ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Update flatmm related kernels (#3022)

Browse Source 
---------

Co-authored-by: Ding, Yi <yi.ding@amd.com>
Co-authored-by: felix <felix.li@amd.com>
This commit is contained in:
lalala-sh
2025-10-22 22:36:11 +08:00
committed by GitHub
parent cbd1279ae6
commit 211d64e18a
39 changed files with 11183 additions and 739 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
example/ck_tile/18_flatmm/CMakeLists.txt
View File

@@ -1,6 +1,32 @@
add_executable(tile_example_flatmm_basic EXCLUDE_FROM_ALL flatmm_basic.cpp)
set(SUPPORTED_GPUS gfx908 gfx90a gfx942 gfx950)
set(has_supported_gpu FALSE)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST SUPPORTED_GPUS)
set(has_supported_gpu TRUE)
break()
endif()
endforeach()
if(has_supported_gpu)
add_executable(tile_example_flatmm_basic EXCLUDE_FROM_ALL flatmm_basic.cpp)
add_executable(tile_example_mixed_prec_flatmm EXCLUDE_FROM_ALL mixed_prec/mixed_prec_flatmm.cpp)
add_executable(tile_example_moe_flatmm EXCLUDE_FROM_ALL moe_flatmm.cpp)
add_executable(tile_example_a16w4_moe_flatmm EXCLUDE_FROM_ALL mixed_prec/a16w4_moe_flatmm.cpp)
add_executable(tile_example_grouped_flatmm EXCLUDE_FROM_ALL grouped_flatmm.cpp)
set(EXAMPLE_FLATMM_COMPILE_OPTIONS)
set(EXAMPLE_MOE_FLATMM_COMPILE_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
endif()
target_compile_options(tile_example_flatmm_basic PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_mixed_prec_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_moe_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_a16w4_moe_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_grouped_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
endif()
set(EXAMPLE_FLATMM_COMPILE_OPTIONS)
# list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
# list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -Wno-unused-variable -Wno-unused-parameter)
target_compile_options(tile_example_flatmm_basic PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})

334
example/ck_tile/18_flatmm/flatmm_basic.cpp
View File

@@ -11,7 +11,102 @@
#include "ck_tile/host.hpp"
#include "flatmm_basic.hpp"
#include "run_flatmm_example.inc"
#include <type_traits>
template <typename T>
constexpr const char* DataTypeToString()
{
if constexpr(std::is_same_v<T, ck_tile::half_t>)
{
return "fp16";
}
else if constexpr(std::is_same_v<T, ck_tile::fp8_t>)
{
return "fp8";
}
else if constexpr(std::is_same_v<T, ck_tile::bf8_t>)
{
return "bf8";
}
else if constexpr(std::is_same_v<T, ck_tile::bf16_t>)
{
return "bf16";
}
else
{
return "unknown";
}
}
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>>{};
}
// mfma_type, 0:32x32, 1:16x16
template <typename FlatmmConfig, 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 MaxVecSize = 16 / sizeof(T);
constexpr int KLane = ck_tile::get_warp_size() / FlatmmConfig::N_Warp_Tile;
constexpr int ItemsPerAccess = std::min(MaxVecSize, FlatmmConfig::K_Warp_Tile / KLane);
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
k_ / ItemsPerAccess,
ItemsPerAccess});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 1, 3});
}
template <typename FlatmmConfig, typename T>
auto shuffle_b_v1(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 MaxVecSize = 16 / sizeof(T);
constexpr int KLane = ck_tile::get_warp_size() / FlatmmConfig::N_Warp_Tile;
constexpr int ItemsPerAccess = std::min(MaxVecSize, FlatmmConfig::K_Warp_Tile / KLane);
constexpr int NRepeat = FlatmmConfig::N_Tile / FlatmmConfig::N_Warp_Tile / FlatmmConfig::N_Warp;
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Tile,
FlatmmConfig::N_Warp,
FlatmmConfig::N_Warp_Tile,
NRepeat,
k_ / ItemsPerAccess,
ItemsPerAccess});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 3, 1, 4, 2, 5});
}
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 FlatmmConfig,
typename ADataType,
@@ -23,9 +118,12 @@ template <typename FlatmmConfig,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ScaleM,
typename ScaleN,
bool persistent,
typename CDEElementWise>
float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_config& s)
float flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
const ck_tile::stream_config& s)
{
using CodegenFlatmmShape = ck_tile::TileGemmShape<
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
@@ -80,14 +178,14 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
constexpr auto scheduler = FlatmmConfig::Scheduler;
constexpr auto memory_operation = memory_operation_.value;
using CodegenPipelineProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using CodegenPipelineProblem = ck_tile::FlatmmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using CodegenFlatmmPipeline =
ck_tile::FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
@@ -110,7 +208,10 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
FlatmmConfig::K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation,
FlatmmConfig::NumWaveGroups>>;
FlatmmConfig::NumWaveGroups,
false,
1,
FlatmmConfig::TiledMMAPermuteN>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
@@ -118,8 +219,8 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -167,40 +268,145 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
hipGetErrorString(hipMemsetAsync(
args.e_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
};
return ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
return ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
{
return Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
return Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
};
return ave_time = BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
return ave_time;
}
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout,
typename ScaleM,
typename ScaleN,
bool UsePersistentKernel = false,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
ck_tile::DeviceMem& b_shuffle_dev_buf,
ck_tile::DeviceMem& c_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,
ScaleM scale_m,
ScaleN scale_n,
int n_warmup,
int n_repeat)
{
ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN> args = {a_dev_buf.GetDeviceBuffer(),
b_shuffle_dev_buf.GetDeviceBuffer(),
{},
c_dev_buf.GetDeviceBuffer(),
kbatch,
M,
N,
K,
stride_A,
stride_B,
{},
stride_C,
scale_m,
scale_n};
float ave_time = flatmm_calc<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
ScaleM,
ScaleN,
UsePersistentKernel,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
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 Flatmm kernel with DataType = " << DataTypeToString<ADataType>()
<< " M =" << M << " N =" << N << " K =" << K << " StrideA =" << stride_A
<< " StrideB =" << stride_B << " StrideC =" << stride_C << " : " << ave_time
<< " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
return ave_time;
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "256", "m dimension")
.insert("n", "256", "n dimension")
.insert("k", "128", "k dimension")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "C", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("prec", "fp8", "data type. fp16/bf16/fp8/bf8")
.insert("wave_tile", "16", "only support 16(16x16) or 32(32x32)")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "splitK value")
.insert("init", "0", "0:random, 1:linear, 2:constant(1)")
.insert("scale", "0", "0:without scale, 1:per-token/channel scale, only for fp8/bf8")
.insert("persistent", "0", "0: no persistent, 1: persistent kernel")
.insert("warp_tile",
"0",
"0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
#include "run_flatmm_example.inc"
template <template <typename PreType> typename FlatmmConfig>
int run_flatmm_example(int argc, char* argv[])
{
@@ -214,20 +420,10 @@ int run_flatmm_example(int argc, char* argv[])
std::string data_type = arg_parser.get_str("prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
int k = arg_parser.get_int("k");
int stride_b = arg_parser.get_int("stride_b");
if(b_layout == "C" && stride_b > k)
{
throw std::runtime_error(
"For ColumnMajor layout, StrideB must be smaller than or equal to K (" +
std::to_string(k) + ")");
}
int scale_opt = arg_parser.get_int("scale");
int persistent_opt = arg_parser.get_int("persistent");
if(a_layout == "R" && b_layout == "C")
{
if(data_type == "fp16")
{
run_flatmm_example_with_layouts<ck_tile::half_t, FlatmmConfig<ck_tile::half_t>>(
@@ -240,13 +436,53 @@ int run_flatmm_example(int argc, char* argv[])
}
else if(data_type == "fp8")
{
run_flatmm_example_with_layouts<ck_tile::fp8_t, FlatmmConfig<ck_tile::fp8_t>>(
argc, argv, Row{}, Col{}, Row{});
if(scale_opt == 0)
{
if(persistent_opt == 0)
{
run_flatmm_example_with_layouts<ck_tile::fp8_t, FlatmmConfig<ck_tile::fp8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
run_flatmm_example_with_layouts<ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>,
-1,
-1,
true>(argc, argv, Row{}, Col{}, Row{});
}
}
else
{
if(persistent_opt == 0)
{
run_flatmm_example_with_layouts<ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>,
1,
1>(argc, argv, Row{}, Col{}, Row{});
}
else
{
run_flatmm_example_with_layouts<ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>,
1,
1,
true>(argc, argv, Row{}, Col{}, Row{});
}
}
}
else if(data_type == "bf8")
{
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>>(
argc, argv, Row{}, Col{}, Row{});
if(scale_opt == 0)
{
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>, 1, 1>(
argc, argv, Row{}, Col{}, Row{});
}
}
else
{
@@ -268,9 +504,6 @@ int main(int argc, char* argv[])
try
{
#if defined(CK_TILE_USE_WMMA)
return !run_flatmm_example<FlatmmConfig16_Wmma>(argc, argv);
#else
int warp_tile = arg_parser.get_int("warp_tile");
if(warp_tile == 0)
{
@@ -288,7 +521,6 @@ int main(int argc, char* argv[])
{
return !run_flatmm_example<FlatmmConfig32_950>(argc, argv);
}
#endif
}
catch(const std::runtime_error& e)
{

66
example/ck_tile/18_flatmm/flatmm_basic.hpp
View File

@@ -35,12 +35,13 @@ struct FlatmmConfig32
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 2;
static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool DoubleSmemBuffer = false;
static constexpr bool TiledMMAPermuteN = false; // disable PermuteN when NWarpTile != 16
};
template <typename DataType>
@@ -72,26 +73,28 @@ struct FlatmmConfig16
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 2;
static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool DoubleSmemBuffer = false;
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
static constexpr bool TiledMMAPermuteN = N_Repeat % 4 == 0;
};
template <typename DataType>
struct FlatmmConfig16_950 : public FlatmmConfig16<DataType>
{
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(DataType);
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 32 : 128;
};
static constexpr int kBlockPerCu = 1;
template <typename DataType>
struct FlatmmConfig16_Wmma : public FlatmmConfig16<DataType>
{
static constexpr ck_tile::index_t M_Tile = 64;
static constexpr ck_tile::index_t K_Tile = 64;
static constexpr ck_tile::index_t K_Warp_Tile = 16;
static constexpr int N_Repeat =
N_Tile / FlatmmConfig16<DataType>::N_Warp_Tile / FlatmmConfig16<DataType>::N_Warp;
static constexpr bool TiledMMAPermuteN = N_Repeat % 4 == 0;
};
template <typename ADataType>
@@ -172,42 +175,19 @@ struct is_8bit_type
{
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "256", "m dimension")
.insert("n", "256", "n dimension")
.insert("k", "128", "k dimension")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "C", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "splitK value")
.insert("init", "0", "0:random, 1:linear, 2:constant(1)")
#if !defined(CK_TILE_USE_WMMA)
.insert(
"warp_tile", "0", "0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)")
#endif
.insert("json", "0", "0: No Json, 1: Dump Results in Json format")
.insert("jsonfile", "flatmm_basic.json", "json file name to dump results");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
// host API
template <typename ADataType,
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename FlatmmConfig,
typename ALayout,
typename BLayout,
typename CLayout>
float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_config& s);
typename DsLayout,
typename ELayout,
typename ScaleM,
typename ScaleN,
bool persistent,
typename CDEElementWise>
float flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
const ck_tile::stream_config& s);

