ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 10:37:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add MoE & FP8 Blockscale WP Kernels for GFX950 (#2297)

Browse Source 
* [fix] align v3 gufusion pipeline

* fix device kernel selection.

* Add .co direct asm support by CK_USE_ASM_MOE_STAGE2_BLOCKSCALE

* experimental optimization for scale load in blkscale gemm

* Add asm for no-loop v3_128x128x128

* fix bugs

* tune fp8 example

* Update v1_128x128x128 to 2x2 instead of 4x1

* wip

* add warmup to asm launch

* wip2

* 16x16 function merged to moe

* temp save, a performant version.

* wip3

* Update .co binary to 16x16

* 16x16x128 correct; 64x64x128 failed

* update

* use mem_op::set when topk=1

* add mx fp8 b_preshuffle support, function not yet tested.

* Spilt the fp4 target. Fix the known bugs. 128x128x128 sanity checked; remove prints

* some fixes

* fix update

* remove some unnecessary hacky; enable 256x256x256 tilesize

* update for function debug

* Add pipeline v3. Have some runtime issue and register spill

* Fix pipe v3 correctness issue

* remove unnecessary hacky

* clang format

* fix a bug

* fix the bug, functional test passed

* tempsave; buggy at passed 4 e8m0 to scaled mfma

* added fp4_bpreshuffle example, build failures

* fixed some bugs

* implement shuffled scale mxfp4gemm, blocker: opsel not effect

* hotfix

* fix bugs, build passed

* (M, N, K)=(128, 128, 128) function failed.

* temp save for gemm1. Function not ready

* fix compile error. Gemm2 pass. Gemm1 WIP

* fix bug for a lds read

* update moe

* Compile pass. Gemm1 function WIP

* update moe

* fix fp8; fix even/odd

* tempsave

* update moe

* Revert "update"

This reverts commit c7d79dcb672616d9bc0fd9958f714fc80e7c84fd.

* Revert "use mem_op::set when topk=1"

This reverts commit 8c7772860735001a51421e7b6d0a28f6676d6c40.

* Add v3 128x128x128_4x4_16x16.co for gfx950

* temp cmake flag suppression  for aiter test

* add code for mxfp4 gemm, blockscale not supported yet

* gemm1 up-only pass. GU WIP

* function pass with inline asm hacky

* revert unexpected file change

* updated and build passed

* update CE elementOP

* added code for debug

* Gemm1 GUFusion function pass. Perf WIP

* Fix fp8/bf8; remove duplicated code

* disable the scheduler in v3; bring it back when compiler feature ready.

* update moe v1 pipeline

* Add gemm1 v1 32x128x128

* remove schedule barrier

* updated

* Fix fp8/bf8 B-row

* mfma using asm, device result correct, host result need to check

* gemm1 v3 64x128x128 debug

* fix cpu ref

* a/b thread_desc stride fix

* Use random scale for init1

* 16x16x128 input size blockscale function passed

* fix blockscale gemm bug

* tempsave. Almost all instances passed.

* v1 fix for mi350.

* temp save

* debug save

* update debug

* fix the bug, 128x128x256 tile function passed

* v3

* rename moe block selector and pipeline

* Add gemm1 v1

* Add gemm1 v1 to selector

* added mx moe block v3 support, function passed

* compile error fix

* Improve the pipeline

* Pack e8m0 as int32_t

* v1 compile pass. Function not ready

* debug synchronize issue over different GPU/ROCm

* minor fix

* Add profiler filter

* Add f4 ckProfiler

* Fix example compile error

* Add f4 profiler examples

* tempsave

* v1 function pass.

* v3 function pass

* align file and function name

* mx_moe_fp4 ready for aiter with clang-format.

* modify the way we represent fp4

* generalize the pipeline scheduling.

* init moe mx f4 scale shuffle

* Cmakelist diable compiler-bound flags

* mx_fp4 default parameter change

* Moe blockscale gemm1&gemm2 asm support for aiter. Suppression cmkae flag til new compler.

* update code

* tempsave; modify the way we represent fp4

* generalize the pipeline scheduling.

* Add gemm1 gfx942 .co support

* updated code, build passed.

* Update gemm2 asm with latest compiler flag

* Fix mx f4 ckProfiler

* Fix blockwise gemm mx v1

* lds conflict free + buffer load lds

* Add gemm2 v3 64x128x128

* fix a, b scale loading bugs, a, b scale loading now correctly

* Add gemm2 v3 64x128x128

* commit with debug info

* fix fp4 profiler

* Add mx fp4 pileline v1 instances

* Fix v2 topk_weight cal. Add silu asm.

* v2 tok_weight WIP

* init mx fp4 B no preshuffle version

* tempsave. compile pass, function wrong

* enable fp4 moe no weigth preshuffle, function pass

* update the TFlops calculation in the example

* Add gemm2 64x128x128 asm. Fix BF16 ref.

* fix 2 typos in fp4_preshuffle

* Better kernel selection in device classes

* correct preShuffleBuffer

we should used packed k to do shuffle.

* lds conflict free + buffer load lds

* optimize offset math in dma

* Fix fp4 ckProfiler

* Fix MX MFMA tests

* fix f4 pipeline issues

* gemm1 func pass

* update mx moe gemm1_bns tile size to 64x128x256

* update mx moe gemm1 gemm2 TF and BW calculation

* fix typo

* temp save

* Fix example_gemm_mx build

* rename the block pipeline

* correct a typo in tail

* Add rotating to mx examples

* fix the correctness issue

* Fix v1; use M padding

* Add NT flag to B/BScale buffer

* Merge gemm_mx_common.hpp

* temp save, 4.4~4.5

* Fix 'Merge gemm_mx_common.hpp'

* refactor the pipeline

* Pad the M for scale buffer unconditionaly

* update MX moe GEMM1 hotloopscheduling

* change the gemm1 tile from 64x128x128 to 128x64x128

* Unconditional Ascale padding

* Pad shuffled a scale only

* pad ascale

* add vmcnt guard for async copy

* Profiler add f4 wp

* Merge preshuffle device

* Add more fp4 wp instances

* Fix do_weight in gemm1. Fix cshuffle_datatype. Clang-format

* Clang-format after 2 merges

* Remove rocm6.3 workaround flags and macro

* Fix fp8 config

* Fix bf8 config

* flag and barrier fix for copmiler branch MainOpSelV3

* Add fp8 profiler instances

* Remove debug infos; Enable flags for blockscale f8

* No asm ver. for merging moe blocksale fp8 into mainline

* update the flag name for f8blockscale

* recover example

* fix performance bug of bpreshuffle f8 gemm

* clang format, remove  single rate mfma restriction for f8

* remove single rate mfma restriction for f8 blockscale gemm

* Fix moe blockscale gemm1 barrier 0x800 for new compiler

* add pipeline v1 for MOE Gemm2

* Use v1 pipeline for example_moe_gemm2_xdl_mx_fp4_bns

* Fix OOB; add MB96 instances

* remove unnecessary files

* fix the cmake issue

* Enable splitk for mxfp4; clang format;

* Generate random tensor values with multiple threads

* Use packed_size_v for A/BPackedSize

* Fix warning

* Fix target_compile_options for disabled target on gfx942

* fix moe pki4 on gfx950

* doc the kGroup definition

* Fix ThreadwiseTensorSliceTransfer_v4::Run (Fuse scale)

* Refactor thread_copy_lds_direct_load; fix gfx942 direct lds load example; fix f16_pki4 example

* Fix unknown compiler flag

* fix two failed examples.

* fix some failure tile size in gfx950 universal gemm. fix test_gemm_fp16

* workaround fix for test_gemm_f32; * We have very limited support for lds direct load if input matrix is not K major

* fix test_gemm_splitk;

* Fix compile for mx_mfma_op

* add mfma selection logic for multipled_v3

* Clean up

* Fix device gemm mx link error

* improve the global atomic pattern

* Revert unnecessary copyright updates

* restore minimum_occupancy logic

* Avoid data race in moe gemm2 ref

* Build fp8 gemm_multiply_multiply and moe only on gfx94/95

* update the instance in device_mx_gemm

* Resolve comments

* Copyright 2025

* Remove unused code

* fix library linking issue

---------

Co-authored-by: OscarXu <huaiguxu@amd.com>
Co-authored-by: lalala-sh <Jiaxing.Wen@amd.com>
Co-authored-by: mtgu0705 <mtgu@amd.com>
Co-authored-by: aska-0096 <haocwang@amd.com>
Co-authored-by: Your Name <you@example.com>
Co-authored-by: valarLip <340077269@qq.com>
Co-authored-by: feifei14119 <feiw@amd.com>
Co-authored-by: Lin, Qun <qlin@amd.com>
Co-authored-by: Andriy Roshchenko <andriy.roshchenko@amd.com>
Co-authored-by: joye <joye@amd.com>
Co-authored-by: asleepzzz <hanwen.chang@amd.com>

[ROCm/composable_kernel commit: 37554c31e8]
This commit is contained in:
Yi DING
2025-06-12 09:25:59 +08:00
committed by GitHub
parent e98451e42f
commit b40267f7a6
85 changed files with 32508 additions and 431 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/01_gemm/gemm_xdl_lds_direct_load_fp16.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
@@ -38,7 +38,7 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>;
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 0, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, 4>;
// clang-format on
#else
// clang-format off

