ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 12:41:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

introducing ck_tile! (#1216)

Browse Source 
* enable gfx940

* switch between intrinsic mfma routines on mi100/200 and mi300

* fix mfma_int8 on MI300

* disable 2 int8 examples on MI300

* Update cmake-ck-dev.sh

* restore gitignore file

* modify Jenkinsfile to the internal repo

* Bump rocm-docs-core from 0.24.0 to 0.29.0 in /docs/sphinx

Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.24.0 to 0.29.0.
- [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases)
- [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/RadeonOpenCompute/rocm-docs-core/compare/v0.24.0...v0.29.0)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

* initial enablement of gfx950

* fix clang format

* disable examples 31 and 41 int8 on gfx950

* add code

* fix build wip

* fix xx

* now can build

* naming

* minor fix

* wip fix

* fix macro for exp2; fix warpgemm a/b in transposedC

* unify as tuple_array

* Update the required Python version to 3.9

* Update executable name in test scripts

* re-structure tuple/array to avoid spill

* Merge function templates

* Fix format

* Add constraint to array<> ctor

* Re-use function

* Some minor changes

* remove wrong code in store_raw()

* fix compile issue in transpose

* Rename enum
Rename 'cood_transform_enum' to 'coord_transform_enum'

* let more integral_constant->constant, and formating

* make sure thread_buffer can be tuple/array

* temp fix buffer_store spill

* not using custom data type by default, now we can have ISA-level same code as opt_padding

* fix compile error, fp8 not ready now

* fix fp8 duplicated move/shift/and/or problem

* Default use CK_TILE_FLOAT_TO_FP8_STOCHASTIC rounding mode

* fix scratch in fp8 kernel

* update some readme

* fix merge from upstream

* sync with upstream

* sync upstream again

* sync 22

* remove unused

* fix clang-format

* update README of ck_tile example

* fix several issue

* let python version to be 3.8 as minimal

* remove ck_tile example from default cmake target like all/install/check

* remove mistake

* 1).support receipe in generate.py 2).use simplified mask type 3).change left/right to pass into karg

* fix some bug in group-mode masking and codegen. update README

* F8 quantization for FMHA forward (#1224)

* Add SAccElementFunction, PComputeElementFunction, OAccElementFunction in pipeline

* Add element function to fmha api

* Adjust P elementwise function

* Fix bug of elementwise op, our elementwise op is not inout

* Add some elementwise op, prepare to quantization

* Let generate.py can generate different elementwise function

* To prevent compiler issue, remove the elementwise function we have not used.

* Remove f8 pipeline, we should share the same pipeline even in f8

* Remove remove_cvref_t

* Avoid warning

* Fix wrong fp8 QK/KV block gemm setting

* Check fp8 rounding error in check_err()

* Set fp8 rounding error for check_err()

* Use CK_TILE_FLOAT_TO_FP8_STANDARD as default fp8 rounding mode

* 1. codgen the f8 api and kernel
2. f8 host code

* prevent warning in filter mode

* Remove not-in-use elementwise function kargs

* Remove more not-in-use elementwise function kargs

* Small refinements in C++ source files

* Use conditional_t<> to simplify code

* Support heterogeneous argument for binary function types

* Re-use already-existing scales<> functor template

* Fix wrong value produced by saturating

* Generalize the composes<> template

* Unify saturates<> implementation

* Fix type errors in composes<>

* Extend less_equal<>

* Reuse the existing template less_equal<> in check_err()

* Add equal<float> & equal<double>

* Rename check_err() parameter

* Rename check_err() parameter

* Add FIXME comment for adding new macro in future

* Remove unnecessary cast to void

* Eliminate duplicated code

* Avoid dividing api pool into more than 2 groups

* Use more clear variable names

* Use affirmative condition in if stmt

* Remove blank lines

* Donot perfect forwarding in composes<>

* To fix compile error, revert generate.py back to 4439cc107d

* Fix bug of p element function

* Add compute element op to host softmax

* Remove element function in api interface

* Extract user parameter

* Rename pscale and oscale variable

* rename f8 to fp8

* rename more f8 to fp8

* Add pipeline::operator() without element_functor

* 1. Remove deprecated pipeline enum
2. Refine host code parameter

* Use quantization range as input

* 1. Rename max_dtype to dtype_max.
2. Rename scale to scale_s
3.Add init description

* Refine description

* prevent early return

* unify _squant kernel name in cpp, update README

* Adjust the default range.

* Refine error message and bias range

* Add fp8 benchmark and smoke test

* fix fp8 swizzle_factor=4 case

---------

Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: Jing Zhang <jizha@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com>
Co-authored-by: rocking <ChunYu.Lai@amd.com>
This commit is contained in:
carlushuang
2024-04-16 08:27:12 +08:00
committed by GitHub
parent dd34ab6e64
commit db376dd8a4
141 changed files with 30623 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

184
include/ck_tile/host/arg_parser.hpp Normal file
View File

@@ -0,0 +1,184 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include <iomanip>
#include <iostream>
#include <stdio.h>
#include <stdlib.h>
#include <string>
#include <unordered_map>
#include <vector>
namespace ck_tile {
/*
* a host side utility, arg parser for
* -[key0]=[value0] -[key1]=[value1] ...
*/
class ArgParser
{
public:
class Arg
{
public:
std::string name;
std::string value;
std::string help_text;
};
ArgParser() {}
ArgParser& insert(const std::string& _name,
const std::string& _default_value,
const std::string& _help_text)
{
Arg in;
in.name = _name;
in.value = _default_value;
in.help_text = _help_text;
if(input_map.count(_name) != 0)
{
printf("arg:%s already exist\n", _name.c_str());
}
else
{
input_map[_name] = in;
keys.push_back(_name);
}
return *this;
}
void print()
{
printf("args:\n");
for(auto& key : keys)
{
auto value = input_map[key];
std::vector<std::string> help_text_lines;
size_t pos = 0;
for(size_t next_pos = value.help_text.find('\n', pos); next_pos != std::string::npos;)
{
help_text_lines.push_back(std::string(value.help_text.begin() + pos,
value.help_text.begin() + next_pos++));
pos = next_pos;
next_pos = value.help_text.find('\n', pos);
}
help_text_lines.push_back(
std::string(value.help_text.begin() + pos, value.help_text.end()));
std::string default_value = std::string("(default:") + value.value + std::string(")");
std::cout << std::setw(2) << std::setw(12 - value.name.length()) << "-" << key
<< std::setw(4) << " " << help_text_lines[0] << " " << default_value
<< std::endl;
for(auto help_next_line = std::next(help_text_lines.begin());
help_next_line != help_text_lines.end();
++help_next_line)
{
std::cout << std::setw(17) << " " << *help_next_line << std::endl;
}
}
}
bool parse(int argc, char* argv[], int start_index = 1)
{
if(argc < start_index)
{
printf("not enough args\n");
return false;
}
for(int i = start_index; i < argc; i++)
{
char* cur_arg = argv[i];
if(cur_arg[0] != '-')
{
printf("illegal input\n");
print();
return false;
}
else
{
std::string text(cur_arg + 1);
if(text == "?")
{
print();
return false;
}
auto pos = text.find('=');
if(pos == std::string::npos)
{
printf("arg should be [key]=[value] pair, here:%s\n", text.c_str());
return false;
}
if(pos >= (text.size() - 1))
{
printf("cant find value after \"=\", here:%s\n", text.c_str());
return false;
}
auto key = text.substr(0, pos);
auto value = text.substr(pos + 1);
if(input_map.count(key) == 0)
{
printf("no such arg:%s\n", key.c_str());
return false;
}
input_map[key].value = value;
}
}
return true;
}
std::string get_str(const std::string& name) const
{
std::string value = input_map.at(name).value;
return value;
}
int get_int(const std::string& name) const
{
int value = atoi(input_map.at(name).value.c_str());
return value;
}
uint32_t get_uint32(const std::string& name) const
{
uint32_t value = strtoul(input_map.at(name).value.c_str(), nullptr, 10);
return value;
}
uint64_t get_uint64(const std::string& name) const
{
uint64_t value = strtoull(input_map.at(name).value.c_str(), nullptr, 10);
return value;
}
bool get_bool(const std::string& name) const
{
auto v = input_map.at(name).value;
if(v.compare("t") == 0 || v.compare("true") == 0)
return true;
if(v.compare("f") == 0 || v.compare("false") == 0)
return false;
int value = atoi(v.c_str());
return value == 0 ? false : true;
}
float get_float(const std::string& name) const
{
double value = atof(input_map.at(name).value.c_str());
return static_cast<float>(value);
}
double get_double(const std::string& name) const
{
double value = atof(input_map.at(name).value.c_str());
return value;
}
private:
std::unordered_map<std::string, Arg> input_map;
std::vector<std::string> keys;
};
} // namespace ck_tile

