Skip to content

GitLab

"include/vscode:/vscode.git/clone" did not exist on "b238662a8de337b7a6cc830f8a4564c8ec901d5f"
Commit 0475a327 authored by dummycoderfe's avatar dummycoderfe
Browse files

Merge branch 'ck_tile/layernorm2d_fwd_optimize' into ck_tile/ln_add_cache_clear

parents c9b961ab 27ff3dec
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 838 additions and 240 deletions
CMakeLists.txt
View file @ 0475a327
...@@ -137,7 +137,7 @@ if(GPU_TARGETS) ...@@ -137,7 +137,7 @@ if(GPU_TARGETS)
else() else()
set(USER_GPU_TARGETS 0) set(USER_GPU_TARGETS 0)
endif() endif()
find_package(hip) find_package(hip REQUIRED)
# No assumption that HIP kernels are launched with uniform block size for backward compatibility # No assumption that HIP kernels are launched with uniform block size for backward compatibility
# SWDEV-413293 and https://reviews.llvm.org/D155213 # SWDEV-413293 and https://reviews.llvm.org/D155213
math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}") math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}")
...@@ -170,7 +170,10 @@ else() ...@@ -170,7 +170,10 @@ else()
set(CK_GPU_TARGETS ${GPU_TARGETS}) set(CK_GPU_TARGETS ${GPU_TARGETS})
endif() endif()
endif() endif()
#if the user did not set GPU_TARGETS, delete whatever was set by HIP package
if(NOT USER_GPU_TARGETS)
set(GPU_TARGETS "")
endif()
#make sure all the targets on the list are actually supported by the current compiler #make sure all the targets on the list are actually supported by the current compiler
rocm_check_target_ids(SUPPORTED_GPU_TARGETS rocm_check_target_ids(SUPPORTED_GPU_TARGETS
TARGETS ${CK_GPU_TARGETS}) TARGETS ${CK_GPU_TARGETS})
...@@ -187,6 +190,10 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1 ...@@ -187,6 +190,10 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1
add_definitions(-DCK_USE_WMMA) add_definitions(-DCK_USE_WMMA)
set(CK_USE_WMMA "ON") set(CK_USE_WMMA "ON")
endif() endif()
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"))
add_definitions(-DCK_USE_FP8_ON_UNSUPPORTED_ARCH)
endif()
# CK config file to record supported datatypes, etc. # CK config file to record supported datatypes, etc.
configure_file(include/ck/config.h.in ${CMAKE_CURRENT_BINARY_DIR}/include/ck/config.h) configure_file(include/ck/config.h.in ${CMAKE_CURRENT_BINARY_DIR}/include/ck/config.h)
...@@ -314,7 +321,6 @@ link_libraries(${OpenMP_gomp_LIBRARY}) ...@@ -314,7 +321,6 @@ link_libraries(${OpenMP_gomp_LIBRARY})
link_libraries(${OpenMP_pthread_LIBRARY}) link_libraries(${OpenMP_pthread_LIBRARY})
## HIP ## HIP
find_package(HIP REQUIRED)
# Override HIP version in config.h, if necessary. # Override HIP version in config.h, if necessary.
# The variables set by find_package() can't be overwritten, # The variables set by find_package() can't be overwritten,
# therefore let's use intermediate variables. # therefore let's use intermediate variables.
......
README.md
View file @ 0475a327
...@@ -137,12 +137,11 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa ...@@ -137,12 +137,11 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa
You can find instructions for running ckProfiler in [profiler](/profiler). You can find instructions for running ckProfiler in [profiler](/profiler).
Note the `-j` option for building with multiple threads in parallel. This speeds up the build significantly. Note the `-j` option for building with multiple threads in parallel, which speeds up the build significantly.
However, `-j` launches unlimited number of threads, which can cause the build to run out of memory and
crash. On average, you should expect each thread to use ~2Gb of RAM.
Depending on the number of CPU cores and the amount of RAM on your system, you may want to Depending on the number of CPU cores and the amount of RAM on your system, you may want to
limit the number of threads. For example, if you have a 128-core CPU and 64 Gb of RAM. limit the number of threads. For example, if you have a 128-core CPU and 128 Gb of RAM it's advisable to use `-j32`.
By default, `-j` launches one thread per CPU core, which can cause the build to run out of memory and
crash. In such cases, you can reduce the number of threads to 32 by using `-j32`.
Additional cmake flags can be used to significantly speed-up the build: Additional cmake flags can be used to significantly speed-up the build:
...@@ -154,6 +153,11 @@ Additional cmake flags can be used to significantly speed-up the build: ...@@ -154,6 +153,11 @@ Additional cmake flags can be used to significantly speed-up the build:
`batched_gemm_multi_d_dl`. These instances are useful on architectures like the NAVI2x, as most `batched_gemm_multi_d_dl`. These instances are useful on architectures like the NAVI2x, as most
other platforms have faster instances, such as `xdl` or `wmma`, available. other platforms have faster instances, such as `xdl` or `wmma`, available.
* `CK_USE_FP8_ON_UNSUPPORTED_ARCH` (default is OFF) must be set to ON in order to build instances,
such as `gemm_universal` and `gemm_multiply_multiply` for fp8 data type for GPU targets which do not
have native support for fp8 data type, such as gfx908 or gfx90a. These instances are useful on
architectures like the MI100/MI200 for the functional support only.
## Using sccache for building ## Using sccache for building
The default CK Docker images come with a pre-installed version of sccache, which supports clang The default CK Docker images come with a pre-installed version of sccache, which supports clang
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View file @ 0475a327
...@@ -21,6 +21,14 @@ DTYPE_BITS = { ...@@ -21,6 +21,14 @@ DTYPE_BITS = {
"bf8" : 8 "bf8" : 8
} }
K0_MAX_SUBMAX_MAP = {
32 : 32,
64 : 64,
96 : 128,
128: 128,
256: 256
}
TILE_PARTITIONER_MAP = { TILE_PARTITIONER_MAP = {
"shb" : "ck_tile::FmhaFwdTilePartitioner_SHB", "shb" : "ck_tile::FmhaFwdTilePartitioner_SHB",
"hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS", "hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS",
...@@ -35,7 +43,7 @@ FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT ...@@ -35,7 +43,7 @@ FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
FMHA_FWD_KERNEL_BODY=""" 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_bk0blen}>; 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_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},
...@@ -87,7 +95,7 @@ using fmha_kernel_{F_idx} = ...@@ -87,7 +95,7 @@ using fmha_kernel_{F_idx} =
fmha_pipeline_{F_idx}, fmha_pipeline_{F_idx},
fmha_epilogue_{F_idx}>; fmha_epilogue_{F_idx}>;
using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
{F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>; {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
#include <iostream> #include <iostream>
...@@ -125,7 +133,7 @@ FMHA_FWD_API_PER_HDIM_CASE=""" {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v < ...@@ -125,7 +133,7 @@ FMHA_FWD_API_PER_HDIM_CASE=""" {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <
FMHA_FWD_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.has_lse == {F_lse}) && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) && FMHA_FWD_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.has_lse == {F_lse}) && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{ ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>; using trait_ = fmha_fwd_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}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
return fmha_fwd_<trait_>(s, a); return fmha_fwd_<trait_>(s, a);
}} }}
""" """
...@@ -142,7 +150,7 @@ class FmhaFwdApiTrait: ...@@ -142,7 +150,7 @@ class FmhaFwdApiTrait:
bk0 : int # tile size along qk gemm unroll bk0 : int # tile size along qk gemm unroll
bn1 : int # tile size along v head_dim bn1 : int # tile size along v head_dim
bk1 : int # tile size along kv gemm unroll bk1 : int # tile size along kv gemm unroll
bk0blen : int bk0max : int
vlayout : str vlayout : str
mask : str mask : str
bias : str # bias : str #
...@@ -156,7 +164,7 @@ class FmhaFwdApiTrait: ...@@ -156,7 +164,7 @@ class FmhaFwdApiTrait:
@property @property
def name(self) -> str: def name(self) -> str:
return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0blen}-'+\ return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\
f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}' f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}'
@property @property
...@@ -188,8 +196,9 @@ class FmhaFwdApiTrait: ...@@ -188,8 +196,9 @@ class FmhaFwdApiTrait:
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']:
if self.dpad == 't': return f'true /*a.hdim_q % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly) bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
else : return f'a.hdim_q % {self.bk0blen} == 0' 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: assert False else: assert False
@property @property
