refactor: Improve builtins compilation

Related-To: NEO-7621
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
This commit is contained in:
Kamil Kopryk 2024-01-14 23:28:43 +00:00 committed by Compute-Runtime-Automation
parent 9b568c6ba9
commit cc1e3965ce
1 changed files with 50 additions and 57 deletions

View File

@ -16,17 +16,12 @@ set(GENERATED_BUILTINS_AUX_TRANSLATION ${GENERATED_BUILTINS_AUX_TRANSLATION} PAR
set(GENERATED_BUILTINS_IMAGES ${GENERATED_BUILTINS_IMAGES} PARENT_SCOPE)
set(GENERATED_BUILTINS_STATELESS ${GENERATED_BUILTINS_STATELESS} PARENT_SCOPE)
set(BUILTIN_OPTIONS_STATELESS
"-cl-intel-greater-than-4GB-buffer-required"
)
set(BUILTIN_OPTIONS_STATELESS "-cl-intel-greater-than-4GB-buffer-required")
set(BUILTIN_OPTIONS_STATEFUL "-force_stos_opt")
set(bindless_OPTIONS
-internal_options "-cl-intel-use-bindless-mode -cl-intel-use-bindless-advanced-mode"
)
set(bindful_OPTIONS
""
)
set(stateless_INTERNAL_OPTIONS "")
set(bindful_INTERNAL_OPTIONS "")
set(bindless_INTERNAL_OPTIONS -internal_options "-cl-intel-use-bindless-mode -cl-intel-use-bindless-advanced-mode")
if("${CMAKE_BUILD_TYPE}" STREQUAL "Debug")
list(APPEND __ocloc__options__ "-D DEBUG")
@ -41,13 +36,27 @@ function(get_bits_for_stateless core_type platform_it_lower)
endif()
endfunction()
function(get_builtin_options core_type neo_arch)
function(get_builtin_options core_type neo_arch mode_type)
if("${neo_arch}" STREQUAL "x32")
set(BUILTIN_OPTIONS "-cl-intel-greater-than-4GB-buffer-required" PARENT_SCOPE)
elseif("${core_type}" STREQUAL "XE_HPC_CORE")
set(BUILTIN_OPTIONS "" PARENT_SCOPE)
set(BUILTIN_OPTIONS "${BUILTIN_OPTIONS_STATELESS}" PARENT_SCOPE)
elseif("${mode_type}" STREQUAL "stateless")
set(BUILTIN_OPTIONS "${BUILTIN_OPTIONS_STATELESS}" PARENT_SCOPE)
else()
set(BUILTIN_OPTIONS "-force_stos_opt" PARENT_SCOPE)
set(BUILTIN_OPTIONS "${BUILTIN_OPTIONS_STATEFUL}" PARENT_SCOPE)
endif()
endfunction()
function(get_generated_builtins core_type mode_type)
if("${core_type}" STREQUAL "XE_HPC_CORE")
# Compile all builtin kernels as stateless for XeHpc
set(generated_builtins "${GENERATED_BUILTINS};${GENERATED_BUILTINS_STATELESS}" PARENT_SCOPE)
set(generated_builtins_images "${GENERATED_BUILTINS_IMAGES};${GENERATED_BUILTINS_IMAGES_STATELESS}" PARENT_SCOPE)
elseif(${mode_type} STREQUAL "stateless")
set(generated_builtins "${GENERATED_BUILTINS_STATELESS}" PARENT_SCOPE)
set(generated_builtins_images "${GENERATED_BUILTINS_IMAGES_STATELESS}" PARENT_SCOPE)
else()
set(generated_builtins "${GENERATED_BUILTINS}" PARENT_SCOPE)
set(generated_builtins_images "${GENERATED_BUILTINS_IMAGES}" PARENT_SCOPE)
endif()
endfunction()
@ -66,7 +75,7 @@ function(compile_builtin core_type platform_it builtin bits builtin_options mode
${OUTPUTDIR}/${mode}_${BASENAME}_${platform_it_lower}.spv
)
list(APPEND __ocloc__options__ "-cl-kernel-arg-info")
set(INTERNAL_OPTIONS "${${mode}_OPTIONS}")
set(INTERNAL_OPTIONS "${${mode}_INTERNAL_OPTIONS}")
add_custom_command(
OUTPUT ${OUTPUT_FILE_SPV}
COMMAND ${ocloc_cmd_prefix} -q -file ${absolute_filepath} -spv_only -device ${platform_it_lower} ${builtin_options} -${bits} -output ${mode}_${BASENAME} -out_dir ${OUTPUTDIR} ${INTERNAL_OPTIONS} -options "$<JOIN:${__ocloc__options__}, >"
@ -175,11 +184,11 @@ macro(macro_for_each_platform)
CORE_CONTAINS_PLATFORM("SUPPORTED_IMAGES" ${CORE_TYPE} ${PLATFORM_IT} IMAGE_SUPPORT)
CORE_CONTAINS_PLATFORM("SUPPORTED_AUX_TRANSLATION" ${CORE_TYPE} ${PLATFORM_IT} AUX_TRANSLATION_SUPPORT)
get_bits_for_stateless(${CORE_TYPE} ${PLATFORM_IT})
get_builtin_options(${CORE_TYPE} ${NEO_ARCH})
set(target_name builtins_${PLATFORM_IT_LOWER})
