Skip to content

GitLab

Unverified Commit e6bb1dd7 authored by Po Yen Chen's avatar Po Yen Chen Committed by GitHub
Browse files

Merge branch 'develop' into feature/check-window-lengths

parents 9d6a3704 ab250afd
Hide whitespace changes
Inline Side-by-side
Showing with 702 additions and 119 deletions
codegen/test/grouped_conv_fwd_multiple_d_v4.cpp 0 → 100644
View file @ e6bb1dd7
#include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
#include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_problem.hpp"
#include "ck/host/headers.hpp"
#include "ck/host/stringutils.hpp"
#include "ck/host/utils.hpp"
#include "ck/tensor_operation/gpu/device/helper.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
#include "common.hpp"
#include <test.hpp>
#include <rtc/compile_kernel.hpp>
#include <rtc/hip.hpp>
#include <fstream>
// need this for verification
/**struct Epilogue
{
Epilogue(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename D>
__host__ __device__ constexpr void operator()(E& e, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, ck::half_t>(ck::half_t& e,
const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * e + beta_ * ck::type_convert<float>(d));
}
float alpha_;
float beta_;
};**/
const std::string conv_compile_check = R"__ck__(
#include <${include}>
${template};
)__ck__";
TEST_CASE(test_problem_kernel)
{
// set up problem specification
ck::host::conv::Problem_Conv_Fwd prob;
prob.NumDim = 2;
prob.G = 32;
prob.N = 256;
prob.C = 32;
prob.K = 64;
prob.Y = 3;
prob.X = 3;
prob.Hi = 28;
prob.Wi = 28;
prob.Ho = 28;
prob.Wo = 28;
check_all<ck::half_t> check;
// user provided fusion operations
std::string epilogue = R"(
struct Epilogue
{
__host__ __device__ Epilogue(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename D>
__host__ __device__ constexpr void operator()(E& e, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, ck::half_t>(ck::half_t& e,
const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * e + beta_ * ck::type_convert<float>(d));
}
float alpha_;
float beta_;
};
)";
std::string prologue = "";
// length+stride arrays
ck::Array<ck::index_t, 5> in_lengths{static_cast<int>(prob.G),
static_cast<int>(prob.N),
static_cast<int>(prob.C),
static_cast<int>(prob.Hi),
static_cast<int>(prob.Wi)};
ck::Array<ck::index_t, 5> out_lengths{static_cast<int>(prob.G),
static_cast<int>(prob.N),
static_cast<int>(prob.K),
static_cast<int>(prob.Ho),
static_cast<int>(prob.Wo)};
ck::Array<ck::index_t, 5> wei_lengths{static_cast<int>(prob.G),
static_cast<int>(prob.K),
static_cast<int>(prob.C),
static_cast<int>(prob.Y),
static_cast<int>(prob.X)};
ck::Array<ck::index_t, 5> d_lengths = {};
ck::Array<ck::index_t, 5> in_strides{static_cast<int>(prob.C),
static_cast<int>(prob.Hi * prob.Wi * prob.G * prob.C),
1,
static_cast<int>(prob.Wi * prob.G * prob.C),
static_cast<int>(prob.G * prob.C)};
ck::Array<ck::index_t, 5> out_strides{static_cast<int>(prob.K),
static_cast<int>(prob.Ho * prob.Wo * prob.G * prob.K),
1,
static_cast<int>(prob.Wo * prob.G * prob.K),
static_cast<int>(prob.G * prob.K)};
ck::Array<ck::index_t, 5> wei_strides{static_cast<int>(prob.K * prob.Y * prob.X * prob.C),
static_cast<int>(prob.Y * prob.X * prob.C),
1,
static_cast<int>(prob.X * prob.C),
static_cast<int>(prob.C)};
ck::Array<ck::index_t, 5> d_strides = {};
ck::Array<ck::index_t, 2> conv_filter_strides = {1, 1};
ck::Array<ck::index_t, 2> conv_filter_dilations = {1, 1};
ck::Array<ck::index_t, 2> input_left_pads = {1, 1};
