Skip to content

GitLab

"magic_pdf/git@developer.sourcefind.cn:wangsen/mineru.git" did not exist on "34a13a898bc0cec7e7751f4ce8547699d8514c3f"
Unverified Commit 1837040a authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Navi3 rel (#1176) 



* wmma_op + unit test

* add arch limitation to wmma test

* change arch limitation

* Refactor + Add all type unit test(int4 compile failed)

* Add f32_16x16x16_bf16 unit test

* tempsave

* tempsave

* tempsave

* runtime bug, cannot find symbol

* workaround for incorrect HIP warpSize return value

* debugging

* tempsave

* Correctness OK, waiting for optimization

* Tidy up + format

* temp save

* temp save, reproduce the v_bfi_b32 issue

* add inline asm for wmmaop test

* tidy up

* clean some debug purpose code

* discard some codes

* clang format

* clang format

* compiler issue fixed + increase tile size

* navi3x_multipleD+example

* temp save

* workable

* batchedgemm[OK], groupconv[debug]

* groupconv: Sanity check[OK], Performance[Bad]

* navi3x_groupconv_need_optimization

* create necessary files

* save progress

* Add Inter-Row thread transfer

* save progress

* save debugging progress

* sanity check pass

* fix a host tensor bug and clean up flash-attn code

* format

* cancel unnecessary change

* cancel unnecessary change

* cancel unnecessary change

* temp save, add asm backend flag to amd_wmma

* Mat-A LDS Bypass sanity pass

* temp save

* gemm sanity fix

* Porting new blockwise gemm to flash attention

* Example branch provide to compiler team

* tempsave

* Fix a bug

* batched gemm ported

* conv A-skip lds ported

* Skip B-Lds real gemm

* Skip B Lds Gemm + MulD

* batched gemm, conv, skip b lds

* format

* Attn, skip b lds

* Change GridwiseOp nam

* fix a typo caused bug

* Skip A_Lds sanity pass, Skip B_Lds scratch occured

* Bug found, intra-row permute off caused

* bug found

* a fix

* disable buffer load due to incorrect 3rd dword

* update fmha config, no scratch generated

* update 3rd dword

* fmha config update

* FMHA, add support to gfx1101/gfx1102

* Merge origin dev (#2)

* [Navi3x] Fix Gridwise_multiple_d operation (#649)

* Add CMake Option "USE_OPT_NAVI3X"

* fix bug

* standardize docs (#655)

* Separate bibtex requirement from rocm-docs-core (#656)

* separate bibtex requirement from rocm-docs-core

* point requirements to source rocm-docs-core repo

* Add CMake Option "USE_OPT_NAVI3X" (#647)

* Add CMake Option "USE_OPT_NAVI3X"

* remove navi3x opt compile option from cmake script

* Conv + quantization + tanh  (#645)

* Rename file. Prepare to support another activation

* Add comment for quantization

* Extract out_elementop

* Add tanh example

* Add conv + bias + tanh quantization instance

* Add missing parameter

* Refine cmake

* Add external api and client example

* Extract variable in example

* Fix the comment

---------
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>

* Add a denorm test fix (#603)

* Add type_convert implementations for bf16

* Add the fix for conv_fwd

* Add the fix for conv_bwd_data

* Add the fix for conv_bwd_weight

* Format

* Format

* Another format

* Add a macro to use workaround on MI200 only

* Format

---------
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>

* simplify karg in device/grid of split-k op (#644)

* simplify karg in device/grid split-k op

* fix mk_kn_mn instances

* add more instances

* use name from tensor layout

* fix 3rd dword of buffer source descriptor (#659)

* add fp64 instances (#658)
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>

* Issue #666: Revert "simplify karg in device/grid of split-k op (#644)" (#665)

This reverts commit bb5530af

.

* Groupnorm + swish external api (#668)

* Rename to proper naming

* Add example of groupnorm + swish

* Extract duplicate code in example

* Add groupnorm + swish instances

* Ractor instance generation, split into multiple cpp file

* Add external api and client example

* Refine profiler message

* Use ck math version of exp

* Refine problem size in example

* Add host version of exp

* add a marco to turn on/off denorm fix (off by default) (#673)

* add a marco to turn off denorm fix by default

* expose the marco

---------
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>

* fixed quant example (#672)
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>

* Add dependabot config and pin rocm-docs-core (#663)

* [gtest] suppress unsafe buffer warn (#670)

ref: https://github.com/ROCmSoftwarePlatform/MIOpen/pull/1912



* Add memory index guard in wmma device ops (#667)

* Add more macros to turn on/off denorm fix (#678)
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>

* Fix a typo (#676)

* Add (#677)

* Allow using ROCm release candidate compilers. (#679)

* enable use of rocm5.5 release candidate 4

* upgrade to ROCM5.5 RC5

* try fix the PUB_KEY error, remove the cmake-data package

* upgrade to latest cmake version

* use private dockerhub repo for rocm5.5 rc5

* add missing bracket

* add vector load check

* solve conflicts

---------
Co-authored-by: default avatarSam Wu <sjwu@ualberta.ca>
Co-authored-by: default avatarSam Wu <sam.wu2@amd.com>
Co-authored-by: default avatarrocking5566 <ChunYu.Lai@amd.com>
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>
Co-authored-by: default avatarRostyslav Geyyer <46627076+geyyer@users.noreply.github.com>
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: default avatarcarlushuang <carlus.huang@amd.com>
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>
Co-authored-by: default avatarJun Liu <Liu.Jun@amd.com>
Co-authored-by: default avatarIllia Silin <98187287+illsilin@users.noreply.github.com>

* Disable SkipLDS & Align AIT api (#3)

* fix layernorm, reduction Ops (#4)

* [Navi3x] Fix Gridwise_multiple_d operation (#649)

* Add CMake Option "USE_OPT_NAVI3X"

* fix bug

* standardize docs (#655)

* Separate bibtex requirement from rocm-docs-core (#656)

* separate bibtex requirement from rocm-docs-core

* point requirements to source rocm-docs-core repo

* Add CMake Option "USE_OPT_NAVI3X" (#647)

* Add CMake Option "USE_OPT_NAVI3X"

* remove navi3x opt compile option from cmake script

* Conv + quantization + tanh  (#645)

* Rename file. Prepare to support another activation

* Add comment for quantization

* Extract out_elementop

* Add tanh example

* Add conv + bias + tanh quantization instance

* Add missing parameter

* Refine cmake

* Add external api and client example

* Extract variable in example

* Fix the comment

---------
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>

* Add a denorm test fix (#603)

* Add type_convert implementations for bf16

* Add the fix for conv_fwd

* Add the fix for conv_bwd_data

* Add the fix for conv_bwd_weight

* Format

* Format

* Another format

* Add a macro to use workaround on MI200 only

* Format

---------
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>

* simplify karg in device/grid of split-k op (#644)

* simplify karg in device/grid split-k op

* fix mk_kn_mn instances

* add more instances

* use name from tensor layout

* fix 3rd dword of buffer source descriptor (#659)

* add fp64 instances (#658)
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>

* Issue #666: Revert "simplify karg in device/grid of split-k op (#644)" (#665)

This reverts commit bb5530af

.

* Groupnorm + swish external api (#668)

* Rename to proper naming

* Add example of groupnorm + swish

* Extract duplicate code in example

* Add groupnorm + swish instances

* Ractor instance generation, split into multiple cpp file

* Add external api and client example

* Refine profiler message

* Use ck math version of exp

* Refine problem size in example

* Add host version of exp

* add a marco to turn on/off denorm fix (off by default) (#673)

* add a marco to turn off denorm fix by default

* expose the marco

---------
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>

* fixed quant example (#672)
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>

* Add dependabot config and pin rocm-docs-core (#663)

* [gtest] suppress unsafe buffer warn (#670)

ref: https://github.com/ROCmSoftwarePlatform/MIOpen/pull/1912



* Add memory index guard in wmma device ops (#667)

* Add more macros to turn on/off denorm fix (#678)
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>

* Fix a typo (#676)

* Add (#677)

* Allow using ROCm release candidate compilers. (#679)

* enable use of rocm5.5 release candidate 4

* upgrade to ROCM5.5 RC5

* try fix the PUB_KEY error, remove the cmake-data package

* upgrade to latest cmake version

* use private dockerhub repo for rocm5.5 rc5

* add missing bracket

* Disable SkipLDS & Align AIT api

* Update dependabot config (#682)
Co-authored-by: default avatarsamjwu <samjwu@users.noreply.github.com>

* update attn api

* solve type_convert bug + enable

---------
Co-authored-by: default avatarSam Wu <sjwu@ualberta.ca>
Co-authored-by: default avatarSam Wu <sam.wu2@amd.com>
Co-authored-by: default avatarrocking5566 <ChunYu.Lai@amd.com>
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>
Co-authored-by: default avatarRostyslav Geyyer <46627076+geyyer@users.noreply.github.com>
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: default avatarcarlushuang <carlus.huang@amd.com>
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>
Co-authored-by: default avatarJun Liu <Liu.Jun@amd.com>
Co-authored-by: default avatarIllia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: default avatarsamjwu <samjwu@users.noreply.github.com>
Co-authored-by: default avatarhaocwang <Haocong.WANG@amd.com>

* fix typo

* Fix attention with causal mask

* multiple fix, try ait compile

* Add A/B not use LDS pipeline

* Clang format, Add gfx1101, gfx1102 support of FMHA example

* cancel change of format script

* 1. Enable 2-stage global Prefetch ( May cause VGPR spilling)
2. Enable FP16 accumulator blockwise_gemm

* clang-format

* 1. change blockwise gemm loopover direction from kmn to mnk ( ~1% improvement)
2. change kernel timing mode to 50 warmup + 50 timed repeat

* Update low level abstration of blockwise gemm wmma

* (2/5) bilinear gemm pass, perf bug: skip a lds has lower performance than skip b lds

* (3/5) batched gemm pass, perf bug: skip a lds has lower performance than skip b lds

* (4/5) grouped conv pass

* (5/5) attention pass, todo: debug lds perf bug

* AIT Attention API refactor (#8)

* sanity pass

* sanity pass 2

* confirm significant performance regression.

* turn on all instances

* turn off instance format

* Fix bug & tunning & format

* DML meta, self_attn+cross_attn

* sanity pass

* remove useless flag

* update tile and problem size used in AIT attention

* bug fix in grouped conv supporting check

* deprecate inline asm wmma

* Bug fix: double lds skip

* clang-format

* Fix errors in
1. example, fmha
2. gridwise pipeline
3. deviceop, fmha, change some containers from vector to array

* part2 of previous commit

* clang format

* API fix of gridwisegemmpipeline

* separate array base and vector base attention tensor transformation

* fix gemm

* clang format

* add gemm fp16 instances

* Temp save

* fpAintB kernel compile pass

* Sanity pass.

* Temp save

* debug code enabled

* Fp16AInt8B_GEMM sanity

* MQA implementation

* GQA-4 example

* tempsave

* Compile pass

* New implementation of fp16Aint8B Gemm, Acheieve similar math throughput with native fp16 Gemm

* format

* Todo: fix gemm_bilinear_wmma instances compilation bug

* Solve a bug when K1=16

* remove unnecessary changes

* Remove tensor layout limitation to LDS usage in tesnor contraction

* update self-attention and cross-attention

* fix a typo of name

* Add arch limiter for fp8 gemm

* enable fp8 gemm_xdl for all gfx9 targets

* temporarily disable gemm_xdl_fp16_fp8 on MI100/200

* fix the cmake logic for gemm_xdl_fp16_fp8

* re-enable the gemm_xdl_fp16_fp8 on MI100/200

---------
Co-authored-by: default avataraska-0096 <haocwang@amd.com>
Co-authored-by: default avatarSam Wu <sjwu@ualberta.ca>
Co-authored-by: default avatarSam Wu <sam.wu2@amd.com>
Co-authored-by: default avatarrocking5566 <ChunYu.Lai@amd.com>
Co-authored-by: default avatarRostyslav Geyyer <46627076+geyyer@users.noreply.github.com>
Co-authored-by: default avatarRosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: default avatarcarlushuang <carlus.huang@amd.com>
Co-authored-by: default avatarroot <root@ctr-ubbsmc15.amd.com>
Co-authored-by: default avatarJun Liu <Liu.Jun@amd.com>
Co-authored-by: default avatarIllia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: default avatarsamjwu <samjwu@users.noreply.github.com>
Co-authored-by: default avatarhaocwang <Haocong.WANG@amd.com>
Co-authored-by: default avatarillsilin <Illia.Silin@amd.com>
parent 363feb48
Hide whitespace changes
Inline Side-by-side
Showing with 3736 additions and 148 deletions
example/01_gemm/CMakeLists.txt
View file @ 1837040a
......@@ -27,7 +27,7 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
if(GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_example_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
......@@ -53,12 +53,6 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
add_example_executable(example_gemm_xdl_fp8 gemm_xdl_fp8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
add_example_executable(example_gemm_xdl_fp8_bf8 gemm_xdl_fp8_bf8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
list(APPEND gpu_list gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
......@@ -72,5 +66,12 @@ foreach(gpu IN LISTS GPU_TARGETS)
endif()
endforeach()
add_example_executable(example_gemm_xdl_fp8 gemm_xdl_fp8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
add_example_executable(example_gemm_xdl_fp8_bf8 gemm_xdl_fp8_bf8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
add_example_executable(example_gemm_xdl_fp16_fp8 gemm_xdl_fp16_fp8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)
example/01_gemm/gemm_wmma_fp16.cpp
View file @ 1837040a
......@@ -19,15 +19,50 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer|MRepeat|NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN|MWmmaPerWave|NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmMNKPadding, 256, 128, 256, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, 1>;
< ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CElementOp,
GemmDefault,
1, // Prefetch stage
128, // BlockSize
64, // MPerBlock
128, // NPerBlock
64, // KPerBlock
8, // K1
16, // MPerWmma
16, // NPerWmma
2, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
4, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
8>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
......
example/01_gemm/run_gemm_example.inc
View file @ 1837040a
......@@ -150,6 +150,22 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
break;
case 2:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break;
case 3:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
break;
case 4:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{1.f, 1.f}(b_k_n);
break;
case 5:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-2.f, 2.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-2.f, 2.f}(b_k_n);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-0.1f, 0.1f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-0.1f, 0.1f}(b_k_n);
......
example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
View file @ 1837040a
......@@ -65,48 +65,49 @@ using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<ALayout,
BLayout,
ck::Tuple<DLayout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
256,
128,
256,
8,
8,
16,
16,
4,
4,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1,
1,
S<1, 32, 1, 8>,
8>;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<
ALayout,
BLayout,
ck::Tuple<DLayout>,
ELayout,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
2, // Prefetch stage
128, // BlockSize
128, // MPerBlock
64, // NPerBlock
64, // KPerBlock
8, // K1
16, // MPerWmma
16, // NPerWmma
4, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
2, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
8>;
int main(int argc, char* argv[])
{
......@@ -264,7 +265,7 @@ int main(int argc, char* argv[])
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;
<< device_op.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
......
example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
View file @ 1837040a
......@@ -55,7 +55,7 @@ using DDataType = I8;
using EDataType = I8;
using ALayout = Row;
using BLayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
......@@ -65,48 +65,49 @@ using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<ALayout,
BLayout,
ck::Tuple<DLayout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
32,
16,
16,
4,
16,
16,
16,
1,
1,
S<2, 16, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
16,
16,
1,
S<4, 1, 8>,
S<0, 2, 1>,
S<0, 2, 1>,
1,
16,
2,
1,
1,
1,
S<1, 16, 1, 2>,
8>;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<
ALayout,
BLayout,
ck::Tuple<DLayout>,
ELayout,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
2, // Prefetch stage
128, // BlockSize
128, // MPerBlock
64, // NPerBlock
64, // KPerBlock
8, // K1
16, // MPerWmma
16, // NPerWmma
4, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
2, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
8>;
int main(int argc, char* argv[])
{
......
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @ 1837040a
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
if(GPU_TARGETS MATCHES "gfx11")
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
endif()
example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
View file @ 1837040a
......@@ -43,9 +43,10 @@ using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Add;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
static constexpr auto ASpec = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto BSpec = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceOpInstanceKKNN =
......@@ -55,43 +56,44 @@ using DeviceOpInstanceKKNN =
NumDimK,
ADataType,
BDataType,
DsDataType,
EDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
ABSpec,
ABSpec,
ASpec,
BSpec,
DESpec,
256,
1,
128,
256,
8,
8,
64,
64,
64,
4,
16,
16,
1,
4,
4,
S<4, 64, 1>,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
4,
4,
true,
S<4, 64, 1>,
S<4, 32, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
4,
4,
true,
1,
1,
S<1, 32, 1, 8>,
S<1, 64, 1, 2>,
8>;
using DeviceOpInstance = DeviceOpInstanceKKNN;
......@@ -251,21 +253,6 @@ int main(int argc, char* argv[])
ck::index_t K0 = 2048;
// A[G0, G1, M0, M1, K0]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
// B[G0, G1, N0, N1, K0]
std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
// D[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
// E[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> e_gs_ms_ns_strides{
G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
if(argc == 1)
{
// use default case
......@@ -276,13 +263,43 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
G0 = std::stoi(argv[4]);
G1 = std::stoi(argv[5]);
M0 = std::stoi(argv[6]);
M1 = std::stoi(argv[7]);
N0 = std::stoi(argv[8]);
N1 = std::stoi(argv[9]);
K0 = std::stoi(argv[10]);
}
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("arg4-10: G0, G1, M0, M1, N0, N1, K0\n");
exit(0);
}
// A[G0, G1, M0, M1, K0]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
// B[G0, G1, N0, N1, K0]
std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
// D[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
// E[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> e_gs_ms_ns_strides{
G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
......
example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
View file @ 1837040a
......@@ -42,41 +42,42 @@ using DeviceConvFwdInstance =
OutputLayout<NDimSpatial>,
InKernelDataType,
WeiKernelDataType,
ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
OutKernelDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
1, // Prefetch stage
128, // BlockSize
64, // MPerBlock
64, // NPerBlock
64, // KPerBlock
8, // K1
16, // MPerWMMA
16, // NPerWMMA
4, // MRepeat
2, // NRepeat
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
1, // NRepeat
S<4, 32, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
true, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<4, 32, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
true, // BBlockLdsExtraN
4,
2,
S<1, 32, 1, 8>,
1,
1,
S<1, 16, 1, 8>,
8>;
template <ck::index_t NDimSpatial>
......@@ -277,9 +278,9 @@ bool run_grouped_conv_fwd_bias_relu_add_example(int argc, char* argv[])
switch(conv_param.num_dim_spatial_)
{
case 1: return run_grouped_conv_fwd_bias_relu_add<1>(config, conv_param);
// case 1: return run_grouped_conv_fwd_bias_relu_add<1>(config, conv_param);
case 2: return run_grouped_conv_fwd_bias_relu_add<2>(config, conv_param);
case 3: return run_grouped_conv_fwd_bias_relu_add<3>(config, conv_param);
// case 3: return run_grouped_conv_fwd_bias_relu_add<3>(config, conv_param);
}
return false;
......
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ 1837040a
if(GPU_TARGETS MATCHES "gfx11")
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_wmma_fp16 batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp)
add_example_executable(example_self_attention_forward_wmma_fp16 self_attention_forward_wmma_fp16.cpp)
add_example_executable(example_cross_attention_forward_wmma_fp16 cross_attention_forward_wmma_fp16.cpp)
add_example_executable(example_multi_query_attention_forward_wmma_fp16 multi_query_attention_forward_wmma_fp16.cpp)
add_example_executable(example_grouped_query_attention_forward_wmma_fp16 grouped_query_attention_forward_wmma_fp16.cpp)
endif()
add_custom_target(example_gemm_scale_softmax_gemm)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
......@@ -20,4 +29,3 @@ add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_sc
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g_k_l) * B1_g_l_n
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#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/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using Acc0DataType = F32;
using Acc1DataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceMHAFactory =
std::tuple<ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc0DataType,
Acc1BiasDataType,
Acc1DataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
// Gemm 0
128, // MPerBlock
64, // LPerBlock
64, // KPerBlock
8, // AK1
8, // BK1
// Gemm 1
64, // NPerBlock
64, // LTilePerBlock
8, // L1
16, // MPerWMMA
16, // LPerWMMA
16, // NPerWMMA
// Per repeat = wave_m = wave_num, wave_n = 1
1, // MRepeat
4, // LRepeat
4, // NRepeat
S<4, 64, 1>, // ABlockTransfer MK -> K0 M K1
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // B0BlockTransfer LK -> K0 L K1
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 8, 8>, // B1BlockTransfer NL -> L0 N L1
S<0, 2, 1>,
S<0, 2, 1>,
1,
8,
1,
false,
1, // CShuffleMWmmaPerWavePerShuffle
2, // CShuffleNWmmaPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec> // MaskingSpecialization
>;
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
Acc0DataType,
Acc1DataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<Acc0DataType, ADataType, Acc0DataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
Acc1DataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_batched_gemm_scale_softmax_gemm_permute_wmma.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g_k_l) * B1_g_l_n
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#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/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using Acc0DataType = F32;
using Acc1DataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
// #define CK_MHA_USE_WAVE_1
// #define CK_MHA_USE_WAVE_2
// #define CK_MHA_USE_WAVE_4
#define CK_MHA_USE_WAVE_8
using DeviceMHAFactory =
std::tuple<
#ifdef CK_MHA_USE_WAVE_1
// 1 wave, mrepeat = 1, nrepeat = 2, k/o repeat = 1~5
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_2
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_4
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_8
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>
#endif
>;
// clang-format on
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
Acc0DataType,
Acc1DataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<Acc0DataType, ADataType, Acc0DataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
Acc1DataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_batched_gemm_scale_softmax_gemm_permute_wmma.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g_k_l) * B1_g_l_n
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#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/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using Acc0DataType = F32;
using Acc1DataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
#define CK_MHA_USE_WAVE_1
#define CK_MHA_USE_WAVE_2
#define CK_MHA_USE_WAVE_4
#define CK_MHA_USE_WAVE_8
using DeviceMHAFactory =
std::tuple<
#ifdef CK_MHA_USE_WAVE_1
// 1 wave, mrepeat = 1, nrepeat = 2, k/o repeat = 1~5
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 32, 160, 8, 8,
// Gemm 1
80, 32, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 2, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 64, 80, 8, 8,
// Gemm 1
80, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 64, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_2
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 64, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 64, 80, 8, 8,
// Gemm 1
80, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 32, 160, 8, 8,
// Gemm 1
80, 32, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 2, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_4
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 128, 80, 8, 8,
// Gemm 1
80, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 192, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 12, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 64, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_8
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 192, 48, 8,4,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 12, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 64, 48, 8,4,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>
#endif
>;
// clang-format on
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
Acc0DataType,
Acc1DataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<Acc0DataType, ADataType, Acc0DataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
Acc1DataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_cross_attention_wmma.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/grouped_query_attention_forward_wmma_fp16.cpp 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Grouped Query Attention,
Ainslie, Joshua, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit
Sanghai. “GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints.”
arXiv, May 22, 2023. https://doi.org/10.48550/arXiv.2305.13245.
Example is GQA-4
*/
#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/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_query_attention_forward_wmma.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using Acc0DataType = F32;
using Acc1DataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
static constexpr ck::index_t QueryGroupNumber = 4;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
// #define CK_MHA_USE_WAVE_1
// #define CK_MHA_USE_WAVE_2
// #define CK_MHA_USE_WAVE_4
#define CK_MHA_USE_WAVE_8
using DeviceMHAFactory =
std::tuple<
#ifdef CK_MHA_USE_WAVE_1
// 1 wave, mrepeat = 1, nrepeat = 2, k/o repeat = 1~5
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
32,
// Gemm 0
16, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
32,
// Gemm 0
16, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_2
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
64,
// Gemm 0
32, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
64,
// Gemm 0
32, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_4
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
128,
// Gemm 0
64, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
128,
// Gemm 0
64, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_8
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
256,
// Gemm 0
128, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceGroupedQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
QueryGroupNumber,
256,
// Gemm 0
128, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>
#endif
>;
// clang-format on
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance =
ck::tensor_operation::host::ReferenceBatchedGemm_GQA<ADataType,
B0DataType,
Acc0DataType,
Acc1DataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
QueryGroupNumber>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<Acc0DataType, ADataType, Acc0DataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance =
ck::tensor_operation::host::ReferenceBatchedGemm_GQA<ADataType,
B1DataType,
CDataType,
Acc1DataType,
AElementOp,
B1ElementOp,
CElementOp,
QueryGroupNumber>;
#include "run_grouped_query_attention_forward_wmma.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/multi_query_attention_forward_wmma_fp16.cpp 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Multi-Query Attention
Shazeer, Noam. “Fast Transformer Decoding: One Write-Head Is All You Need.” arXiv.org, November 6,
2019. https://arxiv.org/abs/1911.02150v1.
*/
#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/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_multi_query_attention_forward_wmma.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using Acc0DataType = F32;
using Acc1DataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
// #define CK_MHA_USE_WAVE_1
// #define CK_MHA_USE_WAVE_2
// #define CK_MHA_USE_WAVE_4
#define CK_MHA_USE_WAVE_8
using DeviceMHAFactory =
std::tuple<
#ifdef CK_MHA_USE_WAVE_1
// 1 wave, mrepeat = 1, nrepeat = 2, k/o repeat = 1~5
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_2
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_4
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 64, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_8
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceMultiQueryAttentionForward_Wmma<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 128, 64, 8, 8,
// Gemm 1
64, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 4,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>
#endif
>;
// clang-format on
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm_MQA<ADataType,
B0DataType,
Acc0DataType,
Acc1DataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<Acc0DataType, ADataType, Acc0DataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm_MQA<ADataType,
B1DataType,
CDataType,
Acc1DataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_multi_query_attention_forward_wmma.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute_wmma.inc 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 120;
ck::index_t N = 1000;
ck::index_t K = 64;
ck::index_t O = 128;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
float alpha = 1;
bool input_permute = false;
bool output_permute = true;
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 == 13)
{
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]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
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 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
case 4: // A, B0, B1 1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5: // Rand: b1 b0; unit: a
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 6: // Rand: a b0 ; unit: B1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 7: // Rand: a b1 ; unit: b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 8: // Rand: a ; unit: b0 b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 9: // Rand: b0 ; unit: a b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 10: // Rand: b1 ; unit: a b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
float best_perf = .0;
float best_time = .0;
int not_pass = 0;
std::string best_kernel = "";
printf("Verification: %s\n", do_verification ? "ON" : "OFF");
// TODO ANT: replace array with vector?
ck::static_for<0, std::tuple_size_v<DeviceMHAFactory>, 1>{}([&](auto i) -> void {
const auto device_mha_instance = std::get<i>(DeviceMHAFactory{});
using DeviceMHAInstance = ck::remove_cvref_t<decltype(device_mha_instance)>;
auto gemm = DeviceMHAInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
G0,
G1,
alpha,
input_permute,
output_permute);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
// return 0;
}
ck::index_t BatchCount = G0 * G1;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
if(tflops > best_perf)
{
best_perf = tflops;
best_time = ave_time * 1000;
best_kernel = gemm.GetTypeString();
}
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({BatchCount, M, K});
Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
Tensor<Acc0DataType> acc0_g_m_n({BatchCount, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = typename DeviceMHAInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
bool this_run_verification = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol);
printf("Verification: %s, Pass: %s\n",
do_verification ? "ON" : "OFF",
this_run_verification ? "YES" : "NO");
if(!this_run_verification)
{
not_pass = 1;
printf("%d th MHA instance verification Failed \n", i.value);
}
}
});
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Problem Size: BatchCount: " << G0 << ", HeadNum: " << G1 << ", M: " << M
<< ", N: " << N << ", K: " << K << ", O: " << O << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
<< " us" << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
return not_pass;
}
example/32_batched_gemm_scale_softmax_gemm/run_cross_attention_wmma.inc 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t q_sequence_length = 256;
ck::index_t kv_sequence_length = 64;
ck::index_t head_dim = 80;
// Output shape C[batch_size, q_sequence_length, head_num, head_dim]. Batch dim, outer dim,
// inner dim must match GEMM shape C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o]) C_g0_m_g1_o =
// permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t batch_size = 2;
ck::index_t head_num = 8;
float alpha = 1;
bool input_permute = true;
bool output_permute = true;
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 == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
q_sequence_length = std::stoi(argv[4]);
kv_sequence_length = std::stoi(argv[5]);
head_dim = std::stoi(argv[6]);
batch_size = std::stoi(argv[7]);
head_num = std::stoi(argv[8]);
alpha = std::stof(argv[9]);
}
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 8: q_sequence_length, kv_sequence_length, head_dim, batch_size, head_num\n");
printf("arg9: scale (alpha)\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{batch_size, head_num, q_sequence_length, head_dim};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute ? std::vector<ck::index_t>{q_sequence_length * head_num * head_dim,
head_dim,
head_num * head_dim,
1}
// A layout [batch_size, q_sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * q_sequence_length * head_dim,
q_sequence_length * head_dim,
head_dim,
1}; // A layout [batch_size, head_num, q_sequence_length, head_dim]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{
batch_size, head_num, kv_sequence_length, head_dim};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute ? std::vector<ck::index_t>{kv_sequence_length * head_num * head_dim,
head_dim,
head_num * head_dim,
1}
// B0 layout [batch_size, kv_sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * kv_sequence_length * head_dim,
kv_sequence_length * head_dim,
head_dim,
1}; // B0 layout [batch_size, head_num, kv_sequence_length, head_dim]
std::vector<ck::index_t> b1_gs_os_ns_lengths{
batch_size, head_num, head_dim, kv_sequence_length};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{kv_sequence_length * head_num * head_dim,
head_dim,
1,
head_num * head_dim}
// B1 layout [batch_size, kv_sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * kv_sequence_length * head_dim,
kv_sequence_length * head_dim,
1,
head_dim}; // B1 layout [batch_size, head_num, kv_sequence_length, head_dim]
std::vector<ck::index_t> c_gs_ms_os_lengths{batch_size, head_num, q_sequence_length, head_dim};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute ? std::vector<ck::index_t>{q_sequence_length * head_num * head_dim,
head_dim,
head_num * head_dim,
1}
// C layout [batch_size, q_sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * q_sequence_length * head_dim,
q_sequence_length * head_dim,
head_dim,
1}; // C layout [batch_size, head_num, q_sequence_length, head_dim]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
case 4: // A, B0, B1 1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5: // Rand: b1 b0; unit: a
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 6: // Rand: a b0 ; unit: B1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 7: // Rand: a b1 ; unit: b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 8: // Rand: a ; unit: b0 b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 9: // Rand: b0 ; unit: a b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 10: // Rand: b1 ; unit: a b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
std::vector<ck::index_t> kv_gs_ns_ks_lengths{
batch_size, head_num, kv_sequence_length, 2, head_dim};
std::vector<ck::index_t> kv_gs_ns_ks_strides = std::vector<ck::index_t>{
kv_sequence_length * head_num * 2 * head_dim,
2 * head_dim,
head_num * 2 * head_dim,
head_dim,
1}; // kv layout [batch_size, q_sequence_length, head_num, 2, head_dim]
Tensor<ADataType> kv_gs_ns_ks(kv_gs_ns_ks_lengths, kv_gs_ns_ks_strides);
// merge kv into a packed pointer send to device
b0_gs_ns_ks.ForEach(
[&](auto& self, auto idx) { kv_gs_ns_ks(idx[0], idx[1], idx[2], 0, idx[3]) = self(idx); });
b1_gs_os_ns.ForEach(
[&](auto& self, auto idx) { kv_gs_ns_ks(idx[0], idx[1], idx[3], 1, idx[2]) = self(idx); });
DeviceMem q_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem kv_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize() +
sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
q_device_buf.ToDevice(a_gs_ms_ks.mData.data());
kv_device_buf.ToDevice(kv_gs_ns_ks.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
float best_perf = .0;
float best_time = .0;
int not_pass = 0;
std::string best_kernel = "";
printf("Verification: %s\n", do_verification ? "ON" : "OFF");
// TODO ANT: replace array with vector?
ck::static_for<0, std::tuple_size_v<DeviceMHAFactory>, 1>{}([&](auto i) -> void {
const auto device_mha_instance = std::get<i>(DeviceMHAFactory{});
using DeviceMHAInstance = ck::remove_cvref_t<decltype(device_mha_instance)>;
auto gemm = DeviceMHAInstance{};
auto invoker = gemm.MakeCrossAttnInvoker();
auto argument =
gemm.MakeCrossAttnArgument(static_cast<ADataType*>(q_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(kv_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
batch_size,
q_sequence_length,
kv_sequence_length,
head_num,
head_dim,
alpha);
// if(!gemm.IsSupportedArgument(argument))
// {
// std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
// return 0;
// }
ck::index_t BatchCount = batch_size * head_num;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(q_sequence_length) * kv_sequence_length * head_dim * 2 +
size_t(q_sequence_length) * kv_sequence_length * head_dim * 2) *
BatchCount;
std::size_t num_btype = (sizeof(ADataType) * q_sequence_length * head_dim +
sizeof(B0DataType) * head_dim * kv_sequence_length +
sizeof(B1DataType) * kv_sequence_length * head_dim +
sizeof(CDataType) * q_sequence_length * head_dim) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
if(tflops > best_perf)
{
best_perf = tflops;
best_time = ave_time * 1000;
best_kernel = gemm.GetTypeString();
}
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({BatchCount, q_sequence_length, head_dim});
Tensor<B0DataType> b0_g_k_n({BatchCount, head_dim, kv_sequence_length});
Tensor<B1DataType> b1_g_n_o({BatchCount, kv_sequence_length, head_dim});
Tensor<Acc0DataType> acc0_g_m_n(
{BatchCount, q_sequence_length, kv_sequence_length}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount,
q_sequence_length,
kv_sequence_length}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result(
{BatchCount, q_sequence_length, head_dim}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * head_num + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * head_num + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * head_num + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = typename DeviceMHAInstance::C0MatrixMask(kv_sequence_length);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * head_num + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
bool this_run_verification = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol);
printf("Verification: %s, Pass: %s\n",
do_verification ? "ON" : "OFF",
this_run_verification ? "YES" : "NO");
if(!this_run_verification)
{
not_pass = 1;
printf("%d th MHA instance verification Failed \n", i.value);
}
}
});
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Problem Size: BatchCount: " << batch_size << ", HeadNum: " << head_num
<< ", q_sequence_length: " << q_sequence_length
<< ", kv_sequence_length: " << kv_sequence_length << ", head_dim: " << head_dim
<< std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
<< " us" << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
return not_pass;
}
example/32_batched_gemm_scale_softmax_gemm/run_grouped_query_attention_forward_wmma.inc 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 64;
ck::index_t O = 64;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 4;
ck::index_t G1 = 16;
ck::index_t KV_head = QueryGroupNumber;
float alpha = 1;
bool input_permute = false;
bool output_permute = true;
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 == 13)
{
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]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
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 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, KV_head, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * KV_head * K, K, KV_head * K, 1}
// B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{KV_head * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, KV_head, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * KV_head * O, O, 1, KV_head * O}
// B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{KV_head * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
case 4: // A, B0, B1 1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5: // Rand: b1 b0; unit: a
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 6: // Rand: a b0 ; unit: B1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 7: // Rand: a b1 ; unit: b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 8: // Rand: a ; unit: b0 b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 9: // Rand: b0 ; unit: a b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 10: // Rand: b1 ; unit: a b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
float best_perf = .0;
float best_time = .0;
int not_pass = 0;
std::string best_kernel = "";
printf("Verification: %s\n", do_verification ? "ON" : "OFF");
// TODO ANT: replace array with vector?
ck::static_for<0, std::tuple_size_v<DeviceMHAFactory>, 1>{}([&](auto i) -> void {
const auto device_mha_instance = std::get<i>(DeviceMHAFactory{});
using DeviceMHAInstance = ck::remove_cvref_t<decltype(device_mha_instance)>;
auto gemm = DeviceMHAInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
G0,
G1,
alpha,
input_permute,
output_permute);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
// return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
std::size_t num_btype =
(sizeof(ADataType) * M * K + sizeof(CDataType) * M * O) * G0 * G1 +
(sizeof(B0DataType) * K * N + sizeof(B1DataType) * N * O) * G0 * QueryGroupNumber;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
if(tflops > best_perf)
{
best_perf = tflops;
best_time = ave_time * 1000;
best_kernel = gemm.GetTypeString();
}
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g0_g1_m_k({G0, G1, M, K});
Tensor<B0DataType> b0_g0_gq_k_n({G0, QueryGroupNumber, K, N});
Tensor<B1DataType> b1_g0_gq_n_o({G0, QueryGroupNumber, N, O});
Tensor<Acc0DataType> acc0_g0_g1_m_n({G0, G1, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g0_g1_m_n({G0, G1, M, N}); // scratch object after softmax
Tensor<CDataType> c_g0_g1_m_o_host_result({G0, G1, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g0_g1_m_k(idx[0], idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g0_gq_k_n(idx[0], idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g0_gq_n_o(idx[0], idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(a_g0_g1_m_k,
b0_g0_gq_k_n,
acc0_g0_g1_m_n,
a_element_op,
b0_element_op,
acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = typename DeviceMHAInstance::C0MatrixMask(N);
acc0_g0_g1_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[2], idx[3]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument =
ref_softmax.MakeArgument(acc0_g0_g1_m_n, a1_g0_g1_m_n, 1, 0, {3});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g0_g1_m_n,
b1_g0_gq_n_o,
c_g0_g1_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach(
[&](auto& self, auto idx) { self(idx) = c_g0_g1_m_o_host_result(idx); });
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
bool this_run_verification = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol);
printf("Verification: %s, Pass: %s\n",
do_verification ? "ON" : "OFF",
this_run_verification ? "YES" : "NO");
if(!this_run_verification)
{
not_pass = 1;
printf("%d th MQA instance verification Failed \n", i.value);
}
}
});
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Problem Size: BatchCount: " << G0 << ", HeadNum: " << G1 << ", M: " << M
<< ", N: " << N << ", K: " << K << ", O: " << O << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
<< " us" << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
return not_pass;
}
example/32_batched_gemm_scale_softmax_gemm/run_multi_query_attention_forward_wmma.inc 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 120;
ck::index_t N = 1000;
ck::index_t K = 64;
ck::index_t O = 128;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
ck::index_t KV_head = 1;
float alpha = 1;
bool input_permute = false;
bool output_permute = true;
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 == 13)
{
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]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
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 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, KV_head, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * KV_head * K, K, KV_head * K, 1}
// B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{KV_head * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, KV_head, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * KV_head * O, O, 1, KV_head * O}
// B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{KV_head * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
case 4: // A, B0, B1 1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5: // Rand: b1 b0; unit: a
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 6: // Rand: a b0 ; unit: B1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 7: // Rand: a b1 ; unit: b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 8: // Rand: a ; unit: b0 b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 9: // Rand: b0 ; unit: a b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 10: // Rand: b1 ; unit: a b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
float best_perf = .0;
float best_time = .0;
int not_pass = 0;
std::string best_kernel = "";
printf("Verification: %s\n", do_verification ? "ON" : "OFF");
// TODO ANT: replace array with vector?
ck::static_for<0, std::tuple_size_v<DeviceMHAFactory>, 1>{}([&](auto i) -> void {
const auto device_mha_instance = std::get<i>(DeviceMHAFactory{});
using DeviceMHAInstance = ck::remove_cvref_t<decltype(device_mha_instance)>;
auto gemm = DeviceMHAInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
G0,
G1,
alpha,
input_permute,
output_permute);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
// return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(CDataType) * M * O) * G0 * G1 +
(sizeof(B0DataType) * K * N + sizeof(B1DataType) * N * O) * G0;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
if(tflops > best_perf)
{
best_perf = tflops;
best_time = ave_time * 1000;
best_kernel = gemm.GetTypeString();
}
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g0_g1_m_k({G0, G1, M, K});
Tensor<B0DataType> b0_g0_1_k_n({G0, 1, K, N});
Tensor<B1DataType> b1_g0_1_n_o({G0, 1, N, O});
Tensor<Acc0DataType> acc0_g0_g1_m_n({G0, G1, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g0_g1_m_n({G0, G1, M, N}); // scratch object after softmax
Tensor<CDataType> c_g0_g1_m_o_host_result({G0, G1, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g0_g1_m_k(idx[0], idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g0_1_k_n(idx[0], idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g0_1_n_o(idx[0], idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(a_g0_g1_m_k,
b0_g0_1_k_n,
acc0_g0_g1_m_n,
a_element_op,
b0_element_op,
acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = typename DeviceMHAInstance::C0MatrixMask(N);
acc0_g0_g1_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[2], idx[3]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument =
ref_softmax.MakeArgument(acc0_g0_g1_m_n, a1_g0_g1_m_n, 1, 0, {3});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g0_g1_m_n,
b1_g0_1_n_o,
c_g0_g1_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach(
[&](auto& self, auto idx) { self(idx) = c_g0_g1_m_o_host_result(idx); });
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
bool this_run_verification = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol);
printf("Verification: %s, Pass: %s\n",
do_verification ? "ON" : "OFF",
this_run_verification ? "YES" : "NO");
if(!this_run_verification)
{
not_pass = 1;
printf("%d th MQA instance verification Failed \n", i.value);
}
}
});
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Problem Size: BatchCount: " << G0 << ", HeadNum: " << G1 << ", M: " << M
<< ", N: " << N << ", K: " << K << ", O: " << O << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
<< " us" << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
return not_pass;
}
example/32_batched_gemm_scale_softmax_gemm/run_self_attention_wmma.inc 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t sequence_length = 256;
ck::index_t head_dim = 80;
// Output shape C[batch_size, sequence_length, head_num, head_dim]. Batch dim, outer dim, inner
// dim must match GEMM shape C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o]) C_g0_m_g1_o =
// permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t batch_size = 2;
ck::index_t head_num = 8;
float alpha = 1;
bool input_permute = true;
bool output_permute = true;
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 == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
sequence_length = std::stoi(argv[4]);
head_dim = std::stoi(argv[5]);
batch_size = std::stoi(argv[6]);
head_num = std::stoi(argv[7]);
alpha = std::stof(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 7: sequence_length, head_dim, batch_size, head_num\n");
printf("arg8: scale (alpha)\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{batch_size, head_num, sequence_length, head_dim};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute ? std::vector<ck::index_t>{sequence_length * head_num * head_dim,
head_dim,
head_num * head_dim,
1}
// A layout [batch_size, sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * sequence_length * head_dim,
sequence_length * head_dim,
head_dim,
1}; // A layout [batch_size, head_num, sequence_length, head_dim]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{batch_size, head_num, sequence_length, head_dim};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute ? std::vector<ck::index_t>{sequence_length * head_num * head_dim,
head_dim,
head_num * head_dim,
1}
// B0 layout [batch_size, sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * sequence_length * head_dim,
sequence_length * head_dim,
head_dim,
1}; // B0 layout [batch_size, head_num, sequence_length, head_dim]
std::vector<ck::index_t> b1_gs_os_ns_lengths{batch_size, head_num, head_dim, sequence_length};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{sequence_length * head_num * head_dim,
head_dim,
1,
head_num * head_dim}
// B1 layout [batch_size, sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * sequence_length * head_dim,
sequence_length * head_dim,
1,
head_dim}; // B1 layout [batch_size, head_num, sequence_length, head_dim]
std::vector<ck::index_t> c_gs_ms_os_lengths{batch_size, head_num, sequence_length, head_dim};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute ? std::vector<ck::index_t>{sequence_length * head_num * head_dim,
head_dim,
head_num * head_dim,
1}
// C layout [batch_size, sequence_length, head_num, head_dim]
: std::vector<ck::index_t>{
head_num * sequence_length * head_dim,
sequence_length * head_dim,
head_dim,
1}; // C layout [batch_size, head_num, sequence_length, head_dim]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
case 4: // A, B0, B1 1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5: // Rand: b1 b0; unit: a
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 6: // Rand: a b0 ; unit: B1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 7: // Rand: a b1 ; unit: b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 8: // Rand: a ; unit: b0 b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 9: // Rand: b0 ; unit: a b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 10: // Rand: b1 ; unit: a b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
std::vector<ck::index_t> qkv_gs_ms_ks_lengths{
batch_size, head_num, sequence_length, 3, head_dim};
std::vector<ck::index_t> qkv_gs_ms_ks_strides = std::vector<ck::index_t>{
sequence_length * head_num * 3 * head_dim,
3 * head_dim,
head_num * 3 * head_dim,
head_dim,
1}; // qkv layout [batch_size, sequence_length, head_num, 3, head_dim]
Tensor<ADataType> qkv_gs_ms_ks(qkv_gs_ms_ks_lengths, qkv_gs_ms_ks_strides);
// merge qkv into a packed pointer send to device
a_gs_ms_ks.ForEach(
[&](auto& self, auto idx) { qkv_gs_ms_ks(idx[0], idx[1], idx[2], 0, idx[3]) = self(idx); });
b0_gs_ns_ks.ForEach(
[&](auto& self, auto idx) { qkv_gs_ms_ks(idx[0], idx[1], idx[2], 1, idx[3]) = self(idx); });
b1_gs_os_ns.ForEach(
[&](auto& self, auto idx) { qkv_gs_ms_ks(idx[0], idx[1], idx[3], 2, idx[2]) = self(idx); });
DeviceMem qkv_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize() +
sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize() +
sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
qkv_device_buf.ToDevice(qkv_gs_ms_ks.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
float best_perf = .0;
float best_time = .0;
int not_pass = 0;
std::string best_kernel = "";
printf("Verification: %s\n", do_verification ? "ON" : "OFF");
// TODO ANT: replace array with vector?
ck::static_for<0, std::tuple_size_v<DeviceMHAFactory>, 1>{}([&](auto i) -> void {
const auto device_mha_instance = std::get<i>(DeviceMHAFactory{});
using DeviceMHAInstance = ck::remove_cvref_t<decltype(device_mha_instance)>;
auto gemm = DeviceMHAInstance{};
auto invoker = gemm.MakeSelfAttnInvoker();
auto argument =
gemm.MakeSelfAttnArgument(static_cast<ADataType*>(qkv_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
batch_size,
sequence_length,
head_num,
head_dim,
alpha);
// if(!gemm.IsSupportedArgument(argument))
// {
// std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
// return 0;
// }
ck::index_t BatchCount = batch_size * head_num;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(sequence_length) * sequence_length * head_dim * 2 +
size_t(sequence_length) * sequence_length * head_dim * 2) *
BatchCount;
std::size_t num_btype = (sizeof(ADataType) * sequence_length * head_dim +
sizeof(B0DataType) * head_dim * sequence_length +
sizeof(B1DataType) * sequence_length * head_dim +
sizeof(CDataType) * sequence_length * head_dim) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
if(tflops > best_perf)
{
best_perf = tflops;
best_time = ave_time * 1000;
best_kernel = gemm.GetTypeString();
}
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({BatchCount, sequence_length, head_dim});
Tensor<B0DataType> b0_g_k_n({BatchCount, head_dim, sequence_length});
Tensor<B1DataType> b1_g_n_o({BatchCount, sequence_length, head_dim});
Tensor<Acc0DataType> acc0_g_m_n(
{BatchCount, sequence_length, sequence_length}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n(
{BatchCount, sequence_length, sequence_length}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result(
{BatchCount, sequence_length, head_dim}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * head_num + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * head_num + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * head_num + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = typename DeviceMHAInstance::C0MatrixMask(sequence_length);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * head_num + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
bool this_run_verification = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol);
printf("Verification: %s, Pass: %s\n",
do_verification ? "ON" : "OFF",
this_run_verification ? "YES" : "NO");
if(!this_run_verification)
{
not_pass = 1;
printf("%d th MHA instance verification Failed \n", i.value);
}
}
});
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Problem Size: BatchCount: " << batch_size << ", HeadNum: " << head_num
<< ", sequence_length: " << sequence_length << ", head_dim: " << head_dim
<< std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
<< " us" << std::endl;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
return not_pass;
}
example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp 0 → 100644
View file @ 1837040a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g_k_l) * B1_g_l_n
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#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/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using Acc0DataType = F32;
using Acc1DataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
#define CK_MHA_USE_WAVE_1
#define CK_MHA_USE_WAVE_2
#define CK_MHA_USE_WAVE_4
#define CK_MHA_USE_WAVE_8
using DeviceMHAFactory =
std::tuple<
#ifdef CK_MHA_USE_WAVE_1
// 1 wave, mrepeat = 1, nrepeat = 2, k/o repeat = 1~5
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 32, 160, 8, 8,
// Gemm 1
80, 32, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 2, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 64, 80, 8, 8,
// Gemm 1
80, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
32,
// Gemm 0
16, 64, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 2, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 16, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_2
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 64, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 64, 80, 8, 8,
// Gemm 1
80, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
64,
// Gemm 0
32, 32, 160, 8, 8,
// Gemm 1
80, 32, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 2, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 4, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 32, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_4
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 128, 80, 8, 8,
// Gemm 1
80, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 8, 5,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 192, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 12, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
128,
// Gemm 0
64, 64, 48, 8, 8,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 4, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 64, 1, 2>, 8,
MaskingSpec>,
#endif
#ifdef CK_MHA_USE_WAVE_8
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
GemmSpec, TensorSpecA, TensorSpecB0, TensorSpecB1, TensorSpecC, 1,
256,
// Gemm 0
128, 192, 48, 8,4,
// Gemm 1
48, 64, 8,
16, 16, 16,
// Per repeat = wave_m = wave_num, wave_n = 1
1, 12, 3,
// ABlockTransfer MK -> K0 M K1
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
// B0BlockTransfer LK -> K0 L K1
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true,
// B1BlockTransfer NL -> L0 N L1
S<2, 16, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, false,
// CShuffleBlockTransfer MN
1, 1, S<1, 128, 1, 2>, 8,
MaskingSpec>
#endif
>;
// clang-format on
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
Acc0DataType,
Acc1DataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<Acc0DataType, ADataType, Acc0DataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
Acc1DataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_self_attention_wmma.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment