Extend XDL kernel to Support RDNA3/4 - Part 5 (#2725)

* Enable xdl in gfx11 & gfx12

* update cmake file

* fix all instance build (cmake)

* fix batched_gemm_gemm(cmake)

* rebase cmake files

* fix cmake build error

* remve CK_ENABLE_DYNAMIC_WARP_SIZE

* update cmake build error2

* fix gfx11 build

CK_USE_XDL is enabled on gfx11 and gfx12

* fix gfx10 build

* fix gfx11 error

---------

Co-authored-by: Lin, Qun <Quentin.Lin+amdeng@amd.com>

[ROCm/composable_kernel commit: f22740df82]
This commit is contained in:
linqunAMD
2025-09-16 01:59:25 +08:00
committed by GitHub
parent 8cbf571d53
commit a9e6cb0ec0
33 changed files with 243 additions and 397 deletions

View File

@@ -54,7 +54,7 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
# Do not build XDL instances if gfx9 targets are not on the target list
if(NOT INST_TARGETS MATCHES "gfx9" AND source_name MATCHES "_xdl")
if(NOT INST_TARGETS MATCHES "gfx9" AND NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12" AND source_name MATCHES "_xdl")
message(DEBUG "removing xdl instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
@@ -73,13 +73,13 @@ function(add_instance_library INSTANCE_NAME)
message(DEBUG "removing mha instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
# Do not build XDL gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94
# Do not build XDL gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94, gfx95 and gfx12
if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
if(NOT INST_TARGETS MATCHES "gfx94|gfx95|gfx12" AND source_name MATCHES "gemm_multiply_multiply" AND source_name MATCHES "_f8_")
message(DEBUG "removing gemm_multiply_multiply_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source_name MATCHES "gemm_xdl_universal" AND source_name MATCHES "_f8_")
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND NOT INST_TARGETS MATCHES "gfx12" AND source_name MATCHES "gemm_xdl_universal" AND source_name MATCHES "_f8_")
message(DEBUG "removing gemm_universal_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
@@ -89,8 +89,8 @@ function(add_instance_library INSTANCE_NAME)
message(DEBUG "removing gemm_universal_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
# Do not build gemm_universal_preshuffle_f8 for any targets except gfx94
if(NOT (INST_TARGETS MATCHES "gfx942" OR INST_TARGETS MATCHES "gfx950") AND (source_name MATCHES "gemm_universal_preshuffle" OR source_name MATCHES "gemm_xdl_universal_preshuffle") AND (source_name MATCHES "_f8_f8_f16" OR source_name MATCHES "_f8_f8_bf16"))
# Do not build gemm_universal_preshuffle_f8 for any targets except gfx94, gfx95 and gfx12
if(NOT (INST_TARGETS MATCHES "gfx942" OR INST_TARGETS MATCHES "gfx950" OR INST_TARGETS MATCHES "gfx12") AND (source_name MATCHES "gemm_universal_preshuffle" OR source_name MATCHES "gemm_xdl_universal_preshuffle") AND (source_name MATCHES "_f8_f8_f16" OR source_name MATCHES "_f8_f8_bf16"))
message(DEBUG "removing gemm_universal_preshuffle_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
@@ -106,7 +106,7 @@ function(add_instance_library INSTANCE_NAME)
set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
if(source_name MATCHES "_xdl")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx10-3-generic)
elseif(source_name MATCHES "_wmma")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx950)
elseif(source_name MATCHES "mha")
@@ -120,29 +120,29 @@ function(add_instance_library INSTANCE_NAME)
#only build the fp8 gemm instances for gfx90a if the build argument is set, otherwise only build for gfx942/gfx950 and gfx1200/gfx1201
if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
if(source_name MATCHES "gemm_xdl_universal" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
if(source_name MATCHES "gemm_multiply_multiply" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
if(source_name MATCHES "gemm_universal_preshuffle" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
if(source_name MATCHES "gemm_xdl_universal_preshuffle" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
else()
if(source_name MATCHES "gemm_xdl_universal" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
if(source_name MATCHES "gemm_multiply_multiply" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
if(source_name MATCHES "gemm_universal_preshuffle" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