ck::Array<ck::index_t, 2> input_right_pads = {1, 1};
// move the data onto the device
auto in_dev =
to_gpu(generate_buffer<ck::half_t, ck::Array<ck::index_t, 5>>(in_lengths, in_strides, 0));
auto wei_dev =
to_gpu(generate_buffer<ck::half_t, ck::Array<ck::index_t, 5>>(wei_lengths, wei_strides, 1));
auto out_dev =
to_gpu(generate_buffer<ck::half_t, ck::Array<ck::index_t, 5>>(out_lengths, out_strides, 2));
// CK Verficiation: Reference Kernel
/**bool pass = true;
Tensor<ck::half_t> in_host(in_lengths, in_strides);
in_host.GenerateTensorValue(GeneratorTensor_1<ck::half_t>{1});
Tensor<ck::half_t> wei_host(wei_lengths, wei_strides);
wei_host.GenerateTensorValue(GeneratorTensor_1<ck::half_t>{1});
Tensor<ck::half_t> out_host(out_lengths, out_strides);
std::vector<ck::index_t> conv_filter_strides_ = {1, 1};
std::vector<ck::index_t> conv_filter_dilations_ = {1, 1};
std::vector<ck::index_t> input_left_pads_ = {1, 1};
std::vector<ck::index_t> input_right_pads_ = {1, 1};
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<
2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
Epilogue>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_host,
wei_host,
out_host,
conv_filter_strides_,
conv_filter_dilations_,
input_left_pads_,
input_right_pads_,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
Epilogue{1.0f, 1.0f});
out_host.SetZero();
ref_invoker.Run(ref_argument);**/
for(auto solution : prob.GetSolutions("gfx908", prologue, epilogue))
{
// substitute instance values into the template
auto src = ck::host::InterpolateString(
conv_compile_check,
{{"include", prob.GetIncludeHeader()}, {"template", solution.ToTemplateString()}});
auto srcs = get_headers_for_test();
srcs.push_back({"main.cpp", src});
rtc::compile_options options;
auto name = solution.GetTemplateParameter<std::string>("name");
options.kernel_name = "run_" + name;
auto k = rtc::compile_kernel(srcs, options);
// Grid size calculation
auto block_size = solution.GetTemplateParameter<ck::index_t>("BlockSize");
auto tmp = get_launch_params(solution, out_lengths, out_strides);
auto grid_size = tmp * in_lengths[1];
// launch the kernel with arguments needed for the argument pointer
k.launch(nullptr, grid_size * block_size, block_size)(in_dev.data(),
wei_dev.data(),
out_dev.data(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
out_lengths,
out_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
// auto res = rtc::from_gpu(out_dev);
// pass &= ck::utils::check_err(res, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
// assert(pass);
// Simple check: this checks that the output from each instance matches the output from the
// first instance
CHECK(report(solution, check(rtc::from_gpu(out_dev))));
}
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }
codegen/test/rtc/src/compile_kernel.cpp
View file @ e6bb1dd7
......@@ -56,6 +56,8 @@ void write_string(const std::string& filename, const std::string_view& buffer)
}
std::string compiler() { return "/opt/rocm/llvm/bin/clang++ -x hip --cuda-device-only"; }
// 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)
{
......@@ -89,6 +91,12 @@ kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options
auto obj = read_buffer(out_path.string());
std::ofstream ofh("obj.o", std::ios::binary);
for(auto i : obj)
ofh << i;
ofh.close();
// int s = std::system(("/usr/bin/cp " + out_path.string() + " codeobj.bin").c_str());
// assert(s == 0);
return kernel{obj.data(), options.kernel_name};
}
......