394
include/ck_tile/host/check_err.hpp Normal file
View File

@@ -0,0 +1,394 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <algorithm>
#include <cmath>
#include <cstdlib>
#include <iostream>
#include <iomanip>
#include <iterator>
#include <limits>
#include <type_traits>
#include <vector>
#include "ck_tile/core.hpp"
#include "ck_tile/host/ranges.hpp"
namespace ck_tile {
template <typename T>
std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
{
using size_type = typename std::vector<T>::size_type;
os << "[";
for(size_type idx = 0; idx < v.size(); ++idx)
{
if(0 < idx)
{
os << ", ";
}
os << v[idx];
}
return os << "]";
}
template <typename Range, typename RefRange>
typename std::enable_if<
std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_floating_point_v<ranges::range_value_t<Range>> &&
!std::is_same_v<ranges::range_value_t<Range>, half_t>,
bool>::type CK_TILE_HOST
check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-5,
double atol = 3e-6,
bool allow_infinity_ref = false)
{
if(out.size() != ref.size())
{
std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
<< std::endl;
return false;
}
const auto is_infinity_error = [=](auto o, auto r) {
const bool either_not_finite = !std::isfinite(o) || !std::isfinite(r);
const bool both_infinite_and_same = std::isinf(o) && std::isinf(r) && (o == r);
return either_not_finite && !(allow_infinity_ref && both_infinite_and_same);
};
bool res{true};
int err_count = 0;
double err = 0;
double max_err = std::numeric_limits<double>::min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const double o = *std::next(std::begin(out), i);
const double r = *std::next(std::begin(ref), i);
err = std::abs(o - r);
if(err > atol + rtol * std::abs(r) || is_infinity_error(o, r))
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
}
res = false;
}
}
if(!res)
{
const float error_percent =
static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
std::cerr << "max err: " << max_err;
std::cerr << ", number of errors: " << err_count;
std::cerr << ", " << error_percent << "% wrong values" << std::endl;
}
return res;
}
template <typename Range, typename RefRange>
typename std::enable_if<
std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_same_v<ranges::range_value_t<Range>, bf16_t>,
bool>::type CK_TILE_HOST
check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-3,
double atol = 1e-3,
bool allow_infinity_ref = false)
{
if(out.size() != ref.size())
{
std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
<< std::endl;
return false;
}
const auto is_infinity_error = [=](auto o, auto r) {
const bool either_not_finite = !std::isfinite(o) || !std::isfinite(r);
const bool both_infinite_and_same = std::isinf(o) && std::isinf(r) && (o == r);
return either_not_finite && !(allow_infinity_ref && both_infinite_and_same);
};
bool res{true};
int err_count = 0;
double err = 0;
// TODO: This is a hack. We should have proper specialization for bf16_t data type.
double max_err = std::numeric_limits<float>::min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const double o = type_convert<float>(*std::next(std::begin(out), i));
const double r = type_convert<float>(*std::next(std::begin(ref), i));
err = std::abs(o - r);
if(err > atol + rtol * std::abs(r) || is_infinity_error(o, r))
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
}
res = false;
}
}
if(!res)
{
const float error_percent =
static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
std::cerr << "max err: " << max_err;
std::cerr << ", number of errors: " << err_count;
std::cerr << ", " << error_percent << "% wrong values" << std::endl;
}
return res;
}
template <typename Range, typename RefRange>
typename std::enable_if<
std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_same_v<ranges::range_value_t<Range>, half_t>,
bool>::type CK_TILE_HOST
check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-3,
double atol = 1e-3,
bool allow_infinity_ref = false)
{
if(out.size() != ref.size())
{
std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
<< std::endl;
return false;
}
const auto is_infinity_error = [=](auto o, auto r) {
const bool either_not_finite = !std::isfinite(o) || !std::isfinite(r);
const bool both_infinite_and_same = std::isinf(o) && std::isinf(r) && (o == r);
return either_not_finite && !(allow_infinity_ref && both_infinite_and_same);
};
bool res{true};
int err_count = 0;
double err = 0;
double max_err = static_cast<double>(std::numeric_limits<ranges::range_value_t<Range>>::min());
for(std::size_t i = 0; i < ref.size(); ++i)
{
const double o = type_convert<float>(*std::next(std::begin(out), i));
const double r = type_convert<float>(*std::next(std::begin(ref), i));
err = std::abs(o - r);
if(err > atol + rtol * std::abs(r) || is_infinity_error(o, r))
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
}
res = false;
}
}
if(!res)
{
const float error_percent =
static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
std::cerr << "max err: " << max_err;
std::cerr << ", number of errors: " << err_count;
std::cerr << ", " << error_percent << "% wrong values" << std::endl;
}
return res;
}
template <typename Range, typename RefRange>
std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_integral_v<ranges::range_value_t<Range>> &&
!std::is_same_v<ranges::range_value_t<Range>, bf16_t>)
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| std::is_same_v<ranges::range_value_t<Range>, int4_t>
#endif
,
bool>
CK_TILE_HOST check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
double = 0,
double atol = 0)
{
if(out.size() != ref.size())
{
std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
<< std::endl;
return false;
}
bool res{true};
int err_count = 0;
int64_t err = 0;
int64_t max_err = std::numeric_limits<int64_t>::min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const int64_t o = *std::next(std::begin(out), i);
const int64_t r = *std::next(std::begin(ref), i);
err = std::abs(o - r);
if(err > atol)
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
{
std::cerr << msg << " out[" << i << "] != ref[" << i << "]: " << o << " != " << r
<< std::endl;
}
res = false;
}
}
if(!res)
{
const float error_percent =
static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
std::cerr << "max err: " << max_err;
std::cerr << ", number of errors: " << err_count;
std::cerr << ", " << error_percent << "% wrong values" << std::endl;
}
return res;
}
template <typename Range, typename RefRange>
std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_same_v<ranges::range_value_t<Range>, fp8_t>),
bool>
CK_TILE_HOST check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
unsigned max_rounding_point_distance = 1,
double atol = 1e-1,
bool allow_infinity_ref = false)
{
if(out.size() != ref.size())
{
std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
<< std::endl;
return false;
}
const auto is_infinity_error = [=](auto o, auto r) {
const bool either_not_finite = !std::isfinite(o) || !std::isfinite(r);
const bool both_infinite_and_same = std::isinf(o) && std::isinf(r) && (o == r);
return either_not_finite && !(allow_infinity_ref && both_infinite_and_same);
};
static const auto get_rounding_point_distance = [](fp8_t o, fp8_t r) -> unsigned {
static const auto get_sign_bit = [](fp8_t v) -> bool {
return 0x80 & bit_cast<uint8_t>(v);
};
if(get_sign_bit(o) ^ get_sign_bit(r))
{
return std::numeric_limits<unsigned>::max();
}
else
{
return std::abs(bit_cast<int8_t>(o) - bit_cast<int8_t>(r));
}
};
bool res{true};
int err_count = 0;
double err = 0;
double max_err = std::numeric_limits<float>::min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const fp8_t o_fp8 = *std::next(std::begin(out), i);
const fp8_t r_fp8 = *std::next(std::begin(ref), i);
const double o_fp64 = type_convert<float>(o_fp8);
const double r_fp64 = type_convert<float>(r_fp8);
err = std::abs(o_fp64 - r_fp64);
if(!(less_equal<double>{}(err, atol) ||
get_rounding_point_distance(o_fp8, r_fp8) <= max_rounding_point_distance) ||
is_infinity_error(o_fp64, r_fp64))
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o_fp64 << " != " << r_fp64 << std::endl;
}
res = false;
}
}
if(!res)
{
std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
}
return res;
}
template <typename Range, typename RefRange>
std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_same_v<ranges::range_value_t<Range>, bf8_t>),
bool>
CK_TILE_HOST check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-3,
double atol = 1e-3,
bool allow_infinity_ref = false)
{
if(out.size() != ref.size())
{
std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
<< std::endl;
return false;
}
const auto is_infinity_error = [=](auto o, auto r) {
const bool either_not_finite = !std::isfinite(o) || !std::isfinite(r);
const bool both_infinite_and_same = std::isinf(o) && std::isinf(r) && (o == r);
return either_not_finite && !(allow_infinity_ref && both_infinite_and_same);
};
bool res{true};
int err_count = 0;
double err = 0;
double max_err = std::numeric_limits<float>::min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const double o = type_convert<float>(*std::next(std::begin(out), i));
const double r = type_convert<float>(*std::next(std::begin(ref), i));
err = std::abs(o - r);
if(err > atol + rtol * std::abs(r) || is_infinity_error(o, r))
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
}
res = false;
}
}
if(!res)
{
std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
}
return res;
}
} // namespace ck_tile