add_custom_target(${target_name})
add_dependencies(builtins ${target_name})
set_target_properties(${target_name} PROPERTIES FOLDER "${SHARED_SOURCE_PROJECTS_FOLDER}/${SHARED_BUILTINS_PROJECTS_FOLDER}/${PLATFORM_IT_LOWER}")
foreach(MODE ${ADDRESSING_MODES})
unset(BUILTINS_COMMANDS)
@ -187,48 +196,31 @@ macro(macro_for_each_platform)
continue()
endif()
if(${AUX_TRANSLATION_SUPPORT})
foreach(GENERATED_BUILTIN_AUX_TRANSLATION ${GENERATED_BUILTINS_AUX_TRANSLATION})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTIN_AUX_TRANSLATION}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
# XeHpc supports only stateless addressing mode
if(NOT ${MODE} STREQUAL "stateless" AND "${CORE_TYPE}" STREQUAL "XE_HPC_CORE")
continue()
endif()
if(${MODE} STREQUAL "stateless")
foreach(GENERATED_BUILTIN_STATELESS ${GENERATED_BUILTINS_STATELESS})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTIN_STATELESS}.builtin_kernel ${BITS} "${BUILTIN_OPTIONS_STATELESS}" ${MODE})
get_builtin_options(${CORE_TYPE} ${NEO_ARCH} ${MODE})
get_generated_builtins(${CORE_TYPE} ${MODE})
if(${AUX_TRANSLATION_SUPPORT})
# Stateless aux translation are only supported for XeHpc
if(NOT ${MODE} STREQUAL "stateless" OR "${CORE_TYPE}" STREQUAL "XE_HPC_CORE")
foreach(GENERATED_BUILTIN_AUX_TRANSLATION ${GENERATED_BUILTINS_AUX_TRANSLATION})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTIN_AUX_TRANSLATION}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
endif()
endif()
foreach(generated_builtin ${generated_builtins})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${generated_builtin}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
if(${IMAGE_SUPPORT})
foreach(generated_builtin_images ${generated_builtins_images})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${generated_builtin_images}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
if(${IMAGE_SUPPORT})
foreach(GENERATED_BUILTIN_IMAGES_STATELESS ${GENERATED_BUILTINS_IMAGES_STATELESS})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTIN_IMAGES_STATELESS}.builtin_kernel ${BITS} "${BUILTIN_OPTIONS_STATELESS}" ${MODE})
endforeach()
endif()
if("${CORE_TYPE}" STREQUAL "XE_HPC_CORE")
# Compile stateful versions as stateless
foreach(GENERATED_BUILTIN ${GENERATED_BUILTINS})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTIN}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
if(${IMAGE_SUPPORT})
foreach(GENERATED_BUILTINS_IMAGES ${GENERATED_BUILTINS_IMAGES})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTINS_IMAGES}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
endif()
endif()
else()
if("${CORE_TYPE}" STREQUAL "XE_HPC_CORE")
continue()
endif()
foreach(GENERATED_BUILTIN ${GENERATED_BUILTINS})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTIN}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
if(${IMAGE_SUPPORT})
foreach(GENERATED_BUILTINS_IMAGES ${GENERATED_BUILTINS_IMAGES})
compile_builtin(${CORE_TYPE} ${PLATFORM_IT} ${GENERATED_BUILTINS_IMAGES}.builtin_kernel ${NEO_BITS} "${BUILTIN_OPTIONS}" ${MODE})
endforeach()
endif()
endif()
get_property(GENERATED_BUILTINS_CPPS_${MODE} GLOBAL PROPERTY GENERATED_BUILTINS_CPPS_${MODE})
@ -240,6 +232,7 @@ macro(macro_for_each_platform)
add_dependencies(${target_name} ${target_name}_${MODE})
set_target_properties(${target_name}_${MODE} PROPERTIES FOLDER "${SHARED_SOURCE_PROJECTS_FOLDER}/${SHARED_BUILTINS_PROJECTS_FOLDER}/${PLATFORM_IT_LOWER}")
endforeach()
endmacro()
macro(macro_for_each_core_type)
@ -272,4 +265,4 @@ if(NOT "${OUTPUT_LIST_CPP_FILES}" STREQUAL "")
)
endif()
apply_macro_for_each_core_type("SUPPORTED")
apply_macro_for_each_core_type("SUPPORTED")