \
}
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAITN(name) \
template <class... Ts> \
struct name : bool_constant<__##name(Ts...)> \
{ \
}
CK_BUILTIN_TYPE_TRAIT1(is_class);
CK_BUILTIN_TYPE_TRAIT1(is_pointer);
CK_BUILTIN_TYPE_TRAIT1(is_reference);
CK_BUILTIN_TYPE_TRAIT1(is_trivially_copyable);
CK_BUILTIN_TYPE_TRAIT1(is_unsigned);
CK_BUILTIN_TYPE_TRAIT2(is_base_of);
template <class T>
struct remove_cv
{
using type = T;
};
template <class T>
struct remove_cv<const T> : remove_cv<T>
{
};
template <class T>
struct remove_cv<volatile T> : remove_cv<T>
{
};
template <class T>
struct remove_reference
{
typedef T type;
};
template <class T>
struct remove_reference<T&>
{
typedef T type;
};
template <class T>
struct remove_reference<T&&>
{
typedef T type;
};
template <class T>
struct remove_pointer
{
typedef T type;
};
template <class T>
struct remove_pointer<T*>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* const>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* volatile>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* const volatile>
{
typedef T type;
};
template <typename T>
constexpr T&& forward(typename remove_reference<T>::type& t_) noexcept
{
return static_cast<T&&>(t_);
}
template <typename T>
constexpr T&& forward(typename remove_reference<T>::type&& t_) noexcept
{
return static_cast<T&&>(t_);
}
template <class T>
struct is_const : public integral_constant<bool, false>
{
};
template <class T>
struct is_const<const T> : public integral_constant<bool, true>
{
};
template <class T>
inline constexpr bool is_const_v = is_const<T>::value;
template <typename T>
inline constexpr bool is_reference_v = is_reference<T>::value;
template <class T>
struct remove_const
{
typedef T type;
};
template <class T>
struct remove_const<const T>
{
typedef T type;
};
template <class T>
using remove_const_t = typename remove_const<T>::type;
template <class T>
inline constexpr bool is_class_v = is_class<T>::value;
template <class T>
inline constexpr bool is_trivially_copyable_v = is_trivially_copyable<T>::value;
// template <typename T>
// T&& declval() noexcept;
template <class T, class U = T&&>
U private_declval(int);
template <class T>
T private_declval(long);
template <class T>
auto declval() noexcept -> decltype(private_declval<T>(0));
template <class...>
using void_t = void;
#else
#include <utility>
#include <type_traits>
using std::declval;
using std::forward;
using std::is_base_of;
using std::is_class;
using std::is_class_v;
using std::is_const_v;
using std::is_pointer;
using std::is_reference;
using std::is_reference_v;
using std::is_trivially_copyable;
using std::is_trivially_copyable_v;
using std::is_unsigned;
using std::remove_const_t;
using std::remove_cv;
using std::remove_pointer;
using std::remove_reference;
using std::void_t;
#endif
template <typename X, typename Y>
struct is_same : public integral_constant<bool, false>
{
};
template <typename X>
struct is_same<X, X> : public integral_constant<bool, true>
{
};
template <typename X>
struct is_floating_point : public integral_constant<bool, false>
{
};
template <>
struct is_floating_point<float> : public integral_constant<bool, true>
{
};
template <>
struct is_floating_point<double> : public integral_constant<bool, true>
{
};
template <>
struct is_floating_point<long double> : public integral_constant<bool, true>
{
};
template <typename X>
struct is_integral : public integral_constant<bool, false>
{
};
template <>
struct is_integral<int> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned int> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<short> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned short> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<long long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned long long> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<char> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<signed char> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<unsigned char> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<wchar_t> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<char16_t> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<char32_t> : public integral_constant<bool, true>
{
};
template <>
struct is_integral<bool> : public integral_constant<bool, true>
{
};
template <typename X, typename Y>
inline constexpr bool is_same_v = is_same<X, Y>::value;
template <typename X, typename Y>
inline constexpr bool is_base_of_v = is_base_of<X, Y>::value;
template <typename T>
inline constexpr bool is_unsigned_v = is_unsigned<T>::value;
template <typename T>
using remove_reference_t = typename remove_reference<T>::type;
template <typename T>
using remove_cv_t = typename remove_cv<T>::type;
template <typename T>
using remove_cvref_t = remove_cv_t<remove_reference_t<T>>;
template <typename T>
using remove_pointer_t = typename remove_pointer<T>::type;
template <typename T>
inline constexpr bool is_pointer_v = is_pointer<T>::value;
template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
__host__ __device__ constexpr Y bit_cast(const X& x)
{
static_assert(__has_builtin(__builtin_bit_cast), "");
static_assert(sizeof(X) == sizeof(Y), "Do not support cast between different size of type");
return __builtin_bit_cast(Y, x);
}
} // namespace ck

4
include/ck/utility/type_convert.hpp
View File

@@ -279,7 +279,6 @@ inline __host__ __device__ f8_fnuz_t f8_convert_sr<f8_fnuz_t, half_t>(half_t x)
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 1254739;
#ifndef CK_CODE_GEN_RTC
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
#else
@@ -344,7 +343,6 @@ inline __host__ __device__ bf8_fnuz_t f8_convert_sr<bf8_fnuz_t, half_t>(half_t x
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 1254739;
#ifndef CK_CODE_GEN_RTC
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
#else
@@ -1981,7 +1979,7 @@ inline __host__ __device__ float32_t type_convert<float32_t, bf6x32_t>(bf6x32_t
#endif
}
#ifndef CK_CODE_GEN_RTC
#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
template <typename Y, typename X, size_t NumElems>
inline __host__ __device__ void array_convert(std::array<Y, NumElems>& y,
const std::array<X, NumElems>& x)