37
example/65_gemm_multiply_multiply/CMakeLists.txt
View File

@@ -1,11 +1,20 @@
add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp16_bpreshuffle gemm_multiply_multiply_xdl_fp16_bpreshuffle.cpp)
add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)
set(EXAMPLE_COMPILE_OPTIONS)
# Open it when SGBPack branch landed on mainline
# list(APPEND EXAMPLE_COMPILE_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --schedmodel=0 -mllvm -misched=gcn-iterative-max-occupancy-experimental")
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
add_example_executable(example_moe_gemm1_xdl_fp8 moe_gemm1_xdl_fp8.cpp)
add_example_executable(example_moe_gemm2_xdl_fp8 moe_gemm2_xdl_fp8.cpp)
add_example_executable(example_moe_gemm2_xdl_fp8_blockscale moe_gemm2_xdl_fp8_blockscale.cpp)
add_example_executable(example_moe_gemm1_xdl_fp8_blockscale moe_gemm1_xdl_fp8_blockscale.cpp)
list(APPEND gpu_list gfx942 gfx950)
set(target 0)
@@ -19,14 +28,32 @@ foreach(gpu IN LISTS GPU_TARGETS)
if(HAS_MAX_ILP_SCHEDULING_STRATEGY)
list(APPEND EXAMPLE_COMPILE_OPTIONS -mllvm --amdgpu-enable-max-ilp-scheduling-strategy=1)
endif()
target_compile_options(example_moe_gemm1_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
target_compile_options(example_moe_gemm2_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
example_compile_options(example_moe_gemm1_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
example_compile_options(example_moe_gemm2_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
endif()
set(GEMM_OPTIONS)
list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
example_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
example_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
set(target 1)
endif()
endforeach()
set(GEMM_OPTIONS)
list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
set(BLOCKSCALE_GEMM_OPTIONS)
list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --schedmodel=0 -mllvm --misched-bottomup=1")
check_cxx_compiler_flag("-mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental " HAS_MAX_OCCUPANCY_EXPERIMENTAL)
if(HAS_MAX_OCCUPANCY_EXPERIMENTAL)
list(APPEND BLOCKSCALE_GEMM_OPTIONS -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental)
endif()
# list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --misched-bottomup=1")
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
example_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
example_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})
example_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})
example_compile_options(example_moe_gemm2_xdl_fp8_blockscale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})
example_compile_options(example_moe_gemm1_xdl_fp8_blockscale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS})

16
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_ab_scale.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -65,14 +65,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_ABScale_
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
16, 128,
256, 16, 16,
128, 128,
128, 16, 16,
16, 16,
1, 2,
S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 2, S<1, 16, 1, 16>, S<8>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>;
4, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 2, S<1, 32, 1, 8>, S<8>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
// clang-format on
int main(int argc, char* argv[])

372
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp Normal file
View File

@@ -0,0 +1,372 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using FP8 = ck::f8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = FP8;
using A1DataType = F32;
using B0DataType = FP8;
using B1DataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using A1Layout = Col;
using B0Layout = Col;
using D0Layout = Row;
using D1Layout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
void preShuffleBuffer(const FP8* src, FP8* dst, int N, int K, int NXdl)
{
int KPack = 16;
int NLane = NXdl;
int KLane = 64 / NLane;
int K0 = K / (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; ++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 + k];
}
}
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t Scale_Block_M = 1;
static constexpr ck::index_t Scale_Block_N = 128;
static constexpr ck::index_t Scale_Block_K = 128;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle
// clang-format off
<Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
128, 128,
128, 16, 16,
16, 16,
8, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 1, S<1, 32, 1, 8>, S<8>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool flush_cache = true;
// GEMM shape
ck::index_t M = 128;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 8)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
flush_cache = std::stoi(argv[7]);
StrideA = K;
StrideB = K;
StrideE = N;
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: M, N, K\n");
printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n");
exit(0);
}
// Transpose the AScale tensor for better performance
ck::index_t Scale_Stride_AK = (M + Scale_Block_M - 1) / Scale_Block_M;
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<A1DataType> a1_m_k(f_host_tensor_descriptor((M + Scale_Block_M - 1) / Scale_Block_M,
(K + Scale_Block_K - 1) / Scale_Block_K,
Scale_Stride_AK,
A1Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B0DataType> b0_preshuffled(
f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K,
(N + Scale_Block_N - 1) / Scale_Block_N,
Scale_Stride_BN,
B0Layout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
case 2:
a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 3:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 4:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
a1_device_buf.ToDevice(a1_m_k.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumDTensor = DsDataType::Size();
// do GEMM
auto device_op = DeviceOpInstance{};
int NPerXdl = device_op.GetPreShuffleParameters();
preShuffleBuffer(b0_k_n.mData.data(), b0_preshuffled.mData.data(), N, K, NPerXdl);
b0_device_buf.ToDevice(b0_preshuffled.mData.data());
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, NumDTensor>{},
StrideE,
a1_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float ave_time = 0.0f;
if(flush_cache)
{
int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype;
ave_time = invoker.Run(argument,
StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf});
}
else
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100});
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
Tensor<float> a_m_k({M, K});
Tensor<float> b_k_n({K, N});
for(int m = 0; m < M; m++)
{
for(int k = 0; k < K; k++)
{
a_m_k(m, k) = ck::type_convert<float>(a0_m_k(m, k)) *
a1_m_k(m / Scale_Block_M, k / Scale_Block_K);
}
}
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
b_k_n(k, n) = ck::type_convert<float>(b0_k_n(k, n)) *
b1_k_n(k / Scale_Block_K, n / Scale_Block_N);
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<float,
float,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
#if 1
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
e_m_n_host_result(m, n) = ck::type_convert<EDataType>(c_m_n(m, n));
}
}
#endif
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2)
? 0
: 1;
}
return 0;
}

10
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -139,13 +139,13 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
// clang-format off
< Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec, 256,
128, 128, 128,
256, 256, 128,
16, 16,
32, 32,
4, 1,
16, 16,
16, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 1, S<1, 32, 1, 8>, S<8, 8, 1>,
2, 1, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
// clang-format on

50
example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp
View File

@@ -158,21 +158,22 @@ using BElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t MPerBlock = 128;
static constexpr ck::index_t MXDLPerWave = 4;
static constexpr ck::index_t NXDLPerWave = 2;
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t NPerBlock = 64;
static constexpr ck::index_t MNPerXDL = 16;
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
static constexpr ck::index_t Nswizzle = false;
static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType);
static constexpr ck::index_t EVec = 16 / sizeof(EDataType);
static constexpr ck::index_t D0Vec = 1;
static constexpr ck::index_t D1Vec = 1;
static constexpr ck::index_t ActOP = 1; // 0: gelu_and_mul, 1: silu_and_mul
static constexpr bool MulRoutedWeight = false;
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t MNPerXDL = 16;
static constexpr ck::index_t MXDLPerWave = MPerBlock / (MNPerXDL * 1);
static constexpr ck::index_t NXDLPerWave = NPerBlock / (MNPerXDL * 4);
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
static constexpr ck::index_t Nswizzle = false;
static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType);
static constexpr ck::index_t EVec = 16 / sizeof(EDataType);
static constexpr ck::index_t D0Vec = 1;
static constexpr ck::index_t D1Vec = 1;
static constexpr ck::index_t ActOP = 1; // 0: gelu_and_mul, 1: silu_and_mul
static constexpr bool MulRoutedWeight = false;
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm
// clang-format off
< Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
@@ -183,15 +184,15 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceM
// mn_perxdl
MNPerXDL, MNPerXDL,
// mn_xdlperwave
MXDLPerWave, NXDLPerWave,
MXDLPerWave, NXDLPerWave,
// a,b: loadtranfer cluster, cluster order, srcorder,VECDIM, srcpervec, dstpervec, lds_extra
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
// CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
2, 2, S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, MulRoutedWeight, true, int32_t, A0DataType>;
2, 2, S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, MulRoutedWeight, true, int32_t, A0DataType>;
// clang-format on
@@ -205,9 +206,9 @@ int main(int argc, char* argv[])
ck::index_t N = 4096;
ck::index_t K = 6144;
ck::index_t experts = 8;
ck::index_t sorted_tile_num = 16;
ck::index_t valid_tile_num = 13;
ck::index_t tokens = 64;
ck::index_t sorted_tile_num = 256;
ck::index_t valid_tile_num = 256;
ck::index_t tokens = 16384;
ck::index_t topk = 2;
if(argc == 1)
@@ -263,11 +264,12 @@ int main(int argc, char* argv[])
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1 + sorted_tile_num}));
max_token_id.mData = {valid_size};
int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3};
// int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3};
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = eids[i];
expert_ids.mData[i] = i / (valid_tile_num / experts);
}
int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
int tokenid = 0;
@@ -307,7 +309,7 @@ int main(int argc, char* argv[])
case 0: break;
case 1:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.1, 0.1});
d0_t_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_e_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});

548
example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8_blockscale.cpp Normal file
View File