if(source_name MATCHES "gemm_xdl_universal_preshuffle" AND source_name MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx10-3-generic gfx11-generic)
endif()
endif()
if(source_name MATCHES "gemm_wmma_universal" AND source_name MATCHES "f8")
@@ -266,7 +266,7 @@ FOREACH(subdir_path ${dir_list})
message(DEBUG "Found only dl instances, but DL_KERNELS is not set. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY XDL_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9"))
if(("${cmake_instance}" MATCHES "ONLY XDL_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12"))
message(DEBUG "Found only xdl instances, but gfx9 is not on the targets list. Skipping.")
set(add_inst 0)
endif()
@@ -278,33 +278,36 @@ FOREACH(subdir_path ${dir_list})
message(DEBUG "Found only wmma instances, but gfx11 is not on the targets list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_DL_KERNELS") AND (NOT DEFINED DL_KERNELS) AND (NOT INST_TARGETS MATCHES "gfx9"))
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_DL_KERNELS") AND (NOT DEFINED DL_KERNELS) AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12"))
message(DEBUG "Found only xdl and dl instances, but gfx9 is not on the targets listand DL_KERNELS is not set. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9"))
if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12"))
message(DEBUG "Found only xdl and wmma instances, but gfx11 and gfx9 are not on the targets list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "XDL_DL_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9") AND (NOT DEFINED DL_KERNELS))
if(("${cmake_instance}" MATCHES "XDL_DL_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9|gfx11|gfx12") AND (NOT DEFINED DL_KERNELS))
message(DEBUG "Found xdl, dl, and wmma instances, but none of those meet the target list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "gemm_multiply_multiply" AND "${cmake_instance}" MATCHES "_f8_" ) AND (NOT INST_TARGETS MATCHES "gfx94|gfx95|gfx11|gfx12") AND (NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH))
message(DEBUG "Found gemm_multiply_multiply_f8 instances, but gfx94/gfx95 not on the target list. Skipping.")
if("${cmake_instance}" MATCHES "gemm_multiply_multiply_wp" AND (NOT INST_TARGETS MATCHES "gfx94|gfx95|gfx12"))
message(DEBUG "Found gemm_multiply_multiply_wp instances, but gfx94/gfx95/gfx12 not on the target list. Skipping. ${cmake_instance}")
set(add_inst 0)
elseif("${cmake_instance}" MATCHES "gemm_multiply_multiply" AND (NOT INST_TARGETS MATCHES "gfx94|gfx95|gfx11|gfx12"))
message(DEBUG "Found gemm_multiply_multiply instances, but gfx94/gfx95/gfx11/gfx12 not on the target list. Skipping. ${cmake_instance}")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "gemm_universal_preshuffle" AND "${cmake_instance}" MATCHES "_f8_" ) AND (NOT INST_TARGETS MATCHES "gfx94") AND (NOT INST_TARGETS MATCHES "gfx95") AND (NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH))
if(("${cmake_instance}" MATCHES "gemm_universal_preshuffle" AND "${cmake_instance}" MATCHES "_f8_" ) AND (NOT INST_TARGETS MATCHES "gfx94|gfx95|gfx12") AND (NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH))
message(DEBUG "Found gemm_universal_preshuffle_f8 instances, but gfx94/gfx95 not on the target list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "gemm_xdl_universal_preshuffle" AND "${cmake_instance}" MATCHES "_f8_" ) AND (NOT INST_TARGETS MATCHES "gfx94") AND (NOT INST_TARGETS MATCHES "gfx95") AND (NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH))
if(("${cmake_instance}" MATCHES "gemm_xdl_universal_preshuffle" AND "${cmake_instance}" MATCHES "_f8_" ) AND (NOT INST_TARGETS MATCHES "gfx94|gfx95|gfx12") AND (NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH))
message(DEBUG "Found gemm_xdl_universal_preshuffle_f8_f8_bf16 instances, but gfx94/gfx95 not on the target list. Skipping.")
set(add_inst 0)
endif()
if ("${cmake_instance}" MATCHES "gemm_bilinear")
set(add_inst 0)
if((SUPPORTED_GPU_TARGETS MATCHES "gfx9") AND (DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES))
if((SUPPORTED_GPU_TARGETS MATCHES "gfx9|gfx11|gfx12") AND (DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES))
set(add_inst 1)
endif()
if((SUPPORTED_GPU_TARGETS MATCHES "gfx1[12]") AND (DTYPES MATCHES "int8" OR NOT DEFINED DTYPES))