Merge remote-tracking branch 'rccl/develop' into 2.19.4

[ROCm/rccl commit: 1a134b283b]
This commit is contained in:
Wenkai Du
2024-01-31 11:08:54 -06:00
14 zmienionych plików z 31 dodań i 30 usunięć
+6 -9
Wyświetl plik
@@ -154,7 +154,7 @@ function(gen_device_table)
foreach(func IN LISTS FUNC_LIST)
string(FIND "${func}" "LL128" IS_LL128)
if(NOT IS_LL128 EQUAL -1)
file(APPEND ${DEVICE_TABLE_H_FILE} "#if defined(__gfx90a__)\n")
file(APPEND ${DEVICE_TABLE_H_FILE} "#if defined(__gfx90a__) && defined(ENABLE_LL128)\n")
endif()
if(ENABLE_IFC)
file(APPEND ${DEVICE_TABLE_H_FILE} "__device__ void ${func}();\n")
@@ -173,7 +173,7 @@ function(gen_device_table)
foreach(func ${FUNC_LIST})
string(FIND "${func}" "LL128" IS_LL128)
if(NOT IS_LL128 EQUAL -1)
file(APPEND ${DEVICE_TABLE_H_FILE} "#if defined(__gfx90a__)\n")
file(APPEND ${DEVICE_TABLE_H_FILE} "#if defined(__gfx90a__) && defined(ENABLE_LL128)\n")
file(APPEND ${DEVICE_TABLE_H_FILE} " ${func},\n")
file(APPEND ${DEVICE_TABLE_H_FILE} "#else\n")
string(REPLACE "LL128" "LL" func "${func}")
@@ -284,11 +284,9 @@ endfunction()
## Function to generate MSCCL Kernels
###########################################################################################################
function(gen_msccl_kernels)
foreach(REDOP_CURRENT IN LISTS ALL_REDOPS)
set(MSCCL_REDOP Sum Prod MinMax)
foreach(REDOP_CURRENT IN LISTS MSCCL_REDOP)
foreach(DATA_TYPE ${ALL_TYPES})
if (REDOP_CURRENT STREQUAL "SumPostDiv" AND DATA_TYPE IN_LIST FLOATS_LIST)
continue() # Skip the iteration for FLOATS_LIST when REDOP_CURRENT is SumPostDiv
endif()
set(FILE_NAME "${HIPIFY_DIR}/src/device/msccl_kernel_${REDOP_CURRENT}_${DATA_TYPE}.cpp")
message(STATUS "Generating ${FILE_NAME}")
file(WRITE ${FILE_NAME}
@@ -296,8 +294,7 @@ function(gen_msccl_kernels)
#include \"primitives.h\"
#include \"nccl_common.h\"
#include \"device.h\"
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(${REDOP_CURRENT}, ${DATA_TYPE}, false);
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(${REDOP_CURRENT}, ${DATA_TYPE}, true);")
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(${REDOP_CURRENT}, ${DATA_TYPE}, false);")
list(APPEND HIP_SOURCES ${FILE_NAME})
endforeach()
endforeach()
@@ -365,7 +362,7 @@ function(gen_collectives)
file(WRITE ${FILE_PATH} "#include \"${COLL_LOWER}.h\"\n#include \"common.h\"\n#include \"collectives.h\"\n")
string(FIND "${list_name}" "LL128" IS_LL128)
if(NOT IS_LL128 EQUAL -1)
file(APPEND ${FILE_PATH} "#if defined(__gfx90a__)\n")
file(APPEND ${FILE_PATH} "#if defined(__gfx90a__) && defined(ENABLE_LL128)\n")
endif()
foreach(IMPL IN LISTS IMPL_LIST)
file(APPEND ${FILE_PATH} "${IMPL}")