364
example/ck_tile/18_flatmm/grouped_flatmm.cpp Normal file
View File

@@ -0,0 +1,364 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include "flatmm_basic.hpp"
#include "ck_tile/host.hpp"
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>>{};
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("Ms", "1,1,1", "m dimension")
.insert("Ns", "5120,5120,5120", "n dimension")
.insert("Ks", "6144,6144,6144", "k dimension")
.insert("group_count", "3", "group count")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "C", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("prec", "fp8", "data type. fp16/bf16/fp8/bf8")
.insert("mode",
"masked",
"grouped gemm mode: [general | contiguous | masked], general by default")
.insert("wave_tile", "16", "only support 16(16x16) or 32(32x32)")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "splitK value")
.insert("init", "0", "0:random, 1:linear, 2:constant(1)")
.insert("scale", "0", "0:without scale, 1:per-token/channel scale, only for fp8/bf8")
.insert("warp_tile",
"0",
"0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
bool persistent,
typename CDEElementWise,
typename KernelArguments>
float grouped_flatmm(const KernelArguments& args, const ck_tile::stream_config& s)
{
using CodegenFlatmmShape = ck_tile::TileGemmShape<
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
ck_tile::sequence<FlatmmConfig::M_Warp, FlatmmConfig::N_Warp, FlatmmConfig::K_Warp>,
ck_tile::sequence<FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile>>;
using TilePartitioner =
ck_tile::GemmSpatiallyLocalTilePartitioner<CodegenFlatmmShape,
FlatmmConfig::TileParitionerGroupNum,
FlatmmConfig::TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
ALayout,
BLayout,
ELayout,
FlatmmConfig::NumWaveGroups>;
using CodegenGemmTraits = ck_tile::TileGemmUniversalTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
FlatmmConfig::DoubleSmemBuffer,
ALayout,
BLayout,
ELayout,
FlatmmConfig::TransposeC,
FlatmmConfig::UseStructuredSparsity,
persistent,
FlatmmConfig::NumWaveGroups,
true>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenFlatmmShape, Traits>;
using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_,
const auto tail_number_,
const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = FlatmmConfig::Scheduler;
constexpr auto memory_operation = memory_operation_.value;
using CodegenPipelineProblem = ck_tile::FlatmmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using CodegenFlatmmPipeline =
ck_tile::FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
DsLayout,
ELayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
FlatmmConfig::M_Warp,
FlatmmConfig::N_Warp,
FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation,
FlatmmConfig::NumWaveGroups>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel =
ck_tile::GroupedFlatmmKernel<TilePartitioner, CodegenFlatmmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.flush_cache_)
{
std::cout << "Flushing cache..." << std::endl;
static constexpr ck_tile::index_t APackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
static constexpr ck_tile::index_t BPackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
args.group_count * args.M, args.K, args.stride_A, is_row_major(ALayout{})));
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.group_count * args.N, args.stride_B, is_row_major(BLayout{})));
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
kargs.a_ptr, kargs.b_shuffle_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck_tile::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(args.k_batch > 1)
hipGetErrorString(
hipMemsetAsync(args.e_ptr,
0,
args.group_count * args.M * args.N * sizeof(CDataType),
s.stream_id_));
};
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
};
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
return ave_time;
}
#include "run_grouped_flatmm_example.inc"
template <template <typename PreType> typename FlatmmConfig>
int run_grouped_flatmm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string data_type = arg_parser.get_str("prec");
std::string mode = arg_parser.get_str("mode");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "C")
{
if(mode == "contiguous")
{
if(data_type == "fp16")
{
run_contiguous_grouped_flatmm_example_with_layouts<ck_tile::half_t,
FlatmmConfig<ck_tile::half_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
run_contiguous_grouped_flatmm_example_with_layouts<ck_tile::bf16_t,
FlatmmConfig<ck_tile::bf16_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
run_contiguous_grouped_flatmm_example_with_layouts<ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
run_contiguous_grouped_flatmm_example_with_layouts<ck_tile::bf8_t,
FlatmmConfig<ck_tile::bf8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else if(mode == "masked")
{
if(data_type == "fp16")
{
run_masked_grouped_flatmm_example_with_layouts<ck_tile::half_t,
FlatmmConfig<ck_tile::half_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
run_masked_grouped_flatmm_example_with_layouts<ck_tile::bf16_t,
FlatmmConfig<ck_tile::bf16_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
run_masked_grouped_flatmm_example_with_layouts<ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
run_masked_grouped_flatmm_example_with_layouts<ck_tile::bf8_t,
FlatmmConfig<ck_tile::bf8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else
{
throw std::runtime_error("Unsupported mode!");
}
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
return -1;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return EXIT_FAILURE;
try
{
int warp_tile = arg_parser.get_int("warp_tile");
if(warp_tile == 0)
{
return !run_grouped_flatmm_example<FlatmmConfig16>(argc, argv);
}
// else if(warp_tile == 1)
// {
// return !run_grouped_flatmm_example<FlatmmConfig32>(argc, argv);
// }
// else if(warp_tile == 2)
// {
// return !run_grouped_flatmm_example<FlatmmConfig16_950>(argc, argv);
// }
// else
// {
// return !run_grouped_flatmm_example<FlatmmConfig32_950>(argc, argv);
// }
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

50
example/ck_tile/18_flatmm/mixed_prec/a16w4_flatmm.hpp Normal file
View File

@@ -0,0 +1,50 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
// GEMM config with 16x16 warp tile
struct A16W4_FlatmmConfig16
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 256;
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = 32;
static constexpr bool kPadM = false;
static constexpr bool kPadN = false;
static constexpr bool kPadK = false;
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool DoubleSmemBuffer = false;
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
static constexpr bool TiledMMAPermuteN = false;
};
struct A16W4_FlatmmConfig16_950 : public A16W4_FlatmmConfig16
{
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr int kBlockPerCu = 1;
static constexpr int N_Repeat =
N_Tile / A16W4_FlatmmConfig16::N_Warp_Tile / A16W4_FlatmmConfig16::N_Warp;
static constexpr bool TiledMMAPermuteN = false;
};

511
example/ck_tile/18_flatmm/mixed_prec/a16w4_moe_flatmm.cpp Normal file
View File

@@ -0,0 +1,511 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include <memory>
#include "a16w4_moe_flatmm.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/ops/flatmm.hpp"
#include "ck_tile/ops/moe_flatmm.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/host/reference/reference_moe_gemm.hpp"
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>>{};
}
// gemm1
// operand-A = [num_token, d_model]
// operand-B = [num_expert, hidden, d_model]
// operand-C = [num_token, topk, hidden]
// gemm2
// operand-A = [num_token, topk, hidden]
// operand-B = [num_expert, d_model, hidden]
// operand-C = [num_token, d_model]
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ck_tile::MoeFlatmmKind moe_kind = ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_only,
typename CDEElementWise = ck_tile::element_wise::PassThrough,
typename MoeFlatmmHostArgs>
float a16w4_moe_gemm(const MoeFlatmmHostArgs& args, const ck_tile::stream_config& s)
{
using CodegenFlatmmShape = ck_tile::TileGemmShape<
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
ck_tile::sequence<FlatmmConfig::M_Warp, FlatmmConfig::N_Warp, FlatmmConfig::K_Warp>,
ck_tile::sequence<FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile>>;
using TilePartitioner =
ck_tile::GemmSpatiallyLocalTilePartitioner<CodegenFlatmmShape,
FlatmmConfig::TileParitionerGroupNum,
FlatmmConfig::TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
ALayout,
BLayout,
ELayout,
FlatmmConfig::NumWaveGroups>;
using CodegenGemmTraits = ck_tile::TileGemmUniversalTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
FlatmmConfig::DoubleSmemBuffer,
ALayout,
BLayout,
ELayout,
FlatmmConfig::TransposeC,
FlatmmConfig::UseStructuredSparsity,
false, // UsePersistentKernel_
FlatmmConfig::NumWaveGroups,
true>; // Preshuffle_
constexpr bool MXFP4_Pipeline = std::is_same_v<BDataType, ck_tile::pk_fp4_t>;
if constexpr(!MXFP4_Pipeline && moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up)
{
static_assert(
FlatmmConfig::N_Tile % (FlatmmConfig::N_Warp * FlatmmConfig::N_Warp_Tile * 2) == 0,
"requires NRepeat is multiple of 2 for FFN_gemm1_gate_up");
}
using ComputeDataType = ADataType;
static_assert(sizeof(ComputeDataType) >= sizeof(BDataType),
"mixed_prec_flatmm requires ADataType is a wider type than BDataType");
using GemmPipelineProblem = ck_tile::GemmPipelineProblem<ComputeDataType,
ComputeDataType,
AccDataType,
CodegenFlatmmShape,
Traits>;
using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_,
const auto tail_number_,
const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = FlatmmConfig::Scheduler;
constexpr auto memory_operation = memory_operation_.value;
using CodegenPipelineProblem =
std::conditional_t<MXFP4_Pipeline,
ck_tile::F16xMXF4FlatmmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>,
ck_tile::FlatmmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>>;
constexpr int BlockedXDLN_PerWarp = 2; // determined by scale shuffle pattern
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ComputeDataType,
ComputeDataType,
DsDatatype,
AccDataType,
CDataType,
DsLayout,
ELayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
FlatmmConfig::M_Warp,
FlatmmConfig::N_Warp,
FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation,
FlatmmConfig::NumWaveGroups,
false,
1,
FlatmmConfig::TiledMMAPermuteN,
BlockedXDLN_PerWarp>>;
using CodegenFlatmmPipeline = std::conditional_t<
MXFP4_Pipeline,
ck_tile::F16xMXF4FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>,
ck_tile::MoeFlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>>;
using FusedAct =
std::conditional_t<MXFP4_Pipeline, ck_tile::moe::Swiglu, ck_tile::moe::MoeSilu>;
using Kernel = ck_tile::MoeFlatmmKernel<TilePartitioner,
CodegenFlatmmPipeline,
GemmEpilogue,
moe_kind,
FusedAct>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:" << CodegenFlatmmShape::GetName() << "\n"
<< "Shape: " << CodegenFlatmmShape::GetName() << "\n"
<< "problem: " << CodegenPipelineProblem::GetName() << "\n"
<< "pipeline: " << CodegenFlatmmPipeline::GetName() << "\n"
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
if(s.flush_cache_)
{
std::cout << "Flushing cache..." << std::endl;
static constexpr ck_tile::index_t APackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ||
std::is_same_v<BDataType, ck_tile::pk_fp4_t>
? 2
: 1;
static constexpr ck_tile::index_t BPackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ||
std::is_same_v<BDataType, ck_tile::pk_fp4_t>
? 2
: 1;
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm2 ? args.NumTokens * args.TopK
: args.NumTokens,
args.K,
args.stride_A,
is_row_major(ALayout{})));
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.N * args.NumExperts, args.stride_B, is_row_major(BLayout{})));
const int outputN =
moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up ? args.N / 2 : args.N;
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
kargs.a_ptr, kargs.b_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck_tile::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm2)
hipGetErrorString(hipMemsetAsync(
args.e_ptr, 0, args.NumTokens * args.N * sizeof(CDataType), s.stream_id_));
else if(args.k_batch > 1)
hipGetErrorString(
hipMemsetAsync(args.e_ptr,
0,
args.NumTokens * args.TopK * outputN * sizeof(CDataType),
s.stream_id_));
};
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
};
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
return ave_time;
}
template <class FlatmmConfig, ck_tile::MoeFlatmmKind moe_kind, class IterSrc, class IterDst>
void shuffle_mxfp4_weight(const IterSrc src, IterDst dst, int experts_cnt, int N, int K)
{
int KPack = 16;
int NLane = FlatmmConfig::N_Warp_Tile;
int KLane = 64 / NLane;
int K_pk = K / 2;
int K0 = K_pk / (KLane * KPack);
// K -> K0 KLane KPack
// N -> N0 NLane
// N, K -> N0 K0 KLane NLane KPack
int tempk;
if constexpr(moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up)
{
int up_stride = N / 2 / NLane;
for(long eid = 0; eid < experts_cnt; ++eid)
{
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K_pk; ++k)
{
int n0 = n / NLane;
int n1 = n % NLane;
// interleave gate and up part with granularity is 16.
int n0_interleave = n >= N / 2 ? (n0 - up_stride) * 2 + 1 : // up part
n0 * 2; // gate part
int k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
int k1 = tempk / KPack;
int k2 = tempk % KPack;
long outputIndex = eid * N * K_pk + n0_interleave * KPack * NLane * KLane * K0 +
k0 * KPack * NLane * KLane + k1 * KPack * NLane +
n1 * KPack + k2;
dst[outputIndex] = src[eid * N * K_pk + n * K_pk + k];
}
}
}
}
else
{
for(long eid = 0; eid < experts_cnt; ++eid)
{
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K_pk; ++k)
{
int n0 = n / NLane;
int n1 = n % NLane;
int k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
int k1 = tempk / KPack;
int k2 = tempk % KPack;
long outputIndex = eid * N * K_pk + n0 * KPack * NLane * KLane * K0 +
k0 * KPack * NLane * KLane + k1 * KPack * NLane +
n1 * KPack + k2;
dst[outputIndex] = src[eid * N * K_pk + n * K_pk + k];
}
}
}
}
}
template <typename FlatmmConfig, ck_tile::MoeFlatmmKind moe_kind, typename T>
auto shuffle_mxfp4_scale(const ck_tile::HostTensor<T>& scale, int experts_cnt)
{
assert(scale.get_lengths().size() == 2);
int n_ = scale.get_lengths()[1];
int k_ = scale.get_lengths()[0];
int k_per_expert = k_ / experts_cnt;
constexpr int K_Pack = 2; // fixed for mxfp4
constexpr int N_Pack = 2; // fixed for mxfp4
constexpr int GranularityK = 32; // fixed for mxfp4
constexpr int K_Lane = 64 / FlatmmConfig::N_Warp_Tile; // 4
static_assert(FlatmmConfig::N_Warp_Tile == 16, "only support XDL_N == 16");
static_assert(FlatmmConfig::N_Repeat % N_Pack == 0);
static_assert(FlatmmConfig::K_Tile % (K_Pack * K_Lane * GranularityK) == 0);
if constexpr(moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up)
{
ck_tile::HostTensor<T> shfl_scale({
experts_cnt,
k_per_expert / K_Pack / K_Lane,
K_Pack,
K_Lane,
N_Pack, // N_Pack = 2 is composed of Gate + Up.
n_ / FlatmmConfig::N_Warp_Tile / N_Pack,
FlatmmConfig::N_Warp_Tile,
});
std::copy(scale.begin(), scale.end(), shfl_scale.begin());
return ck_tile::reference_permute(shfl_scale, {0, 5, 1, 3, 6, 2, 4});
}
else
{
ck_tile::HostTensor<T> shfl_scale({
experts_cnt,
k_per_expert / K_Pack / K_Lane,
K_Pack,
K_Lane,
n_ / FlatmmConfig::N_Warp_Tile / N_Pack,
N_Pack,
FlatmmConfig::N_Warp_Tile,
});
std::copy(scale.begin(), scale.end(), shfl_scale.begin());
return ck_tile::reference_permute(shfl_scale, {0, 4, 1, 3, 6, 2, 5});
}
}
#include "run_a16w4_moe_flatmm_example.inc"
template <typename FlatmmConfig>
int run_a16w4_moe_flatmm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
{
return -1;
}
const std::string a_layout = arg_parser.get_str("a_layout");
const std::string b_layout = arg_parser.get_str("b_layout");
const std::string mixed_prec = arg_parser.get_str("mixed_prec");
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
if(a_layout == "R" && b_layout == "C")
{
const std::string gemm_kind = arg_parser.get_str("gemm_kind");
if(gemm_kind == "gemm1_gate_up")
{
if(mixed_prec == "fp16xfp4")
{
return run_a16w4_moe_gemm_example_with_layouts<
ck_tile::half_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up>(argc, argv, Row{}, Col{}, Row{});
}
else if(mixed_prec == "bf16xfp4")
{
return run_a16w4_moe_gemm_example_with_layouts<
ck_tile::bfloat16_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported precision type for gemm1_gate_up!");
}
}
else if(gemm_kind == "gemm2")
{
if(mixed_prec == "fp16xfp4")
{
return run_a16w4_moe_gemm_example_with_layouts<ck_tile::half_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
ck_tile::MoeFlatmmKind::kFFN_gemm2>(
argc, argv, Row{}, Col{}, Row{});
}
else if(mixed_prec == "bf16xfp4")
{
return run_a16w4_moe_gemm_example_with_layouts<ck_tile::bfloat16_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
ck_tile::MoeFlatmmKind::kFFN_gemm2>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported precision type for gemm2!");
}
}
else
{
throw std::runtime_error("Unrecoginized gemm_kind parameter, only accept value "
"[gemm1_gate_up | gemm2]");
}
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
return -1;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return EXIT_FAILURE;
try
{
int warp_tile = arg_parser.get_int("warp_tile");
if(warp_tile == 0)
{
return !run_a16w4_moe_flatmm_example<A16W4_FlatmmConfig16>(argc, argv);
}
// else if(warp_tile == 1)
// {
// return !run_a16w4_moe_flatmm_example<A16W4_FlatmmConfig16_950>(argc, argv);
// }
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

87
example/ck_tile/18_flatmm/mixed_prec/a16w4_moe_flatmm.hpp Normal file
View File

@@ -0,0 +1,87 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include <tuple>
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/moe_flatmm.hpp"
// GEMM config with 16x16 warp tile
struct A16W4_FlatmmConfig16
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 256;
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = 32;
static constexpr bool kPadM = false;
static constexpr bool kPadN = false;
static constexpr bool kPadK = false;
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool DoubleSmemBuffer = false;
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
static constexpr bool TiledMMAPermuteN = false;
};
struct A16W4_FlatmmConfig16_950 : public A16W4_FlatmmConfig16
{
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr int kBlockPerCu = 1;
static constexpr int N_Repeat =
N_Tile / A16W4_FlatmmConfig16::N_Warp_Tile / A16W4_FlatmmConfig16::N_Warp;
static constexpr bool TiledMMAPermuteN = false;
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("experts", "8", "Num of experts - 8 by default")
.insert("NumTokens", "128", "M dimensions - 128 by default.")
.insert("TopK", "3", "Top K - 3 by default.")
.insert("N", "4096", "N dimensions - 4096 by default.")
.insert("K", "4096", "K dimensions - 4096 by default.")
.insert("stride_A", "", "Tensor A strides - it is empty by default.")
.insert("stride_B", "", "Tensor B strides - it is empty by default.")
.insert("stride_C", "", "Tensor C strides - it is empty by default.")
.insert("a_layout", "R", "A tensor data layout - Row by default.")
.insert("b_layout", "C", "B tensor data layout - Col by default.")
.insert("c_layout", "R", "C tensor data layout - Row by default.")
.insert("gemm_kind",
"gemm1_gate_up",
"Gemm kind in FFN network [gemm1_gate_up | gemm2] - "
"gemm1_gate_up by default.")
.insert("validate", "1", "0. No validation, 1. Validation on CPU.")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("mixed_prec",
"bf16xfp4",
"data type for activation and weight, support: bf16xfp4, fp16xfp4")
.insert("init", "0", "0:random, 1:constant(1)")
.insert("warp_tile",
"0",
"0: 16x16, 1: 16x16 (950 only, may use a larger tile than warp_tile=0)")
.insert("repeat", "10", "number of iterations to benchmark the kernel.");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}

482
example/ck_tile/18_flatmm/mixed_prec/mixed_prec_flatmm.cpp Normal file
View File

@@ -0,0 +1,482 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include <type_traits>
#include "ck_tile/host.hpp"
#include "mixed_prec_flatmm.hpp"
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 FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ScaleM,
typename ScaleN,
bool persistent,
typename CDEElementWise>
float mixed_prec_flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
const ck_tile::stream_config& s)
{
using CodegenFlatmmShape = ck_tile::TileGemmShape<
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
ck_tile::sequence<FlatmmConfig::M_Warp, FlatmmConfig::N_Warp, FlatmmConfig::K_Warp>,
ck_tile::sequence<FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile>>;
using TilePartitioner =
ck_tile::GemmSpatiallyLocalTilePartitioner<CodegenFlatmmShape,
FlatmmConfig::TileParitionerGroupNum,
FlatmmConfig::TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
ALayout,
BLayout,
ELayout,
FlatmmConfig::NumWaveGroups>;
using CodegenGemmTraits = ck_tile::TileGemmUniversalTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
FlatmmConfig::DoubleSmemBuffer,
ALayout,
BLayout,
ELayout,
FlatmmConfig::TransposeC,
FlatmmConfig::UseStructuredSparsity,
persistent,
FlatmmConfig::NumWaveGroups,
true>;
using ComputeDataType = ADataType;
static_assert(sizeof(ComputeDataType) >= sizeof(BDataType),
"mixed_prec_flatmm requires ADataType is a wider type than BDataType");
using GemmPipelineProblem = ck_tile::GemmPipelineProblem<ComputeDataType,
ComputeDataType,
AccDataType,
CodegenFlatmmShape,
Traits>;
using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_,
const auto tail_number_,
const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = FlatmmConfig::Scheduler;
constexpr auto memory_operation = memory_operation_.value;
constexpr int BlockedXDLN_PerWarp = 2; // determined by scale shuffle pattern
using CodegenPipelineProblem = ck_tile::F16xMXF4FlatmmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using CodegenFlatmmPipeline =
ck_tile::F16xMXF4FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ComputeDataType,
ComputeDataType,
DsDatatype,
AccDataType,
CDataType,
DsLayout,
ELayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
FlatmmConfig::M_Warp,
FlatmmConfig::N_Warp,
FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation,
FlatmmConfig::NumWaveGroups,
false, // FixedVectorSize
1, // VectorSizeC
FlatmmConfig::TiledMMAPermuteN,
BlockedXDLN_PerWarp>>;
using Kernel =
ck_tile::F16xMXF4FlatmmKernel<TilePartitioner, CodegenFlatmmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:" << CodegenFlatmmShape::GetName() << "\n"
<< "Shape: " << CodegenFlatmmShape::GetName() << "\n"
<< "problem: " << CodegenPipelineProblem::GetName() << "\n"
<< "pipeline: " << CodegenFlatmmPipeline::GetName() << "\n"
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
if(s.flush_cache_)
{
std::cout << "Flushing cache..." << std::endl;
constexpr ck_tile::index_t APackedSize = ck_tile::numeric_traits<ADataType>::PackedSize;
constexpr ck_tile::index_t BPackedSize = ck_tile::numeric_traits<BDataType>::PackedSize;
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
kargs.a_ptr, kargs.b_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck_tile::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(args.k_batch > 1)
hipGetErrorString(hipMemsetAsync(
args.e_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
};
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
};
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
return ave_time;
}
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout,
typename ScaleN,
bool UsePersistentKernel = false,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float invoke_mixed_prec_flatmm(ck_tile::DeviceMem& a_dev_buf,
ck_tile::DeviceMem& b_shuffle_dev_buf,
ck_tile::DeviceMem& c_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,
ScaleN dequant_scale_n,
int n_warmup,
int n_repeat)
{
// Activation has no scale
using ActScaleType = ck_tile::FlatmmScalePointer<-1>;
ck_tile::ScaleFlatmmHostArgs<ActScaleType, ScaleN> args = {a_dev_buf.GetDeviceBuffer(),
b_shuffle_dev_buf.GetDeviceBuffer(),
{},
c_dev_buf.GetDeviceBuffer(),
kbatch,
M,
N,
K,
stride_A,
stride_B,
{},
stride_C,
{},
dequant_scale_n};
float ave_time = mixed_prec_flatmm_calc<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
ActScaleType,
ScaleN,
UsePersistentKernel,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
constexpr int PackedSize = ck_tile::numeric_traits<BDataType>::PackedSize;
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * N * K / PackedSize +
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 A16W4_Flatmm kernel " << " M =" << M << " N =" << N << " K =" << K
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
<< " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
return ave_time;
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "256", "m dimension")
.insert("n", "256", "n dimension")
.insert("k", "512", "k dimension")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "C", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "1", "0. No validation, 1. Validation on GPU")
.insert("mixed_prec",
"bf16xfp4",
"data type for activation and weight, support: bf16xfp4, fp16xfp4")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "splitK value")
.insert("init", "0", "0:random, 1:constant(1)")
.insert("persistent", "0", "0: no persistent, 1: persistent kernel")
.insert("warp_tile",
"0",
"0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <class FlatmmConfig, class IterSrc, class IterDst>
void preShuffleWeight(const IterSrc src, IterDst dst, int N, int K)
{
int KPack = 16;
int NLane = FlatmmConfig::N_Warp_Tile;
int KLane = 64 / NLane;
int K_pk = K / 2;
int K0 = K_pk / (KLane * KPack);
// K -> K0 KLane KPack
// N -> N0 NLane
// N, K -> N0 K0 KLane NLane KPack
int tempk;
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K_pk; ++k)
{
int n0 = n / NLane;
int n1 = n % NLane;
int k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
int k1 = tempk / KPack;
int k2 = tempk % KPack;
int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
k1 * KPack * NLane + n1 * KPack + k2;
dst[outputIndex] = src[n * K_pk + k];
}
}
}
template <class FlatmmConfig, class T>
auto preShuffleScale(const ck_tile::HostTensor<T>& scale)
{
assert(scale.get_lengths().size() == 2);
int n_ = scale.get_lengths()[1];
int k_ = scale.get_lengths()[0];
constexpr int K_Pack = 2; // fixed for mxfp4
constexpr int N_Pack = 2; // fixed for mxfp4
constexpr int GranularityK = 32; // fixed for mxfp4
constexpr int K_Lane = 64 / FlatmmConfig::N_Warp_Tile; // 4
static_assert(FlatmmConfig::N_Warp_Tile == 16, "only support XDL_N == 16");
static_assert(FlatmmConfig::N_Repeat % N_Pack == 0);
static_assert(FlatmmConfig::K_Tile % (K_Pack * K_Lane * GranularityK) == 0);
ck_tile::HostTensor<T> shfl_scale({
k_ / K_Pack / K_Lane,
K_Pack,
K_Lane,
n_ / FlatmmConfig::N_Warp_Tile / N_Pack,
N_Pack,
FlatmmConfig::N_Warp_Tile,
});
std::copy(scale.begin(), scale.end(), shfl_scale.begin());
return ck_tile::reference_permute(shfl_scale, {3, 0, 2, 5, 1, 4});
}
#include "run_mixed_prec_flatmm.inc"
template <typename FlatmmConfig>
int run_mixed_prec_flatmm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string mixed_prec = arg_parser.get_str("mixed_prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
int persistent_opt = arg_parser.get_int("persistent");
if(a_layout == "R" && b_layout == "C")
{
if(mixed_prec == "bf16xfp4")
{
if(persistent_opt == 0)
{
run_mixed_prec_flatmm_with_layouts<ck_tile::bf16_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
false>(argc, argv, Row{}, Col{}, Row{});
}
else
{
run_mixed_prec_flatmm_with_layouts<ck_tile::bf16_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
true>(argc, argv, Row{}, Col{}, Row{});
}
}
else if(mixed_prec == "fp16xfp4")
{
if(persistent_opt == 0)
{
run_mixed_prec_flatmm_with_layouts<ck_tile::fp16_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
false>(argc, argv, Row{}, Col{}, Row{});
}
else
{
run_mixed_prec_flatmm_with_layouts<ck_tile::fp16_t,
ck_tile::pk_fp4_t,
FlatmmConfig,
true>(argc, argv, Row{}, Col{}, Row{});
}
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
return -1;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return EXIT_FAILURE;
try
{
int warp_tile = arg_parser.get_int("warp_tile");
if(warp_tile == 0)
{
return !run_mixed_prec_flatmm_example<A16W4_FlatmmConfig16>(argc, argv);
}
else if(warp_tile == 1)
{
return !run_mixed_prec_flatmm_example<A16W4_FlatmmConfig16_950>(argc, argv);
}
else
{
throw std::runtime_error("Unsupported warp_tile!");
}
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

15
example/ck_tile/18_flatmm/mixed_prec/mixed_prec_flatmm.hpp Normal file
View File

@@ -0,0 +1,15 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/flatmm.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "a16w4_flatmm.hpp"

353
example/ck_tile/18_flatmm/mixed_prec/run_a16w4_moe_flatmm_example.inc Normal file
View File

@@ -0,0 +1,353 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ck_tile::MoeFlatmmKind kind,
typename CDEElementWise = ck_tile::element_wise::PassThrough,
typename MoeHostArgs>
float invoke_a16w4_moe_gemm(int n_warmup, int n_repeat, const MoeHostArgs& args)
{
float ave_time = a16w4_moe_gemm<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
ELayout,
kind,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
std::string op_name{"Moe Gemm"};
constexpr int PackedSize = ck_tile::numeric_traits<BDataType>::PackedSize;
std::size_t flop = std::size_t(2) * args.M * args.N * args.K;
std::size_t num_byte = sizeof(ADataType) * args.M * args.K +
sizeof(BDataType) * args.N * args.K / PackedSize +
sizeof(CDataType) * args.M * args.N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
return ave_time;
}
template <typename PrecActType,
typename PrecWeightType,
typename FlatmmConfig,
ck_tile::MoeFlatmmKind kind,
typename ALayout,
typename BLayout,
typename CLayout>
int run_a16w4_moe_gemm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
{
return -1;
};
using ADataType = PrecActType;
using BDataType = PrecWeightType;
using CDataType = PrecActType;
using AccDataType = float;
using ScaleType = ck_tile::e8m0_t;
constexpr int ScaleGranularityN = 1;
constexpr int ScaleGranularityK = 32;
const ck_tile::index_t N = arg_parser.get_int("N");
const 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 init_method = arg_parser.get_int("init");
const ck_tile::index_t num_tokens = arg_parser.get_int("NumTokens");
const ck_tile::index_t topk = arg_parser.get_int("TopK");
const ck_tile::index_t warmup = arg_parser.get_int("warmup");
const ck_tile::index_t repeat = arg_parser.get_int("repeat");
const ck_tile::index_t experts = arg_parser.get_int("experts");
// TODO: replace the magic declaration
const ck_tile::index_t MPerBlock = FlatmmConfig::M_Tile;
ck_tile::index_t sorted_tile_num = (num_tokens + MPerBlock - 1) / MPerBlock * MPerBlock * topk;
ck_tile::index_t valid_tile_num = sorted_tile_num;
ck_tile::index_t sorted_size = sorted_tile_num * MPerBlock;
const ck_tile::index_t M = sorted_tile_num * MPerBlock;
const ck_tile::index_t outputN = kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up ? N / 2 : N;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
constexpr bool IsInputGemm = kind != ck_tile::MoeFlatmmKind::kFFN_gemm2;
stride_A = ck_tile::get_default_stride(
IsInputGemm ? num_tokens : num_tokens * topk, 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(
IsInputGemm ? num_tokens * topk : num_tokens, outputN, stride_C, is_row_major(CLayout{}));
auto a_m_k_tensor = ck_tile::HostTensor<ADataType>(ck_tile::host_tensor_descriptor(
IsInputGemm ? num_tokens : num_tokens * topk, K, stride_A, is_row_major(a_layout)));
auto b_k_n_tensor = ck_tile::HostTensor<BDataType>(
is_row_major(b_layout)
? ck_tile::host_tensor_descriptor(experts * N, K, stride_B, is_row_major(b_layout))
: ck_tile::host_tensor_descriptor(K, experts * N, stride_B, is_row_major(b_layout)));
auto c_m_n_tensor = ck_tile::HostTensor<CDataType>(ck_tile::host_tensor_descriptor(
IsInputGemm ? num_tokens * topk : num_tokens, outputN, stride_C, is_row_major(CLayout{})));
ck_tile::HostTensor<ScaleType> scale_b(ck_tile::HostTensorDescriptor(
{K * experts / ScaleGranularityK, N / ScaleGranularityN}, {N / ScaleGranularityN, 1}));
if(init_method == 0)
{
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_m_k_tensor);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_k_n_tensor);
ck_tile::FillUniformDistribution<ScaleType>{0.f, 1.f}(scale_b);
}
else
{
ck_tile::FillUniformDistribution<ADataType>{1.0f, 1.0f}(a_m_k_tensor);
ck_tile::FillUniformDistribution<BDataType>{1.0f, 1.0f}(b_k_n_tensor);
ck_tile::FillUniformDistribution<ScaleType>{1.0f, 1.0f}(scale_b);
}
ck_tile::HostTensor<BDataType> b_shuffle_host(
ck_tile::host_tensor_descriptor(K, experts * N, stride_B, is_row_major(b_layout)));
shuffle_mxfp4_weight<FlatmmConfig, kind>(
b_k_n_tensor.begin(), b_shuffle_host.begin(), experts, N, K);
ck_tile::HostTensor<ScaleType> scale_b_shuffle =
shuffle_mxfp4_scale<FlatmmConfig, kind>(scale_b, experts);
ck_tile::DeviceMem scale_b_shuffle_dev_buf(scale_b_shuffle.get_element_space_size_in_bytes());
std::cout << "moe_flatmm:" << "\n num_experts: " << experts << "\n num_tokens: " << num_tokens
<< "\n topk: " << topk << "\n sorted_tile_num: " << sorted_tile_num
<< "\n problem_n: " << N << "\n problem_k: " << K
<< "\n a_m_k: " << a_m_k_tensor.mDesc << "\n b_k_n: " << b_k_n_tensor.mDesc
<< "\n b_shuffle: " << b_shuffle_host.mDesc << "\n c_m_n: " << c_m_n_tensor.mDesc
<< std::endl;
ck_tile::HostTensor<ck_tile::index_t> expert_ids(
ck_tile::HostTensorDescriptor({sorted_tile_num}, {1}));
ck_tile::HostTensor<ck_tile::index_t> sorted_token_ids(
ck_tile::HostTensorDescriptor({sorted_size}, {1}));
ck_tile::HostTensor<AccDataType> expert_weight(
ck_tile::HostTensorDescriptor({sorted_size}, {1}));
ck_tile::HostTensor<ck_tile::index_t> max_token_id(
ck_tile::HostTensorDescriptor({1 + sorted_tile_num}));
ck_tile::HostTensor<AccDataType> expert_bias(ck_tile::HostTensorDescriptor({experts * N}, {1}));
if(init_method == 0)
{
// for verification only, no need to satify weight normalization
ck_tile::FillUniformDistribution<AccDataType>{0.0f, 1.0f}(expert_weight);
ck_tile::FillUniformDistribution<AccDataType>{-1.0f, 1.0f}(expert_bias);
}
else
{
ck_tile::FillUniformDistribution<AccDataType>{1.0f, 1.0f}(expert_weight);
ck_tile::FillUniformDistribution<AccDataType>{0.0f, 0.0f}(expert_bias);
}
max_token_id.mData = {valid_tile_num * MPerBlock, 0, 1, 2, 3, 4, 6, 7, 8, 8};
// int eids[] = {0, 1, 2, 3, 4, 4, 5, 6, 3, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ((valid_tile_num + experts - 1) / experts);
}
int token_per_tile = (num_tokens * topk + valid_tile_num - 1) / valid_tile_num;
// int token_per_tile = num_tokens * topk / valid_tile_num;
int tokenid = 0;
// sorted_token_ids.mData[0] = 0;
for(int i = 0; i < sorted_tile_num * MPerBlock; i++)
{
int tile_off = i % MPerBlock;
if(tile_off < token_per_tile && tokenid < num_tokens * topk)
{
sorted_token_ids.mData[i] = (tokenid % num_tokens) | ((tokenid / num_tokens) << 24);
tokenid++;
}
else
{
sorted_token_ids.mData[i] = num_tokens;
}
}
ck_tile::DeviceMem a_m_k_dev_buf{a_m_k_tensor.get_element_space_size_in_bytes()};
ck_tile::DeviceMem b_origin_dev_buf{b_k_n_tensor.get_element_space_size_in_bytes()};
ck_tile::DeviceMem b_shuffle_dev_buf{b_shuffle_host.get_element_space_size_in_bytes()};
ck_tile::DeviceMem c_m_n_dev_buf{c_m_n_tensor.get_element_space_size_in_bytes()};
a_m_k_dev_buf.ToDevice(a_m_k_tensor.data());
b_origin_dev_buf.ToDevice(b_k_n_tensor.data());
b_shuffle_dev_buf.ToDevice(b_shuffle_host.data());
c_m_n_dev_buf.SetZero();
c_m_n_tensor.SetZero();
ck_tile::DeviceMem sorted_token_ids_dev{sorted_token_ids.get_element_space_size_in_bytes()};
ck_tile::DeviceMem expert_ids_dev{expert_ids.get_element_space_size_in_bytes()};
ck_tile::DeviceMem max_token_id_dev{max_token_id.get_element_space_size_in_bytes()};
ck_tile::DeviceMem expert_weight_dev{expert_weight.get_element_space_size_in_bytes()};
ck_tile::DeviceMem expert_bias_dev{expert_bias.get_element_space_size_in_bytes()};
sorted_token_ids_dev.ToDevice(sorted_token_ids.data());
expert_ids_dev.ToDevice(expert_ids.data());
max_token_id_dev.ToDevice(max_token_id.data());
expert_weight_dev.ToDevice(expert_weight.data());
expert_bias_dev.ToDevice(expert_bias.data());
scale_b_shuffle_dev_buf.ToDevice(scale_b_shuffle.data());
const ck_tile::index_t* p_sorted_token_ids_dev =
static_cast<ck_tile::index_t*>(sorted_token_ids_dev.GetDeviceBuffer());
const ck_tile::index_t* p_expert_ids_dev =
static_cast<ck_tile::index_t*>(expert_ids_dev.GetDeviceBuffer());
const ck_tile::index_t* p_max_token_id_dev =
static_cast<ck_tile::index_t*>(max_token_id_dev.GetDeviceBuffer());
const AccDataType* p_sorted_expert_weight_dev =
static_cast<AccDataType*>(expert_weight_dev.GetDeviceBuffer());
auto scale_b_shuffle_dev_ptr =
ck_tile::FlatmmScalePointer<ScaleGranularityN, ScaleGranularityK>{
static_cast<float*>(scale_b_shuffle_dev_buf.GetDeviceBuffer()), N / ScaleGranularityN};
auto exp_bias_dev_ptr = ck_tile::FlatmmScalePointer<1>{
static_cast<float*>(expert_bias_dev.GetDeviceBuffer()), experts * N};
using MoeFlatmmArgs = ck_tile::MoeFlatmmHostArgs<
ck_tile::FlatmmScalePointer<-1>,
ck_tile::FlatmmScalePointer<ScaleGranularityN, ScaleGranularityK>,
ck_tile::FlatmmScalePointer<1>>;
MoeFlatmmArgs gemm_desc{p_sorted_token_ids_dev,
p_sorted_expert_weight_dev,
p_expert_ids_dev,
p_max_token_id_dev,
a_m_k_dev_buf.GetDeviceBuffer(),
b_shuffle_dev_buf.GetDeviceBuffer(),
c_m_n_dev_buf.GetDeviceBuffer(),
num_tokens,
experts,
topk,
1, // k_batch
M,
N,
K,
stride_A,
stride_B,
stride_C,
nullptr,
scale_b_shuffle_dev_ptr,
exp_bias_dev_ptr};
invoke_a16w4_moe_gemm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
kind>(warmup, repeat, gemm_desc);
c_m_n_dev_buf.FromDevice(c_m_n_tensor.data());
bool pass{true};
if(arg_parser.get_int("validate"))
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
ck_tile::host_tensor_descriptor(IsInputGemm ? num_tokens * topk : num_tokens,
outputN,
stride_C,
is_row_major(CLayout{})));
c_m_n_host_ref.SetZero();
ck_tile::HostTensor<AccDataType> scale_A(
ck_tile::HostTensorDescriptor({1, K / ScaleGranularityK}, {1, 1}));
// scaleA = 1 has no effect on the result
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(scale_A);
ck_tile::DeviceMem scale_A_dev_buf(scale_A.get_element_space_size_in_bytes());
scale_A_dev_buf.ToDevice(scale_A.data());
// convert scale_b from e8m0 to float
ck_tile::HostTensor<AccDataType> scale_b_float(ck_tile::HostTensorDescriptor(
{K * experts / ScaleGranularityK, N / ScaleGranularityN}, {N / ScaleGranularityN, 1}));
std::copy(scale_b.begin(), scale_b.end(), scale_b_float.begin());
ck_tile::DeviceMem scale_b_float_dev_buf(scale_b_float.get_element_space_size_in_bytes());
scale_b_float_dev_buf.ToDevice(scale_b_float.data());
std::unique_ptr<ck_tile::DeviceMem> c_m_n_ref_buf =
std::make_unique<ck_tile::DeviceMem>(c_m_n_tensor.get_element_space_size_in_bytes());
c_m_n_ref_buf->SetZero();
ck_tile::reference_moe_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout,
static_cast<int>(kind),
ck_tile::moe::Swiglu>(
p_sorted_token_ids_dev,
p_expert_ids_dev,
p_max_token_id_dev,
static_cast<const ADataType*>(a_m_k_dev_buf.GetDeviceBuffer()),
static_cast<const BDataType*>(b_origin_dev_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_ref_buf->GetDeviceBuffer()),
p_sorted_expert_weight_dev,
num_tokens,
MPerBlock,
topk,
M,
N,
K,
stride_A,
stride_B,
stride_C,
M,
1,
ScaleGranularityK,
static_cast<float*>(scale_A_dev_buf.GetDeviceBuffer()),
static_cast<float*>(scale_b_float_dev_buf.GetDeviceBuffer()),
static_cast<float*>(expert_bias_dev.GetDeviceBuffer()));
c_m_n_ref_buf->FromDevice(c_m_n_host_ref.data());
const float rtol = std::is_same_v<ADataType, ck_tile::half_t> && IsInputGemm ? 1e-3 : 1e-2;
const float atol = std::is_same_v<ADataType, ck_tile::half_t> && IsInputGemm ? 1e-3 : 1e-2;
pass = ck_tile::check_err(
c_m_n_tensor, c_m_n_host_ref, "Error: Incorrect results!", rtol, atol);
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
<< std::endl;
std::cout << "The CPU verification result is:" << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}

