ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
4d041837ade7ae01900a0442d939f80b723b1631
composable_kernel/example/ck_tile/03_gemm/run_gemm_example.inc
rahjain-amd 4d041837ad Add json dump support to output details from CK/CKTile Examples. (#2551) 
* Adding RapidJson Library

* Adding Json Dumps in all CK_Tile Examples

Not verified yet

* Adding json to cktile Batched Transpose

* adding json dumps to layernorm2d_fwd

* Adding  json dump to flatmm_basic

* Adding RapidJson Library

* Adding Json Dumps in all CK_Tile Examples

Not verified yet

* Adding json to cktile Batched Transpose

* adding json dumps to layernorm2d_fwd

* Adding  json dump to flatmm_basic

* Adding json in 03_gemm

* Add json dump to 16_batched_gemm

* Add json dump to gemm_multi_d_fp16

* Add json dump to grouped_gemm

* fix fmha_bwd/fwd

* Fix clang-format errors

exclude include/rapidjson in jenkins as its a third-party library

* Saparating function and defination.

* Update Documentation of 03_gemm

* Refactoring as per code review

* Disable fp8 instances on unsupported targets (#2592)

* Restrict building of gemm_universal_preshuffle_f8 instances to specific targets in CMakeLists.txt

* Add condition to skip gemm_xdl_universal_preshuffle_f8 instances for unsupported targets in CMakeLists.txt

* Add conditions to skip unsupported targets for gemm_universal_preshuffle_f8 and gemm_xdl_universal_preshuffle_f8 instances in CMakeLists.txt

* Refine conditions to exclude gemm_universal_preshuffle_f8 instances for unsupported targets in CMakeLists.txt

---------

Co-authored-by: AviralGoelAMD <aviralgoel@amd.com>

* fix clang format

* remove duplicate lines of code from library/src/tensor_operation_instance/gpu/CMakeLists.txt

* Fixing Readme and unifying jsondumps

* adding moe_smoothquant

* adding fused_moe

* Fixing Readme for batched_gemm

* Fixing Readme for grouped_gemm

* adding flatmm

* adding gemm_multi_d_fp16

* adding elementwise

* adding File name when json is dumped

* Fixing Reduce after merge

* adding batched_transpose

* Adding Warptile in Gemm

* Fixing Clang Format

---------

Co-authored-by: Aviral Goel <aviral.goel@amd.com>
Co-authored-by: AviralGoelAMD <aviralgoel@amd.com>
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
2025-09-02 23:31:29 -07:00