codegen/test/rtc/src/hip.cpp
View file @ e6bb1dd7
......@@ -2,6 +2,7 @@
#include <rtc/manage_ptr.hpp>
#include <stdexcept>
#include <cassert>
#include <iostream>
namespace rtc {
......@@ -49,7 +50,10 @@ std::size_t get_available_gpu_memory()
size_t total;
auto status = hipMemGetInfo(&free, &total);
if(status != hipSuccess)
throw std::runtime_error("Failed getting available memory: " + hip_error(status));
{
std::cerr << "Failed getting available memory: " + hip_error(status) << std::endl;
return (8ull * 1024ull * 1024ull * 1024ull);
}
return free;
}
......
docs/sphinx/requirements.in
View file @ e6bb1dd7
rocm-docs-core==0.38.1
rocm-docs-core==1.5.0
sphinxcontrib-bibtex==2.6.2
docs/sphinx/requirements.txt
View file @ e6bb1dd7
#
# This file is autogenerated by pip-compile with Python 3.8
# This file is autogenerated by pip-compile with Python 3.10
# by the following command:
#
# pip-compile requirements.in
......@@ -48,12 +48,6 @@ idna==3.4
# via requests
imagesize==1.4.1
# via sphinx
importlib-metadata==6.8.0
# via
# sphinx
# sphinxcontrib-bibtex
importlib-resources==6.1.0
# via rocm-docs-core
jinja2==3.1.2
# via
# myst-parser
......@@ -99,8 +93,6 @@ pyjwt[crypto]==2.6.0
# via pygithub
pynacl==1.5.0
# via pygithub
pytz==2023.3.post1
# via babel
pyyaml==6.0
# via
# myst-parser
......@@ -111,7 +103,7 @@ requests==2.31.0
# via
# pygithub
# sphinx
rocm-docs-core==0.38.1
rocm-docs-core==1.5.0
# via -r requirements.in
six==1.16.0
# via
......@@ -165,7 +157,3 @@ urllib3==1.26.18
# via requests
wrapt==1.15.0
# via deprecated
zipp==3.17.0
# via
# importlib-metadata
# importlib-resources
example/01_gemm/CMakeLists.txt
View file @ e6bb1dd7
......@@ -22,6 +22,8 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
add_example_executable(example_gemm_xdl_fp16_v2 gemm_xdl_fp16_v2.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v2)
add_example_executable(example_gemm_xdl_fp16_streamk_v3 gemm_xdl_fp16_streamk_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_streamk_v3)
add_example_executable(example_gemm_xdl_fp16_v3 gemm_xdl_fp16_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v3)
add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
......
example/01_gemm/README.md
View file @ e6bb1dd7
......@@ -8,16 +8,20 @@
./bin/example_gemm_xdl 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
# Instructions for ```example_gemm_xdl_fp16_streamk_v3```
## Run ```example_gemm_xdl_fp16_streamk_v3```
```bash
arg1: verification (0=no, 1=yes)
arg2: initialization (0=no init, 1=integer value, 2=decimal value)
arg3: time kernel (0=no, 1=yes)
arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
arg10: stream-k select (-1: default config, 0: all DP, 1: 1-tile SK, 2: 2-tile SK)
arg11: Grid_size(-1 for max occupancy)
bin/example_gemm_xdl_fp16_streamk_v3 1 2 1 3840 4096 4096 4096 4096 4096 1 -1
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
problem {M:3840, N:4096, K:4096, SA:4096, SB:4096, SC:4096, MP:4032, NP:4096, KRead:4096, KP:4096, AK0:512, BK0:2048, MBlock: 18, NBlock: 16, Stream-K Selection:1, Grid size:-1}
Perf: 0.292022 ms, 441.23 TFlops, 330.348 GB/s, DeviceGemmXdlUniversal<MNPadding, RRR> BlkSize: 256, BlkTile: 224x256x64, WaveTile: 16x16, WaveMap: 7x8, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2
```
example/01_gemm/common.hpp
View file @ e6bb1dd7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -45,6 +45,19 @@ struct ProblemSizeStreamK final
ck::index_t NumSKBlocks = -1;
};
struct ProblemSizeStreamK_universal final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t Grid_size = -1; // defaults to max occupancy
ck::index_t Streamk_sel = 1; // defaults to 1-tile SK
};
struct ProblemSizeSplitK final
{
......@@ -123,6 +136,57 @@ bool parse_cmd_args<ProblemSize>(int argc,
return true;
}
template <>
bool parse_cmd_args<ProblemSizeStreamK_universal>(int argc,
char* argv[],
ProblemSizeStreamK_universal& problem_size,
ExecutionConfig& config)
{
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc >= 10)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
problem_size.StrideA = std::stoi(argv[7]);
problem_size.StrideB = std::stoi(argv[8]);
problem_size.StrideC = std::stoi(argv[9]);
if(argc >= 11)
{
problem_size.Streamk_sel = std::stoi(argv[10]);
problem_size.Grid_size = std::stoi(argv[11]);
}
}
else
{
std::cerr
<< "arg1: verification (0=no, 1=yes)" << std::endl
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)" << std::endl
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
<< "arg10: stream-k select (-1: default config, 0: all DP, 1: 1-tile SK, 2: 2-tile SK)"
<< "\narg11: Grid_size(-1 for max occupancy)" << std::endl;
return false;
}
return true;
}
template <>
bool parse_cmd_args<ProblemSizeStreamK>(int argc,
char* argv[],
......@@ -165,7 +229,8 @@ bool parse_cmd_args<ProblemSizeStreamK>(int argc,
<< std::endl
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
<< "arg10: NumSKBlocks(optional)" << std::endl;
<< "arg10: stream-k select (0: all DP, 1: 1-tile SK, 2: 2-tile SK)"
<< "\narg11: Grid_size(-1 for max occupancy)" << std::endl;
return false;
}
......
example/01_gemm/gemm_wmma_fp16.cpp
View file @ e6bb1dd7
......@@ -23,45 +23,45 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
< ALayout,
BLayout,
CLayout,
ADataType,
< ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CElementOp,
GemmDefault,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CElementOp,
GemmDefault,
1, // Prefetch stage
128, // BlockSize
64, // MPerBlock
128, // NPerBlock
64, // KPerBlock
8, // K1
2, // K1
16, // MPerWmma
16, // NPerWmma
2, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
4, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
2,
2,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
2,
2,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
S<1, 32, 1, 4>,
8>;
// clang-format on
......
example/01_gemm/gemm_xdl_fp16_streamk_v3.cpp 0 → 100644
View file @ e6bb1dd7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CDataType = ck::half_t;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmV2_Streamk_Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_Streamk_V3<
ALayout, BLayout, CLayout,
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault,
256,
224, 256,
64, 8, 2,
16, 16,
7, 8,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 2, 0,
1, 2, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example_streamk_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_universal_streamk_example(argc, argv); }
example/01_gemm/run_gemm_example.inc
View file @ e6bb1dd7
......@@ -159,7 +159,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
break;
case 4:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{1.f, 1.f}(b_k_n);
break;
case 5:
......
example/01_gemm/run_gemm_example_streamk_v2.inc 0 → 100644
View file @ e6bb1dd7
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename DataType>
inline __host__ __device__ constexpr double get_rtol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 1e-1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 1.5e-1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename DataType>
inline __host__ __device__ constexpr double get_atol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 16.1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 8192.1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
#endif
using namespace ck::literals;
auto M = problem_size.M;
auto N = problem_size.N;
auto K = problem_size.K;
auto StrideA = problem_size.StrideA;
auto StrideB = problem_size.StrideB;
auto StrideC = problem_size.StrideC;
auto Grid_size = problem_size.Grid_size;
auto Streamk_sel = problem_size.Streamk_sel;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<std::size_t>(col);
}
else
{
return static_cast<std::size_t>(row);
}
}
else
return static_cast<std::size_t>(stride);
};
auto f_get_default_streamk_policy = [](ck::index_t streamk_sel) {
if(streamk_sel == -1)
{
return static_cast<std::size_t>(4);
}
else
return static_cast<std::size_t>(streamk_sel);
};
StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
Streamk_sel = f_get_default_streamk_policy(Streamk_sel);
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
switch(config.init_method)
{
case 0:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
break;
case 3:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
#ifdef BUILD_INT4_EXAMPLE
DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
c_m_n_device_result.mDesc.GetElementSpaceSize());
const Tensor<KernelADataType> a_m_k_converted(a_m_k);
const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
#else
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
#endif
DeviceMem workspace;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmV2_Streamk_Instance{};
auto invoker = gemm.MakeInvoker();
float ave_time = 0;
auto argument = gemm.MakeArgument(
#ifdef BUILD_INT4_EXAMPLE
static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#endif
M,
N,
K,
StrideA,
StrideB,
StrideC,
Streamk_sel,
Grid_size,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
return true;
}
bool pass = true;
if(config.do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 1});
#ifdef BUILD_INT4_EXAMPLE
Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
#else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
#endif
}
if(config.time_kernel)
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
}
return pass;
}
bool run_gemm_universal_streamk_example(int argc, char* argv[])
{
ProblemSizeStreamK_universal problem_size;
ExecutionConfig config;
return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
}
example/02_gemm_bilinear/README.md
View file @ e6bb1dd7
......@@ -9,20 +9,3 @@
#arg11 to 12: alpha, beta
./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
```
Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
arg.c0_grid_desc_m_n_{ 3840, 4096}
arg.c_grid_desc_m_n_{ 3840, 4096}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Perf: 0.936965 ms, 137.517 TFlops, 102.959 GB/s
error: 0
max_diff: 0, 558.5, 558.5
```
example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
View file @ e6bb1dd7
......@@ -17,6 +17,7 @@
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/host_utility/device_prop.hpp"
struct AlphaBetaAdd
{
......@@ -175,6 +176,14 @@ int main(int argc, char* argv[])
exit(0);
}
bool is_supported = ck::is_gfx11_supported();
if(!is_supported)
{
std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
<< std::endl;
return 0;
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
......
example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
View file @ e6bb1dd7
......@@ -17,6 +17,7 @@
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/host_utility/device_prop.hpp"
struct AlphaBetaAdd
{
......@@ -175,6 +176,14 @@ int main(int argc, char* argv[])
exit(0);
}
bool is_supported = ck::is_gfx11_supported();
if(!is_supported)
{
std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
<< std::endl;
return 0;
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
......
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View file @ e6bb1dd7
......@@ -24,4 +24,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_lds_direct_load_fp32)
set(target 1)
endif()
endforeach()
\ No newline at end of file
endforeach()
example/04_gemm_add_add_fastgelu/README.md
View file @ e6bb1dd7
......@@ -8,16 +8,3 @@
#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
```
example/09_convnd_fwd/CMakeLists.txt
View file @ e6bb1dd7
......@@ -3,8 +3,7 @@ add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp8 convnd_fwd_xdl_fp8.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
add_example_executable(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
add_example_executable(example_convnd_fwd_xdl_bf8 convnd_fwd_xdl_bf8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16_comp_fp8 convnd_fwd_xdl_fp16_comp_fp8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp8_bf8 convnd_fwd_xdl_fp8_bf8.cpp)
......
example/09_convnd_fwd/README.md
View file @ e6bb1dd7
......@@ -16,17 +16,3 @@
# <right padding>, (ie RightPy, RightPx for 2D)
./bin/example_convnd_fwd_xdl 0 1 100
```
Result (MI100 @ 1087Mhz, 33.4TFlops peak FP32)
```
input: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
weights: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
output: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
arg.a_grid_desc_k0_m_k1_{432, 165888, 4}
arg.b_grid_desc_k0_n_k1_{432, 256, 4}
arg.c_grid_desc_m_n_{ 165888, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 100 times...
Perf: 4.43736 ms, 33.0753 TFlops, 150.357 GB/s
```
example/15_grouped_gemm/README.md
View file @ e6bb1dd7
......@@ -7,19 +7,3 @@
#arg3: run kernel # of times (>1)
./bin/example_grouped_gemm_xdl_fp16 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
```
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment