Skip to content

GitLab

"src/tl_templates/vscode:/vscode.git/clone" did not exist on "eab47249c0573771784f8f3b6659b9e720c0ad81"
Commit 400cac28 authored by coderfeli's avatar coderfeli
Browse files

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8

parents 7cec63a6 4c2eff02
Hide whitespace changes
Inline Side-by-side
Showing with 238 additions and 142 deletions
.github/CONTRIBUTING.md 0 → 100644
View file @ 400cac28
We'd love for you to contribute to our source code!
Some helpful links:
- [Code of Conduct guidelines](https://www.contributor-covenant.org/version/2/1/code_of_conduct/code_of_conduct.txt)
- [New issue guidelines](https://github.com/rocm/composable_kernel/blob/develop/.github/ISSUE_TEMPLATE.md)
- [Submitting a pull request guidelines](https://github.com/rocm/composable_kernel/blob/develop/.github/PULL_REQUEST_TEMPLATE.md)
- [Maintainers](https://github.com/rocm/composable_kernel/blob/develop/CONTRIBUTORS.md)
- [General information](https://github.com/rocm/composable_kernel/blob/develop/README.md)
- [ROCm documentation](https://rocm.docs.amd.com/en/latest/how-to/llm-fine-tuning-optimization/optimizing-with-composable-kernel.html)
\ No newline at end of file
.github/ISSUE_TEMPLATE.md 0 → 100644
View file @ 400cac28
When creating an issue, please check if a similar issue already exists.
### When reporting a bug, please include:
- [ ] A descriptive title
- [ ] An isolated way to reproduce the behavior (preferably a docker container with a repro)
- [ ] ROCm version, clang version, Composable Kernel commit pin
- [ ] Environment variables
- [ ] The behavior you expect to see, and the behavior you actually see
### When requesting a feature, please include:
- [ ] A descriptive title
- [ ] A detailed description of the problem you are trying to solve
- [ ] An overview of the suggested solution
- [ ] Explanation why the solution is an improvement
\ No newline at end of file
.github/PULL_REQUEST_TEMPLATE.md 0 → 100644
View file @ 400cac28
## Proposed changes
Please describe the motivation behind the pull request, whether it enables a new feature or fixes a bug. If there are associated pull requests or issues, please link them to the pull request.
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and the unit tests are passing locally
- [ ] I have added inline documentation which enables the maintainers with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant after this pull request
- [ ] (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
- [ ] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged
## Discussion
If this is a relatively large or complex change, feel free to start a discussion by explaining why you chose the solution you did and what alternatives you considered
CMakeLists.txt
View file @ 400cac28
...@@ -183,14 +183,17 @@ message("Building CK for the following targets: ${SUPPORTED_GPU_TARGETS}") ...@@ -183,14 +183,17 @@ message("Building CK for the following targets: ${SUPPORTED_GPU_TARGETS}")
if (SUPPORTED_GPU_TARGETS MATCHES "gfx9") if (SUPPORTED_GPU_TARGETS MATCHES "gfx9")
message("Enabling XDL instances") message("Enabling XDL instances")
add_definitions(-DCK_USE_XDL) add_definitions(-DCK_USE_XDL)
set(CK_USE_XDL "ON")
endif() endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx94") if (SUPPORTED_GPU_TARGETS MATCHES "gfx94")
message("Enabling FP8 gemms on native architectures") message("Enabling FP8 gemms on native architectures")
add_definitions(-DCK_USE_GFX94) add_definitions(-DCK_USE_GFX94)
set(CK_USE_GFX94 "ON")
endif() endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
message("Enabling WMMA instances") message("Enabling WMMA instances")
add_definitions(-DCK_USE_WMMA) add_definitions(-DCK_USE_WMMA)
set(CK_USE_WMMA "ON")
endif() endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx12") if (SUPPORTED_GPU_TARGETS MATCHES "gfx12")
add_definitions(-DCK_USE_OCP_FP8) add_definitions(-DCK_USE_OCP_FP8)
...@@ -204,6 +207,7 @@ endif() ...@@ -204,6 +207,7 @@ endif()
option(CK_USE_FP8_ON_UNSUPPORTED_ARCH "Enable FP8 GEMM instances on older architectures" OFF) option(CK_USE_FP8_ON_UNSUPPORTED_ARCH "Enable FP8 GEMM instances on older architectures" OFF)
if(CK_USE_FP8_ON_UNSUPPORTED_ARCH AND (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx908")) if(CK_USE_FP8_ON_UNSUPPORTED_ARCH AND (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx908"))
add_definitions(-DCK_USE_FP8_ON_UNSUPPORTED_ARCH) add_definitions(-DCK_USE_FP8_ON_UNSUPPORTED_ARCH)
set(CK_USE_FP8_ON_UNSUPPORTED_ARCH "ON")
endif() endif()
# CK config file to record supported datatypes, etc. # CK config file to record supported datatypes, etc.
......
Dockerfile
View file @ 400cac28
...@@ -64,6 +64,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -64,6 +64,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
nano \ nano \
zlib1g-dev \ zlib1g-dev \
zip \ zip \
libzstd-dev \
openssh-server \ openssh-server \
clang-format-12 \ clang-format-12 \
kmod && \ kmod && \
...@@ -93,7 +94,7 @@ RUN pip install --upgrade cmake==3.27.5 && \ ...@@ -93,7 +94,7 @@ RUN pip install --upgrade cmake==3.27.5 && \
dpkg -i dumb-init_*.deb && rm dumb-init_*.deb && \ dpkg -i dumb-init_*.deb && rm dumb-init_*.deb && \
# Install packages for processing the performance results # Install packages for processing the performance results
pip3 install --upgrade pip && \ pip3 install --upgrade pip && \
pip3 install sqlalchemy==1.4.46 pymysql pandas==2.0.3 setuptools-rust sshtunnel==0.4.0 && \ pip3 install sqlalchemy==2.0.36 pymysql pandas==2.2.3 setuptools-rust sshtunnel==0.4.0 && \
# Add render group # Add render group
groupadd -f render && \ groupadd -f render && \
# Install the new rocm-cmake version # Install the new rocm-cmake version
......
Jenkinsfile
View file @ 400cac28
...@@ -566,11 +566,9 @@ def Build_CK(Map conf=[:]){ ...@@ -566,11 +566,9 @@ def Build_CK(Map conf=[:]){
ls -ltr ls -ltr
CC=hipcc CXX=hipcc cmake -Bbuild . -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install" CC=hipcc CXX=hipcc cmake -Bbuild . -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install"
cmake --build build -- -j cmake --build build -- -j
ctest --test-dir build
""" """
} }
dir("hipTensor-${params.hipTensor_branch}/build"){
sh 'ctest'
}
} }
} }
} }
......
docs/sphinx/requirements.in
View file @ 400cac28
rocm-docs-core==1.11.0 rocm-docs-core==1.12.0
sphinxcontrib-bibtex==2.6.3 sphinxcontrib-bibtex==2.6.3
docs/sphinx/requirements.txt
View file @ 400cac28
...@@ -103,7 +103,7 @@ requests==2.32.3 ...@@ -103,7 +103,7 @@ requests==2.32.3
# via # via
# pygithub # pygithub
# sphinx # sphinx
rocm-docs-core==1.11.0 rocm-docs-core==1.12.0
# via -r requirements.in # via -r requirements.in
six==1.16.0 six==1.16.0
# via pybtex # via pybtex
......
example/24_batched_gemm/batched_gemm_xdl_bf16_v3.cpp
View file @ 400cac28
...@@ -78,14 +78,14 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD ...@@ -78,14 +78,14 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD
2, // ABlockTransferSrcVectorDim 2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector 8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1 8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM 0, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1 S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim 2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector 8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1 8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN 0, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle 1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle 1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
......
example/ck_tile/01_fmha/codegen/cpp_symbol_map.py
View file @ 400cac28
...@@ -119,6 +119,7 @@ PIPELINE_MAP = { ...@@ -119,6 +119,7 @@ PIPELINE_MAP = {
PIPELINE_ENUM_MAP = { PIPELINE_ENUM_MAP = {
"qr" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS", "qr" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS",
"qr_async" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC", "qr_async" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC",
"qr_nwarp_sshuffle" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS",
} }
BOOL_MAP = { BOOL_MAP = {
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View file @ 400cac28
...@@ -44,13 +44,12 @@ FMHA_FWD_KERNEL_BODY=""" ...@@ -44,13 +44,12 @@ FMHA_FWD_KERNEL_BODY="""
using fmha_dtype_{F_idx} = {F_dtype}; using fmha_dtype_{F_idx} = {F_dtype};
using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>; using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>;
using fmha_warp_tile_{F_idx} = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx}, using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx},
ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>, ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>,
fmha_warp_tile_{F_idx}, ck_tile::sequence<{F_wm0}, {F_wn0}, {F_wk0}>,
ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>, ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>,
fmha_warp_tile_{F_idx}, ck_tile::sequence<{F_wm1}, {F_wn1}, {F_wk1}>,
{F_vlayout}>; {F_vlayout}>;
using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad}, using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
...@@ -306,15 +305,19 @@ class FmhaFwdTileSize: ...@@ -306,15 +305,19 @@ class FmhaFwdTileSize:
F_rm1 : int # number of warps for gemm1 along q seqlen F_rm1 : int # number of warps for gemm1 along q seqlen
F_rn1 : int # number of warps for gemm1 along head dim v F_rn1 : int # number of warps for gemm1 along head dim v
F_rk1 : int # number of warps for gemm1 along k seqlen (not used) F_rk1 : int # number of warps for gemm1 along k seqlen (not used)
F_wm : int # warp size along m (warp size) F_wm0 : int # gemm0 warp size along m
F_wn : int # warp size along n F_wn0 : int # gemm0 warp size along n
F_wk : int # warp size along k F_wk0 : int # gemm0 warp size along k
F_wm1 : int # gemm1 warp size along m
F_wn1 : int # gemm1 warp size along n
F_wk1 : int # gemm1 warp size along k
F_occupancy : int # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy F_occupancy : int # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
@property @property
def name(self) -> str: def name(self) -> str:
return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0max}" +\ return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0max}" +\
f"_r{self.F_rm0}x{self.F_rn0}x{self.F_rk0}_r{self.F_rm1}x{self.F_rn1}x{self.F_rk1}" +\ f"_r{self.F_rm0}x{self.F_rn0}x{self.F_rk0}_r{self.F_rm1}x{self.F_rn1}x{self.F_rk1}" +\
f"_w{self.F_wm}x{self.F_wn}x{self.F_wk}" + ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}") f"_w{self.F_wm0}x{self.F_wn0}x{self.F_wk0}_w{self.F_wm1}x{self.F_wn1}x{self.F_wk1}" +\
("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
@dataclass @dataclass
class FmhaFwdKernel: class FmhaFwdKernel:
...@@ -352,9 +355,12 @@ class FmhaFwdKernel: ...@@ -352,9 +355,12 @@ class FmhaFwdKernel:
F_rm1 = self.F_tile.F_rm1, F_rm1 = self.F_tile.F_rm1,
F_rn1 = self.F_tile.F_rn1, F_rn1 = self.F_tile.F_rn1,
F_rk1 = self.F_tile.F_rk1, F_rk1 = self.F_tile.F_rk1,
F_wm = self.F_tile.F_wm, F_wm0 = self.F_tile.F_wm0,
F_wn = self.F_tile.F_wn, F_wn0 = self.F_tile.F_wn0,
F_wk = self.F_tile.F_wk, F_wk0 = self.F_tile.F_wk0,
F_wm1 = self.F_tile.F_wm1,
F_wn1 = self.F_tile.F_wn1,
F_wk1 = self.F_tile.F_wk1,
F_vlayout = LAYOUT_MAP[self.F_pipeline.F_vlayout], F_vlayout = LAYOUT_MAP[self.F_pipeline.F_vlayout],
F_spad = BOOL_MAP[self.F_pipeline.F_spad], F_spad = BOOL_MAP[self.F_pipeline.F_spad],
F_skpad = BOOL_MAP[self.F_pipeline.F_skpad], F_skpad = BOOL_MAP[self.F_pipeline.F_skpad],
...@@ -409,17 +415,17 @@ class FmhaFwdKernel: ...@@ -409,17 +415,17 @@ class FmhaFwdKernel:
def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]: def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
if dtype == 'fp16' or dtype == 'bf16': if dtype == 'fp16' or dtype == 'bf16':
return { return {
'32' : FmhaFwdTileSize(128, 64, 16, 32, 32, 32, 2, 1, 1, 2, 1, 1, 32, 32, 16, -1), '32' : FmhaFwdTileSize(128, 64, 16, 32, 32, 32, 2, 1, 1, 2, 1, 1, 32, 32, 16, 32, 32, 16, -1),
'64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1), '64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1),
## '96' : FmhaFwdTileSize(128, 128, 32, 128, 32, 96, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1), ### '96' : FmhaFwdTileSize(128, 128, 32, 128, 32, 96, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1),
'128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1), '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1),
'256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1), '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1),
} }
elif dtype == 'fp8' or dtype == 'bf8': elif dtype == 'fp8' or dtype == 'bf8':
return { return {
'64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 2, 1, 1, 2, 1, 1, 32, 32, 32, -1), '64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 2, 1, 1, 2, 1, 1, 32, 32, 32, 32, 32, 32, -1),
'128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 32, -1), '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 32, 32, 32, 32, -1),
'256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 32, -1) '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 32, 32, 32, 32, -1),
} }
else: else:
return None return None
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View file @ 400cac28
...@@ -39,6 +39,7 @@ K0_MAX_SUBMAX_MAP = { ...@@ -39,6 +39,7 @@ K0_MAX_SUBMAX_MAP = {
FMHA_FWD_SPLITKV_PIPELINE_MAP = { FMHA_FWD_SPLITKV_PIPELINE_MAP = {
"qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS", "qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS",
"qr_nwarp_sshuffle" : "ck_tile::BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS",
"qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync", "qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync",
} }
...@@ -50,13 +51,12 @@ namespace {{ ...@@ -50,13 +51,12 @@ namespace {{
template <bool kHasUnevenSplits> template <bool kHasUnevenSplits>
struct kernel_runner {{ struct kernel_runner {{
using fmha_block_tile = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>; using fmha_block_tile = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>;
using fmha_warp_tile = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
using fmha_shape = ck_tile::TileFmhaShape<fmha_block_tile, using fmha_shape = ck_tile::TileFmhaShape<fmha_block_tile,
ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>, ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>,
fmha_warp_tile, ck_tile::sequence<{F_wm0}, {F_wn0}, {F_wk0}>,
ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>, ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>,
fmha_warp_tile, ck_tile::sequence<{F_wm1}, {F_wn1}, {F_wk1}>,
{F_vlayout}>; {F_vlayout}>;
using fmha_trait = ck_tile::TileFmhaFwdSplitKVTraits<{F_spad}, using fmha_trait = ck_tile::TileFmhaFwdSplitKVTraits<{F_spad},
...@@ -161,9 +161,8 @@ using fmha_pipeline_problem = ck_tile::BlockFmhaSplitKVCombinePipelineProblem< ...@@ -161,9 +161,8 @@ using fmha_pipeline_problem = ck_tile::BlockFmhaSplitKVCombinePipelineProblem<
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType, typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType, typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
{F_hdim}, {F_hdim},
{F_bm0},
{F_bn1},
{F_mode}, {F_mode},
{F_bn1},
fmha_trait>; fmha_trait>;
using fmha_pipeline = ck_tile::BlockFmhaFwdSplitKVCombinePipeline< using fmha_pipeline = ck_tile::BlockFmhaFwdSplitKVCombinePipeline<
...@@ -177,9 +176,11 @@ using fmha_epilogue = ...@@ -177,9 +176,11 @@ using fmha_epilogue =
false, false>>; false, false>>;
using fmha_kernel = using fmha_kernel =
ck_tile::FmhaFwdSplitKVCombineKernel<ck_tile::FmhaFwdSplitKVCombineTilePartitioner<{F_bm0}, {F_bn1}>, ck_tile::FmhaFwdSplitKVCombineKernel<
fmha_pipeline, ck_tile::FmhaFwdSplitKVCombineTilePartitioner<
fmha_epilogue>; fmha_pipeline_problem::kM0, fmha_pipeline_problem::kN1>,
fmha_pipeline,
fmha_epilogue>;
static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a) static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
{{ {{
...@@ -192,7 +193,7 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a) ...@@ -192,7 +193,7 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
}}; }};
}} }}
using trait_{F_idx} = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn1}, using trait_{F_idx} = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bn1},
{F_lse}, {F_squant}, {F_spad}, {F_dvpad}>; {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
#include <iostream> #include <iostream>
...@@ -250,16 +251,25 @@ float fmha_fwd_splitkv(fmha_fwd_splitkv_traits t, fmha_fwd_splitkv_args a, const ...@@ -250,16 +251,25 @@ float fmha_fwd_splitkv(fmha_fwd_splitkv_traits t, fmha_fwd_splitkv_args a, const
FMHA_FWD_SPLITKV_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.do_fp8_static_quant == {F_squant}) && FMHA_FWD_SPLITKV_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.do_fp8_static_quant == {F_squant}) &&
((a.block_table_ptr != nullptr) == {F_pagedkv}) && ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{ ((a.block_table_ptr != nullptr) == {F_pagedkv}) && ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, true, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>; using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, true, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
// get combine kernel tile sizes
using OaccDataType = typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType;
constexpr ck_tile::index_t kM0 = ck_tile::BlockFmhaSplitKVCombinePipelineTileSizes<OaccDataType, /*F_bn1=*/32>::kM0;
// make sure we can reuse the padding flags in combine kernels
static_assert({F_bm0} % kM0 == 0);
static_assert({F_bn1} % 32 == 0);
if (t.has_lse) {{ if (t.has_lse) {{
if constexpr (std::is_same_v<{F_dtype}, ck_tile::fp8_t>) {{ if constexpr (std::is_same_v<{F_dtype}, FmhaFwdFp8>) {{
return -1; return -1;
}} else {{ }} else {{
using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, true, {F_squant}, {F_spad}, {F_dvpad}>; using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, /*F_bn1=*/32, true, {F_squant}, {F_spad}, {F_dvpad}>;
return fmha_fwd_splitkv_<traits_, traits2_>(s, a); return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
}} }}
}} else {{ }} else {{
using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, false, {F_squant}, {F_spad}, {F_dvpad}>; using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, /*F_bn1=*/32, false, {F_squant}, {F_spad}, {F_dvpad}>;
return fmha_fwd_splitkv_<traits_, traits2_>(s, a); return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
}} }}
...@@ -302,7 +312,7 @@ class FmhaFwdSplitKVApiTrait: ...@@ -302,7 +312,7 @@ class FmhaFwdSplitKVApiTrait:
if self.pipeline_tag == 'qr_async': if self.pipeline_tag == 'qr_async':
if self.spad == 't' : return 'true' # always support if self.spad == 't' : return 'true' # always support
else : return 'true' else : return 'true'
elif self.pipeline_tag in ['qr']: elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
if self.spad == 't' : return f'true /*a.seqlen_q % {self.bm0} != 0*/' # TODO: order of get_pipelines() matters! (ugly) if self.spad == 't' : return f'true /*a.seqlen_q % {self.bm0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.seqlen_q % {self.bm0} == 0' else : return f'a.seqlen_q % {self.bm0} == 0'
else: assert False else: assert False
...@@ -313,7 +323,7 @@ class FmhaFwdSplitKVApiTrait: ...@@ -313,7 +323,7 @@ class FmhaFwdSplitKVApiTrait:
if self.pipeline_tag == 'qr_async': if self.pipeline_tag == 'qr_async':
if self.skpad == 't' : return f'a.seqlen_k == 0 || a.seqlen_k % {self.bn0} != 0' if self.skpad == 't' : return f'a.seqlen_k == 0 || a.seqlen_k % {self.bn0} != 0'
else : return f'a.seqlen_k != 0 && a.seqlen_k % {self.bn0} == 0' else : return f'a.seqlen_k != 0 && a.seqlen_k % {self.bn0} == 0'
elif self.pipeline_tag in ['qr', 'qr_fp8']: elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
if self.skpad == 't' : return f'true /*a.seqlen_k % {self.bn0} != 0*/' # TODO: order of get_pipelines() matters! (ugly) if self.skpad == 't' : return f'true /*a.seqlen_k % {self.bn0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.seqlen_k % {self.bn0} == 0' else : return f'a.seqlen_k % {self.bn0} == 0'
else: assert False else: assert False
...@@ -324,7 +334,7 @@ class FmhaFwdSplitKVApiTrait: ...@@ -324,7 +334,7 @@ class FmhaFwdSplitKVApiTrait:
vec = int((32 * 4) / DTYPE_BITS[self.dtype]) vec = int((32 * 4) / DTYPE_BITS[self.dtype])
if self.dpad == 't': return f'a.hdim_q % {vec} == 0' if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
else : assert False else : assert False
elif self.pipeline_tag in ['qr']: elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max] bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
if self.dpad == 't': return f'true /*a.hdim_q % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly) if self.dpad == 't': return f'true /*a.hdim_q % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.hdim_q % {bk0submax} == 0' else : return f'a.hdim_q % {bk0submax} == 0'
...@@ -336,7 +346,7 @@ class FmhaFwdSplitKVApiTrait: ...@@ -336,7 +346,7 @@ class FmhaFwdSplitKVApiTrait:
vec = int((32 * 4) / DTYPE_BITS[self.dtype]) vec = int((32 * 4) / DTYPE_BITS[self.dtype])
if self.dvpad == 't': return f'a.hdim_v % {vec} == 0' if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
else : assert False else : assert False
elif self.pipeline_tag in ['qr']: elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max] bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
if self.dvpad == 't': return f'true /*a.hdim_v % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly) if self.dvpad == 't': return f'true /*a.hdim_v % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.hdim_v % {bk0submax} == 0' else : return f'a.hdim_v % {bk0submax} == 0'
...@@ -447,12 +457,11 @@ class FmhaFwdSplitKVApiPool: ...@@ -447,12 +457,11 @@ class FmhaFwdSplitKVApiPool:
@dataclass @dataclass
class FmhaFwdSplitKVCombineTileSize: class FmhaFwdSplitKVCombineTileSize:
F_bm0 : int # tile size along q seqlen
F_bn1 : int # tile size along v head_dim F_bn1 : int # tile size along v head_dim
F_occupancy : int # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy F_occupancy : int # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
@property @property
def name(self) -> str: def name(self) -> str:
return f"b{self.F_bm0}x{self.F_bn1}" +\ return f"b{self.F_bn1}" +\
("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}") ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
@dataclass @dataclass
...@@ -485,9 +494,12 @@ class FmhaFwdSplitKVKernel: ...@@ -485,9 +494,12 @@ class FmhaFwdSplitKVKernel:
F_rm1 = self.F_tile.F_rm1, F_rm1 = self.F_tile.F_rm1,
F_rn1 = self.F_tile.F_rn1, F_rn1 = self.F_tile.F_rn1,
F_rk1 = self.F_tile.F_rk1, F_rk1 = self.F_tile.F_rk1,
F_wm = self.F_tile.F_wm, F_wm0 = self.F_tile.F_wm0,
F_wn = self.F_tile.F_wn, F_wn0 = self.F_tile.F_wn0,
F_wk = self.F_tile.F_wk, F_wk0 = self.F_tile.F_wk0,
F_wm1 = self.F_tile.F_wm1,
F_wn1 = self.F_tile.F_wn1,
F_wk1 = self.F_tile.F_wk1,
F_vlayout = LAYOUT_MAP[self.F_pipeline.F_vlayout], F_vlayout = LAYOUT_MAP[self.F_pipeline.F_vlayout],
F_spad = BOOL_MAP[self.F_pipeline.F_spad], F_spad = BOOL_MAP[self.F_pipeline.F_spad],
F_skpad = BOOL_MAP[self.F_pipeline.F_skpad], F_skpad = BOOL_MAP[self.F_pipeline.F_skpad],
...@@ -553,7 +565,6 @@ class FmhaFwdSplitKVCombineKernel: ...@@ -553,7 +565,6 @@ class FmhaFwdSplitKVCombineKernel:
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = FWD_DTYPE_MAP[self.F_dtype], F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0,
F_bn1 = self.F_tile.F_bn1, F_bn1 = self.F_tile.F_bn1,
F_spad = BOOL_MAP[self.F_pipeline.F_spad], F_spad = BOOL_MAP[self.F_pipeline.F_spad],
F_dvpad = BOOL_MAP[self.F_pipeline.F_dvpad], F_dvpad = BOOL_MAP[self.F_pipeline.F_dvpad],
...@@ -577,17 +588,17 @@ class FmhaFwdSplitKVCombineKernel: ...@@ -577,17 +588,17 @@ class FmhaFwdSplitKVCombineKernel:
def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]: def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
if dtype == 'fp16' or dtype == 'bf16': if dtype == 'fp16' or dtype == 'bf16':
return { return {
'32' : FmhaFwdTileSize(32, 64, 16, 32, 32, 32, 2, 1, 1, 2, 1, 1, 16, 16, 16, -1), '32' : FmhaFwdTileSize(32, 64, 16, 32, 32, 32, 2, 1, 1, 2, 1, 1, 16, 16, 16, 16, 16, 16, -1),
'64' : FmhaFwdTileSize(64, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 16, 16, 16, -1), '64' : FmhaFwdTileSize(64, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 16, 16, 16, 16, 16, 16, -1),
## '96' : FmhaFwdTileSize(64, 128, 32, 128, 32, 96, 4, 1, 1, 4, 1, 1, 16, 16, 16, -1), ### '96' : FmhaFwdTileSize(64, 128, 32, 128, 32, 96, 4, 1, 1, 4, 1, 1, 16, 16, 16, 16, 16, 16, -1),
'128' : FmhaFwdTileSize(64, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 16, 16, 16, -1), '128' : FmhaFwdTileSize(64, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 16, 16, 16, 16, 16, 16, -1),
'256' : FmhaFwdTileSize(64, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 16, 16, 16, -1), '256' : FmhaFwdTileSize(64, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 16, 16, 16, 16, 16, 16, -1),
} }
elif dtype == 'fp8' or dtype == 'bf8': elif dtype == 'fp8' or dtype == 'bf8':
return { return {
'64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 2, 1, 1, 2, 1, 1, 32, 32, 32, -1), '64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 2, 1, 1, 2, 1, 1, 32, 32, 32, 32, 32, 32, -1),
'128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 32, -1), '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 4, 1, 1, 32, 32, 32, 32, 32, 32, -1),
'256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 32, -1) '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 32, 32, 32, 32, -1),
} }
else: else:
return None return None
...@@ -595,17 +606,17 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]: ...@@ -595,17 +606,17 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[dict]: def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
if dtype == 'fp16' or dtype == 'bf16': if dtype == 'fp16' or dtype == 'bf16':
return { return {
'32' : FmhaFwdSplitKVCombineTileSize(16, 16, -1), '32' : FmhaFwdSplitKVCombineTileSize(32, -1),
'64' : FmhaFwdSplitKVCombineTileSize(32, 32, -1), '64' : FmhaFwdSplitKVCombineTileSize(32, -1),
## '96' : FmhaFwdSplitKVCombineTileSize(32, 64, -1), ### '96' : FmhaFwdSplitKVCombineTileSize(32, -1),
'128' : FmhaFwdSplitKVCombineTileSize(32, 64, -1), '128' : FmhaFwdSplitKVCombineTileSize(32, -1),
'256' : FmhaFwdSplitKVCombineTileSize(32, 128, -1), '256' : FmhaFwdSplitKVCombineTileSize(32, -1),
} }
elif dtype == 'fp8' or dtype == 'bf8': elif dtype == 'fp8' or dtype == 'bf8':
return { return {
'64' : FmhaFwdSplitKVCombineTileSize(64, 32, -1), '64' : FmhaFwdSplitKVCombineTileSize(32, -1),
'128' : FmhaFwdSplitKVCombineTileSize(64, 64, -1), '128' : FmhaFwdSplitKVCombineTileSize(32, -1),
'256' : FmhaFwdSplitKVCombineTileSize(64, 128, -1), '256' : FmhaFwdSplitKVCombineTileSize(32, -1),
} }
else: else:
return None return None
......
example/ck_tile/01_fmha/fmha_fwd.hpp
View file @ 400cac28
...@@ -709,7 +709,6 @@ std::string fmha_fwd_splitkv_get_name_(); ...@@ -709,7 +709,6 @@ std::string fmha_fwd_splitkv_get_name_();
template <ck_tile::index_t HDim_, template <ck_tile::index_t HDim_,
typename DataType_, typename DataType_,
bool kIsGroupMode_, bool kIsGroupMode_,
ck_tile::index_t kM0_,
ck_tile::index_t kN1_, ck_tile::index_t kN1_,
bool kStoreLse_, bool kStoreLse_,
bool kDoFp8StaticQuant_, bool kDoFp8StaticQuant_,
...@@ -720,7 +719,6 @@ struct fmha_fwd_splitkv_combine_traits_ ...@@ -720,7 +719,6 @@ struct fmha_fwd_splitkv_combine_traits_
static constexpr ck_tile::index_t HDim = HDim_; static constexpr ck_tile::index_t HDim = HDim_;
using DataType = ck_tile::remove_cvref_t<DataType_>; using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool kIsGroupMode = kIsGroupMode_; static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr ck_tile::index_t kM0 = kM0_;
static constexpr ck_tile::index_t kN1 = kN1_; static constexpr ck_tile::index_t kN1 = kN1_;
static constexpr bool kStoreLse = kStoreLse_; static constexpr bool kStoreLse = kStoreLse_;
static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_; static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
......
example/ck_tile/03_gemm/gemm_basic.cpp
View file @ 400cac28
...@@ -15,7 +15,7 @@ ...@@ -15,7 +15,7 @@
#include "gemm_basic.hpp" #include "gemm_basic.hpp"
template <typename ALayout, typename BLayout, typename CLayout> template <typename ALayout, typename BLayout, typename CLayout>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s) float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{ {
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part. // The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false; constexpr bool kPadM = false;
...@@ -79,17 +79,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s) ...@@ -79,17 +79,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy. // Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>; using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args.p_a, auto kargs = Kernel::MakeKernelArgs(args);
args.p_b,
args.p_c, const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
args.M,
args.N,
args.K,
args.stride_A,
args.stride_B,
args.stride_C);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize(); constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs)) if(!Kernel::IsSupportedArgument(kargs))
......
example/ck_tile/03_gemm/gemm_basic.hpp
View file @ 400cac28
...@@ -51,20 +51,6 @@ using BDataType = Types::BDataType; ...@@ -51,20 +51,6 @@ using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType; using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType; using CDataType = Types::CDataType;
struct gemm_basic_args
{
const void* p_a;
const void* p_b;
void* p_c;
ck_tile::index_t kbatch;
ck_tile::index_t M;
ck_tile::index_t N;
ck_tile::index_t K;
ck_tile::index_t stride_A;
ck_tile::index_t stride_B;
ck_tile::index_t stride_C;
};
auto create_args(int argc, char* argv[]) auto create_args(int argc, char* argv[])
{ {
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
...@@ -89,4 +75,4 @@ auto create_args(int argc, char* argv[]) ...@@ -89,4 +75,4 @@ auto create_args(int argc, char* argv[])
} }
// host API // host API
float gemm_calc(gemm_basic_args args, const ck_tile::stream_config& s); float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s);
example/ck_tile/03_gemm/run_gemm_example.inc
View file @ 400cac28
...@@ -16,11 +16,11 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf, ...@@ -16,11 +16,11 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
int n_warmup, int n_warmup,
int n_repeat) int n_repeat)
{ {
gemm_basic_args args; ck_tile::GemmHostArgs args;
args.p_a = a_m_k_dev_buf.GetDeviceBuffer(); args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.p_b = b_k_n_dev_buf.GetDeviceBuffer(); args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.p_c = c_m_n_dev_buf.GetDeviceBuffer(); args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.kbatch = kbatch; args.k_batch = kbatch;
args.M = M; args.M = M;
args.N = N; args.N = N;
args.K = K; args.K = K;
...@@ -161,14 +161,39 @@ int run_gemm_example_with_layouts(int argc, ...@@ -161,14 +161,39 @@ int run_gemm_example_with_layouts(int argc,
c_m_n_gpu_ref.SetZero(); c_m_n_gpu_ref.SetZero();
c_m_n_gpu_buf_ref.SetZero(); c_m_n_gpu_buf_ref.SetZero();
ADataType* d_A;
BDataType* d_B;
CDataType* d_C;
ck_tile::hip_check_error(hipMalloc(&d_A, M * K * sizeof(ADataType)));
ck_tile::hip_check_error(hipMalloc(&d_B, N * K * sizeof(BDataType)));
ck_tile::hip_check_error(hipMalloc(&d_C, M * N * sizeof(CDataType)));
ck_tile::hip_check_error(hipMemcpy(d_A,
a_m_k_dev_buf.GetDeviceBuffer(),
M * K * sizeof(ADataType),
hipMemcpyHostToDevice));
ck_tile::hip_check_error(hipMemcpy(d_B,
b_k_n_dev_buf.GetDeviceBuffer(),
N * K * sizeof(BDataType),
hipMemcpyHostToDevice));
ck_tile::reference_gemm_gpu<ADataType, ck_tile::reference_gemm_gpu<ADataType,
BDataType, BDataType,
AccDataType, AccDataType,
CDataType, CDataType,
ALayout, ALayout,
BLayout, BLayout,
CLayout>( CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
a_m_k_dev_buf, b_k_n_dev_buf, c_m_n_gpu_buf_ref, M, N, K, stride_A, stride_B, stride_C);
ck_tile::hip_check_error(hipMemcpy(c_m_n_gpu_buf_ref.GetDeviceBuffer(),
d_C,
M * N * sizeof(CDataType),
hipMemcpyDeviceToHost));
ck_tile::hip_check_error(hipFree(d_A));
ck_tile::hip_check_error(hipFree(d_B));
ck_tile::hip_check_error(hipFree(d_C));
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data()); c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref); pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref);
......
example/ck_tile/13_moe_sorting/moe_sorting_api.cpp
View file @ 400cac28
...@@ -3,18 +3,42 @@ ...@@ -3,18 +3,42 @@
#include "moe_sorting_api.hpp" #include "moe_sorting_api.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \ #define MOE_SORTING_DISPATCH_ETILE(unroll_num_, expert_tile_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \ constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \ constexpr ck_tile::index_t expert_tile = expert_tile_; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \ using ms_problem = \
auto kargs = kernel::MakeKargs(a); \ ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num, expert_tile>; \
const dim3 grids = kernel::GridSize(a); \ using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
const dim3 blocks = kernel::BlockSize(a); \ auto kargs = kernel::MakeKargs(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \ const dim3 grids = kernel::GridSize(a); \
float ave_time = ck_tile::launch_kernel( \ const dim3 blocks = kernel::BlockSize(a); \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \ const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time; return ave_time;
#define MOE_SORTING_DISPATCH(unroll_num_) \
if(a.num_experts <= 8) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 8) \
} \
else if(a.num_experts <= 16) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 16) \
} \
else if(a.num_experts <= 32) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 32) \
} \
else if(a.num_experts <= 64) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 64) \
} \
else \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 0) \
}
float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s) float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s)
{ {
if(t.weight_type == "fp32" && t.index_type == "int32") if(t.weight_type == "fp32" && t.index_type == "int32")
...@@ -49,21 +73,12 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi ...@@ -49,21 +73,12 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
case(6): { case(6): {
MOE_SORTING_DISPATCH(6); MOE_SORTING_DISPATCH(6);
} }
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): { case(8): {
MOE_SORTING_DISPATCH(8); MOE_SORTING_DISPATCH(8);
} }
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): { case(10): {
MOE_SORTING_DISPATCH(10); MOE_SORTING_DISPATCH(10);
} }
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: { default: {
MOE_SORTING_DISPATCH(4); MOE_SORTING_DISPATCH(4);
} }
......
example/ck_tile/13_moe_sorting/script/smoke_test.sh
View file @ 400cac28
...@@ -16,4 +16,5 @@ $EXE -t=127 -e=99 -k=19 ...@@ -16,4 +16,5 @@ $EXE -t=127 -e=99 -k=19
$EXE -t=71 -e=11 -k=11 $EXE -t=71 -e=11 -k=11
$EXE -t=1 -e=1 -k=1 $EXE -t=1 -e=1 -k=1
$EXE -t=99 -e=2 -k=1 $EXE -t=99 -e=2 -k=1
$EXE -t=333 -e=99 -k=13 $EXE -t=333 -e=99 -k=13
\ No newline at end of file $EXE -t=128 -e=32 -k=5 -moe_buf_size=262144
example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp
View file @ 400cac28
...@@ -3,18 +3,42 @@ ...@@ -3,18 +3,42 @@
#include "fused_moesorting.hpp" #include "fused_moesorting.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \ #define MOE_SORTING_DISPATCH_ETILE(unroll_num_, expert_tile_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \ constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \ constexpr ck_tile::index_t expert_tile = expert_tile_; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \ using ms_problem = \
auto kargs = kernel::MakeKargs(a); \ ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num, expert_tile>; \
const dim3 grids = kernel::GridSize(a); \ using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
const dim3 blocks = kernel::BlockSize(a); \ auto kargs = kernel::MakeKargs(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \ const dim3 grids = kernel::GridSize(a); \
float ave_time = ck_tile::launch_kernel( \ const dim3 blocks = kernel::BlockSize(a); \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \ const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time; return ave_time;
#define MOE_SORTING_DISPATCH(unroll_num_) \
if(a.num_experts <= 8) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 8) \
} \
else if(a.num_experts <= 16) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 16) \
} \
else if(a.num_experts <= 32) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 32) \
} \
else if(a.num_experts <= 64) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 64) \
} \
else \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 0) \
}
float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s) float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s)
{ {
if(t.weight_type == "fp32" && t.index_type == "int32") if(t.weight_type == "fp32" && t.index_type == "int32")
...@@ -49,21 +73,12 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til ...@@ -49,21 +73,12 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til
case(6): { case(6): {
MOE_SORTING_DISPATCH(6); MOE_SORTING_DISPATCH(6);
} }
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): { case(8): {
MOE_SORTING_DISPATCH(8); MOE_SORTING_DISPATCH(8);
} }
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): { case(10): {
MOE_SORTING_DISPATCH(10); MOE_SORTING_DISPATCH(10);
} }
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: { default: {
MOE_SORTING_DISPATCH(4); MOE_SORTING_DISPATCH(4);
} }
......
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View file @ 400cac28
...@@ -16,7 +16,7 @@ ...@@ -16,7 +16,7 @@
#include "batched_gemm.hpp" #include "batched_gemm.hpp"
template <typename ALayout, typename BLayout, typename CLayout> template <typename ALayout, typename BLayout, typename CLayout>
float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config& s) float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s)
{ {
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part. // The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false; constexpr bool kPadM = false;
...@@ -79,9 +79,9 @@ float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config& ...@@ -79,9 +79,9 @@ float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config&
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy. // Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>; using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args); auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args); const dim3 grids = Kernel::GridSize(args.M, args.N, args.batch_count);
constexpr dim3 blocks = Kernel::BlockSize(); constexpr dim3 blocks = Kernel::BlockSize();
if(s.log_level_ > 0) if(s.log_level_ > 0)
......
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