...@@ -199,8 +208,9 @@ class FmhaFwdApiTrait: ...@@ -199,8 +208,9 @@ class FmhaFwdApiTrait:
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']:
if self.dvpad == 't': return f'true /*a.hdim_v % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly) bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
else : return f'a.hdim_v % {self.bk0blen} == 0' 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: assert False else: assert False
@dataclass @dataclass
...@@ -271,7 +281,7 @@ class FmhaFwdApiPool: ...@@ -271,7 +281,7 @@ class FmhaFwdApiPool:
F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] , F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] ,
F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen, F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype]) F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if' if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners) per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
...@@ -289,7 +299,7 @@ class FmhaFwdTileSize: ...@@ -289,7 +299,7 @@ class FmhaFwdTileSize:
F_bk0 : int # tile size along qk gemm unroll F_bk0 : int # tile size along qk gemm unroll
F_bn1 : int # tile size along v head_dim F_bn1 : int # tile size along v head_dim
F_bk1 : int # tile size along kv gemm unroll F_bk1 : int # tile size along kv gemm unroll
F_bk0blen : int # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile) F_bk0max : int # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
F_rm0 : int # number of warps for gemm0 along q seqlen F_rm0 : int # number of warps for gemm0 along q seqlen
F_rn0 : int # number of warps for gemm0 along k seqlen F_rn0 : int # number of warps for gemm0 along k seqlen
F_rk0 : int # number of warps for gemm0 along head dim q (not used) F_rk0 : int # number of warps for gemm0 along head dim q (not used)
...@@ -302,7 +312,7 @@ class FmhaFwdTileSize: ...@@ -302,7 +312,7 @@ class FmhaFwdTileSize:
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_bk0blen}" +\ 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_wm}x{self.F_wn}x{self.F_wk}" + ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
...@@ -335,7 +345,7 @@ class FmhaFwdKernel: ...@@ -335,7 +345,7 @@ class FmhaFwdKernel:
F_bk0 = self.F_tile.F_bk0, F_bk0 = self.F_tile.F_bk0,
F_bn1 = self.F_tile.F_bn1, F_bn1 = self.F_tile.F_bn1,
F_bk1 = self.F_tile.F_bk1, F_bk1 = self.F_tile.F_bk1,
F_bk0blen = self.F_tile.F_bk0blen, F_bk0max = self.F_tile.F_bk0max,
F_rm0 = self.F_tile.F_rm0, F_rm0 = self.F_tile.F_rm0,
F_rn0 = self.F_tile.F_rn0, F_rn0 = self.F_tile.F_rn0,
F_rk0 = self.F_tile.F_rk0, F_rk0 = self.F_tile.F_rk0,
...@@ -382,7 +392,7 @@ class FmhaFwdKernel: ...@@ -382,7 +392,7 @@ class FmhaFwdKernel:
bk0=self.F_tile.F_bk0, bk0=self.F_tile.F_bk0,
bn1=self.F_tile.F_bn1, bn1=self.F_tile.F_bn1,
bk1=self.F_tile.F_bk1, bk1=self.F_tile.F_bk1,
bk0blen=self.F_tile.F_bk0blen, bk0max=self.F_tile.F_bk0max,
vlayout=self.F_pipeline.F_vlayout, vlayout=self.F_pipeline.F_vlayout,
mask=self.F_pipeline.F_mask, mask=self.F_pipeline.F_mask,
bias=self.F_pipeline.F_bias, bias=self.F_pipeline.F_bias,
...@@ -401,6 +411,7 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]: ...@@ -401,6 +411,7 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
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, -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, -1),
## '96' : FmhaFwdTileSize(128, 128, 32, 128, 32, 96, 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, -1), '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 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, -1), '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 32, 32, 16, -1),
} }
...@@ -510,4 +521,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im ...@@ -510,4 +521,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im
_, kernels = get_fwd_blobs(kernel_filter, receipt, mask_impl) _, kernels = get_fwd_blobs(kernel_filter, receipt, mask_impl)
for kernel in kernels: for kernel in kernels:
f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n") f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n") f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n")
\ No newline at end of file
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View file @ 0475a327
...@@ -29,6 +29,14 @@ DTYPE_BITS = { ...@@ -29,6 +29,14 @@ DTYPE_BITS = {
"bf8" : 8 "bf8" : 8
} }
K0_MAX_SUBMAX_MAP = {
32 : 32,
64 : 64,
96 : 128,
128: 128,
256: 256
}
FMHA_FWD_SPLITKV_PIPELINE_MAP = { FMHA_FWD_SPLITKV_PIPELINE_MAP = {
"qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS", "qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS",
"qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync", "qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync",
...@@ -41,7 +49,7 @@ using fmha_mask_{F_idx} = {F_mask}; ...@@ -41,7 +49,7 @@ using fmha_mask_{F_idx} = {F_mask};
namespace {{ 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_bk0blen}>; 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_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,
...@@ -103,7 +111,7 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a) ...@@ -103,7 +111,7 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
}}; }};
}} }}
using trait_{F_idx} = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, using trait_{F_idx} = 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}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad},
{F_dvpad}>; {F_dvpad}>;
...@@ -241,7 +249,7 @@ float fmha_fwd_splitkv(fmha_fwd_splitkv_traits t, fmha_fwd_splitkv_args a, const ...@@ -241,7 +249,7 @@ 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.has_lse == {F_lse}) && (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.has_lse == {F_lse}) && (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_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {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}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>; using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
return fmha_fwd_splitkv_<traits_, traits2_>(s, a); return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
...@@ -260,7 +268,7 @@ class FmhaFwdSplitKVApiTrait: ...@@ -260,7 +268,7 @@ class FmhaFwdSplitKVApiTrait:
bk0 : int # tile size along qk gemm unroll bk0 : int # tile size along qk gemm unroll
bn1 : int # tile size along v head_dim bn1 : int # tile size along v head_dim
bk1 : int # tile size along kv gemm unroll bk1 : int # tile size along kv gemm unroll
bk0blen : int bk0max : int
vlayout : str vlayout : str
mask : str mask : str
bias : str # bias : str #
...@@ -270,11 +278,11 @@ class FmhaFwdSplitKVApiTrait: ...@@ -270,11 +278,11 @@ class FmhaFwdSplitKVApiTrait:
skpad : str skpad : str
dpad : str dpad : str
dvpad : str dvpad : str
pagedkv : str pagedkv : str
@property @property
def name(self) -> str: def name(self) -> str:
return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0blen}-'+\ return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\
f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-'+\ f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-'+\
f'{self.dvpad}-{self.pagedkv}' f'{self.dvpad}-{self.pagedkv}'
...@@ -307,8 +315,9 @@ class FmhaFwdSplitKVApiTrait: ...@@ -307,8 +315,9 @@ class FmhaFwdSplitKVApiTrait:
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']:
if self.dpad == 't': return f'true /*a.hdim_q % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly) bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
else : return f'a.hdim_q % {self.bk0blen} == 0' 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: assert False else: assert False
@property @property
...@@ -318,8 +327,9 @@ class FmhaFwdSplitKVApiTrait: ...@@ -318,8 +327,9 @@ class FmhaFwdSplitKVApiTrait:
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']:
if self.dvpad == 't': return f'true /*a.hdim_v % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly) bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
else : return f'a.hdim_v % {self.bk0blen} == 0' 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: assert False else: assert False
@dataclass @dataclass
...@@ -414,7 +424,7 @@ class FmhaFwdSplitKVApiPool: ...@@ -414,7 +424,7 @@ class FmhaFwdSplitKVApiPool:
F_lse=BOOL_MAP[trait.lse], F_squant=BOOL_MAP[trait.squant], F_pagedkv=BOOL_MAP[trait.pagedkv], F_lse=BOOL_MAP[trait.lse], F_squant=BOOL_MAP[trait.squant], F_pagedkv=BOOL_MAP[trait.pagedkv],
F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen, F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype]) F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if' if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners) per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
...@@ -458,7 +468,7 @@ class FmhaFwdSplitKVKernel: ...@@ -458,7 +468,7 @@ class FmhaFwdSplitKVKernel:
F_bk0 = self.F_tile.F_bk0, F_bk0 = self.F_tile.F_bk0,
F_bn1 = self.F_tile.F_bn1, F_bn1 = self.F_tile.F_bn1,
F_bk1 = self.F_tile.F_bk1, F_bk1 = self.F_tile.F_bk1,
F_bk0blen = self.F_tile.F_bk0blen, F_bk0max = self.F_tile.F_bk0max,
F_rm0 = self.F_tile.F_rm0, F_rm0 = self.F_tile.F_rm0,
F_rn0 = self.F_tile.F_rn0, F_rn0 = self.F_tile.F_rn0,
F_rk0 = self.F_tile.F_rk0, F_rk0 = self.F_tile.F_rk0,
...@@ -504,7 +514,7 @@ class FmhaFwdSplitKVKernel: ...@@ -504,7 +514,7 @@ class FmhaFwdSplitKVKernel:
bk0=self.F_tile.F_bk0, bk0=self.F_tile.F_bk0,
bn1=self.F_tile.F_bn1, bn1=self.F_tile.F_bn1,
bk1=self.F_tile.F_bk1, bk1=self.F_tile.F_bk1,
bk0blen=self.F_tile.F_bk0blen, bk0max=self.F_tile.F_bk0max,
vlayout=self.F_pipeline.F_vlayout, vlayout=self.F_pipeline.F_vlayout,
mask=self.F_pipeline.F_mask, mask=self.F_pipeline.F_mask,
bias=self.F_pipeline.F_bias, bias=self.F_pipeline.F_bias,
...@@ -559,6 +569,7 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]: ...@@ -559,6 +569,7 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
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, -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, -1),
## '96' : FmhaFwdTileSize(64, 128, 32, 128, 32, 96, 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, -1), '128' : FmhaFwdTileSize(64, 128, 32, 128, 32, 128, 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, -1), '256' : FmhaFwdTileSize(64, 128, 32, 256, 32, 256, 4, 1, 1, 4, 1, 1, 16, 16, 16, -1),
} }
...@@ -576,6 +587,7 @@ def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[d ...@@ -576,6 +587,7 @@ def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[d
return { return {
'32' : FmhaFwdSplitKVCombineTileSize(16, 16, -1), '32' : FmhaFwdSplitKVCombineTileSize(16, 16, -1),
'64' : FmhaFwdSplitKVCombineTileSize(32, 32, -1), '64' : FmhaFwdSplitKVCombineTileSize(32, 32, -1),
## '96' : FmhaFwdSplitKVCombineTileSize(32, 64, -1),
'128' : FmhaFwdSplitKVCombineTileSize(32, 64, -1), '128' : FmhaFwdSplitKVCombineTileSize(32, 64, -1),
'256' : FmhaFwdSplitKVCombineTileSize(32, 128, -1), '256' : FmhaFwdSplitKVCombineTileSize(32, 128, -1),
} }
...@@ -604,7 +616,7 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -604,7 +616,7 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
if dtype in ['fp16', 'bf16']: if dtype in ['fp16', 'bf16']:
for mask, bias, lse, pagedkv in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]): for mask, bias, lse, pagedkv in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]):
# TODO: use async pipeline when compiler is more stable # TODO: use async pipeline when compiler is more stable
if hdim == 256 or hdim in [32, 64, 128]: if hdim == 256 or hdim in [32, 64, 128]: ### [32, 64, 96, 128]:
# if True: # if True:
pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask)) pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask)) pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
...@@ -743,4 +755,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im ...@@ -743,4 +755,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im
_, kernels = get_fwd_splitkv_blobs(kernel_filter, receipt, mask_impl) _, kernels = get_fwd_splitkv_blobs(kernel_filter, receipt, mask_impl)
for kernel in kernels: for kernel in kernels:
f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n") f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_SPLITKV_API_FILENAME) + "\n") f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_SPLITKV_API_FILENAME) + "\n")
\ No newline at end of file
example/ck_tile/02_layernorm2d/CMakeLists.txt
View file @ 0475a327
set(LAYERNORM2D_FWD_KNOWN_APIS "fwd;bwd")
set(LAYERNORM2D_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${LAYERNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
if(LAYERNORM2D_FWD_ENABLE_APIS STREQUAL "all")
set(LAYERNORM2D_FWD_ENABLE_APIS ${LAYERNORM2D_FWD_KNOWN_APIS})
endif()
# generate a list of kernels, but not actually emit files at config sta
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
RESULT_VARIABLE ret
)
if(ret AND NOT ret EQUAL 0)
message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
endif()
file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/layernorm2d_fwd_blobs.txt LAYERNORM2D_FWD_GEN_BLOBS)
add_custom_command(
OUTPUT ${LAYERNORM2D_FWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
)
set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd") set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding example ${EXAMPLE_LAYERNORM2D_FWD}") message("adding example ${EXAMPLE_LAYERNORM2D_FWD}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${EXAMPLE_LAYERNORM2D_FWD} EXCLUDE_FROM_ALL layernorm2d_fwd.cpp) add_executable(${EXAMPLE_LAYERNORM2D_FWD} EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR}) target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${INSTANCE_SRCS}) target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${LAYERNORM2D_FWD_GEN_BLOBS})
set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS) set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS)
......
example/ck_tile/02_layernorm2d/README.md
View file @ 0475a327
# Layernorm2D forward # Layernorm2D forward
This folder contains example for Layernorm2D forward using ck_tile tile-programming implementation. This folder contains example for Layernorm2D forward using `ck_tile` tile-programming implementation.
# Implementation and feature support
## welford online algorithm
We use welfold algorithm to update `mean`/`variance` block by block. For `N <=4096` case we can compute `mean`/`var`/`normalization` within one loop, we call it `one-pass`. For large N case, it is hard to keep `mean`/`var` inside register/LDS and then computation `normalization`, so we need to load input twice, first time to compute `mean`/`var` block-by-block, then load input another time to compute the `normalization`. We call it `two-pass`.
## mean/variance save
In training case the mean/variance need to store out (TBD, not supported yet)
## prenorm/postnorm
![](misc/pnorm.png)
since [prenorm/postnorm](https://arxiv.org/pdf/1906.01787) is quite common in LLM blocks, this example boosts this feature by kernel fusion. Note that `prenorm`/`postnorm` always need to do elementwise-add a `shortcut` before the actual layernorm computation, and optionally store out the result to global. You can use `-fadd=1` to test `pre-add+store`, or `-fadd=2` to test `pre-add` without store out (not codegen by default).
## smooth-quant/dynamic-quant
we support smooth/dynamic quantization for `int8` output, by setting `-fquant=1` and `-prec_o=int8`. In this case the output will doing a rowwise dynamic quantization like below. Note that smooth-quant require input a `(1*N)` size per-channel scale(in fp32 in our example, though this is customizable), then elememt-wise multiply the tensor for each row, then compute the rowwise dynamic quant. if set `-fquant=2` will have the input per-channel scale stage, only the dynamic quant. This case is supported in our kernel but by default not generated (TBD: add some filter in generate.py support on-demand codegen)
![](misc/dquant.png)
```
# assume output int8, hidden_states is [m, n] shape and in fp16/bf16
# [m, 1]
per_token_amax, _ = torch.max(
input=torch.abs(hidden_states),
dim=-1,
keepdim=True
)
per_token_scale = per_token_amax.to(dtype=torch.float32) / 127.0
# quant hidden_states
hidden_states = (hidden_states / per_token_scale).to(dtype=torch.int8)
return hidden_states, per_token_scale
# hidden_states now is int8 will feed to next layer as intput
# per_token_scale will be used as dequant factor later layer
```
## build ## build
``` ```
...@@ -15,8 +51,35 @@ This will result in an executable `build/bin/tile_example_layernorm2d_fwd` ...@@ -15,8 +51,35 @@ This will result in an executable `build/bin/tile_example_layernorm2d_fwd`
``` ```
args: args:
-m m dimension (default:3328) -m m dimension (default:3328)
-n m dimension (default:4096) -n n dimension (default:4096)
-stride stride per row, if -1 then equal to n (default:-1)
-e epsilon (default:1e-5) -e epsilon (default:1e-5)
-save_mv save mean/variance(invstd) or not. set to 1 in training case (default:0)
-v cpu validation or not (default:1) -v cpu validation or not (default:1)
-prec precision (default:fp16) -kname print kernel name or not (default:1)
-prec_i input precision (default:fp16)
-prec_o output precision, set auto will be the same as input (default:auto)
-prec_sx output quant scale type, set auto will be the same as input. used when fquant=1 (default:auto)
-prec_sy output quant scale type, set auto will be the same as input. used when fquant=1 or 2 (default:auto)
-fadd fused-add, 0:no fused add, 1:preadd+store, 2:preadd only (default:0)
-fquant fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant (default:0)
-warmup cold iter (default:5)
-repeat hot iter (default:20)
```
## limitations
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
```
# some case
# standard fp16 layernorm 2d, m=10. n=1024
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024
# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant, output in int8
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024 -prec_o=int8 -fquant=1
# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant+fused-add-store, output in int8
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024 -prec_o=int8 -fquant=1 -fadd=1
``` ```
example/ck_tile/02_layernorm2d/generate.py 0 → 100644
View file @ 0475a327
This diff is collapsed.
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n1024_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
#if 0
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n1536_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n2048_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n3072_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n4096_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n4096_tp_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n512_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n1024_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
#if 0
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n1536_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n2048_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n3072_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 128, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n4096_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n4096_tp_instance.cpp deleted 100644 → 0
View file @ c9b961ab
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd mv 2p
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on
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