112
include/ck_tile/host/device_memory.hpp Normal file
View File

@@ -0,0 +1,112 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <hip/hip_runtime.h>
#include <stdint.h>
#include <stdexcept>
#include "ck_tile/host/hip_check_error.hpp"
namespace ck_tile {
template <typename T>
__global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size)
{
for(uint64_t i = threadIdx.x; i < buffer_element_size; i += blockDim.x)
{
p[i] = x;
}
}
/**
* @brief Container for storing data in GPU device memory
*
*/
struct DeviceMem
{
DeviceMem() : mpDeviceBuf(nullptr), mMemSize(0) {}
DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
{
HIP_CHECK_ERROR(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void Realloc(std::size_t mem_size)
{
if(mpDeviceBuf)
{
HIP_CHECK_ERROR(hipFree(mpDeviceBuf));
}
mMemSize = mem_size;
HIP_CHECK_ERROR(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void* GetDeviceBuffer() const { return mpDeviceBuf; }
std::size_t GetBufferSize() const { return mMemSize; }
void ToDevice(const void* p) const
{
if(mpDeviceBuf)
{
HIP_CHECK_ERROR(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
}
else
{
throw std::runtime_error("ToDevice with an empty pointer");
}
}
void ToDevice(const void* p, const std::size_t cpySize) const
{
HIP_CHECK_ERROR(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), cpySize, hipMemcpyHostToDevice));
}
void FromDevice(void* p) const
{
if(mpDeviceBuf)
{
HIP_CHECK_ERROR(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
}
else
{
throw std::runtime_error("FromDevice with an empty pointer");
}
}
void FromDevice(void* p, const std::size_t cpySize) const
{
HIP_CHECK_ERROR(hipMemcpy(p, mpDeviceBuf, cpySize, hipMemcpyDeviceToHost));
}
void SetZero() const
{
if(mpDeviceBuf)
{
HIP_CHECK_ERROR(hipMemset(mpDeviceBuf, 0, mMemSize));
}
}
template <typename T>
void SetValue(T x) const
{
if(mMemSize % sizeof(T) != 0)
{
throw std::runtime_error("wrong! not entire DeviceMem will be set");
}
// TODO: call a gpu kernel to set the value (?)
set_buffer_value<T><<<1, 1024>>>(static_cast<T*>(mpDeviceBuf), x, mMemSize / sizeof(T));
}
~DeviceMem()
{
if(mpDeviceBuf)
{
try
{
HIP_CHECK_ERROR(hipFree(mpDeviceBuf));
}
catch(std::runtime_error& re)
{
std::cerr << re.what() << std::endl;
}
}
}
void* mpDeviceBuf;
std::size_t mMemSize;
};
} // namespace ck_tile

232
include/ck_tile/host/fill.hpp Normal file
View File

@@ -0,0 +1,232 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <algorithm>
#include <cmath>
#include <iterator>
#include <optional>
#include <random>
#include <type_traits>
#include <utility>
#include "ck_tile/core.hpp"
namespace ck_tile {
template <typename T>
struct FillUniformDistribution
{
float a_{-5.f};
float b_{5.f};
std::optional<uint32_t> seed_{11939};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}());
std::uniform_real_distribution<float> dis(a_, b_);
std::generate(first, last, [&dis, &gen]() { return ck_tile::type_convert<T>(dis(gen)); });
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const
-> std::void_t<decltype(std::declval<const FillUniformDistribution&>()(
std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
template <typename T>
struct FillNormalDistribution
{
float mean_{0.f};
float variance_{1.f};
std::optional<uint32_t> seed_{11939};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}());
std::normal_distribution<float> dis(mean_, std::sqrt(variance_));
std::generate(first, last, [&dis, &gen]() { return ck_tile::type_convert<T>(dis(gen)); });
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const
-> std::void_t<decltype(std::declval<const FillNormalDistribution&>()(
std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
// Normally FillUniformDistributionIntegerValue should use std::uniform_int_distribution as below.
// However this produces segfaults in std::mt19937 which look like inifite loop.
// template <typename T>
// struct FillUniformDistributionIntegerValue
// {
// int a_{-5};
// int b_{5};
//
// template <typename ForwardIter>
// void operator()(ForwardIter first, ForwardIter last) const
// {
// std::mt19937 gen(11939);
// std::uniform_int_distribution<int> dis(a_, b_);
// std::generate(
// first, last, [&dis, &gen]() { return ck_tile::type_convert<T>(dis(gen)); });
// }
// };
// Workaround for uniform_int_distribution not working as expected. See note above.<
template <typename T>
struct FillUniformDistributionIntegerValue
{
float a_{-5.f};
float b_{5.f};
std::optional<uint32_t> seed_{11939};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}());
std::uniform_real_distribution<float> dis(a_, b_);
std::generate(
first, last, [&dis, &gen]() { return ck_tile::type_convert<T>(std::round(dis(gen))); });
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const
-> std::void_t<decltype(std::declval<const FillUniformDistributionIntegerValue&>()(
std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
template <typename T>
struct FillNormalDistributionIntegerValue
{
float mean_{0.f};
float variance_{1.f};
std::optional<uint32_t> seed_{11939};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}());
std::normal_distribution<float> dis(mean_, std::sqrt(variance_));
std::generate(
first, last, [&dis, &gen]() { return ck_tile::type_convert<T>(std::round(dis(gen))); });
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const
-> std::void_t<decltype(std::declval<const FillNormalDistributionIntegerValue&>()(
std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
template <typename T>
struct FillMonotonicSeq
{
T init_value_{0};
T step_{1};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::generate(first, last, [=, n = init_value_]() mutable {
auto tmp = n;
n += step_;
return tmp;
});
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const
-> std::void_t<decltype(std::declval<const FillMonotonicSeq&>()(
std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
template <typename T>
struct FillConstant
{
T value_{0};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::fill(first, last, value_);
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const -> std::void_t<
decltype(std::declval<const FillConstant&>()(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
template <typename T, bool UseCos = true, bool UseAbs = false>
struct FillTrigValue
{
template <typename T_, bool UseCos_ = true, bool UseAbs_ = false>
struct LinearTrigGen
{
int i{0};
auto operator()()
{
float v = 0;
if constexpr(UseCos_)
{
v = cos(i);
}
else
{
v = sin(i);
}
if constexpr(UseAbs_)
v = abs(v);
i++;
return ck_tile::type_convert<T_>(v);
}
};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
LinearTrigGen<T, UseCos, UseAbs> gen;
std::generate(first, last, gen);
}
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const -> std::void_t<
decltype(std::declval<const FillTrigValue&>()(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
};
} // namespace ck_tile

36
include/ck_tile/host/hip_check_error.hpp Normal file
View File

@@ -0,0 +1,36 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core/config.hpp"
#include <sstream>
#include <stdexcept>
#include <hip/hip_runtime.h>
namespace ck_tile {
// To be removed, which really does not tell the location of failed HIP functional call
CK_TILE_HOST void hip_check_error(hipError_t x)
{
if(x != hipSuccess)
{
std::ostringstream ss;
ss << "HIP runtime error: " << hipGetErrorString(x) << ". " << __FILE__ << ": " << __LINE__
<< "in function: " << __func__;
throw std::runtime_error(ss.str());
}
}
} // namespace ck_tile
#define HIP_CHECK_ERROR(retval_or_funcall) \
do \
{ \
hipError_t _tmpVal = retval_or_funcall; \
if(_tmpVal != hipSuccess) \
{ \
std::ostringstream ostr; \
ostr << "HIP Function Failed (" << __FILE__ << "," << __LINE__ << ") " \
<< hipGetErrorString(_tmpVal); \
throw std::runtime_error(ostr.str()); \
} \
} while(0)