180
example/ck_tile/18_flatmm/mixed_prec/run_mixed_prec_flatmm.inc Normal file
View File

@@ -0,0 +1,180 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
template <typename PrecActType,
typename PrecWeightType,
typename FlatmmConfig,
bool UsePersistentKernel = false,
typename ALayout,
typename BLayout,
typename CLayout>
int run_mixed_prec_flatmm_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using ADataType = PrecActType;
using BDataType = PrecWeightType;
using CDataType = PrecActType;
using AccDataType = float;
using ScaleType = ck_tile::e8m0_t;
constexpr int DequantGranularityN = 1;
constexpr int DequantGranularityK = 32;
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");
ck_tile::index_t init_method = arg_parser.get_int("init");
ck_tile::index_t n_warmup = arg_parser.get_int("warmup");
ck_tile::index_t n_repeat = arg_parser.get_int("repeat");
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_host(
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_origin_host(
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
ck_tile::HostTensor<CDataType> c_rslt_host(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
ck_tile::HostTensor<ScaleType> scale_b(ck_tile::HostTensorDescriptor(
{K / DequantGranularityK, N / DequantGranularityN}, {N / DequantGranularityN, 1}));
if(init_method == 0)
{
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
ck_tile::FillUniformDistribution<ScaleType>{-2.f, 2.f}(scale_b);
}
else if(init_method == 1)
{
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
ck_tile::FillUniformDistribution<ScaleType>{1.f, 1.f}(scale_b);
}
ck_tile::HostTensor<BDataType> b_shuffle_host(
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
preShuffleWeight<FlatmmConfig>(b_origin_host.begin(), b_shuffle_host.begin(), N, K);
ck_tile::HostTensor<ScaleType> scale_b_shuffle = preShuffleScale<FlatmmConfig>(scale_b);
ck_tile::DeviceMem a_dev_buf(a_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_shuffle_dev_buf(b_shuffle_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_dev_buf(c_rslt_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem scale_b_dev_buf(scale_b_shuffle.get_element_space_size_in_bytes());
a_dev_buf.ToDevice(a_host.data());
b_shuffle_dev_buf.ToDevice(b_shuffle_host.data());
c_rslt_host.SetZero();
scale_b_dev_buf.ToDevice(scale_b_shuffle.data());
auto scale_b_dev_ptr = ck_tile::FlatmmScalePointer<DequantGranularityN, DequantGranularityK>{
static_cast<float*>(scale_b_dev_buf.GetDeviceBuffer()), N / DequantGranularityN};
invoke_mixed_prec_flatmm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
decltype(scale_b_dev_ptr),
UsePersistentKernel>(a_dev_buf,
b_shuffle_dev_buf,
c_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
scale_b_dev_ptr,
n_warmup,
n_repeat);
c_dev_buf.FromDevice(c_rslt_host.data());
bool pass = true;
if(arg_parser.get_int("v") == 1)
{
ck_tile::DeviceMem b_origin_dev_buf(b_origin_host.get_element_space_size_in_bytes());
b_origin_dev_buf.ToDevice(b_origin_host.data());
ck_tile::HostTensor<CDataType> c_gpu_ref_host(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
ck_tile::DeviceMem c_gpu_ref_dev_buf(c_gpu_ref_host.get_element_space_size_in_bytes());
ck_tile::HostTensor<AccDataType> scale_A(
ck_tile::HostTensorDescriptor({1, K / DequantGranularityK}, {1, 1}));
// scaleA = 1 has no effect on the result
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(scale_A);
ck_tile::DeviceMem scale_A_dev_buf(scale_A.get_element_space_size_in_bytes());
scale_A_dev_buf.ToDevice(scale_A.data());
// convert scale_b from e8m0 to float
ck_tile::HostTensor<AccDataType> scale_b_float(ck_tile::HostTensorDescriptor(
{K / DequantGranularityK, N / DequantGranularityN}, {N / DequantGranularityN, 1}));
std::copy(scale_b.begin(), scale_b.end(), scale_b_float.begin());
ck_tile::DeviceMem scale_b_float_dev_buf(scale_b_float.get_element_space_size_in_bytes());
scale_b_float_dev_buf.ToDevice(scale_b_float.data());
c_gpu_ref_dev_buf.SetZero();
ck_tile::reference_blockwise_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
static_cast<ADataType*>(a_dev_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_origin_dev_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_gpu_ref_dev_buf.GetDeviceBuffer()),
M,
N,
K,
stride_A,
stride_B,
stride_C,
M,
DequantGranularityN,
DequantGranularityK,
static_cast<float*>(scale_A_dev_buf.GetDeviceBuffer()),
static_cast<float*>(scale_b_float_dev_buf.GetDeviceBuffer()));
c_gpu_ref_dev_buf.FromDevice(c_gpu_ref_host.data());
const float rtol = std::is_same_v<ADataType, ck_tile::half_t> ? 1e-3 : 1e-2;
const float atol = std::is_same_v<ADataType, ck_tile::half_t> ? 1e-3 : 1e-2;
pass = ck_tile::check_err(
c_rslt_host, c_gpu_ref_host, "Error: Incorrect results!", rtol, atol);
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
<< std::endl;
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}

470
example/ck_tile/18_flatmm/moe_flatmm.cpp Normal file
View File

@@ -0,0 +1,470 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include <memory>
#include "moe_flatmm.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/ops/flatmm.hpp"
#include "ck_tile/ops/moe_flatmm.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/host/reference/reference_moe_gemm.hpp"
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 FlatmmConfig, 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 MaxVecSize = 16 / sizeof(T);
constexpr int KLane = ck_tile::get_warp_size() / FlatmmConfig::N_Warp_Tile;
constexpr int ItemsPerAccess = std::min(MaxVecSize, FlatmmConfig::K_Warp_Tile / KLane);
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
k_ / ItemsPerAccess,
ItemsPerAccess});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 1, 3});
}
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));
}
// gemm1
// operand-A = [num_token, d_model]
// operand-B = [num_expert, hidden, d_model]
// operand-C = [num_token, topk, hidden]
// gemm2
// operand-A = [num_token, topk, hidden]
// operand-B = [num_expert, d_model, hidden]
// operand-C = [num_token, d_model]
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ck_tile::MoeFlatmmKind moe_kind = ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_only,
typename CDEElementWise = ck_tile::element_wise::PassThrough,
typename ScaleM,
typename ScaleN>
float moe_gemm(const ck_tile::MoeFlatmmHostArgs<ScaleM, ScaleN>& args,
const ck_tile::stream_config& s)
{
using CodegenFlatmmShape = ck_tile::TileGemmShape<
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
ck_tile::sequence<FlatmmConfig::M_Warp, FlatmmConfig::N_Warp, FlatmmConfig::K_Warp>,
ck_tile::sequence<FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile>>;
using TilePartitioner =
ck_tile::GemmSpatiallyLocalTilePartitioner<CodegenFlatmmShape,
FlatmmConfig::TileParitionerGroupNum,
FlatmmConfig::TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
ALayout,
BLayout,
ELayout,
FlatmmConfig::NumWaveGroups>;
using CodegenGemmTraits = ck_tile::TileGemmUniversalTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
FlatmmConfig::DoubleSmemBuffer,
ALayout,
BLayout,
ELayout,
FlatmmConfig::TransposeC,
FlatmmConfig::UseStructuredSparsity,
false, // UsePersistentKernel_
FlatmmConfig::NumWaveGroups,
true>; // Preshuffle_
if constexpr(moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up)
{
static_assert(
FlatmmConfig::N_Tile % (FlatmmConfig::N_Warp * FlatmmConfig::N_Warp_Tile * 2) == 0,
"requires NRepeat is multiple of 2 for FFN_gemm1_gate_up");
}
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenFlatmmShape, Traits>;
using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_,
const auto tail_number_,
const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = FlatmmConfig::Scheduler;
constexpr auto memory_operation = memory_operation_.value;
using CodegenPipelineProblem = ck_tile::FlatmmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
constexpr int BlockedXDLN_PerWarp = moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up
? 2
: 1; // determined by scale shuffle pattern
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
DsLayout,
ELayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
FlatmmConfig::M_Warp,
FlatmmConfig::N_Warp,
FlatmmConfig::M_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
FlatmmConfig::K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation,
FlatmmConfig::NumWaveGroups,
false,
1,
FlatmmConfig::TiledMMAPermuteN,
BlockedXDLN_PerWarp>>;
using CodegenFlatmmPipeline =
ck_tile::MoeFlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using Kernel = ck_tile::
MoeFlatmmKernel<TilePartitioner, CodegenFlatmmPipeline, GemmEpilogue, moe_kind>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:" << CodegenFlatmmShape::GetName() << "\n"
<< "Shape: " << CodegenFlatmmShape::GetName() << "\n"
<< "problem: " << CodegenPipelineProblem::GetName() << "\n"
<< "pipeline: " << CodegenFlatmmPipeline::GetName() << "\n"
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
if(s.flush_cache_)
{
std::cout << "Flushing cache..." << std::endl;
static constexpr ck_tile::index_t APackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
static constexpr ck_tile::index_t BPackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm2 ? args.NumTokens * args.TopK
: args.NumTokens,
args.K,
args.stride_A,
is_row_major(ALayout{})));
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.N * args.NumExperts, args.stride_B, is_row_major(BLayout{})));
const int outputN =
moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up ? args.N / 2 : args.N;
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
kargs.a_ptr, kargs.b_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck_tile::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(moe_kind == ck_tile::MoeFlatmmKind::kFFN_gemm2)
hipGetErrorString(hipMemsetAsync(
args.e_ptr, 0, args.NumTokens * args.N * sizeof(CDataType), s.stream_id_));
else if(args.k_batch > 1)
hipGetErrorString(
hipMemsetAsync(args.e_ptr,
0,
args.NumTokens * args.TopK * outputN * sizeof(CDataType),
s.stream_id_));
};
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
Run(has_hot_loop_,
tail_number_,
ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
};
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
return ave_time;
}
#include "run_moe_flatmm_example.inc"
template <template <typename PreType> typename FlatmmConfig>
int run_moe_flatmm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
{
return -1;
}
const std::string a_layout = arg_parser.get_str("a_layout");
const std::string b_layout = arg_parser.get_str("b_layout");
const std::string prec_type = arg_parser.get_str("prec");
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
if(a_layout == "R" && b_layout == "C")
{
const std::string gemm_kind = arg_parser.get_str("gemm_kind");
if(gemm_kind == "gemm1_gate_up")
{
if(prec_type == "fp8")
{
return run_moe_gemm_example_with_layouts<
ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up>(argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "bf8")
{
return run_moe_gemm_example_with_layouts<
ck_tile::bf8_t,
FlatmmConfig<ck_tile::bf8_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up>(argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "bf16")
{
return run_moe_gemm_example_with_layouts<
ck_tile::bfloat16_t,
FlatmmConfig<ck_tile::bfloat16_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up>(argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "fp16")
{
return run_moe_gemm_example_with_layouts<
ck_tile::half_t,
FlatmmConfig<ck_tile::half_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported precision type for gemm1_gate_up!");
}
}
else if(gemm_kind == "gemm1_gate_only")
{
if(prec_type == "fp8")
{
return run_moe_gemm_example_with_layouts<
ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_only>(argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "bf8")
{
return run_moe_gemm_example_with_layouts<
ck_tile::bf8_t,
FlatmmConfig<ck_tile::bf8_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_only>(argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "bf16")
{
return run_moe_gemm_example_with_layouts<
ck_tile::bfloat16_t,
FlatmmConfig<ck_tile::bfloat16_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_only>(argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "fp16")
{
return run_moe_gemm_example_with_layouts<
ck_tile::half_t,
FlatmmConfig<ck_tile::half_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_only>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported precision type for gemm1_gate_up!");
}
}
else if(gemm_kind == "gemm2")
{
if(prec_type == "fp8")
{
return run_moe_gemm_example_with_layouts<ck_tile::fp8_t,
FlatmmConfig<ck_tile::fp8_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm2>(
argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "bf8")
{
return run_moe_gemm_example_with_layouts<ck_tile::bf8_t,
FlatmmConfig<ck_tile::bf8_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm2>(
argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "bf16")
{
return run_moe_gemm_example_with_layouts<ck_tile::bfloat16_t,
FlatmmConfig<ck_tile::bfloat16_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm2>(
argc, argv, Row{}, Col{}, Row{});
}
else if(prec_type == "fp16")
{
return run_moe_gemm_example_with_layouts<ck_tile::half_t,
FlatmmConfig<ck_tile::half_t>,
ck_tile::MoeFlatmmKind::kFFN_gemm2>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported precision type for gemm1_gate_up!");
}
}
else
{
throw std::runtime_error("Unrecoginized gemm_kind parameter, only accept value "
"[gemm1_gate_only | gemm1_gate_up | gemm2]");
}
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
return -1;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return EXIT_FAILURE;
try
{
int warp_tile = arg_parser.get_int("warp_tile");
if(warp_tile == 0)
{
return !run_moe_flatmm_example<FlatmmConfig16>(argc, argv);
}
else if(warp_tile == 1)
{
return !run_moe_flatmm_example<FlatmmConfig32>(argc, argv);
}
else if(warp_tile == 2)
{
return !run_moe_flatmm_example<FlatmmConfig16_950>(argc, argv);
}
else
{
return !run_moe_flatmm_example<FlatmmConfig32_950>(argc, argv);
}
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

