ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-17 00:58:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

merge develop

Browse Source
This commit is contained in:
coderfeli
2025-03-18 04:49:40 +00:00
parent 7dbdff9f9f 1342ecf7fb
commit 5f49b91237
53 changed files with 1670 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
example/01_gemm/CMakeLists.txt
View File

@@ -46,9 +46,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
endif()
endforeach()
add_example_executable(example_gemm_xdl_bf16_streamk_v3 gemm_xdl_bf16_streamk_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_streamk_v3)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
@@ -80,6 +77,12 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_example_executable(example_gemm_xdl_lds_direct_load_fp16 gemm_xdl_lds_direct_load_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_lds_direct_load_fp16)
add_example_executable(example_gemm_xdl_bf16_streamk_v3 gemm_xdl_bf16_streamk_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_streamk_v3)
add_example_executable(example_gemm_xdl_fp8_streamk_v3 gemm_xdl_fp8_streamk_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_streamk_v3)
set(target 1)
endif()
endforeach()
@@ -90,9 +93,6 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
add_example_executable(example_gemm_xdl_fp8_bf8 gemm_xdl_fp8_bf8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
add_example_executable(example_gemm_xdl_fp8_streamk_v3 gemm_xdl_fp8_streamk_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_streamk_v3)
add_example_executable(example_gemm_xdl_fp16_fp8 gemm_xdl_fp16_fp8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)

11
example/09_convnd_fwd/CMakeLists.txt
View File

@@ -3,7 +3,6 @@ add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp8 convnd_fwd_xdl_fp8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
add_example_executable(example_convnd_fwd_xdl_bf8 convnd_fwd_xdl_bf8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16_comp_fp8 convnd_fwd_xdl_fp16_comp_fp8.cpp)
add_example_executable(example_convnd_fwd_xdl_fp8_bf8 convnd_fwd_xdl_fp8_bf8.cpp)
@@ -11,3 +10,13 @@ add_example_executable(example_convnd_fwd_xdl_bf8_fp8 convnd_fwd_xdl_bf8_fp8.cpp
add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
# only build fp64 example for the following targets
list(APPEND gpu_list gfx90a gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
set(target 1)
endif()
endforeach()

20
example/15_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -173,8 +173,10 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
std::size_t kargs_size = gemm.GetDeviceKernelArgSize(&argument);
std::size_t hargs_size = gemm.GetHostKernelArgSize(&argument);
DeviceMem gemm_workspace, gemm_kargs;
void* gemm_hargs;
// The following is necessary since TwoStage kernel is using additional memory both
// for Workspace and kernel arguments.
@@ -188,6 +190,11 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
gemm_workspace.Realloc(workspace_size);
gemm.SetWorkSpacePointer(&argument, gemm_workspace.GetDeviceBuffer());
}
if(hargs_size > 0)
{
hip_check_error(hipHostMalloc(&gemm_hargs, hargs_size));
gemm.SetHostKernelArgs(&argument, gemm_hargs);
}
if(!gemm.IsSupportedArgument(argument))
{
@@ -196,7 +203,16 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
"not support this GEMM problem");
}
invoker.Run(argument, StreamConfig{nullptr, false});
hipStream_t stream0 = nullptr;
hip_check_error(hipStreamCreate(&stream0));
hipEvent_t event0 = nullptr;
hip_check_error(hipEventCreate(&event0));
invoker.Run(argument, StreamConfig{nullptr, false}, stream0, event0);
hip_check_error(hipEventSynchronize(event0));
hip_check_error(hipStreamSynchronize(stream0));
bool pass = true;
if(config.do_verification)

7
example/CMakeLists.txt
View File

@@ -104,6 +104,13 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
# Do not build gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94
foreach(source IN LISTS FILE_NAME)
if(NOT EX_TARGETS MATCHES "gfx94" AND NOT EX_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_multiply_multiply_xdl_fp8_bpreshuffle")
message("Skipping ${source} example for current target")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(FILE_NAME)
if(FILE_NAME MATCHES "_xdl")

9
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View File

@@ -118,7 +118,7 @@ FMHA_FWD_API_PER_DTYPE=""" {F_if}(t.data_type.compare(\"{F_dtype}\") == 0){{
{F_hdim_case}
}}
"""
FMHA_FWD_API_PER_HDIM_CASE=""" {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim}) {{
FMHA_FWD_API_PER_HDIM_CASE=""" {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim_v}) {{
{F_inner_dispatch}
}}
"""
@@ -288,7 +288,7 @@ class FmhaFwdApiPool:
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=FWD_DTYPE_MAP[dtype])
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_hdim_v=trait.bn1, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
if not per_dtypes:
@@ -417,6 +417,7 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
'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, 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),
'192' : FmhaFwdTileSize(128, 128, 32, 128, 32, 192, 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, 32, 32, 16, -1),
}
elif dtype == 'fp8' or dtype == 'bf8':
@@ -489,6 +490,10 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
# in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
continue
if hdim == 192 and tile.F_bn1 == 128:
# NOTE: this is used to speedup deepseek prefill case, we don't gen training
if pipeline.F_bias != 'no' or pipeline.F_lse == 't' or pipeline.F_dropout == 't' or (pipeline.F_mask not in ['no', 's_no']):
continue
k = FmhaFwdKernel(F_idx=0,
F_hdim=hdim,
F_dtype=dtype,

2
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_appendkv.py
View File

@@ -181,7 +181,7 @@ class FmhaFwdAppendKVApiPool:
F_pagedkv=BOOL_MAP[trait.pagedkv], 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_rope=ROPE_MAP[trait.rope], F_bs=trait.bs, F_bsk=trait.bsk, F_bd=trait.bd, F_bdv=trait.bdv, F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
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_hdim_v=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_APPENDKV_API.format(F_dispatch = per_dtypes)

2
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View File

@@ -476,7 +476,7 @@ class FmhaFwdSplitKVApiPool:
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=FWD_DTYPE_MAP[dtype])
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_hdim_v=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
if not per_dtypes:

10
example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp
View File

@@ -72,8 +72,14 @@ float fused_moe(fused_moe_traits t, fused_moe_args a, const ck_tile::stream_conf
float r = ck_tile::launch_kernel(
s,
[=, &r0](const ck_tile::stream_config&) { r0 = fused_moesorting(t0, a0, s_sub); },
[=, &r1](const ck_tile::stream_config&) { r1 = fused_moegemm(t1, a1, s_sub); });
[=, &r0](const ck_tile::stream_config&) {
r0 = fused_moesorting(t0, a0, s_sub);
return hipPeekAtLastError() == hipSuccess;
},
[=, &r1](const ck_tile::stream_config&) {
r1 = fused_moegemm(t1, a1, s_sub);
return hipPeekAtLastError() == hipSuccess;
});
// keep unsupported case return negative
if(r0 < 0 || r1 < 0)