@@ -0,0 +1,548 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_moe_gemm_blockscale.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_moe_gemm1_blockscale.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F8 = ck::f8_t;
using F32 = float;
using I64 = int64_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = F8;
using A1DataType = F32;
using B0DataType = F8;
using B1DataType = F32;
// using EDataType = F16;
using EDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = EDataType;
using D2DataType = F32;
using DsDataType = ck::Tuple<D2DataType>;
using A0Layout = Row;
using B0Layout = Col;
using ELayout = Row;
using D0Layout = Row;
using D1Layout = Col;
using D2Layout = ELayout;
using DsLayout = ck::Tuple<D2Layout>;
struct MulABScaleExpertWeight
{
template <typename E, typename C, typename D2>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D2& d2) const;
// for real kernel use
template <>
__host__ __device__ constexpr void
operator()<EDataType, float, float>(EDataType& e, const float& c, const float& d2) const
{
// for real kernel use
(void)d2;
e = ck::type_convert<EDataType>(c);
}
template <>
__host__ __device__ constexpr void
operator()<EDataType, EDataType, float>(EDataType& e, const EDataType& c, const float& d2) const
{
(void)d2;
e = ck::type_convert<EDataType>(c);
}
// for reference cpu
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& e, const float& c, const float& d2) const
{
// for reference cpu
e = ck::type_convert<EDataType>(c * d2);
}
};
void preShuffleBuffer(const B0DataType* src, B0DataType* dst, int N, int K, int NXdl)
{
int KPack = 16 / sizeof(B0DataType);
int NLane = NXdl;
int KLane = 64 / NLane;
int K0 = K / (KLane * KPack);
// K -> K0 KLane KPack
// N -> N0 NLane
// N, K -> N0 K0 KLane NLane KPack
int tempk;
for(I64 n = 0; n < N; ++n)
{
for(I64 k = 0; k < K; ++k)
{
I64 n0 = n / NLane;
I64 n1 = n % NLane;
I64 k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
I64 k1 = tempk / KPack;
I64 k2 = tempk % KPack;
I64 outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
k1 * KPack * NLane + n1 * KPack + k2;
dst[outputIndex] = src[n * static_cast<I64>(K) + k];
}
}
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MulABScaleExpertWeight;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t Scale_Block_M = 1;
static constexpr ck::index_t Scale_Block_N = 128;
static constexpr ck::index_t Scale_Block_K = 128;
static constexpr ck::index_t Nswizzle = false;
static constexpr ck::index_t ActOP = 0; // 0: gelu_and_mul, 1: silu_and_mul
static constexpr bool MulRoutedWeight = true;
#if 0
static constexpr ck::index_t MPerBlock = 32;
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t MNPerXDL = 16;
static constexpr ck::index_t MXDLPerWave = MPerBlock / (MNPerXDL * 1);
static constexpr ck::index_t NXDLPerWave = NPerBlock / (MNPerXDL * 4);
static constexpr ck::index_t CShuffleMXDLPerWave = MXDLPerWave;
static constexpr ck::index_t CShuffleNXDLPerWave = NXDLPerWave;
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType);
static constexpr ck::index_t EVec = 16 / sizeof(EDataType);
static constexpr ck::index_t D0Vec = 1;
static constexpr ck::index_t D1Vec = 1;
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale
// clang-format off
< Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
//threadnum, mblock, nblock, kblock
BLOCKSIZE, Scale_Block_M, Scale_Block_N, Scale_Block_K,
MPerBlock, NPerBlock, KPerBlock,
// ak1, bk1
AK1, BK1,
// mn_perxdl
MNPerXDL, MNPerXDL,
// mn_xdlperwave
MXDLPerWave, NXDLPerWave,
// a,b: loadtranfer cluster, cluster order, srcorder,VECDIM, srcpervec, dstpervec, lds_extra
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
// CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
CShuffleMXDLPerWave, CShuffleNXDLPerWave, S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, MulRoutedWeight, int32_t, A0DataType>;
#else
static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
MPerBlock, 128, 128,
16, 16,
16, 16,
4, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
4, 2, S<1, 32, 1, 8>, S<2, 1, 1, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, MulRoutedWeight, int32_t, A0DataType>;
#endif
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
#if 1
// GEMM shape
ck::index_t N = 4096;
ck::index_t K = 6144;
ck::index_t experts = 8;
ck::index_t topk = 2;
// ck::index_t sorted_tile_num = 515;
// ck::index_t valid_tile_num = 512;
// ck::index_t tokens = 8192;
// ck::index_t sorted_tile_num = 15;
// ck::index_t valid_tile_num = 13;
ck::index_t sorted_tile_num = 259;
ck::index_t valid_tile_num = 256;
ck::index_t tokens = 4096;
#else
// deepseek
ck::index_t N = 2048;
ck::index_t K = 7168;
ck::index_t experts = 256;
ck::index_t topk = 8;
ck::index_t tokens = 4096;
ck::index_t sorted_tile_num = 261;
ck::index_t valid_tile_num = 256;
#endif
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
// use default case
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 7)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
tokens = std::stoi(argv[6]);
}
else if(argc == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
tokens = std::stoi(argv[6]);
sorted_tile_num = std::stoi(argv[7]);
valid_tile_num = std::stoi(argv[8]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: N, K, tokens\n");
exit(0);
}
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
ck::index_t valid_size = valid_tile_num * MPerBlock;
if(tokens * topk > valid_size)
{
printf("err config, tokens * topk > valid_size\n");
exit(-1);
}
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
constexpr ck::index_t NumDTensor = DsDataType::Size();
constexpr auto StrideDs = std::array<ck::index_t, NumDTensor>{0};
ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K;
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
ck::index_t Scale_Stride_B = (N + Scale_Block_N - 1) / Scale_Block_N * 2;
ck::index_t KBatch = 1;
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1 + sorted_tile_num}));
max_token_id.mData = {valid_size};
// int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3};
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);
}
int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
int tokenid = 0;
for(int i = 0; i < sorted_size; i++)
{
int tile_off = i % MPerBlock;
if(tile_off < token_per_tile && tokenid < tokens * topk)
{
sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
tokenid++;
}
else
{
sorted_token_ids.mData[i] = tokens;
}
}
Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
Tensor<A1DataType> a1_t_k(HostTensorDescriptor(
{tokens, (K + Scale_Block_K - 1) / Scale_Block_K}, {Scale_Stride_AM, 1}));
Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}));
Tensor<B1DataType> b1_e_n_k(
HostTensorDescriptor({experts,
(K + Scale_Block_K - 1) / Scale_Block_K,
(N + Scale_Block_N - 1) / Scale_Block_N * 2},
{(Scale_Stride_B * Scale_Stride_BN), 1, Scale_Stride_BN}));
Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}));
Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
Tensor<EDataType> e_t_n_host_result(HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
Tensor<EDataType> e_t_n_device_result(
HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
e_t_n_device_result.SetZero();
std::cout << "a0_t_k: " << a0_t_k.mDesc << std::endl;
std::cout << "a1_t_k: " << a1_t_k.mDesc << std::endl;
std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
break;
case 2:
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
a1_t_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 3:
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
break;
case 4:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
break;
case 5:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
a1_t_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
break;
case 6:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
break;
default:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
}
DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) *
sorted_token_ids.mDesc.GetElementSpaceSize());
DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.mDesc.GetElementSpaceSize());
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_t_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_e_n_k.mDesc.GetElementSpaceSize());
DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.mDesc.GetElementSpaceSize());
sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
expert_ids_dev.ToDevice(expert_ids.mData.data());
max_token_id_dev.ToDevice(max_token_id.mData.data());
a0_device_buf.ToDevice(a0_t_k.mData.data());
a1_device_buf.ToDevice(a1_t_k.mData.data());
b1_device_buf.ToDevice(b1_e_n_k.mData.data());
d2_device_buf.ToDevice(d2_e_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
int NPerXdl = device_op.GetPreShuffleParameters();
preShuffleBuffer(
b0_e_n_k.mData.data(), b0_preshuffled.mData.data(), N * 2 * experts, K, NPerXdl);
b0_device_buf.ToDevice(b0_preshuffled.mData.data());
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(sorted_token_ids_dev.GetDeviceBuffer(),
expert_ids_dev.GetDeviceBuffer(),
max_token_id_dev.GetDeviceBuffer(),
a0_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{d2_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
tokens,
topk,
sorted_size,
N,
K,
StrideA,
StrideB,
StrideDs,
StrideE,
a1_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
KBatch,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * tokens * topk * N * 2 * K;
std::size_t num_btype = sizeof(A0DataType) * valid_tile_num * K +
sizeof(B0DataType) * K * N * 2 * experts +
sizeof(EDataType) * valid_tile_num * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s.\n"
<< device_op.GetTypeString() << std::endl;
}
if(do_verification)
{
invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
Tensor<float> a_t_k({tokens, K});
Tensor<float> b_e_n_k({experts, K, N * 2});
e_device_buf.FromDevice(e_t_n_device_result.mData.data());
Tensor<float> c_t_k_n({tokens, topk, N}, {topk * N, N, 1});
// handle scale before ref.
for(int t = 0; t < tokens; ++t)
{
for(int k = 0; k < K; ++k)
{
a_t_k(t, k) = ck::type_convert<float>(a0_t_k(t, k)) * a1_t_k(t, k / Scale_Block_K);
}
}
for(int e = 0; e < experts; ++e)
{
for(int k = 0; k < K; ++k)
{
for(int n = 0; n < N * 2; ++n)
{
b_e_n_k(e, k, n) = ck::type_convert<float>(b0_e_n_k(e, k, n)) *
b1_e_n_k(e, k / Scale_Block_K, n / Scale_Block_N);
}
}
}
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceMoeGemm1BlockScale<float,
float,
float,
D2DataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough,
ActOP,
MulRoutedWeight>;
auto ref_moe_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_moe_gemm.MakeInvoker();
auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
expert_ids,
max_token_id,
MPerBlock,
a_t_k,
b_e_n_k,
d2_e_n,
c_t_k_n,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < valid_size; ++m)
{
const int fuse_t = sorted_token_ids.mData[m];
const int t = fuse_t & 0xffffff;
const int topk_id = (fuse_t & 0xff000000) >> 24;
if(t >= tokens)
{
continue;
}
for(int n = 0; n < N; ++n)
{
e_t_n_host_result(t, topk_id, n) =
ck::type_convert<EDataType>(c_t_k_n(t, topk_id, n));
}
}
e_device_buf.FromDevice(e_t_n_device_result.mData.data());
auto status =
ck::utils::check_err(
e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-1)
? 0
: 1;
if(status == 0)
{
printf("Validation Pass.\n");
}
return status;
}
return 0;
}