202
example/ck_tile/18_flatmm/moe_flatmm.hpp Normal file
View File

@@ -0,0 +1,202 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include <tuple>
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/moe_flatmm.hpp"
template <typename DataType>
struct FlatmmConfig32
{
static constexpr ck_tile::index_t M_Tile = 64;
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(DataType);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 16 : 32;
static constexpr bool kPadM = false;
static constexpr bool kPadN = false;
static constexpr bool kPadK = false;
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool DoubleSmemBuffer = false;
static constexpr bool TiledMMAPermuteN = false; // disable PermuteN when NWarpTile != 16
};
template <typename DataType>
struct FlatmmConfig32_950 : public FlatmmConfig32<DataType>
{
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 16 : 64;
};
// GEMM config with 16x16 warp tile
template <typename DataType>
struct FlatmmConfig16
{
static constexpr ck_tile::index_t M_Tile = 64;
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(DataType);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 32 : 64;
static constexpr bool kPadM = false;
static constexpr bool kPadN = false;
static constexpr bool kPadK = false;
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool DoubleSmemBuffer = false;
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
static constexpr bool TiledMMAPermuteN = false;
};
template <typename DataType>
struct FlatmmConfig16_950 : public FlatmmConfig16<DataType>
{
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(DataType);
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 32 : 128;
static constexpr int kBlockPerCu = 1;
static constexpr int N_Repeat =
N_Tile / FlatmmConfig16<DataType>::N_Warp_Tile / FlatmmConfig16<DataType>::N_Warp;
static constexpr bool TiledMMAPermuteN = false; // N_Repeat % 2 == 0;
};
template <typename ADataType>
struct GemmBasicTypeConfig;
template <>
struct GemmBasicTypeConfig<ck_tile::half_t>
{
using ADataType = ck_tile::half_t;
using BDataType = ck_tile::half_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
// ToDo: Add more bias config to support different categories of GEMM.
};
template <>
struct GemmBasicTypeConfig<ck_tile::bf16_t>
{
using ADataType = ck_tile::bf16_t;
using BDataType = ck_tile::bf16_t;
using AccDataType = float;
using CDataType = ck_tile::bf16_t;
};
template <>
struct GemmBasicTypeConfig<ck_tile::fp8_t>
{
using ADataType = ck_tile::fp8_t;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
// ToDo: Add more bias config to support different categories of GEMM.
};
template <>
struct GemmBasicTypeConfig<ck_tile::bf8_t>
{
using ADataType = ck_tile::bf8_t;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<ck_tile::fp8_t>
{
static constexpr const char* name = "fp8";
};
template <>
struct DataTypeTraits<ck_tile::bf8_t>
{
static constexpr const char* name = "bf8";
};
template <>
struct DataTypeTraits<float>
{
static constexpr const char* name = "fp32";
};
template <>
struct DataTypeTraits<double>
{
static constexpr const char* name = "fp64";
};
template <>
struct DataTypeTraits<ck_tile::half_t>
{
static constexpr const char* name = "fp16";
};
template <typename T>
struct is_8bit_type
: std::bool_constant<std::is_same_v<T, ck_tile::fp8_t> || std::is_same_v<T, ck_tile::bf8_t>>
{
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("experts", "8", "Num of experts - 8 by default")
.insert("NumTokens", "128", "M dimensions - 128 by default.")
.insert("TopK", "3", "Top K - 3 by default.")
.insert("N", "4096", "N dimensions - 4096 by default.")
.insert("K", "4096", "K dimensions - 4096 by default.")
.insert("stride_A", "", "Tensor A strides - it is empty by default.")
.insert("stride_B", "", "Tensor B strides - it is empty by default.")
.insert("stride_C", "", "Tensor C strides - it is empty by default.")
.insert("a_layout", "R", "A tensor data layout - Row by default.")
.insert("b_layout", "C", "B tensor data layout - Col by default.")
.insert("c_layout", "R", "C tensor data layout - Row by default.")
.insert("gemm_kind",
"gemm1_gate_only",
"Gemm kind in FFN network [gemm1_gate_only | gemm1_gate_up | gemm2] - "
"gemm1_gate_only by default.")
.insert("validate", "1", "0. No validation, 1. Validation on CPU.")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert(
"warp_tile", "0", "0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)")
.insert("repeat", "10", "number of iterations to benchmark the kernel.");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}

332
example/ck_tile/18_flatmm/run_flatmm_example.inc
View File

@@ -1,175 +1,12 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <type_traits>
#include "ck_tile/utility/json_dump.hpp"
template <typename T>
constexpr const char* DataTypeToString()
{
if constexpr(std::is_same_v<T, ck_tile::half_t>)
{
return "fp16";
}
else if constexpr(std::is_same_v<T, ck_tile::fp8_t>)
{
return "fp8";
}
else if constexpr(std::is_same_v<T, ck_tile::bf8_t>)
{
return "bf8";
}
else if constexpr(std::is_same_v<T, ck_tile::bf16_t>)
{
return "bf16";
}
else
{
return "unknown";
}
}
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>>{};
}
// mfma_type, 0:32x32, 1:16x16
template <typename FlatmmConfig, 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];
if(ck_tile::is_gfx12_supported())
{
constexpr int divisor = 2;
constexpr int kABK1PerLane = 8;
constexpr int kABK0PerLane = FlatmmConfig::K_Warp_Tile / divisor / kABK1PerLane;
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
k_ / FlatmmConfig::K_Warp_Tile,
kABK0PerLane,
divisor,
kABK1PerLane});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 4, 1, 3, 5});
}
else
{
int divisor = 1;
if(ck_tile::is_gfx11_supported())
{
divisor = 1;
}
else
{
assert(is_wave32() == false);
divisor = FlatmmConfig::N_Warp_Tile == 32 ? 2 : 4;
}
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
k_ / FlatmmConfig::K_Warp_Tile,
divisor,
FlatmmConfig::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 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 FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
bool persistent,
typename CDEElementWise>
float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_config& s);
template <typename FlatmmConfig,
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_flatmm(ck_tile::DeviceMem& a_dev_buf,
ck_tile::DeviceMem& b_shuffle_dev_buf,
ck_tile::DeviceMem& c_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)
{
ck_tile::FlatmmHostArgs<> args = {a_dev_buf.GetDeviceBuffer(),
b_shuffle_dev_buf.GetDeviceBuffer(),
{},
c_dev_buf.GetDeviceBuffer(),
kbatch,
M,
N,
K,
stride_A,
stride_B,
{},
stride_C};
float ave_time = flatmm_calc<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
false,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
return ave_time;
}
template <typename PrecType,
typename FlatmmConfig,
int ScaleGranularityM = -1,
int ScaleGranularityN = -1,
bool UsePersistentKernel = false,
typename ALayout,
typename BLayout,
typename CLayout>
@@ -213,31 +50,32 @@ int run_flatmm_example_with_layouts(int argc,
ck_tile::HostTensor<CDataType> c_rslt_host(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
ck_tile::HostTensor<AccDataType> per_token_scale(ck_tile::HostTensorDescriptor({M}, {1}));
ck_tile::HostTensor<AccDataType> per_channel_scale(ck_tile::HostTensorDescriptor({N}, {1}));
// TODO: add different init types
if(init_method == 0)
{
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
// ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
// ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_channel_scale);
}
else if(init_method == 1)
{
ck_tile::FillMonotonicSeq<ADataType>{}(a_host);
ck_tile::FillMonotonicSeq<BDataType>{}(b_origin_host);
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_channel_scale);
}
else if(init_method == 2)
{
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
}
else if(init_method == 3)
{
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
}
else if(init_method == 4)
{
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_channel_scale);
}
else
{
@@ -248,52 +86,69 @@ int run_flatmm_example_with_layouts(int argc,
ck_tile::DeviceMem a_dev_buf(a_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_dev_buf(c_rslt_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem per_token_scale_dev_buf(per_token_scale.get_element_space_size_in_bytes());
ck_tile::DeviceMem per_channel_scale_dev_buf(
per_channel_scale.get_element_space_size_in_bytes());
a_dev_buf.ToDevice(a_host.data());
c_rslt_host.SetZero();
per_token_scale_dev_buf.ToDevice(per_token_scale.data());
per_channel_scale_dev_buf.ToDevice(per_channel_scale.data());
// do pre-shuffle
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b<FlatmmConfig>(b_origin_host);
ck_tile::HostTensor<BDataType> b_shuffle_host = [&]() {
if constexpr(FlatmmConfig::TiledMMAPermuteN)
{
return shuffle_b_v1<FlatmmConfig>(b_origin_host);
}
else
{
return shuffle_b<FlatmmConfig>(b_origin_host);
}
}();
ck_tile::DeviceMem b_shuffle_dev_buf(b_shuffle_host.get_element_space_size_in_bytes());
b_shuffle_dev_buf.ToDevice(b_shuffle_host.data());
float ave_time = invoke_flatmm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout>(a_dev_buf,
b_shuffle_dev_buf,
c_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
n_warmup,
n_repeat);
auto per_token_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityM>{
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer())};
auto per_channel_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityN>{
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
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 Flatmm kernel with DataType = " << DataTypeToString<ADataType>()
<< " M =" << M << " N =" << N << " K =" << K << " StrideA =" << stride_A
<< " StrideB =" << stride_B << " StrideC =" << stride_C << " : " << ave_time
<< " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
invoke_flatmm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr),
UsePersistentKernel>(a_dev_buf,
b_shuffle_dev_buf,
c_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
per_token_scale_dev_ptr,
per_channel_scale_dev_ptr,
n_warmup,
n_repeat);
c_dev_buf.FromDevice(c_rslt_host.data());
bool pass = true;
if(arg_parser.get_int("v") == 1)
{
if(ScaleGranularityM != -1 || ScaleGranularityN != -1)
throw std::runtime_error("ScaleAB is not supported for CPU verification!\n");
ck_tile::HostTensor<CDataType> c_ref_host(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
c_ref_host.SetZero();
@@ -341,13 +196,41 @@ int run_flatmm_example_with_layouts(int argc,
N * K * sizeof(BDataType),
hipMemcpyHostToDevice));
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);
if constexpr(ScaleGranularityM == -1 && ScaleGranularityN == -1)
{
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);
}
else
{
ck_tile::reference_blockwise_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
d_A,
d_B,
d_C,
M,
N,
K,
stride_A,
stride_B,
stride_C,
ScaleGranularityM,
ScaleGranularityN,
K,
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()),
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer()));
}
ck_tile::hip_check_error(hipMemcpy(c_gpu_ref_dev_buf.GetDeviceBuffer(),
d_C,
@@ -375,22 +258,5 @@ int run_flatmm_example_with_layouts(int argc,
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
}
if(arg_parser.get_int("json") == 1)
{
dump_flatmm_json_results(arg_parser.get_str("jsonfile"),
DataTypeToString<ADataType>(),
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
pass,
ave_time,
tflops,
gb_per_sec);
}
return pass;
}

605
example/ck_tile/18_flatmm/run_grouped_flatmm_example.inc Normal file
View File

@@ -0,0 +1,605 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
// mfma_type, 0:32x32, 1:16x16
template <typename FlatmmConfig, 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 = FlatmmConfig::N_Warp_Tile == 32 ? 2 : 4;
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
k_ / FlatmmConfig::K_Warp_Tile,
divisor,
FlatmmConfig::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 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 FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout,
typename ScaleM,
typename ScaleN,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float invoke_gemm(int n_warmup,
int n_repeat,
const ck_tile::ContiguousGroupedFlatmmHostArgs<ScaleM, ScaleN>& args)
{
float ave_time = grouped_flatmm<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
false,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::string op_name{"Grouped Gemm"};
std::size_t flop = std::size_t(2) * args.M * args.N * args.K;
std::size_t num_byte = sizeof(ADataType) * args.M * args.K +
sizeof(BDataType) * args.N * args.K +
sizeof(CDataType) * args.M * args.N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
return ave_time;
}
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout,
typename ScaleM,
typename ScaleN,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float invoke_gemm(int n_warmup,
int n_repeat,
int val_m,
const ck_tile::MaskedGroupedFlatmmHostArgs<ScaleM, ScaleN>& args)
{
float ave_time = grouped_flatmm<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
CLayout,
false,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::string op_name{"Grouped Gemm"};
std::size_t flop = std::size_t(2) * val_m * args.N * args.K;
std::size_t num_byte = sizeof(ADataType) * val_m * args.K +
sizeof(BDataType) * args.N * args.K * args.group_count +
sizeof(CDataType) * val_m * args.N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
return ave_time;
}
template <typename PrecType,
typename FlatmmConfig,
int ScaleGranularityM = -1,
int ScaleGranularityN = -1,
typename ALayout,
typename BLayout,
typename CLayout>
int run_contiguous_grouped_flatmm_example_with_layouts(
int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
{
return -1;
};
using ADataType = typename GemmBasicTypeConfig<PrecType>::ADataType;
using BDataType = typename GemmBasicTypeConfig<PrecType>::BDataType;
using CDataType = typename GemmBasicTypeConfig<PrecType>::CDataType;
using AccDataType = typename GemmBasicTypeConfig<PrecType>::AccDataType;
constexpr int BlockM = FlatmmConfig::M_Tile;
const int group_count = arg_parser.get_int("group_count");
const int repeat = arg_parser.get_int("repeat");
const int warmup = arg_parser.get_int("warmup");
std::vector<ck_tile::index_t> Ms = arg_parser.get_int_vec("Ms");
std::vector<ck_tile::index_t> Ns = arg_parser.get_int_vec("Ns");
std::vector<ck_tile::index_t> Ks = arg_parser.get_int_vec("Ks");
if(!(int(Ms.size()) == group_count))
{
std::cout << "Please check the input data." << std::endl;
// padding additional Ms if needed
for(int i = 0; i < group_count; i++)
{
Ms.push_back(256 + 64 * i);
}
}
ck_tile::index_t M =
std::reduce(Ms.begin(), Ms.begin() + group_count, 0, [](auto acc, auto group_m) {
// round up to the multiple of BlockM
return acc + (group_m + BlockM - 1) / BlockM * BlockM;
});
std::cout << "Total M: " << M << std::endl;
ck_tile::index_t N = Ns[0];
ck_tile::index_t K = Ks[0];
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
ck_tile::index_t stride_A = 0;
ck_tile::index_t stride_B = 0;
ck_tile::index_t stride_C = 0;
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 * group_count, stride_B, is_row_major(b_layout));
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(c_layout));
ck_tile::HostTensor<ADataType> a_m_k_tensor(
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_k_n_tensor(ck_tile::HostTensor<BDataType>(
ck_tile::host_tensor_descriptor(K, N * group_count, stride_B, is_row_major(b_layout))));
ck_tile::HostTensor<CDataType> c_m_n_tensor(ck_tile::HostTensor<CDataType>(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(c_layout))));
ck_tile::HostTensor<AccDataType> per_token_scale(ck_tile::HostTensorDescriptor({M}, {1}));
ck_tile::HostTensor<AccDataType> per_channel_scale(ck_tile::HostTensorDescriptor({N}, {1}));
std::vector<ck_tile::index_t> m_indices(M);
int indices_fill_start = 0;
for(int i = 0; i < group_count; ++i)
{
int group_m = Ms[i];
int padded_group_m = (group_m + BlockM - 1) / BlockM * BlockM;
for(int j = 0; j < padded_group_m; j++)
{
m_indices[indices_fill_start + j] = j < group_m ? i : -1; // -1 for padding
}
indices_fill_start += padded_group_m;
}
ck_tile::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k_tensor);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_k_n_tensor);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_channel_scale);
assert(N % N_Warp_Tile == 0 &&
"N must be divisible by N_Warp_Tile for contiguous grouped gemm");
ck_tile::HostTensor<BDataType> b_shuffle_host =
shuffle_b<FlatmmConfig, BDataType>(b_k_n_tensor);
std::unique_ptr<ck_tile::DeviceMem> a_m_k_dev_buf(
std::make_unique<ck_tile::DeviceMem>(a_m_k_tensor.get_element_space_size_in_bytes()));
std::unique_ptr<ck_tile::DeviceMem> b_shfl_dev_buf(
std::make_unique<ck_tile::DeviceMem>(b_shuffle_host.get_element_space_size_in_bytes()));
std::unique_ptr<ck_tile::DeviceMem> c_m_n_dev_buf(
std::make_unique<ck_tile::DeviceMem>(c_m_n_tensor.get_element_space_size_in_bytes()));
ck_tile::DeviceMem per_token_scale_dev_buf(per_token_scale.get_element_space_size_in_bytes());
ck_tile::DeviceMem per_channel_scale_dev_buf(
per_channel_scale.get_element_space_size_in_bytes());
c_m_n_dev_buf->SetZero();
ck_tile::DeviceMem m_indices_dev_buf(M * sizeof(ck_tile::index_t));
m_indices_dev_buf.ToDevice(m_indices.data());
a_m_k_dev_buf->ToDevice(a_m_k_tensor.data());
b_shfl_dev_buf->ToDevice(b_shuffle_host.data());
per_token_scale_dev_buf.ToDevice(per_token_scale.data());
per_channel_scale_dev_buf.ToDevice(per_channel_scale.data());
auto per_token_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityM>{
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer())};
auto per_channel_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityN>{
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
ck_tile::ContiguousGroupedFlatmmHostArgs<decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr)>
kernal_args{static_cast<ck_tile::index_t*>(m_indices_dev_buf.GetDeviceBuffer()),
M,
N,
K,
a_m_k_dev_buf->GetDeviceBuffer(),
stride_A,
b_shfl_dev_buf->GetDeviceBuffer(),
stride_B,
{},
{},
c_m_n_dev_buf->GetDeviceBuffer(),
stride_C,
kbatch,
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()),
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
invoke_gemm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr)>(warmup, repeat, kernal_args);
c_m_n_dev_buf->FromDevice(c_m_n_tensor.data());
bool pass{true};
if(arg_parser.get_int("v") == 1)
{
throw std::runtime_error(
"Not support v=1 host verification in contiguous grouped gemm, use "
"v=2 device verification instead");
}
else if(arg_parser.get_int("v") == 2)
{
BDataType* d_B;
CDataType* d_C;
ck_tile::hip_check_error(hipMalloc(&d_B, N * K * sizeof(BDataType)));
ck_tile::hip_check_error(hipMalloc(&d_C, M * N * sizeof(CDataType)));
ck_tile::hip_check_error(hipMemset(d_C, 0, M * N * sizeof(CDataType)));
ck_tile::HostTensor<CDataType> c_gpu_ref_host(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
ck_tile::index_t acc_m = 0;
for(int i = 0; i < group_count; ++i)
{
ck_tile::index_t padded_M = (Ms[i] + BlockM - 1) / BlockM * BlockM;
ck_tile::hip_check_error(hipMemcpy(d_B,
b_k_n_tensor.data() + i * N * K,
N * K * sizeof(BDataType),
hipMemcpyHostToDevice));
ck_tile::reference_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
static_cast<ADataType*>(a_m_k_dev_buf->GetDeviceBuffer()) + acc_m * K,
d_B,
d_C + acc_m * N,
padded_M,
N,
K,
stride_A,
stride_B,
stride_C);
acc_m += padded_M;
}
ck_tile::hip_check_error(hipMemcpy(
c_gpu_ref_host.data(), d_C, M * N * sizeof(CDataType), hipMemcpyDeviceToHost));
ck_tile::hip_check_error(hipFree(d_B));
ck_tile::hip_check_error(hipFree(d_C));
float rtol = 1e-3;
float atol = 1e-3;
pass = ck_tile::check_err(
c_m_n_tensor, c_gpu_ref_host, "Error: Incorrect results!", rtol, atol);
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
<< std::endl;
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}
template <typename PrecType,
typename FlatmmConfig,
int ScaleGranularityM = -1,
int ScaleGranularityN = -1,
typename ALayout,
typename BLayout,
typename CLayout>
int run_masked_grouped_flatmm_example_with_layouts(
int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
{
return -1;
};
using ADataType = typename GemmBasicTypeConfig<PrecType>::ADataType;
using BDataType = typename GemmBasicTypeConfig<PrecType>::BDataType;
using CDataType = typename GemmBasicTypeConfig<PrecType>::CDataType;
using AccDataType = typename GemmBasicTypeConfig<PrecType>::AccDataType;
constexpr int BlockM = FlatmmConfig::M_Tile;
const int group_count = arg_parser.get_int("group_count");
const int repeat = arg_parser.get_int("repeat");
const int warmup = arg_parser.get_int("warmup");
std::vector<ck_tile::index_t> Ms = arg_parser.get_int_vec("Ms");
std::vector<ck_tile::index_t> Ns = arg_parser.get_int_vec("Ns");
std::vector<ck_tile::index_t> Ks = arg_parser.get_int_vec("Ks");
if(!(int(Ms.size()) == group_count))
{
std::cout << "Please check the input data." << std::endl;
// padding additional Ms if needed
for(int i = 0; i < group_count; i++)
{
Ms.push_back(256 + 64 * i);
}
}
ck_tile::index_t M = 4096; // Ms[0];
ck_tile::index_t N = Ns[0];
ck_tile::index_t K = Ks[0];
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
ck_tile::index_t stride_A = K;
ck_tile::index_t stride_B = K;
ck_tile::index_t stride_C = N;
stride_A = ck_tile::get_default_stride(group_count * M, K, stride_A, is_row_major(a_layout));
stride_B = ck_tile::get_default_stride(K, N * group_count, stride_B, is_row_major(b_layout));
stride_C = ck_tile::get_default_stride(group_count * M, N, stride_C, is_row_major(c_layout));
ck_tile::HostTensor<ADataType> a_m_k_tensor(
ck_tile::host_tensor_descriptor(group_count * M, K, stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_k_n_tensor(ck_tile::HostTensor<BDataType>(
ck_tile::host_tensor_descriptor(K, N * group_count, stride_B, is_row_major(b_layout))));
ck_tile::HostTensor<CDataType> c_m_n_tensor(ck_tile::HostTensor<CDataType>(
ck_tile::host_tensor_descriptor(group_count * M, N, stride_C, is_row_major(c_layout))));
ck_tile::HostTensor<AccDataType> per_token_scale(
ck_tile::HostTensorDescriptor({group_count * M}, {1}));
ck_tile::HostTensor<AccDataType> per_channel_scale(
ck_tile::HostTensorDescriptor({group_count * N}, {1}));
std::vector<ck_tile::index_t> m_indices(group_count);
for(int i = 0; i < group_count; ++i)
{
int group_m = Ms[i];
int padded_group_m = (group_m + BlockM - 1) / BlockM * BlockM;
for(int j = 0; j < padded_group_m; j++)
{
m_indices[i] = group_m;
}
}
ck_tile::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k_tensor);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_k_n_tensor);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_channel_scale);
assert(N % N_Warp_Tile == 0 &&
"N must be divisible by N_Warp_Tile for contiguous grouped gemm");
ck_tile::HostTensor<BDataType> b_shuffle_host =
shuffle_b<FlatmmConfig, BDataType>(b_k_n_tensor);
std::unique_ptr<ck_tile::DeviceMem> a_m_k_dev_buf(
std::make_unique<ck_tile::DeviceMem>(a_m_k_tensor.get_element_space_size_in_bytes()));
std::unique_ptr<ck_tile::DeviceMem> b_shfl_dev_buf(
std::make_unique<ck_tile::DeviceMem>(b_shuffle_host.get_element_space_size_in_bytes()));
std::unique_ptr<ck_tile::DeviceMem> c_m_n_dev_buf(
std::make_unique<ck_tile::DeviceMem>(c_m_n_tensor.get_element_space_size_in_bytes()));
ck_tile::DeviceMem per_token_scale_dev_buf(per_token_scale.get_element_space_size_in_bytes());
ck_tile::DeviceMem per_channel_scale_dev_buf(
per_channel_scale.get_element_space_size_in_bytes());
c_m_n_dev_buf->SetZero();
ck_tile::DeviceMem m_indices_dev_buf(group_count * sizeof(ck_tile::index_t));
m_indices_dev_buf.ToDevice(m_indices.data());
a_m_k_dev_buf->ToDevice(a_m_k_tensor.data());
b_shfl_dev_buf->ToDevice(b_shuffle_host.data());
per_token_scale_dev_buf.ToDevice(per_token_scale.data());
per_channel_scale_dev_buf.ToDevice(per_channel_scale.data());
auto per_token_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityM>{
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer())};
auto per_channel_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityN>{
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
ck_tile::MaskedGroupedFlatmmHostArgs<decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr)>
kernal_args{static_cast<ck_tile::index_t*>(m_indices_dev_buf.GetDeviceBuffer()),
group_count,
M,
N,
K,
a_m_k_dev_buf->GetDeviceBuffer(),
stride_A,
b_shfl_dev_buf->GetDeviceBuffer(),
stride_B,
{},
{},
c_m_n_dev_buf->GetDeviceBuffer(),
stride_C,
kbatch,
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()),
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
int sum_val_m = 0;
for(int gi = 0; gi < group_count; gi++)
{
sum_val_m += m_indices[gi];
}
invoke_gemm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr)>(warmup, repeat, sum_val_m, kernal_args);
c_m_n_dev_buf->FromDevice(c_m_n_tensor.data());
bool pass{true};
if(arg_parser.get_int("v") == 1)
{
throw std::runtime_error(
"Not support v=1 host verification in contiguous grouped gemm, use "
"v=2 device verification instead");
}
else if(arg_parser.get_int("v") == 2)
{
BDataType* d_B;
CDataType* d_C;
ck_tile::hip_check_error(hipMalloc(&d_B, N * K * sizeof(BDataType)));
ck_tile::hip_check_error(hipMalloc(&d_C, group_count * M * N * sizeof(CDataType)));
ck_tile::hip_check_error(hipMemset(d_C, 0, group_count * M * N * sizeof(CDataType)));
ck_tile::HostTensor<CDataType> c_gpu_ref_host(
ck_tile::host_tensor_descriptor(group_count * M, N, stride_C, is_row_major(CLayout{})));
for(int i = 0; i < group_count; ++i)
{
ck_tile::hip_check_error(hipMemcpy(d_B,
b_k_n_tensor.data() + i * N * K,
N * K * sizeof(BDataType),
hipMemcpyHostToDevice));
if constexpr(ScaleGranularityM == -1 && ScaleGranularityN == -1)
{
ck_tile::reference_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
static_cast<ADataType*>(a_m_k_dev_buf->GetDeviceBuffer()) + i * M * K,
d_B,
d_C + i * M * N,
m_indices[i],
N,
K,
stride_A,
stride_B,
stride_C);
}
else
{
ck_tile::reference_blockwise_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
static_cast<ADataType*>(a_m_k_dev_buf->GetDeviceBuffer()) + i * M * K,
d_B,
d_C + i * M * N,
m_indices[i],
N,
K,
stride_A,
stride_B,
stride_C,
ScaleGranularityM,
ScaleGranularityN,
K,
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()) + i * M,
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())) +
i* N;
}
ck_tile::hip_check_error(hipMemcpy(c_gpu_ref_host.data() + i * M * N,
d_C + i * M * N,
M * N * sizeof(CDataType),
hipMemcpyDeviceToHost));
}
ck_tile::hip_check_error(hipFree(d_B));
ck_tile::hip_check_error(hipFree(d_C));
float rtol = 1e-3;
float atol = 1e-3;
pass = ck_tile::check_err(
c_m_n_tensor, c_gpu_ref_host, "Error: Incorrect results!", rtol, atol);
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
<< std::endl;
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}

323
example/ck_tile/18_flatmm/run_moe_flatmm_example.inc Normal file
View File

@@ -0,0 +1,323 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
template <typename FlatmmConfig,
typename ADataType,
typename BDataType,
typename DsDatatype,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ck_tile::MoeFlatmmKind kind,
typename CDEElementWise = ck_tile::element_wise::PassThrough,
typename MoeHostArgs>
float invoke_moe_gemm(int n_warmup, int n_repeat, const MoeHostArgs& args)
{
float ave_time = moe_gemm<FlatmmConfig,
ADataType,
BDataType,
DsDatatype,
AccDataType,
CDataType,
ALayout,
BLayout,
DsLayout,
ELayout,
kind,
CDEElementWise>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
std::string op_name{"Moe Gemm"};
std::size_t flop = std::size_t(2) * args.M * args.N * args.K;
std::size_t num_byte = sizeof(ADataType) * args.M * args.K +
sizeof(BDataType) * args.N * args.K +
sizeof(CDataType) * args.M * args.N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
return ave_time;
}
template <typename PrecType,
typename FlatmmConfig,
ck_tile::MoeFlatmmKind kind,
typename ALayout,
typename BLayout,
typename CLayout>
int run_moe_gemm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
{
return -1;
};
using ADataType = typename GemmBasicTypeConfig<PrecType>::ADataType;
using BDataType = typename GemmBasicTypeConfig<PrecType>::BDataType;
using CDataType = typename GemmBasicTypeConfig<PrecType>::CDataType;
using AccDataType = typename GemmBasicTypeConfig<PrecType>::AccDataType;
constexpr int ScaleGranularityM = 1;
constexpr int ScaleGranularityN = 1;
const ck_tile::index_t N = arg_parser.get_int("N");
const 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");
const ck_tile::index_t num_tokens = arg_parser.get_int("NumTokens");
const ck_tile::index_t topk = arg_parser.get_int("TopK");
const ck_tile::index_t warmup = arg_parser.get_int("warmup");
const ck_tile::index_t repeat = arg_parser.get_int("repeat");
const ck_tile::index_t experts = arg_parser.get_int("experts");
// TODO: replace the magic declaration
const ck_tile::index_t MPerBlock = FlatmmConfig::M_Tile;
ck_tile::index_t sorted_tile_num = (num_tokens + MPerBlock - 1) / MPerBlock * MPerBlock * topk;
ck_tile::index_t valid_tile_num = sorted_tile_num;
ck_tile::index_t sorted_size = sorted_tile_num * MPerBlock;
const ck_tile::index_t M = sorted_tile_num * MPerBlock;
const ck_tile::index_t outputN = kind == ck_tile::MoeFlatmmKind::kFFN_gemm1_gate_up ? N / 2 : N;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
constexpr bool IsInputGemm = kind != ck_tile::MoeFlatmmKind::kFFN_gemm2;
stride_A = ck_tile::get_default_stride(
IsInputGemm ? num_tokens : num_tokens * topk, 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(
IsInputGemm ? num_tokens * topk : num_tokens, outputN, stride_C, is_row_major(CLayout{}));
auto a_m_k_tensor = ck_tile::HostTensor<ADataType>(ck_tile::host_tensor_descriptor(
IsInputGemm ? num_tokens : num_tokens * topk, K, stride_A, is_row_major(a_layout)));
auto b_k_n_tensor = ck_tile::HostTensor<BDataType>(
is_row_major(b_layout)
? ck_tile::host_tensor_descriptor(experts * N, K, stride_B, is_row_major(b_layout))
: ck_tile::host_tensor_descriptor(K, experts * N, stride_B, is_row_major(b_layout)));
auto c_m_n_tensor = ck_tile::HostTensor<CDataType>(ck_tile::host_tensor_descriptor(
IsInputGemm ? num_tokens * topk : num_tokens, outputN, stride_C, is_row_major(CLayout{})));
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_m_k_tensor);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_k_n_tensor);
auto b_shuffle_host = shuffle_b<FlatmmConfig>(b_k_n_tensor);
std::cout << "moe_flatmm:" //
<< "\n num_experts: " << experts << "\n num_tokens: " << num_tokens
<< "\n topk: " << topk << "\n sorted_tile_num: " << sorted_tile_num
<< "\n a_m_k: " << a_m_k_tensor.mDesc << "\n b_k_n: " << b_k_n_tensor.mDesc
<< "\n b_shuffle: " << b_shuffle_host.mDesc << "\n c_m_n: " << c_m_n_tensor.mDesc
<< std::endl;
ck_tile::DeviceMem a_m_k_dev_buf{a_m_k_tensor.get_element_space_size_in_bytes()};
ck_tile::DeviceMem b_origin_dev_buf{b_k_n_tensor.get_element_space_size_in_bytes()};
ck_tile::DeviceMem b_shuffle_dev_buf{b_shuffle_host.get_element_space_size_in_bytes()};
ck_tile::DeviceMem c_m_n_dev_buf{c_m_n_tensor.get_element_space_size_in_bytes()};
a_m_k_dev_buf.ToDevice(a_m_k_tensor.data());
b_origin_dev_buf.ToDevice(b_k_n_tensor.data());
b_shuffle_dev_buf.ToDevice(b_shuffle_host.data());
c_m_n_dev_buf.SetZero();
c_m_n_tensor.SetZero();
const void* p_a = a_m_k_dev_buf.GetDeviceBuffer();
const void* p_b_origin = b_origin_dev_buf.GetDeviceBuffer();
const void* p_b_shuffle = b_shuffle_dev_buf.GetDeviceBuffer();
void* p_c = c_m_n_dev_buf.GetDeviceBuffer();
// TODO: malloc and init sorted tokens and max tokens buffer
ck_tile::HostTensor<ck_tile::index_t> expert_ids(
ck_tile::HostTensorDescriptor({sorted_tile_num}, {1}));
ck_tile::HostTensor<ck_tile::index_t> sorted_token_ids(
ck_tile::HostTensorDescriptor({sorted_size}, {1}));
ck_tile::HostTensor<AccDataType> expert_weight(
ck_tile::HostTensorDescriptor({sorted_size}, {1}));
ck_tile::HostTensor<ck_tile::index_t> max_token_id(
ck_tile::HostTensorDescriptor({1 + sorted_tile_num}));
ck_tile::HostTensor<AccDataType> per_token_scale(
ck_tile::HostTensorDescriptor({IsInputGemm ? num_tokens : M}, {1}));
ck_tile::HostTensor<AccDataType> per_channel_scale(
ck_tile::HostTensorDescriptor({N * experts}, {1}));
ck_tile::FillUniformDistribution<AccDataType>{0.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{0.f, 1.f}(per_channel_scale);
// for verification only, no need to satify weight normalization
ck_tile::FillUniformDistribution<AccDataType>{0.0f, 1.0f}(expert_weight);
ck_tile::DeviceMem sorted_token_ids_dev{sorted_token_ids.get_element_space_size_in_bytes()};
ck_tile::DeviceMem expert_ids_dev{expert_ids.get_element_space_size_in_bytes()};
ck_tile::DeviceMem max_token_id_dev{max_token_id.get_element_space_size_in_bytes()};
ck_tile::DeviceMem expert_weight_dev{expert_weight.get_element_space_size_in_bytes()};
ck_tile::DeviceMem per_token_scale_dev_buf(per_token_scale.get_element_space_size_in_bytes());
ck_tile::DeviceMem per_channel_scale_dev_buf(
per_channel_scale.get_element_space_size_in_bytes());
max_token_id.mData = {valid_tile_num * MPerBlock, 0, 1, 2, 3, 4, 6, 7, 8, 8};
// int eids[] = {0, 1, 2, 3, 4, 4, 5, 6, 3, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ((valid_tile_num + experts - 1) / experts);
}
int token_per_tile = (num_tokens * topk + valid_tile_num - 1) / valid_tile_num;
// int token_per_tile = num_tokens * topk / valid_tile_num;
int tokenid = 0;
// sorted_token_ids.mData[0] = 0;
for(int i = 0; i < sorted_tile_num * MPerBlock; i++)
{
int tile_off = i % MPerBlock;
if(tile_off < token_per_tile && tokenid < num_tokens * topk)
{
sorted_token_ids.mData[i] = (tokenid % num_tokens) | ((tokenid / num_tokens) << 24);
tokenid++;
}
else
{
sorted_token_ids.mData[i] = num_tokens;
}
}
sorted_token_ids_dev.ToDevice(sorted_token_ids.data());
expert_ids_dev.ToDevice(expert_ids.data());
max_token_id_dev.ToDevice(max_token_id.data());
expert_weight_dev.ToDevice(expert_weight.data());
per_token_scale_dev_buf.ToDevice(per_token_scale.data());
per_channel_scale_dev_buf.ToDevice(per_channel_scale.data());
const ck_tile::index_t* p_sorted_token_ids_dev =
static_cast<ck_tile::index_t*>(sorted_token_ids_dev.GetDeviceBuffer());
const ck_tile::index_t* p_expert_ids_dev =
static_cast<ck_tile::index_t*>(expert_ids_dev.GetDeviceBuffer());
const ck_tile::index_t* p_max_token_id_dev =
static_cast<ck_tile::index_t*>(max_token_id_dev.GetDeviceBuffer());
const AccDataType* p_sorted_expert_weight_dev =
static_cast<AccDataType*>(expert_weight_dev.GetDeviceBuffer());
using MoeFlatmmArgs =
ck_tile::MoeFlatmmHostArgs<ck_tile::FlatmmScalePointer<ScaleGranularityM>,
ck_tile::FlatmmScalePointer<ScaleGranularityN>>;
auto per_token_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityM>{
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer())};
auto per_channel_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityN>{
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
MoeFlatmmArgs gemm_desc{p_sorted_token_ids_dev,
p_sorted_expert_weight_dev,
p_expert_ids_dev,
p_max_token_id_dev,
p_a,
p_b_shuffle,
p_c,
num_tokens,
experts,
topk,
1, // k_batch
M,
N,
K,
stride_A,
stride_B,
stride_C,
per_token_scale_dev_ptr,
per_channel_scale_dev_ptr};
invoke_moe_gemm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
kind>(warmup, repeat, gemm_desc);
c_m_n_dev_buf.FromDevice(c_m_n_tensor.data());
bool pass{true};
if(arg_parser.get_int("validate"))
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
ck_tile::host_tensor_descriptor(IsInputGemm ? num_tokens * topk : num_tokens,
outputN,
stride_C,
is_row_major(CLayout{})));
c_m_n_host_ref.SetZero();
std::unique_ptr<ck_tile::DeviceMem> c_m_n_ref_buf =
std::make_unique<ck_tile::DeviceMem>(c_m_n_tensor.get_element_space_size_in_bytes());
c_m_n_ref_buf->SetZero();
ck_tile::reference_moe_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout,
static_cast<int>(kind),
ck_tile::moe::MoeSilu>(
p_sorted_token_ids_dev,
p_expert_ids_dev,
p_max_token_id_dev,
static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b_origin),
static_cast<CDataType*>(c_m_n_ref_buf->GetDeviceBuffer()),
p_sorted_expert_weight_dev,
num_tokens,
MPerBlock,
topk,
M,
N,
K,
stride_A,
stride_B,
stride_C,
1,
1,
K,
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()),
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer()));
const float max_accumulated_value =
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, 1 /*kbatch*/, max_accumulated_value);
c_m_n_ref_buf->FromDevice(c_m_n_host_ref.data());
const float rtol = std::is_same_v<ADataType, ck_tile::half_t> && IsInputGemm ? 1e-3 : 1e-2;
const float atol = std::is_same_v<ADataType, ck_tile::half_t> && IsInputGemm ? 1e-3 : 1e-2;
pass = ck_tile::check_err(
c_m_n_tensor, c_m_n_host_ref, "Error: Incorrect results!", rtol, atol);
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
<< std::endl;
std::cout << "The CPU verification result is:" << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}