523 lines
21 KiB
C++
Raw Blame History

// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
{
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
ck_tile::tensor_layout::gemm::RowMajor>>{};
}
template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
{
using ComputeType =
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
// Calculate thresholds
const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(
ck_tile::integer_divide_ceil(K, kbatch));
const auto atol = ck_tile::get_absolute_threshold<ComputeType, CDataType, AccDataType>(
max_accumulated_value / kbatch, ck_tile::integer_divide_ceil(K, kbatch));
// Calculate error due to split_k accumulation
const auto rtol_split_k =
ck_tile::get_relative_threshold<CDataType, CDataType, CDataType>(kbatch);
const auto atol_split_k = ck_tile::get_absolute_threshold<CDataType, CDataType, CDataType>(
max_accumulated_value, kbatch);
// Use higher threshold
return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
}
template <typename GemmConfig,
typename Tensor,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void permute_tensor_b(Tensor& tensor)
{
using GemmShape = ck_tile::TileGemmShape<
ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
ck_tile::
sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>,
GemmConfig::PermuteA,
GemmConfig::PermuteB>;
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<GemmConfig::kPadM,
GemmConfig::kPadN,
GemmConfig::kPadK,
GemmConfig::DoubleSmemBuffer,
ALayout,
BLayout,
CLayout,
GemmConfig::TransposeC,
GemmConfig::UseStructuredSparsity>;
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
GemmUniversalTraits,
GemmConfig::Scheduler,
true,
ck_tile::TailNumber::Full>;
using GemmPipeline = typename PipelineTypeTraits<GemmConfig::Pipeline>::template GemmPipeline<
UniversalGemmProblem>;
const ck_tile::index_t K = tensor.get_length(0);
const ck_tile::index_t N = tensor.get_length(1);
const ck_tile::index_t K1 = GemmPipeline::GetSmemPackB();
const ck_tile::index_t K0 = K / K1;
Tensor tensor_copy = tensor;
// int K0, N, K1
for(int j = 0; j < K0; j++)
{
for(int i = 0; i < N; i++)
{
for(int jj = 0; jj < K1; jj++)
{
tensor(j * N * K1 + i * K1 + jj) = tensor_copy(i * K + (j * K1 + jj));
}
}
}
}
template <typename Tensor>
void permute_vectors_i4x4_b(Tensor& tensor)
{
const ck_tile::index_t K = tensor.get_length(0);
const ck_tile::index_t N = tensor.get_length(1);
// vector pk_i4x4 permute
for(int i = 0; i < N; i++)
{
for(int j = 0; j < K; j += 8)
{
int8_t input[8];
for(int k = 0; k < 4; k++)
{
int8_t i4x2 = tensor(j + k * 2, i).data;
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
}
// permute 01234567->20643175
{
int8_t hi = input[2];
int8_t lo = input[0];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 0, i) = i4x2;
}
{
int8_t hi = input[6];
int8_t lo = input[4];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 2, i) = i4x2;
}
{
int8_t hi = input[3];
int8_t lo = input[1];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 4, i) = i4x2;
}
{
int8_t hi = input[7];
int8_t lo = input[5];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 6, i) = i4x2;
}
}
}
}
template <typename GemmConfig,
typename ADataType,
typename BDataType,
typename DsDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout,
bool Persistent,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s);
template <typename GemmConfig,
typename ADataType,
typename BDataType,
typename DsDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf,
ck_tile::DeviceMem& c_m_n_dev_buf,
ck_tile::index_t M,
ck_tile::index_t N,
ck_tile::index_t K,
ck_tile::index_t stride_A,
ck_tile::index_t stride_B,
ck_tile::index_t stride_C,
ck_tile::index_t kbatch,
int n_warmup,
int n_repeat,
bool persistent,
bool flush_cache,
int rotating_count)
{
ck_tile::GemmHostArgs args = {a_m_k_dev_buf.GetDeviceBuffer(),
b_k_n_dev_buf.GetDeviceBuffer(),
c_m_n_dev_buf.GetDeviceBuffer(),
kbatch,
M,
N,
K,
stride_A,
stride_B,
stride_C};
float ave_time;
if(persistent)
{
ave_time = gemm<GemmConfig,
ADataType,
BDataType,
DsDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
true,
CDEElementWise>(
args,
ck_tile::stream_config{
nullptr, true, 1, n_warmup, n_repeat, true, flush_cache, rotating_count});
}
else
{
ave_time = gemm<GemmConfig,
ADataType,
BDataType,
DsDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
false,
CDEElementWise>(
args,
ck_tile::stream_config{
nullptr, true, 1, n_warmup, n_repeat, true, flush_cache, rotating_count});
}
return ave_time;
}
template <typename GemmConfig, typename T>
auto shuffle_b(const ck_tile::HostTensor<T>& t)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
constexpr int divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
k_ / GemmConfig::K_Warp_Tile,
divisor,
GemmConfig::K_Warp_Tile / divisor});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
}
template <typename CDataType>
bool do_verify(const ck_tile::HostTensor<CDataType>& c_m_n_dev_result,
const ck_tile::HostTensor<CDataType>& c_m_n_ref,
const ck_tile::tuple<double, double>& rtol_atol,
const char* variant)
{
bool pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{}) << std::endl;
std::cout << "The " << variant << " verification result is:" << (pass ? "correct" : "fail")
<< std::endl;
return pass;
}
template <typename GemmConfig,
typename ADataType,
typename BDataType = ADataType,
typename CDataType = ADataType,
typename ALayout,
typename BLayout,
typename CLayout>
int run_gemm_example_with_layouts(ck_tile::ArgParser& arg_parser,
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
using AccDataType = typename GemmTypeConfig<ADataType, BDataType, CDataType>::AccDataType;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat");
ck_tile::index_t init_method = arg_parser.get_int("init");
bool persistent = arg_parser.get_int("persistent");
bool flush_cache = arg_parser.get_bool("flush_cache");
int rotating_count = arg_parser.get_int("rotating_count");
const bool preshuffle = GemmConfig::Preshuffle;
stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
ck_tile::HostTensor<ADataType> a_m_k(
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_k_n(
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
if(init_method == 0)
{
if constexpr(preshuffle)
{
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_k_n);
}
else
{
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
}
}
else if(init_method == 1)
{
ck_tile::FillMonotonicSeq<ADataType>{}(a_m_k);
ck_tile::FillMonotonicSeq<BDataType>{}(b_k_n);
}
else if(init_method == 2)
{
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_k_n);
}
else
{
a_m_k.SetZero();
b_k_n.SetZero();
}
if(!preshuffle && GemmConfig::UseStructuredSparsity)
{
ck_tile::AdjustToStructuredSparsity<ADataType>{}(a_m_k);
}
ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
static_assert(!GemmConfig::PermuteA, "Not implemented");
if constexpr(preshuffle)
{
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b<GemmConfig>(b_k_n);
// shuffled buffer B for device implementation
b_k_n_dev_buf.ToDevice(b_shuffle_host.data());
}
else
{
if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
{
// Permute vector pk_i4x4 data for device implementation
ck_tile::HostTensor<BDataType> b_k_n_dev = b_k_n;
if constexpr(GemmConfig::PermuteB)
{
permute_tensor_b<GemmConfig,
decltype(b_k_n_dev),
ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(b_k_n_dev);
}
permute_vectors_i4x4_b(b_k_n_dev);
b_k_n_dev_buf.ToDevice(b_k_n_dev.data());
}
else
{
if constexpr(GemmConfig::PermuteB)
{
std::cout << "Permute for this DataType is not implemented." << std::endl;
return false;
}
b_k_n_dev_buf.ToDevice(b_k_n.data());
}
}
a_m_k_dev_buf.ToDevice(a_m_k.data());
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
float ave_time = invoke_gemm<GemmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout>(a_m_k_dev_buf,
b_k_n_dev_buf,
c_m_n_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
n_warmup,
n_repeat,
persistent,
flush_cache,
rotating_count);
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte =
sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Run Gemm kernel with M=" << M << " N=" << N << " K=" << K
<< " StrideA=" << stride_A << " StrideB=" << stride_B << " StrideC=" << stride_C
<< " A_Layout=" << ALayout::name << " B_Layout =" << BLayout::name
<< " C_Layout=" << CLayout::name << " A_Type=" << DataTypeTraits<ADataType>::name
<< " B_Type=" << DataTypeTraits<BDataType>::name
<< " C_Type=" << DataTypeTraits<CDataType>::name
<< " StructuredSparsity=" << (GemmConfig::UseStructuredSparsity ? "on" : "off")
<< " Persistent=" << (persistent ? "on" : "off") << " : " << ave_time << " ms, "
<< tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
bool pass = true;
// memory on host to store gpu reference result
ck_tile::HostTensor<CDataType> c_m_n_ref(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
c_m_n_ref.SetZero();
if(arg_parser.get_int("v") == 1)
{
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k, b_k_n, c_m_n_ref);
const float max_accumulated_value =
*std::max_element(c_m_n_ref.mData.begin(), c_m_n_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, kbatch, max_accumulated_value);
pass = do_verify(c_m_n_dev_result, c_m_n_ref, rtol_atol, "CPU");
}
else if(arg_parser.get_int("v") == 2)
{
if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
{
// Restore input for B for gpu reference
b_k_n_dev_buf.ToDevice(b_k_n.data());
}
if constexpr(GemmConfig::Preshuffle)
{
b_k_n_dev_buf.ToDevice(b_k_n.data());
}
// memory on device to store gpu reference result
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_ref.get_element_space_size_in_bytes());
c_m_n_gpu_buf_ref.SetZero();
ADataType* d_A = static_cast<ADataType*>(a_m_k_dev_buf.GetDeviceBuffer());
BDataType* d_B = static_cast<BDataType*>(b_k_n_dev_buf.GetDeviceBuffer());
CDataType* d_C = static_cast<CDataType*>(c_m_n_gpu_buf_ref.GetDeviceBuffer());
ck_tile::reference_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
c_m_n_gpu_buf_ref.FromDevice(c_m_n_ref.data());
const float max_accumulated_value =
*std::max_element(c_m_n_ref.mData.begin(), c_m_n_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, kbatch, max_accumulated_value);
pass = do_verify(c_m_n_dev_result, c_m_n_ref, rtol_atol, "GPU");
}
if(arg_parser.get_int("json") == 1)
{
dump_gemm_json_results<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
GemmConfig,
DataTypeTraits>(arg_parser.get_str("jsonfile"),
M,
N,
K,
stride_A,
stride_B,
stride_C,
persistent,
pass,
ave_time,
tflops,
gb_per_sec);
}
return pass;
}
Reference in New Issue View Git Blame Copy Permalink