29
example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp
View File

@@ -123,11 +123,11 @@ using BElementOp = PassThrough;
using CDEElementOp = MulABScaleExpertWeight;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t MPerBlock = 128;
static constexpr ck::index_t MPerBlock = 256;
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t MXDLPerWave = 4;
static constexpr ck::index_t MXDLPerWave = 16;
static constexpr ck::index_t NXDLPerWave = 4;
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t NPerBlock = 256;
static constexpr ck::index_t MNPerXDL = 16;
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
@@ -164,12 +164,12 @@ using DeviceOpInstance = ck::tensor_operation::device::Devic
// S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0,
// S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
// CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
4, 2, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, MulRoutedWeight, false, int32_t, A0DataType>;
2, 2, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, MulRoutedWeight, false, int32_t, A0DataType>;
// kernel 2: 128->32x128x128
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, EDataType>;
@@ -186,11 +186,11 @@ int main(int argc, char* argv[])
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t experts = 8;
ck::index_t sorted_tile_num = 16;
ck::index_t valid_tile_num = 13;
ck::index_t sorted_tile_num = 133;
ck::index_t valid_tile_num = 128;
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
ck::index_t valid_size = valid_tile_num * MPerBlock;
ck::index_t tokens = 128;
ck::index_t tokens = 16384;
ck::index_t topk = 2;
if(argc == 1)
@@ -245,13 +245,14 @@ int main(int argc, char* argv[])
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
// max_token_id.mData[0] = valid_size;
// max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
// int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
max_token_id.mData = {valid_size, 0, 1, 2, 3, 4, 5, 6, 7, 8};
// int eids[] = {0, 1, 2, 3, 4, 5, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = eids[i];
expert_ids.mData[i] = i / ((valid_tile_num + experts - 1) / experts);
}
if(tokens * topk > valid_size)
{

541
example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8_blockscale.cpp Normal file
View File

@@ -0,0 +1,541 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_moe_gemm_blockscale.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_moe_gemm2_blockscale.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F8 = ck::f8_t;
using F32 = float;
using I64 = int64_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = F8;
using A1DataType = F32;
using B0DataType = F8;
using B1DataType = F32;
using EDataType = F16;
// using EDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = EDataType;
using D2DataType = F32;
using DsDataType = ck::Tuple<D2DataType>;
using A0Layout = Row;
using B0Layout = Col;
using ELayout = Row;
using D0Layout = Row;
using D1Layout = Col;
using D2Layout = ELayout;
// using DsLayoutGate = ck::Tuple<D0Layout, D1Layout>;
using DsLayout = ck::Tuple<D2Layout>;
// d0: ascale, d1: bscale, d2:expert weight
struct MulABScaleExpertWeight
{
template <typename E, typename C, typename D2>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D2& d2) const;
// for real kernel use
template <>
__host__ __device__ constexpr void
operator()<EDataType, EDataType, float>(EDataType& e, const EDataType& c, const float& d2) const
{
// for real kernel use
(void)d2;
e = ck::type_convert<EDataType>(c);
}
template <>
__host__ __device__ constexpr void
operator()<EDataType, float, float>(EDataType& e, const float& c, const float& d2) const
{
// for real kernel use
(void)d2;
e = ck::type_convert<EDataType>(c);
}
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& e, const float& c, const float& d2) const
{
// for reference cpu
e = ck::type_convert<EDataType>(c * d2);
}
};
void preShuffleBuffer(const B0DataType* src, B0DataType* dst, int N, int K, int NXdl)
{
int KPack = 16 / sizeof(B0DataType);
int NLane = NXdl;
int KLane = 64 / NLane;
int K0 = K / (KLane * KPack);
// K -> K0 KLane KPack
// N -> N0 NLane
// N, K -> N0 K0 KLane NLane KPack
int tempk;
for(I64 n = 0; n < N; ++n)
{
for(I64 k = 0; k < K; ++k)
{
I64 n0 = n / NLane;
I64 n1 = n % NLane;
I64 k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
I64 k1 = tempk / KPack;
I64 k2 = tempk % KPack;
I64 outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
k1 * KPack * NLane + n1 * KPack + k2;
dst[outputIndex] = src[n * static_cast<I64>(K) + k];
}
}
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MulABScaleExpertWeight;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t Scale_Block_M = 1;
static constexpr ck::index_t Scale_Block_N = 128;
static constexpr ck::index_t Scale_Block_K = 128;
static constexpr bool MulRoutedWeight = true;
#if 0
static constexpr ck::index_t MPerBlock = 32;
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t MXDLPerWave = 2;
static constexpr ck::index_t NXDLPerWave = 2;
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t MNPerXDL = 16;
static constexpr ck::index_t KPerBlock = 256 / sizeof(A0DataType);
static constexpr ck::index_t CShuffleNLane = 16;
static constexpr ck::index_t CShuffleMLane = BLOCKSIZE / CShuffleNLane;
static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType);
static constexpr ck::index_t EVec = 2;
static constexpr ck::index_t D0Vec = 1;
static constexpr ck::index_t D1Vec = 1;
static constexpr ck::index_t D2Vec = 1;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
BLOCKSIZE, Scale_Block_M, Scale_Block_N, Scale_Block_K,
MPerBlock, NPerBlock, KPerBlock,
AK1, BK1,
MNPerXDL, MNPerXDL,
MXDLPerWave, NXDLPerWave,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
2, 2, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, MulRoutedWeight, int32_t, A0DataType>;
#else
static constexpr ck::index_t MPerBlock = 64; using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmBlockScale<
Row, Col, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
256, Scale_Block_M, Scale_Block_N, Scale_Block_K,
MPerBlock, 128, 128,
16, 16,
16, 16,
4, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 2, S<1, 32, 1, 8>, S<2, 1, 1, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, MulRoutedWeight, int32_t, A0DataType>;
#endif
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
// tokens = 1
// topk = 1
// experts = 8
// per expert:
constexpr ck::index_t valid_tile_num =
26; // 13 for 128; 52 for 32; 4096 for ds // > token * topk / MPerBlock
constexpr ck::index_t sorted_tile_num = valid_tile_num + 3;
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
ck::index_t valid_size = valid_tile_num * MPerBlock;
#if 1
// GEMM shape
ck::index_t N = 6144;
ck::index_t K = 4096;
ck::index_t experts = 8;
ck::index_t tokens = 832;
ck::index_t topk = 2;
#else
// deepseek
ck::index_t N = 2048;
ck::index_t K = 7160;
ck::index_t experts = 256;
ck::index_t tokens = 1;
ck::index_t topk = 8;
#endif
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
// use default case
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 7)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
tokens = std::stoi(argv[6]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: N, K, tokens\n");
exit(0);
}
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
constexpr ck::index_t NumDTensor = DsDataType::Size();
constexpr auto StrideDs = std::array<ck::index_t, NumDTensor>{0};
ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K;
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
ck::index_t Scale_Stride_B = (N + Scale_Block_N - 1) / Scale_Block_N;
ck::index_t KBatch = 1;
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
max_token_id.mData = {valid_size, 0, 1, 2, 3, 4, 5, 6, 7, 8};
// int eids[] = {0, 1, 3, 3, 3};
// int eids[] = {0, 1, 2, 3, 4, 5, 6, 7}; //, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
// int eids[] = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 3, 3, 3};
// int eids[] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
// 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
// 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
// 3, 3, 3, 3, 3, 3, 3, 3, 4, 4,
// 5, 5, 5, 5, 6, 6, 6, 6, 7, 7,
// 7, 7,
// 3, 3, 3};
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);
}
if(tokens * topk > valid_size)
{
printf("err config, tokens * topk > valid_size\n");
exit(-1);
}
int token_per_tile = tokens * topk / valid_tile_num;
int tokenid = 0;
for(int i = 0; i < sorted_size; i++)
{
int tile_off = i % MPerBlock;
if(tile_off < token_per_tile && tokenid < tokens * topk)
{
sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
tokenid++;
}
else
{
sorted_token_ids.mData[i] = tokens;
}
}
Tensor<A0DataType> a0_t_k_k(HostTensorDescriptor({tokens, topk, K}, {topk * K, K, 1}));
Tensor<A1DataType> a1_t_k_k(
HostTensorDescriptor({tokens, topk, (K + Scale_Block_K - 1) / Scale_Block_K},
{(topk * Scale_Stride_AM), Scale_Stride_AM, 1}));
Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
Tensor<B1DataType> b1_e_n_k(HostTensorDescriptor(
{experts, (K + Scale_Block_K - 1) / Scale_Block_K, (N + Scale_Block_N - 1) / Scale_Block_N},
{(Scale_Stride_B * Scale_Stride_BN), 1, Scale_Stride_BN}));
Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
Tensor<EDataType> e_t_n_host_result(HostTensorDescriptor({tokens, N}, {N, 1}));
Tensor<EDataType> e_t_n_device_result(HostTensorDescriptor({tokens, N}, {N, 1}));
e_t_n_device_result.SetZero();
std::cout << "a0_t_k_k: " << a0_t_k_k.mDesc << std::endl;
std::cout << "a1_t_k_k: " << a1_t_k_k.mDesc << std::endl;
std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-1.0, 1.0});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-1.0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
break;
case 2:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 3:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 4:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
break;
case 5:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
break;
case 6:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{1.0, 1.0});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{1.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{1.0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{1.0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{1.0, 1.0});
for(auto i = 0; i < N * K; i++)
{
b0_e_n_k.mData[i] = ck::type_convert<B0DataType>(static_cast<float>(0.1));
b0_e_n_k.mData[i + N * K] = ck::type_convert<B0DataType>(static_cast<float>(0.2));
}
break;
default:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
}
DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) *
sorted_token_ids.mDesc.GetElementSpaceSize());
DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.mDesc.GetElementSpaceSize());
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k_k.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_t_k_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_e_n_k.mDesc.GetElementSpaceSize());
DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.mDesc.GetElementSpaceSize());
sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
expert_ids_dev.ToDevice(expert_ids.mData.data());
max_token_id_dev.ToDevice(max_token_id.mData.data());
a0_device_buf.ToDevice(a0_t_k_k.mData.data());
a1_device_buf.ToDevice(a1_t_k_k.mData.data());
b1_device_buf.ToDevice(b1_e_n_k.mData.data());
d2_device_buf.ToDevice(d2_e_n.mData.data());
e_device_buf.ToDevice(e_t_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
int NPerXdl = device_op.GetPreShuffleParameters();
preShuffleBuffer(b0_e_n_k.mData.data(), b0_preshuffled.mData.data(), N * experts, K, NPerXdl);
b0_device_buf.ToDevice(b0_preshuffled.mData.data());
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(sorted_token_ids_dev.GetDeviceBuffer(),
expert_ids_dev.GetDeviceBuffer(),
max_token_id_dev.GetDeviceBuffer(),
a0_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{d2_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
tokens,
topk,
sorted_size,
N,
K,
StrideA,
StrideB,
StrideDs,
StrideE,
a1_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
KBatch,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
if(time_kernel)
{
// not result correct here because output buf not setzero
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * tokens * topk * N * K;
std::size_t num_btype = sizeof(A0DataType) * tokens * K * topk +
sizeof(B0DataType) * K * N * experts +
sizeof(EDataType) * tokens * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s.\n"
<< device_op.GetTypeString() << std::endl;
}
if(do_verification)
{
// gemm2 use atomic, so need to reinit outputs
e_device_buf.ToDevice(e_t_n_device_result.mData.data());
invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
Tensor<float> a_t_k_k({tokens, topk, K});
Tensor<float> b_e_n_k({experts, K, N});
Tensor<float> c_t_n({tokens, N});
for(int t = 0; t < tokens; ++t)
{
for(int tk = 0; tk < topk; ++tk)
{
for(int k = 0; k < K; ++k)
{
a_t_k_k(t, tk, k) = ck::type_convert<float>(a0_t_k_k(t, tk, k)) *
a1_t_k_k(t, tk, k / Scale_Block_K);
}
}
}
for(int e = 0; e < experts; ++e)
{
for(int k = 0; k < K; ++k)
{
for(int n = 0; n < N; ++n)
{
b_e_n_k(e, k, n) = ck::type_convert<float>(b0_e_n_k(e, k, n)) *
b1_e_n_k(e, k / Scale_Block_K, n / Scale_Block_N);
}
}
}
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceMoeGemm2BlockScale<float,
float,
float,
D2DataType,
AccDataType,
PassThrough,
PassThrough,
CDEElementOp,
MulRoutedWeight>;
auto ref_moe_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_moe_gemm.MakeInvoker();
auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
expert_ids,
max_token_id,
MPerBlock,
a_t_k_k,
b_e_n_k,
d2_e_n,
c_t_n,
PassThrough{},
PassThrough{},
cde_element_op);
ref_invoker.Run(ref_argument);
for(int t = 0; t < tokens; ++t)
{
for(int n = 0; n < N; ++n)
{
e_t_n_host_result(t, n) = ck::type_convert<EDataType>(c_t_n(t, n));
}
}
e_device_buf.FromDevice(e_t_n_device_result.mData.data());
auto status =
ck::utils::check_err(
e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-2)
? 0
: 1;
if(status == 0)
{
printf("Validation Pass.\n");
}
return status;
}
return 0;
}

30
example/67_gemm_microscaling/CMakeLists.txt
View File

@@ -6,8 +6,9 @@ add_example_dependencies(example_gemm_mx example_gemm_mx_fp8)
add_example_executable(example_gemm_mx_bf8 gemm_mx_bf8.cpp)
add_example_dependencies(example_gemm_mx example_gemm_mx_bf8)
#add_example_executable(example_gemm_mx_fp8_bf8 gemm_mx_fp8_bf8.cpp)
# add_example_dependencies(example_gemm_mx example_gemm_mx_fp8_bf8) TOFO: Fix RRR
# TODO: Fix RRR
# add_example_executable(example_gemm_mx_fp8_bf8 gemm_mx_fp8_bf8.cpp)
# add_example_dependencies(example_gemm_mx example_gemm_mx_fp8_bf8)
add_example_executable(example_gemm_mx_fp4 gemm_mx_fp4.cpp)
add_example_dependencies(example_gemm_mx example_gemm_mx_fp4)
@@ -15,30 +16,23 @@ add_example_dependencies(example_gemm_mx example_gemm_mx_fp4)
add_example_executable(example_gemm_mx_fp4_bpreshuffle gemm_mx_fp4_bpreshuffle.cpp)
add_example_dependencies(example_gemm_mx example_gemm_mx_fp4_bpreshuffle)
#add_example_executable(example_moe_gemm1_xdl_mx_fp4 moe_gemm1_xdl_mx_fp4.cpp)
# add_example_dependencies(example_gemm_mx example_moe_gemm1_xdl_mx_fp4) TODO: Fix
add_example_executable(example_moe_gemm1_xdl_mx_fp4_bns moe_gemm1_xdl_mx_fp4_bns.cpp)
add_example_dependencies(example_gemm_mx example_moe_gemm1_xdl_mx_fp4_bns)
#add_example_executable(example_moe_gemm1_xdl_mx_fp4_bns moe_gemm1_xdl_mx_fp4_bns.cpp)
#add_example_dependencies(example_gemm_mx example_moe_gemm1_xdl_mx_fp4_bns)
#add_example_executable(example_moe_gemm2_xdl_mx_fp4 moe_gemm2_xdl_mx_fp4.cpp)
# add_example_dependencies(example_gemm_mx example_moe_gemm2_xdl_mx_fp4) TODO: Fix
#add_example_executable(example_moe_gemm2_xdl_mx_fp4_bns moe_gemm2_xdl_mx_fp4_bns.cpp)
#add_example_dependencies(example_gemm_mx example_moe_gemm2_xdl_mx_fp4_bns)
add_example_executable(example_moe_gemm2_xdl_mx_fp4_bns moe_gemm2_xdl_mx_fp4_bns.cpp)
add_example_dependencies(example_gemm_mx example_moe_gemm2_xdl_mx_fp4_bns)
set(FP4_MXGEMM_OPTIONS)
list(APPEND FP4_MXGEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --amdgpu-use-amdgpu-trackers=1")
#list(APPEND FP4_MXGEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker -ftemplate-backtrace-limit=0)
example_compile_options(example_gemm_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
example_compile_options(example_gemm_mx_fp4_bpreshuffle PRIVATE ${FP4_MXGEMM_OPTIONS})
# example_compile_options(example_moe_gemm1_xdl_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
# example_compile_options(example_moe_gemm2_xdl_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
# example_compile_options(example_moe_gemm1_xdl_mx_fp4_bns PRIVATE ${FP4_MXGEMM_OPTIONS})
# example_compile_options(example_moe_gemm2_xdl_mx_fp4_bns PRIVATE ${FP4_MXGEMM_OPTIONS})
example_compile_options(example_moe_gemm1_xdl_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
example_compile_options(example_moe_gemm2_xdl_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
example_compile_options(example_moe_gemm1_xdl_mx_fp4_bns PRIVATE ${FP4_MXGEMM_OPTIONS})
example_compile_options(example_moe_gemm2_xdl_mx_fp4_bns PRIVATE ${FP4_MXGEMM_OPTIONS})
set(FP8_MXGEMM_OPTIONS)
list(APPEND FP8_MXGEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
#list(APPEND FP8_MXGEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker -ftemplate-backtrace-limit=0)
example_compile_options(example_gemm_mx_fp8 PRIVATE ${FP8_MXGEMM_OPTIONS})
example_compile_options(example_gemm_mx_bf8 PRIVATE ${FP8_MXGEMM_OPTIONS})