523
include/ck_tile/host/host_tensor.hpp Normal file
View File

@@ -0,0 +1,523 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <algorithm>
#include <cassert>
#include <iostream>
#include <iomanip>
#include <numeric>
#include <thread>
#include <utility>
#include <vector>
#include "ck_tile/core.hpp"
#include "ck_tile/host/ranges.hpp"
namespace ck_tile {
template <typename Range>
CK_TILE_HOST std::ostream& LogRange(std::ostream& os,
Range&& range,
std::string delim,
int precision = std::cout.precision(),
int width = 0)
{
bool first = true;
for(auto&& v : range)
{
if(first)
first = false;
else
os << delim;
os << std::setw(width) << std::setprecision(precision) << v;
}
return os;
}
template <typename T, typename Range>
CK_TILE_HOST std::ostream& LogRangeAsType(std::ostream& os,
Range&& range,
std::string delim,
int precision = std::cout.precision(),
int width = 0)
{
bool first = true;
for(auto&& v : range)
{
if(first)
first = false;
else
os << delim;
os << std::setw(width) << std::setprecision(precision) << static_cast<T>(v);
}
return os;
}
template <typename F, typename T, std::size_t... Is>
CK_TILE_HOST auto call_f_unpack_args_impl(F f, T args, std::index_sequence<Is...>)
{
return f(std::get<Is>(args)...);
}
template <typename F, typename T>
CK_TILE_HOST auto call_f_unpack_args(F f, T args)
{
constexpr std::size_t N = std::tuple_size<T>{};
return call_f_unpack_args_impl(f, args, std::make_index_sequence<N>{});
}
template <typename F, typename T, std::size_t... Is>
CK_TILE_HOST auto construct_f_unpack_args_impl(T args, std::index_sequence<Is...>)
{
return F(std::get<Is>(args)...);
}
template <typename F, typename T>
CK_TILE_HOST auto construct_f_unpack_args(F, T args)
{
constexpr std::size_t N = std::tuple_size<T>{};
return construct_f_unpack_args_impl<F>(args, std::make_index_sequence<N>{});
}
struct HostTensorDescriptor
{
HostTensorDescriptor() = default;
void CalculateStrides()
{
mStrides.clear();
mStrides.resize(mLens.size(), 0);
if(mStrides.empty())
return;
mStrides.back() = 1;
std::partial_sum(mLens.rbegin(),
mLens.rend() - 1,
mStrides.rbegin() + 1,
std::multiplies<std::size_t>());
}
template <typename X, typename = std::enable_if_t<std::is_convertible_v<X, std::size_t>>>
HostTensorDescriptor(const std::initializer_list<X>& lens) : mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename Lengths,
typename = std::enable_if_t<
std::is_convertible_v<ck_tile::ranges::range_value_t<Lengths>, std::size_t>>>
HostTensorDescriptor(const Lengths& lens) : mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename X,
typename Y,
typename = std::enable_if_t<std::is_convertible_v<X, std::size_t> &&
std::is_convertible_v<Y, std::size_t>>>
HostTensorDescriptor(const std::initializer_list<X>& lens,
const std::initializer_list<Y>& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
template <typename Lengths,
typename Strides,
typename = std::enable_if_t<
std::is_convertible_v<ck_tile::ranges::range_value_t<Lengths>, std::size_t> &&
std::is_convertible_v<ck_tile::ranges::range_value_t<Strides>, std::size_t>>>
HostTensorDescriptor(const Lengths& lens, const Strides& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
std::size_t get_num_of_dimension() const { return mLens.size(); }
std::size_t get_element_size() const
{
assert(mLens.size() == mStrides.size());
return std::accumulate(
mLens.begin(), mLens.end(), std::size_t{1}, std::multiplies<std::size_t>());
}
std::size_t get_element_space_size() const
{
std::size_t space = 1;
for(std::size_t i = 0; i < mLens.size(); ++i)
{
if(mLens[i] == 0)
continue;
space += (mLens[i] - 1) * mStrides[i];
}
return space;
}
const std::vector<std::size_t>& get_lengths() const { return mLens; }
const std::vector<std::size_t>& GetStrides() const { return mStrides; }
template <typename... Is>
std::size_t GetOffsetFromMultiIndex(Is... is) const
{
assert(sizeof...(Is) == this->get_num_of_dimension());
std::initializer_list<std::size_t> iss{static_cast<std::size_t>(is)...};
return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
}
std::size_t GetOffsetFromMultiIndex(std::vector<std::size_t> iss) const
{
return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
}
friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
private:
std::vector<std::size_t> mLens;
std::vector<std::size_t> mStrides;
};
template <typename New2Old>
CK_TILE_HOST HostTensorDescriptor transpose_host_tensor_descriptor_given_new2old(
const HostTensorDescriptor& a, const New2Old& new2old)
{
std::vector<std::size_t> new_lengths(a.get_num_of_dimension());
std::vector<std::size_t> new_strides(a.get_num_of_dimension());
for(std::size_t i = 0; i < a.get_num_of_dimension(); i++)
{
new_lengths[i] = a.get_lengths()[new2old[i]];
new_strides[i] = a.GetStrides()[new2old[i]];
}
return HostTensorDescriptor(new_lengths, new_strides);
}
struct joinable_thread : std::thread
{
template <typename... Xs>
joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...)
{
}
joinable_thread(joinable_thread&&) = default;
joinable_thread& operator=(joinable_thread&&) = default;
~joinable_thread()
{
if(this->joinable())
this->join();
}
};
template <typename F, typename... Xs>
struct ParallelTensorFunctor
{
F mF;
static constexpr std::size_t NDIM = sizeof...(Xs);
std::array<std::size_t, NDIM> mLens;
std::array<std::size_t, NDIM> mStrides;
std::size_t mN1d;
ParallelTensorFunctor(F f, Xs... xs) : mF(f), mLens({static_cast<std::size_t>(xs)...})
{
mStrides.back() = 1;
std::partial_sum(mLens.rbegin(),
mLens.rend() - 1,
mStrides.rbegin() + 1,
std::multiplies<std::size_t>());
mN1d = mStrides[0] * mLens[0];
}
std::array<std::size_t, NDIM> GetNdIndices(std::size_t i) const
{
std::array<std::size_t, NDIM> indices;
for(std::size_t idim = 0; idim < NDIM; ++idim)
{
indices[idim] = i / mStrides[idim];
i -= indices[idim] * mStrides[idim];
}
return indices;
}
void operator()(std::size_t num_thread = 1) const
{
std::size_t work_per_thread = (mN1d + num_thread - 1) / num_thread;
std::vector<joinable_thread> threads(num_thread);
for(std::size_t it = 0; it < num_thread; ++it)
{
std::size_t iw_begin = it * work_per_thread;
std::size_t iw_end = std::min((it + 1) * work_per_thread, mN1d);
auto f = [this, iw_begin, iw_end] {
for(std::size_t iw = iw_begin; iw < iw_end; ++iw)
{
call_f_unpack_args(this->mF, this->GetNdIndices(iw));
}
};
threads[it] = joinable_thread(f);
}
}
};
template <typename F, typename... Xs>
CK_TILE_HOST auto make_ParallelTensorFunctor(F f, Xs... xs)
{
return ParallelTensorFunctor<F, Xs...>(f, xs...);
}
template <typename T>
struct HostTensor
{
using Descriptor = HostTensorDescriptor;
using Data = std::vector<T>;
template <typename X>
HostTensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.get_element_space_size())
{
}
template <typename X, typename Y>
HostTensor(std::initializer_list<X> lens, std::initializer_list<Y> strides)
: mDesc(lens, strides), mData(mDesc.get_element_space_size())
{
}
template <typename Lengths>
HostTensor(const Lengths& lens) : mDesc(lens), mData(mDesc.get_element_space_size())
{
}
template <typename Lengths, typename Strides>
HostTensor(const Lengths& lens, const Strides& strides)
: mDesc(lens, strides), mData(get_element_space_size())
{
}
HostTensor(const Descriptor& desc) : mDesc(desc), mData(mDesc.get_element_space_size()) {}
template <typename OutT>
HostTensor<OutT> CopyAsType() const
{
HostTensor<OutT> ret(mDesc);
std::transform(mData.cbegin(), mData.cend(), ret.mData.begin(), [](auto value) {
return ck_tile::type_convert<OutT>(value);
});
return ret;
}
HostTensor() = delete;
HostTensor(const HostTensor&) = default;
HostTensor(HostTensor&&) = default;
~HostTensor() = default;
HostTensor& operator=(const HostTensor&) = default;
HostTensor& operator=(HostTensor&&) = default;
template <typename FromT>
explicit HostTensor(const HostTensor<FromT>& other) : HostTensor(other.template CopyAsType<T>())
{
}
decltype(auto) get_lengths() const { return mDesc.get_lengths(); }
decltype(auto) GetStrides() const { return mDesc.GetStrides(); }
std::size_t get_num_of_dimension() const { return mDesc.get_num_of_dimension(); }
std::size_t get_element_size() const { return mDesc.get_element_size(); }
std::size_t get_element_space_size() const { return mDesc.get_element_space_size(); }
std::size_t get_element_space_size_in_bytes() const
{
return sizeof(T) * get_element_space_size();
}
// void SetZero() { ck_tile::ranges::fill<T>(mData, 0); }
void SetZero() { std::fill(mData.begin(), mData.end(), 0); }
template <typename F>
void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
{
if(rank == mDesc.get_num_of_dimension())
{
f(*this, idx);
return;
}
// else
for(size_t i = 0; i < mDesc.get_lengths()[rank]; i++)
{
idx[rank] = i;
ForEach_impl(std::forward<F>(f), idx, rank + 1);
}
}
template <typename F>
void ForEach(F&& f)
{
std::vector<size_t> idx(mDesc.get_num_of_dimension(), 0);
ForEach_impl(std::forward<F>(f), idx, size_t(0));
}
template <typename F>
void ForEach_impl(const F&& f, std::vector<size_t>& idx, size_t rank) const
{
if(rank == mDesc.get_num_of_dimension())
{
f(*this, idx);
return;
}
// else
for(size_t i = 0; i < mDesc.get_lengths()[rank]; i++)
{
idx[rank] = i;
ForEach_impl(std::forward<const F>(f), idx, rank + 1);
}
}
template <typename F>
void ForEach(const F&& f) const
{
std::vector<size_t> idx(mDesc.get_num_of_dimension(), 0);
ForEach_impl(std::forward<const F>(f), idx, size_t(0));
}
template <typename G>
void GenerateTensorValue(G g, std::size_t num_thread = 1)
{
switch(mDesc.get_num_of_dimension())
{
case 1: {
auto f = [&](auto i) { (*this)(i) = g(i); };
make_ParallelTensorFunctor(f, mDesc.get_lengths()[0])(num_thread);
break;
}
case 2: {
auto f = [&](auto i0, auto i1) { (*this)(i0, i1) = g(i0, i1); };
make_ParallelTensorFunctor(f, mDesc.get_lengths()[0], mDesc.get_lengths()[1])(
num_thread);
break;
}
case 3: {
auto f = [&](auto i0, auto i1, auto i2) { (*this)(i0, i1, i2) = g(i0, i1, i2); };
make_ParallelTensorFunctor(f,
mDesc.get_lengths()[0],
mDesc.get_lengths()[1],
mDesc.get_lengths()[2])(num_thread);
break;
}
case 4: {
auto f = [&](auto i0, auto i1, auto i2, auto i3) {
(*this)(i0, i1, i2, i3) = g(i0, i1, i2, i3);
};
make_ParallelTensorFunctor(f,
mDesc.get_lengths()[0],
mDesc.get_lengths()[1],
mDesc.get_lengths()[2],
mDesc.get_lengths()[3])(num_thread);
break;
}
case 5: {
auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4) {
(*this)(i0, i1, i2, i3, i4) = g(i0, i1, i2, i3, i4);
};
make_ParallelTensorFunctor(f,
mDesc.get_lengths()[0],
mDesc.get_lengths()[1],
mDesc.get_lengths()[2],
mDesc.get_lengths()[3],
mDesc.get_lengths()[4])(num_thread);
break;
}
case 6: {
auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4, auto i5) {
(*this)(i0, i1, i2, i3, i4, i5) = g(i0, i1, i2, i3, i4, i5);
};
make_ParallelTensorFunctor(f,
mDesc.get_lengths()[0],
mDesc.get_lengths()[1],
mDesc.get_lengths()[2],
mDesc.get_lengths()[3],
mDesc.get_lengths()[4],
mDesc.get_lengths()[5])(num_thread);
break;
}
default: throw std::runtime_error("unspported dimension");
}
}
template <typename... Is>
std::size_t GetOffsetFromMultiIndex(Is... is) const
{
return mDesc.GetOffsetFromMultiIndex(is...);
}
template <typename... Is>
T& operator()(Is... is)
{
return mData[mDesc.GetOffsetFromMultiIndex(is...)];
}
template <typename... Is>
const T& operator()(Is... is) const
{
return mData[mDesc.GetOffsetFromMultiIndex(is...)];
}
T& operator()(std::vector<std::size_t> idx)
{
return mData[mDesc.GetOffsetFromMultiIndex(idx)];
}
const T& operator()(std::vector<std::size_t> idx) const
{
return mData[mDesc.GetOffsetFromMultiIndex(idx)];
}
typename Data::iterator begin() { return mData.begin(); }
typename Data::iterator end() { return mData.end(); }
typename Data::pointer data() { return mData.data(); }
typename Data::const_iterator begin() const { return mData.begin(); }
typename Data::const_iterator end() const { return mData.end(); }
typename Data::const_pointer data() const { return mData.data(); }
typename Data::size_type size() const { return mData.size(); }
template <typename U = T>
auto AsSpan() const
{
constexpr std::size_t FromSize = sizeof(T);
constexpr std::size_t ToSize = sizeof(U);
using Element = std::add_const_t<std::remove_reference_t<U>>;
return ck_tile::span<Element>{reinterpret_cast<Element*>(data()),
size() * FromSize / ToSize};
}
template <typename U = T>
auto AsSpan()
{
constexpr std::size_t FromSize = sizeof(T);
constexpr std::size_t ToSize = sizeof(U);
using Element = std::remove_reference_t<U>;
return ck_tile::span<Element>{reinterpret_cast<Element*>(data()),
size() * FromSize / ToSize};
}
Descriptor mDesc;
Data mData;
};
} // namespace ck_tile

166
include/ck_tile/host/kernel_launch.hpp Normal file
View File

@@ -0,0 +1,166 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core/config.hpp"
#include "ck_tile/host/stream_config.hpp"
#include "ck_tile/host/hip_check_error.hpp"
#include <hip/hip_runtime.h>
#include <cstddef>
namespace ck_tile {
template <int MaxThreadPerBlock, int MinBlockPerCu, typename Kernel, typename... Args>
#if CK_TILE_USE_LAUNCH_BOUNDS
__launch_bounds__(MaxThreadPerBlock, MinBlockPerCu)
#endif
__global__ void kentry(Kernel f, Args... args)
{
f(args...);
}
template <typename... Args, typename F>
CK_TILE_HOST float launch_and_time_kernel(const stream_config& s,
F kernel,
dim3 grid_dim,
dim3 block_dim,
std::size_t lds_byte,
Args... args)
{
#if CK_TILE_TIME_KERNEL
if(s.time_kernel_)
{
// warm up
for(int i = 0; i < s.cold_niters_; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
const int nrepeat = s.nrepeat_;
hipEvent_t start, stop;
HIP_CHECK_ERROR(hipEventCreate(&start));
HIP_CHECK_ERROR(hipEventCreate(&stop));
HIP_CHECK_ERROR(hipDeviceSynchronize());
HIP_CHECK_ERROR(hipEventRecord(start, s.stream_id_));
for(int i = 0; i < nrepeat; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
HIP_CHECK_ERROR(hipEventRecord(stop, s.stream_id_));
HIP_CHECK_ERROR(hipEventSynchronize(stop));
float total_time = 0;
HIP_CHECK_ERROR(hipEventElapsedTime(&total_time, start, stop));
return total_time / nrepeat;
}
else
{
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
}
#else
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
#endif
}
template <typename... Args, typename F, typename PreProcessFunc>
CK_TILE_HOST float launch_and_time_kernel_with_preprocess(const stream_config& s,
PreProcessFunc preprocess,
F kernel,
dim3 grid_dim,
dim3 block_dim,
std::size_t lds_byte,
Args... args)
{
#if CK_TILE_TIME_KERNEL
if(s.time_kernel_)
{
#if CK_TILE_DEBUG_LOG
printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
__func__,
grid_dim.x,
grid_dim.y,
grid_dim.z,
block_dim.x,
block_dim.y,
block_dim.z);
printf("Warm up 1 time\n");
#endif
// warm up
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
const int nrepeat = 10;
#if CK_TILE_DEBUG_LOG
printf("Start running %d times...\n", nrepeat);
#endif
hipEvent_t start, stop;
HIP_CHECK_ERROR(hipEventCreate(&start));
HIP_CHECK_ERROR(hipEventCreate(&stop));
HIP_CHECK_ERROR(hipDeviceSynchronize());
HIP_CHECK_ERROR(hipEventRecord(start, s.stream_id_));
for(int i = 0; i < nrepeat; ++i)
{
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
HIP_CHECK_ERROR(hipEventRecord(stop, s.stream_id_));
HIP_CHECK_ERROR(hipEventSynchronize(stop));
float total_time = 0;
HIP_CHECK_ERROR(hipEventElapsedTime(&total_time, start, stop));
return total_time / nrepeat;
}
else
{
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
}
#else
kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
#endif
}
template <int MaxThreadPerBlock = CK_TILE_MAX_THREAD_PER_BLOCK,
int MinBlockPerCu = CK_TILE_MIN_BLOCK_PER_CU,
typename KernelImpl,
typename... Args>
CK_TILE_HOST float launch_kernel(const stream_config& s,
KernelImpl kernel_impl,
dim3 grid_dim,
dim3 block_dim,
std::size_t dynamic_smem_byte,
Args... args)
{
const auto kernel = kentry<MaxThreadPerBlock, MinBlockPerCu, KernelImpl, Args...>;
return launch_and_time_kernel(
s, kernel, grid_dim, block_dim, dynamic_smem_byte, kernel_impl, args...);
}
} // namespace ck_tile

69
include/ck_tile/host/ranges.hpp Normal file
View File

@@ -0,0 +1,69 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iterator>
#include <type_traits>
#include <utility>
// ranges implementation are not intented to be used by user
// TODO: do we need this?
namespace ck_tile {
template <typename T>
using iter_value_t = typename std::iterator_traits<remove_cvref_t<T>>::value_type;
template <typename T>
using iter_reference_t = decltype(*std::declval<T&>());
template <typename T>
using iter_difference_t = typename std::iterator_traits<remove_cvref_t<T>>::difference_type;
namespace ranges {
template <typename R>
using iterator_t = decltype(std::begin(std::declval<R&>()));
template <typename R>
using sentinel_t = decltype(std::end(std::declval<R&>()));
template <typename R>
using range_size_t = decltype(std::size(std::declval<R&>()));
template <typename R>
using range_difference_t = ck_tile::iter_difference_t<ranges::iterator_t<R>>;
template <typename R>
using range_value_t = iter_value_t<ranges::iterator_t<R>>;
template <typename R>
using range_reference_t = iter_reference_t<ranges::iterator_t<R>>;
template <typename T, typename = void>
struct is_range : std::false_type
{
};
template <typename T>
struct is_range<
T,
std::void_t<decltype(std::begin(std::declval<T&>())), decltype(std::end(std::declval<T&>()))>>
: std::true_type
{
};
template <typename T>
inline constexpr bool is_range_v = is_range<T>::value;
template <typename T, typename = void>
struct is_sized_range : std::false_type
{
};
template <typename T>
struct is_sized_range<T, std::void_t<decltype(std::size(std::declval<T&>()))>>
: std::bool_constant<is_range_v<T>>
{
};
} // namespace ranges
} // namespace ck_tile

64
include/ck_tile/host/reference/reference_batched_elementwise.hpp Normal file
View File

@@ -0,0 +1,64 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename AElementOp = ck_tile::identity,
typename BElementOp = ck_tile::identity,
typename BinaryElementOp = ck_tile::plus<AccDataType>>
CK_TILE_HOST void reference_batched_elementwise(const HostTensor<ADataType>& a_b_m_n,
const HostTensor<BDataType>& b_b_m_n,
HostTensor<CDataType>& c_b_m_n,
const AElementOp& a_element_op = {},
const BElementOp& b_element_op = {},
const BinaryElementOp& binary_element_op = {})
{
const ck_tile::index_t N = c_b_m_n.mDesc.get_lengths()[2];
const bool broadcast_a_dim_b = (a_b_m_n.get_lengths()[0] == 1);
const bool broadcast_a_dim_m = (a_b_m_n.get_lengths()[1] == 1);
const bool broadcast_a_dim_n = (a_b_m_n.get_lengths()[2] == 1);
const bool broadcast_b_dim_b = (b_b_m_n.get_lengths()[0] == 1);
const bool broadcast_b_dim_m = (b_b_m_n.get_lengths()[1] == 1);
const bool broadcast_b_dim_n = (b_b_m_n.get_lengths()[2] == 1);
auto f = [&](auto batch, auto m) {
for(ck_tile::index_t n = 0; n < N; ++n)
{
AccDataType v_a{};
{
ck_tile::index_t i_b = (broadcast_a_dim_b ? 0 : batch);
ck_tile::index_t i_m = (broadcast_a_dim_m ? 0 : m);
ck_tile::index_t i_n = (broadcast_a_dim_n ? 0 : n);
v_a = ck_tile::type_convert<AccDataType>(a_element_op(a_b_m_n(i_b, i_m, i_n)));
}
AccDataType v_b{};
{
ck_tile::index_t i_b = (broadcast_b_dim_b ? 0 : batch);
ck_tile::index_t i_m = (broadcast_b_dim_m ? 0 : m);
ck_tile::index_t i_n = (broadcast_b_dim_n ? 0 : n);
v_b = ck_tile::type_convert<AccDataType>(b_element_op(b_b_m_n(i_b, i_m, i_n)));
}
c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(binary_element_op(v_a, v_b));
}
};
make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
std::thread::hardware_concurrency());
}
} // namespace ck_tile

50
include/ck_tile/host/reference/reference_batched_gemm.hpp Normal file
View File

@@ -0,0 +1,50 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename AElementOp = ck_tile::identity,
typename BElementOp = ck_tile::identity,
typename ACCElementOp = ck_tile::identity>
CK_TILE_HOST void reference_batched_gemm(const HostTensor<ADataType>& a_b_m_k,
const HostTensor<BDataType>& b_b_n_k,
HostTensor<CDataType>& c_b_m_n,
const AElementOp& a_element_op = {},
const BElementOp& b_element_op = {},
const ACCElementOp& acc_element_op = {})
{
const int N = b_b_n_k.mDesc.get_lengths()[1];
const int K = b_b_n_k.mDesc.get_lengths()[2];
auto f = [&](auto batch, auto m) {
for(int n = 0; n < N; ++n)
{
AccDataType v_acc = 0;
for(int k = 0; k < K; ++k)
{
ADataType v_a = a_element_op(a_b_m_k(batch, m, k));
BDataType v_b = b_element_op(b_b_n_k(batch, n, k));
v_acc += ck_tile::type_convert<AccDataType>(v_a) *
ck_tile::type_convert<AccDataType>(v_b);
}
c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
}
};
make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
std::thread::hardware_concurrency());
}
} // namespace ck_tile

32
include/ck_tile/host/reference/reference_batched_masking.hpp Normal file
View File

@@ -0,0 +1,32 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename CDataType, typename MaskingType>
CK_TILE_HOST void reference_batched_masking(HostTensor<CDataType>& c_b_m_n, const MaskingType& mask)
{
const int M = c_b_m_n.mDesc.get_lengths()[1];
const int N = c_b_m_n.mDesc.get_lengths()[2];
auto f = [&](auto batch) {
for(int n = 0; n < N; ++n)
{
for(int m = 0; m < M; ++m)
{
if(mask.IsOutOfBound(m, n))
c_b_m_n(batch, m, n) = -ck_tile::numeric<CDataType>::infinity();
}
}
};
make_ParallelTensorFunctor(f,
c_b_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
}
} // namespace ck_tile

71
include/ck_tile/host/reference/reference_batched_softmax.hpp Normal file
View File

@@ -0,0 +1,71 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename ADataType,
typename CompDataType,
typename BDataType,
typename CompElementOp = ck_tile::identity>
CK_TILE_HOST void reference_batched_softmax(
const HostTensor<ADataType>& a_b_m_n,
HostTensor<BDataType>& b_b_m_n,
const CompElementOp& comp_element_op = {},
std::optional<std::reference_wrapper<HostTensor<CompDataType>>> lse_b_m = std::nullopt)
{
const int N = a_b_m_n.mDesc.get_lengths()[2];
auto f = [&](auto batch, auto m) {
CompDataType v_max = -ck_tile::numeric<CompDataType>::infinity();
// max
for(int n = 0; n < N; ++n)
{
const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
v_max = v_max < v_a ? v_a : v_max;
}
CompDataType v_exp_sum = 0;
// validate v_max if all the elements within a row are -INF
if(std::isinf(v_max) && v_max < 0)
{
v_max = ck_tile::type_convert<CompDataType>(0.f);
}
// sum
for(int n = 0; n < N; ++n)
{
const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
v_exp_sum += ck_tile::exp(v_a - v_max);
}
// if sum is zero(masked), or nan/inf(other computation error), don't do divide
CompDataType inv_sum = (v_exp_sum == 0.f ? 1.f : 1.f / v_exp_sum);
// elementwise
for(int n = 0; n < N; ++n)
{
const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
const CompDataType v_b = ck_tile::exp(v_a - v_max) * inv_sum;
b_b_m_n(batch, m, n) = ck_tile::type_convert<BDataType>(comp_element_op(v_b));
}
// lse
if(lse_b_m)
{
lse_b_m->get()(batch, m) = v_max + ck_tile::log(v_exp_sum);
}
};
make_ParallelTensorFunctor(f, b_b_m_n.mDesc.get_lengths()[0], b_b_m_n.mDesc.get_lengths()[1])(
std::thread::hardware_concurrency());
}
} // namespace ck_tile

50
include/ck_tile/host/reference/reference_gemm.hpp Normal file
View File

@@ -0,0 +1,50 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename AElementOp = ck_tile::identity,
typename BElementOp = ck_tile::identity,
typename ACCElementOp = ck_tile::identity>
CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
const HostTensor<BDataType>& b_n_k,
HostTensor<CDataType>& c_m_n,
const AElementOp& a_element_op = {},
const BElementOp& b_element_op = {},
const ACCElementOp& acc_element_op = {})
{
const int N = b_n_k.mDesc.get_lengths()[0];
const int K = b_n_k.mDesc.get_lengths()[1];
auto f = [&](auto m) {
for(int n = 0; n < N; ++n)
{
AccDataType v_acc = 0;
for(int k = 0; k < K; ++k)
{
ADataType v_a = a_element_op(a_m_k(m, k));
BDataType v_b = b_element_op(b_n_k(n, k));
v_acc += ck_tile::type_convert<AccDataType>(v_a) *
ck_tile::type_convert<AccDataType>(v_b);
}
c_m_n(m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
}
};
make_ParallelTensorFunctor(f,
c_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
}
} // namespace ck_tile

61
include/ck_tile/host/reference/reference_im2col.hpp Normal file
View File

@@ -0,0 +1,61 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename T>
CK_TILE_HOST void reference_im2col(HostTensor<T>& in_mtx_host_ref,
const HostTensor<T>& in_host,
int /*N*/,
int /*K*/,
int C,
int /*Y*/,
int X,
int Hi,
int Wi,
int Ho,
int Wo,
int ConvStrideH,
int ConvStrideW,
int ConvDilationH,
int ConvDilationW,
int InLeftPadH,
int InLeftPadW,
int /*InRightPadH*/,
int /*InRightPadW*/)
{
int GemmM = in_mtx_host_ref.get_lengths()[0];
int GemmK = in_mtx_host_ref.get_lengths()[1];
for(int gemm_m = 0; gemm_m < GemmM; ++gemm_m)
{
int mtmp = gemm_m;
int n = mtmp / (Ho * Wo);
mtmp -= n * Ho * Wo;
int ho = mtmp / Wo;
int wo = mtmp - ho * Wo;
for(int gemm_k = 0; gemm_k < GemmK; ++gemm_k)
{
int ktmp = gemm_k;
int y = ktmp / (X * C);
ktmp -= y * X * C;
int x = ktmp / C;
int c = ktmp - x * C;
int hi = y * ConvDilationH + ho * ConvStrideH - InLeftPadH;
int wi = x * ConvDilationW + wo * ConvStrideW - InLeftPadW;
bool inbound = (hi >= 0 && hi < Hi && wi >= 0 && wi < Wi);
in_mtx_host_ref(gemm_m, gemm_k) = inbound ? in_host(n, hi, wi, c) : 0;
}
}
}
} // namespace ck_tile

32
include/ck_tile/host/reference/reference_reduce.hpp Normal file
View File

@@ -0,0 +1,32 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename ADataType, typename AccDataType, typename BDataType>
CK_TILE_HOST void reference_reduce(const HostTensor<ADataType>& a_m_n, HostTensor<BDataType>& b_m)
{
auto f = [&](auto m) {
const int N = a_m_n.mDesc.get_lengths()[1];
AccDataType v_acc = 0;
for(int n = 0; n < N; ++n)
{
const ADataType v_a = a_m_n(m, n);
v_acc += v_a;
}
b_m(m) = ck_tile::type_convert<BDataType>(v_acc);
};
make_ParallelTensorFunctor(f, b_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
}
} // namespace ck_tile

51
include/ck_tile/host/reference/reference_softmax.hpp Normal file
View File

@@ -0,0 +1,51 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename ADataType, typename AccDataType, typename BDataType>
CK_TILE_HOST void reference_softmax(const HostTensor<ADataType>& a_m_n,
HostTensor<BDataType>& b_m_n)
{
auto f = [&](auto m) {
const int N = a_m_n.mDesc.get_lengths()[1];
AccDataType v_max = ck_tile::numeric<ADataType>::Lowest();
// max
for(int n = 0; n < N; ++n)
{
const ADataType v_a = a_m_n(m, n);
v_max = v_max < v_a ? v_a : v_max;
}
AccDataType v_exp_sum = 0;
// sum
for(int n = 0; n < N; ++n)
{
const ADataType v_a = a_m_n(m, n);
v_exp_sum += ck_tile::exp(v_a - v_max);
}
// elementwise
for(int n = 0; n < N; ++n)
{
const ADataType v_a = a_m_n(m, n);
b_m_n(m, n) = ck_tile::exp(v_a - v_max) / v_exp_sum;
}
};
make_ParallelTensorFunctor(f,
b_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
}
} // namespace ck_tile

17
include/ck_tile/host/stream_config.hpp Normal file
View File

@@ -0,0 +1,17 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <hip/hip_runtime.h>
namespace ck_tile {
struct stream_config
{
hipStream_t stream_id_ = nullptr;
bool time_kernel_ = false;
int log_level_ = 0;
int cold_niters_ = 3;
int nrepeat_ = 10;
};
} // namespace ck_tile