34
example/67_gemm_microscaling/gemm_mx_common.hpp
View File

@@ -250,7 +250,7 @@ bool run_mx_gemm(const ProblemSizeSplitK& problem_size, const ExecutionConfig& c
using AScaleLayout = Row;
using BScaleLayout = Col;
auto Scale_Padded_M = (M + ScaleBlockSize - 1) / ScaleBlockSize * ScaleBlockSize;
auto Scale_Padded_M = ck::math::integer_least_multiple(M, ScaleBlockSize);
auto Scale_Stride_AM =
f_get_default_stride(Scale_Padded_M, K / ScaleBlockSize, -1, AScaleLayout{});
auto Scale_Stride_BN = f_get_default_stride(K / ScaleBlockSize, N, -1, BScaleLayout{});
@@ -302,6 +302,8 @@ bool run_mx_gemm(const ProblemSizeSplitK& problem_size, const ExecutionConfig& c
return ck::type_convert<BDataType>(x);
};
using int_distr = std::uniform_int_distribution<int>;
using float_distr = std::uniform_real_distribution<float>;
switch(config.init_method)
{
case 0: // Initializations for development and debugging
@@ -320,22 +322,19 @@ bool run_mx_gemm(const ProblemSizeSplitK& problem_size, const ExecutionConfig& c
break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 6}); // Z[-5,5]
b_k_n->GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 6}); // Z[-5,5]
a_m_k.GenerateTensorDistr(int_distr{-5, 6}); // Z[-5,5]
b_k_n->GenerateTensorDistr(int_distr{-5, 6}); // Z[-5,5]
static_assert(ck::is_same_v<XDataType, ck::e8m0_bexp_t>);
a_m_k_scale.GenerateTensorValue(
GeneratorTensor_2<XDataType>{120, 129}); // scales: {0.25, 0.5, 1, 2}
b_k_n_scale.GenerateTensorValue(
GeneratorTensor_2<XDataType>{125, 129}); // scales: {0.25, 0.5, 1, 2}
a_m_k_scale.GenerateTensorDistr(int_distr{120, 129}); // scales: {0.25, 0.5, 1, 2}
b_k_n_scale.GenerateTensorDistr(int_distr{125, 129}); // scales: {0.25, 0.5, 1, 2}
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-2.0, 2.0});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<XDataType>{powf(2.0f, -125.0f), 1.0f});
a_m_k.GenerateTensorDistr(float_distr{-2.0, 2.0});
a_m_k_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
b_k_n->GenerateTensorValue(GeneratorTensor_3<BDataType>{-2.0, 2.0});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<XDataType>{powf(2.0f, -125.0f), 1.0f});
b_k_n->GenerateTensorDistr(float_distr{-2.0, 2.0});
b_k_n_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
break;
default:
@@ -469,17 +468,6 @@ bool run_mx_gemm(const ProblemSizeSplitK& problem_size, const ExecutionConfig& c
std::cout << "Comparing results..." << std::endl;
}
// if(config.init_method == 0)
// {
// auto expected = static_cast<float>(K);
// auto computed = type_convert<float>(c_m_n_device_result(1, 12));
// res_verified = res_verified && std::abs(expected - computed) <= 0.0f;
// std::cout << "\nExpected vs Computed: " << expected << " vs " << computed
// << ((res_verified) ? " (PASSED!)" : " (FAILED!)") << std::endl
// << std::endl;
// }
res_verified =
res_verified &&
ck::utils::check_err(

2
example/67_gemm_microscaling/gemm_mx_fp4.cpp
View File

@@ -5,8 +5,6 @@
using ADataType = ck::f4x2_pk_t;
using BDataType = ck::f4x2_pk_t;
// using ADataType = ck::f4_t;
// using BDataType = ck::f4_t;
using XDataType = ck::e8m0_bexp_t;
using XPackedDataType = int32_t;

8
example/67_gemm_microscaling/gemm_mx_fp4_bpreshuffle.cpp
View File

@@ -5,8 +5,6 @@
using ADataType = ck::f4x2_pk_t;
using BDataType = ck::f4x2_pk_t;
// using ADataType = ck::f4_t;
// using BDataType = ck::f4_t;
using XDataType = ck::e8m0_bexp_t;
using XPackedDataType = int32_t;
@@ -74,9 +72,9 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMX_Xdl_CShuffle
16, // BBlockTransferDstScalarPerVector_BK1
true, // BBlockLdsExtraN
2, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4, // CShuffleNXdlPerWavePerShuffle
S<1, 8, 1, 32>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlockW
BlkGemmPSched, // BlkGemmPipeSched
BlkGemmPVer, // BlkGemmPipelineVer
ADataType, // ComputeTypeA

545
example/67_gemm_microscaling/moe_gemm1_xdl_mx_fp4_bns.cpp Normal file
View File

@@ -0,0 +1,545 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm_bns.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm1.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F4 = ck::f4x2_pk_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using XDataType = ck::e8m0_bexp_t;
using XPackedDataType = int32_t; // 4 packed e8m0_bexp_t
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = F4;
using A1DataType = XPackedDataType;
using B0DataType = F4;
using B1DataType = XPackedDataType;
using EDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F32;
using D1DataType = F32;
using D2DataType = F32;
using DsDataType = ck::Tuple<D0DataType, D1DataType, D2DataType>;
using A0Layout = Row;
using B0Layout = Col;
using ELayout = Row;
using D0Layout = Row;
using D1Layout = Col;
using D2Layout = ELayout;
using DsLayout = ck::Tuple<D0Layout, D1Layout, D2Layout>;
// d0: ascale, d1: bscale, d2:expert weight
struct MulABScaleExpertWeight
{
template <typename E, typename C, typename D0, typename D1, typename D2>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1, const D2& d2) const;
// for real kernel use
template <>
__host__ __device__ constexpr void operator()<EDataType, float, float, float, float>(
EDataType& e, const float& c, const float& d0, const float& d1, const float& d2) const
{
(void)d0;
(void)d1;
(void)d2;
e = ck::type_convert<EDataType>(c);
}
// for reference cpu
template <>
__host__ __device__ constexpr void operator()<float, float, float, float, float>(
float& e, const float& c, const float& d0, const float& d1, const float& d2) const
{
// for reference cpu
(void)d0;
(void)d1;
(void)d2;
e = ck::type_convert<EDataType>(c);
}
};
using CDEElementOp = MulABScaleExpertWeight;
// A, B Scale preshuffle
template <bool KLast>
void preShuffleScaleBuffer(ck::e8m0_bexp_t* src, ck::e8m0_bexp_t* dst, int MN, int K)
{
int MNXdlPack = 2;
int KXdlPack = 2;
int XdlMNThread = 16;
int XdlKThread = 64 / XdlMNThread;
int K0 = K / KXdlPack / XdlKThread; // KRepeat
// The 4 16x128 building blocks will be packed into 1 32x256 for F4
// The 8 16x16x128 mfma will be packed into 1 32x32x256 for F4
// unfold the MN32xK(256/32) scale buffer
// 4 16 2 2
// To XdlKThread-> XdlMNThread -> KXdlPack -> MNXdlPack
// Then, MNRepeat->KRepeat
for(int n = 0; n < MN; ++n)
{
for(int k = 0; k < K; ++k)
{
int n0 = n / (XdlMNThread * MNXdlPack); // i MNRepeat
int tempn = n % (XdlMNThread * MNXdlPack);
int n1 = tempn % XdlMNThread; // i XdlMNThread
int n2 = tempn / XdlMNThread; // i MNXdlPack
int k0 = k / (XdlKThread * KXdlPack); // i KRepeat
int tempk = k % (XdlKThread * KXdlPack);
int k1 = tempk % XdlKThread; // i XdlKThread
int k2 = tempk / XdlKThread; // i KXdlPack
int outputIndex = n0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread * K0 +
k0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread +
k1 * MNXdlPack * KXdlPack * XdlMNThread + n1 * MNXdlPack * KXdlPack +
k2 * MNXdlPack + n2;
// src[n * K + k] = ck::type_convert<ck::e8m0_bexp_t>(static_cast<float>(powf(2.0f, n2 +
// k2 * MNXdlPack)));
if constexpr(KLast)
dst[outputIndex] = src[n * K + k];
else
dst[outputIndex] = src[k * MN + n];
}
}
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MulABScaleExpertWeight;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
constexpr ck::index_t DataPackedSize = 2; // Packed representation of data
constexpr ck::index_t ScaleBlockSize = 32; // scaling block size
constexpr ck::index_t KPerBlock = 256 / DataPackedSize; // 256 f4 = 128 fp4x2
static constexpr ck::index_t Nswizzle = false;
static constexpr ck::index_t ActOP = 0; // 0: gelu_and_mul, 1: silu_and_mul
static constexpr ck::index_t MPerBlock = 128;
static constexpr ck::index_t NPerBlock = 64;
static constexpr ck::index_t BlockSize = 256;
static constexpr bool MulRoutedWeight = true;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmMXBNS<
A0Layout, B0Layout, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
ScaleBlockSize, BlockSize,
MPerBlock, NPerBlock, KPerBlock,
16, 16,
16, 16,
4, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 2, S<1, 32, 1, 8>, S<8, 1, 1, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3,
ActOP, Nswizzle, true, MulRoutedWeight, ck::index_t, A0DataType>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
// per expert:
// GEMM shape
constexpr ck::index_t sorted_tile_num = 13;
constexpr ck::index_t valid_tile_num = sorted_tile_num;
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
ck::index_t valid_size = valid_tile_num * MPerBlock;
ck::index_t N = 4096;
ck::index_t K = 6144;
ck::index_t experts = 8;
ck::index_t tokens = 832;
ck::index_t topk = 2;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
// use default case
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 7)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
tokens = std::stoi(argv[6]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: N, K, tokens\n");
exit(0);
}
if(K % ScaleBlockSize != 0)
{
throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
};
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
ck::index_t Scale_Stride_AM = (K + ScaleBlockSize - 1) / ScaleBlockSize;
ck::index_t Scale_Stride_BN = (K + ScaleBlockSize - 1) / ScaleBlockSize;
constexpr ck::index_t NumDTensor = DsDataType::Size();
constexpr auto StrideDs = std::array<ck::index_t, NumDTensor>{0, 0, 0};
ck::index_t KBatch = 1;
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({sorted_tile_num + 1}));
max_token_id.mData[0] = valid_size;
if(tokens * topk > valid_size)
{
printf("err config, tokens * topk > valid_size\n");
exit(-1);
}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);
}
int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
int tokenid = 0;
for(int i = 0; i < sorted_size; i++)
{
int tile_off = i % MPerBlock;
if(tile_off < token_per_tile)
{
sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
tokenid++;
}
else
{
sorted_token_ids.mData[i] = tokens;
}
}
Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
Tensor<XDataType> a1_t_k(HostTensorDescriptor(
{tokens, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}));
Tensor<XDataType> b1_e_n_k(
HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N * 2},
{(N * 2 * Scale_Stride_BN), 1, Scale_Stride_BN}));
// A, B Scale preshuffle
Tensor<XDataType> a_scale_sorted(HostTensorDescriptor(
{sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
Tensor<XDataType> a_scale_preshuffled(HostTensorDescriptor(
{sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
Tensor<XDataType> b_scale_preshuffled(
HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N * 2},
{N * 2 * Scale_Stride_BN, 1, Scale_Stride_BN}));
Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
Tensor<EDataType> e_t_k_n_host_result(
HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
Tensor<EDataType> e_t_k_n_device_result(
HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
e_t_k_n_device_result.SetZero();
std::cout << "a0_t_k: " << a0_t_k.mDesc << std::endl;
std::cout << "a1_t_k: " << a1_t_k.mDesc << std::endl;
std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
std::cout << "e_t_k_n: " << e_t_k_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
break;
case 2:
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a1_t_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{0.1f});
break;
case 3:
a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 4:
a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_t_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 5.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 5:
a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{1});
break;
case 6:
a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 7:
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{0.5f});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{1.5f});
a1_t_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{1.0f});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{1.0f});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{0.1f});
break;
default:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
}
DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) * sorted_token_ids.GetElementSpaceSize());
DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.GetElementSpaceSize());
DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.GetElementSpaceSize());
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(XDataType) * a_scale_sorted.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(XDataType) * b1_e_n_k.GetElementSpaceSize());
DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_t_k_n_device_result.GetElementSpaceSize());
// A scale sorted
for(int i = 0; i < sorted_size; i++)
{
int token_id = sorted_token_ids.mData[i] & 0x00FFFFFF;
for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; k++)
{
if(token_id == tokens)
{
a_scale_sorted(i, k) = ck::type_convert<XDataType>(0);
}
else
{
a_scale_sorted(i, k) = a1_t_k(token_id, k);
}
}
}
// A/B scale shuffle
preShuffleScaleBuffer<ck::is_same_v<A0Layout, Row>>(a_scale_sorted.mData.data(),
a_scale_preshuffled.mData.data(),
sorted_size,
K / ScaleBlockSize);
preShuffleScaleBuffer<ck::is_same_v<B0Layout, Col>>(b1_e_n_k.mData.data(),
b_scale_preshuffled.mData.data(),
N * 2 * experts,
K / ScaleBlockSize);
sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
expert_ids_dev.ToDevice(expert_ids.mData.data());
max_token_id_dev.ToDevice(max_token_id.mData.data());
a0_device_buf.ToDevice(a0_t_k.mData.data());
b0_device_buf.ToDevice(b0_e_n_k.mData.data());
a1_device_buf.ToDevice(a_scale_preshuffled.mData.data());
b1_device_buf.ToDevice(b_scale_preshuffled.mData.data());
d2_device_buf.ToDevice(d2_e_n.mData.data());
e_device_buf.ToDevice(e_t_k_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(
sorted_token_ids_dev.GetDeviceBuffer(),
expert_ids_dev.GetDeviceBuffer(),
max_token_id_dev.GetDeviceBuffer(),
a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{nullptr, nullptr, d2_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
tokens,
topk,
sorted_size,
N,
K,
StrideA,
Scale_Stride_AM,
StrideB,
Scale_Stride_BN,
StrideDs,
StrideE,
KBatch,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
if(!(ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950"))
{
std::cout << "This kernel support gfx942 and gfx950 only" << std::endl;
}
if(time_kernel)
{
// not result correct here because output buf not setzero
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop =
// FMA * tokens * N * (Gate+Up) * topk * K +
// FMA * tokens * N * (Gate+Up) * topk * (K/BlockScale)
std::size_t(2) * tokens * N * 2 * topk * K +
std::size_t(2) * tokens * N * 2 * topk * K / ScaleBlockSize;
std::size_t num_btype = sizeof(A0DataType) / 2 * tokens * topk * K +
sizeof(B0DataType) / 2 * K * N * 2 * experts +
sizeof(XDataType) * tokens * topk * K / ScaleBlockSize +
sizeof(XDataType) * K / ScaleBlockSize * N * 2 * experts +
sizeof(EDataType) * tokens * topk * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << device_op.GetTypeString() << std::endl;
}
if(do_verification)
{
// gemm2 use atomic, so need to reinit outputs
e_device_buf.ToDevice(e_t_k_n_device_result.mData.data());
invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
Tensor<CShuffleDataType> c_t_k_n({tokens, topk, N}, {topk * N, N, 1});
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceMoeMXGemm1<A0DataType,
XDataType,
B0DataType,
XDataType,
CShuffleDataType,
D2DataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough,
ActOP,
MulRoutedWeight>;
auto ref_moe_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_moe_gemm.MakeInvoker();
auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
expert_ids,
max_token_id,
MPerBlock,
a0_t_k,
a1_t_k,
b0_e_n_k,
b1_e_n_k,
d2_e_n,
c_t_k_n,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < valid_size; ++m)
{
const int fuse_t = sorted_token_ids.mData[m];
const int t = fuse_t & 0xffffff;
const int topk_id = (fuse_t & 0xff000000) >> 24;
if(t >= tokens)
{
continue;
}
for(int n = 0; n < N; ++n)
{
e_t_k_n_host_result(t, topk_id, n) =
ck::type_convert<EDataType>(c_t_k_n(t, topk_id, n));
}
}
e_device_buf.FromDevice(e_t_k_n_device_result.mData.data());
auto status =
ck::utils::check_err(
e_t_k_n_device_result, e_t_k_n_host_result, "Error: Incorrect results!", 1e-3, 5e-1)
? 0
: 1;
if(status == 0)
{
printf("Validation Pass.\n");
}
return status;
}
return 0;
}

526
example/67_gemm_microscaling/moe_gemm2_xdl_mx_fp4_bns.cpp Normal file
View File

@@ -0,0 +1,526 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm_bns.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm2.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F4 = ck::f4x2_pk_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using XDataType = ck::e8m0_bexp_t;
using XPackedDataType = int32_t; // 4 packed e8m0_bexp_t
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = F4;
using A1DataType = XPackedDataType;
using B0DataType = F4;
using B1DataType = XPackedDataType;
using EDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F32;
using D1DataType = F32;
using D2DataType = F32;
using DsDataType = ck::Tuple<D0DataType, D1DataType, D2DataType>;
using A0Layout = Row;
using B0Layout = Col;
using ELayout = Row;
using D0Layout = Row;
using D1Layout = Col;
using D2Layout = ELayout;
using DsLayout = ck::Tuple<D0Layout, D1Layout, D2Layout>;
// d0: ascale, d1: bscale, d2:expert weight
struct MulABScaleExpertWeight
{
template <typename E, typename C, typename D0, typename D1, typename D2>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1, const D2& d2) const;
// for real kernel use
template <>
__host__ __device__ constexpr void operator()<EDataType, float, float, float, float>(
EDataType& e, const float& c, const float& d0, const float& d1, const float& d2) const
{
(void)d0;
(void)d1;
(void)d2;
e = ck::type_convert<EDataType>(c);
}
// for reference cpu
template <>
__host__ __device__ constexpr void operator()<float, float, float, float, float>(
float& e, const float& c, const float& d0, const float& d1, const float& d2) const
{
// for reference cpu
e = ck::type_convert<EDataType>(c * d0 * d1 * d2);
}
};
using CDEElementOp = MulABScaleExpertWeight;
// A, B Scale preshuffle
template <bool KLast>
void preShuffleScaleBuffer(ck::e8m0_bexp_t* src, ck::e8m0_bexp_t* dst, int MN, int K)
{
int MNXdlPack = 2;
int KXdlPack = 2;
int XdlMNThread = 16;
int XdlKThread = 64 / XdlMNThread;
int K0 = K / KXdlPack / XdlKThread; // KRepeat
// The 4 16x128 building blocks will be packed into 1 32x256 for F4
// The 8 16x16x128 mfma will be packed into 1 32x32x256 for F4
// unfold the MN32xK(256/32) scale buffer
// 4 16 2 2
// To XdlKThread-> XdlMNThread -> KXdlPack -> MNXdlPack
// Then, MNRepeat->KRepeat
for(int n = 0; n < MN; ++n)
{
for(int k = 0; k < K; ++k)
{
int n0 = n / (XdlMNThread * MNXdlPack); // i MNRepeat
int tempn = n % (XdlMNThread * MNXdlPack);
int n1 = tempn % XdlMNThread; // i XdlMNThread
int n2 = tempn / XdlMNThread; // i MNXdlPack
int k0 = k / (XdlKThread * KXdlPack); // i KRepeat
int tempk = k % (XdlKThread * KXdlPack);
int k1 = tempk % XdlKThread; // i XdlKThread
int k2 = tempk / XdlKThread; // i KXdlPack
int outputIndex = n0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread * K0 +
k0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread +
k1 * MNXdlPack * KXdlPack * XdlMNThread + n1 * MNXdlPack * KXdlPack +
k2 * MNXdlPack + n2;
// src[n * K + k] = ck::type_convert<ck::e8m0_bexp_t>(static_cast<float>(powf(2.0f, n2 +
// k2 * MNXdlPack)));
if constexpr(KLast)
dst[outputIndex] = src[n * K + k];
else
dst[outputIndex] = src[k * MN + n];
}
}
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MulABScaleExpertWeight;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
constexpr ck::index_t DataPackedSize = 2; // Packed representation of data
constexpr ck::index_t ScaleBlockSize = 32; // scaling block size
constexpr ck::index_t KPerBlock = 256 / DataPackedSize; // 256 f4 = 128 fp4x2
static constexpr ck::index_t MPerBlock = 128;
static constexpr bool MulRoutedWeight = true;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmMXBNS<
A0Layout, B0Layout, DsLayout, ELayout,
A0DataType, A1DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec,
ScaleBlockSize, 256,
MPerBlock, 128, KPerBlock,
16, 16,
16, 16,
4, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 2, S<1, 32, 1, 8>, S<2, 1, 1, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, MulRoutedWeight, ck::index_t, A0DataType>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
// per expert:
// GEMM shape
constexpr ck::index_t sorted_tile_num = 13;
constexpr ck::index_t valid_tile_num = sorted_tile_num;
ck::index_t sorted_size = sorted_tile_num * MPerBlock;
ck::index_t valid_size = valid_tile_num * MPerBlock;
ck::index_t N = 6144;
ck::index_t K = 4096;
ck::index_t experts = 8;
ck::index_t tokens = 832;
ck::index_t topk = 2;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
// use default case
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 7)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
tokens = std::stoi(argv[6]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: N, K, tokens\n");
exit(0);
}
if(K % ScaleBlockSize != 0)
{
throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
};
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
ck::index_t Scale_Stride_AM = (K + ScaleBlockSize - 1) / ScaleBlockSize;
ck::index_t Scale_Stride_BN = (K + ScaleBlockSize - 1) / ScaleBlockSize;
constexpr ck::index_t NumDTensor = DsDataType::Size();
constexpr auto StrideDs = std::array<ck::index_t, NumDTensor>{0, 0, 0};
ck::index_t KBatch = 1;
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
max_token_id.mData[0] = valid_size;
// int eids[] = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 3, 3, 3};
int eids[sorted_tile_num]{};
for(int i = 0; i < sorted_tile_num; i++)
{
if(i < valid_tile_num)
{
eids[i] = (i * experts) / valid_tile_num;
}
else
{
eids[i] = 3;
}
}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = eids[i];
}
if(tokens * topk > valid_size)
{
printf("err config, tokens * topk > valid_size\n");
exit(-1);
}
int token_per_tile = tokens * topk / valid_tile_num;
int tokenid = 0;
for(int i = 0; i < sorted_size; i++)
{
int tile_off = i % MPerBlock;
if(tile_off < token_per_tile)
{
sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
tokenid++;
}
else
{
sorted_token_ids.mData[i] = tokens;
}
}
Tensor<A0DataType> a0_t_k_k(HostTensorDescriptor({tokens, topk, K}, {topk * K, K, 1}));
Tensor<XDataType> a1_t_k_k(
HostTensorDescriptor({tokens, topk, (K + ScaleBlockSize - 1) / ScaleBlockSize},
{(topk * Scale_Stride_AM), Scale_Stride_AM, 1}));
Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
Tensor<XDataType> b1_e_n_k(
HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N},
{(N * Scale_Stride_BN), 1, Scale_Stride_BN}));
// B preshuffle
Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
// A, B Scale preshuffle
Tensor<XDataType> a_scale_sorted(HostTensorDescriptor(
{sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
Tensor<XDataType> a_scale_preshuffled(HostTensorDescriptor(
{sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
Tensor<XDataType> b_scale_preshuffled(
HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N},
{N * Scale_Stride_BN, 1, Scale_Stride_BN}));
Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
Tensor<EDataType> e_t_n_host_result(HostTensorDescriptor({tokens, N}, {N, 1}));
Tensor<EDataType> e_t_n_device_result(HostTensorDescriptor({tokens, N}, {N, 1}));
e_t_n_device_result.SetZero();
std::cout << "a0_t_k_k: " << a0_t_k_k.mDesc << std::endl;
std::cout << "a1_t_k_k: " << a1_t_k_k.mDesc << std::endl;
std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
break;
case 2:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 3:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 4:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 5.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
case 5:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{1});
break;
case 6:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
break;
default:
a0_t_k_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
}
DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) * sorted_token_ids.GetElementSpaceSize());
DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.GetElementSpaceSize());
DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.GetElementSpaceSize());
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k_k.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(XDataType) * a_scale_sorted.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(XDataType) * b1_e_n_k.GetElementSpaceSize());
DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.GetElementSpaceSize());
// A scale sorted
for(int i = 0; i < sorted_size; i++)
{
int token_id = sorted_token_ids.mData[i] & 0x00FFFFFF;
int topk_id = (sorted_token_ids.mData[i] >> 24) & 0x000000FF;
for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; k++)
{
if(token_id == tokens)
{
a_scale_sorted(i, k) = ck::type_convert<XDataType>(0);
}
else
{
a_scale_sorted(i, k) = a1_t_k_k(token_id, topk_id, k);
}
}
}
preShuffleScaleBuffer<ck::is_same_v<A0Layout, Row>>(a_scale_sorted.mData.data(),
a_scale_preshuffled.mData.data(),
sorted_size,
K / ScaleBlockSize);
preShuffleScaleBuffer<ck::is_same_v<B0Layout, Col>>(
b1_e_n_k.mData.data(), b_scale_preshuffled.mData.data(), N * experts, K / ScaleBlockSize);
sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
expert_ids_dev.ToDevice(expert_ids.mData.data());
max_token_id_dev.ToDevice(max_token_id.mData.data());
a0_device_buf.ToDevice(a0_t_k_k.mData.data());
b0_device_buf.ToDevice(b0_e_n_k.mData.data());
a1_device_buf.ToDevice(a_scale_preshuffled.mData.data());
b1_device_buf.ToDevice(b_scale_preshuffled.mData.data());
d2_device_buf.ToDevice(d2_e_n.mData.data());
e_device_buf.ToDevice(e_t_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(
sorted_token_ids_dev.GetDeviceBuffer(),
expert_ids_dev.GetDeviceBuffer(),
max_token_id_dev.GetDeviceBuffer(),
a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
b1_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{nullptr, nullptr, d2_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
tokens,
topk,
sorted_size,
N,
K,
StrideA,
Scale_Stride_AM,
StrideB,
Scale_Stride_BN,
StrideDs,
StrideE,
KBatch,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
if(!(ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950"))
{
std::cout << "This kernel support gfx942 and gfx950 only" << std::endl;
}
if(time_kernel)
{
// not result correct here because output buf not setzero
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
// FMA * tokens * N * topk * K +
// FMA * tokens * N * topk * (K/BlockScale)
std::size_t flop = std::size_t(2) * tokens * topk * N * K +
std::size_t(2) * tokens * topk * N * K / ScaleBlockSize;
std::size_t num_btype =
sizeof(A0DataType) / 2 * tokens * K * topk + sizeof(B0DataType) / 2 * K * N * experts +
sizeof(XDataType) * tokens * topk * K / ScaleBlockSize +
sizeof(XDataType) * K / ScaleBlockSize * N * experts + sizeof(EDataType) * tokens * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << device_op.GetTypeString() << std::endl;
}
if(do_verification)
{
// gemm2 use atomic, so need to reinit outputs
e_device_buf.ToDevice(e_t_n_device_result.mData.data());
invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
Tensor<CShuffleDataType> c_t_n({tokens, N});
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceMoeMXGemm2<A0DataType,
XDataType,
B0DataType,
XDataType,
D2DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
CDEElementOp,
MulRoutedWeight,
float,
float>;
auto ref_moe_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_moe_gemm.MakeInvoker();
auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
expert_ids,
max_token_id,
MPerBlock,
a0_t_k_k,
a1_t_k_k,
b0_e_n_k,
b1_e_n_k,
d2_e_n, // topk weights
c_t_n,
PassThrough{},
PassThrough{},
cde_element_op);
ref_invoker.Run(ref_argument);
for(int t = 0; t < tokens; ++t)
{
for(int n = 0; n < N; ++n)
{
e_t_n_host_result(t, n) = ck::type_convert<EDataType>(c_t_n(t, n));
}
}
e_device_buf.FromDevice(e_t_n_device_result.mData.data());
return ck::utils::check_err(
e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-2)
? 0
: 1;
}
return 0;
}

10
example/CMakeLists.txt
View File

@@ -104,11 +104,13 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
# Do not build gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94
# Build fp8 gemm_multiply_multiply and moe only on gfx94/95
foreach(source IN LISTS FILE_NAME)
if(NOT EX_TARGETS MATCHES "gfx94" AND NOT EX_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_multiply_multiply_xdl_fp8_bpreshuffle")
message(DEBUG "Skipping ${source} example for current target")
list(REMOVE_ITEM FILE_NAME "${source}")
if(NOT EX_TARGETS MATCHES "gfx94" AND NOT EX_TARGETS MATCHES "gfx95")
if (source MATCHES "fp8" AND source MATCHES "(gemm_multiply_multiply|moe)")
message(DEBUG "Skipping ${source} example for current target")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endif()
endforeach()
#only continue if there are some source files left on the list