WIP: AVX10.1

Author: stefanatwork

CMakeLists.txt

@@ -380,18 +380,21 @@ IF (EMBREE_MAX_ISA STREQUAL "NONE")
       OPTION(EMBREE_ISA_NEON2X "Enables NEON ISA double pumped." OFF)
     ENDIF()
   ELSE()
-    TRY_COMPILE(COMPILER_SUPPORTS_AVX    "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX})
-    TRY_COMPILE(COMPILER_SUPPORTS_AVX2   "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX2})
-    TRY_COMPILE(COMPILER_SUPPORTS_AVX512 "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX512})
+    TRY_COMPILE(COMPILER_SUPPORTS_AVX     "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX})
+    TRY_COMPILE(COMPILER_SUPPORTS_AVX2    "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX2})
+    TRY_COMPILE(COMPILER_SUPPORTS_AVX512  "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX512})
+    TRY_COMPILE(COMPILER_SUPPORTS_AVX10_1 "${CMAKE_BINARY_DIR}" "${PROJECT_SOURCE_DIR}/common/cmake/check_isa.cpp" COMPILE_DEFINITIONS ${FLAGS_AVX10_1})
 
     OPTION(EMBREE_ISA_SSE2 "Enables SSE2 ISA." ON)
     OPTION(EMBREE_ISA_SSE42 "Enables SSE4.2 ISA." ON)
     OPTION(EMBREE_ISA_AVX "Enables AVX ISA." ${COMPILER_SUPPORTS_AVX})
     OPTION(EMBREE_ISA_AVX2 "Enables AVX2 ISA." ${COMPILER_SUPPORTS_AVX2})
     IF (APPLE)
       OPTION(EMBREE_ISA_AVX512 "Enables AVX512 ISA." OFF)
+      OPTION(EMBREE_ISA_AVX10_1 "Enables AVX10.1 ISA." OFF)
     ELSE()
       OPTION(EMBREE_ISA_AVX512 "Enables AVX512 ISA." ${COMPILER_SUPPORTS_AVX512})
+      OPTION(EMBREE_ISA_AVX10_1 "Enables AVX10.1 ISA." ${COMPILER_SUPPORTS_AVX10_1})
     ENDIF()
     # Don't use OPTION, but still set them to OFF, so that embree-config.cmake is consisten with its definitions
     SET(EMBREE_ISA_NEON   OFF)
@@ -406,13 +409,15 @@ ELSEIF (EMBREE_MAX_ISA STREQUAL "DEFAULT")
   UNSET(EMBREE_ISA_AVX CACHE)
   UNSET(EMBREE_ISA_AVX2 CACHE)
   UNSET(EMBREE_ISA_AVX512 CACHE)
+  UNSET(EMBREE_ISA_AVX10_1 CACHE)
   SET(EMBREE_ISA_NEON OFF)
   SET(EMBREE_ISA_NEON2X OFF)
   SET(EMBREE_ISA_SSE2 OFF)
   SET(EMBREE_ISA_SSE42 OFF)
   SET(EMBREE_ISA_AVX OFF)
   SET(EMBREE_ISA_AVX2 OFF)
   SET(EMBREE_ISA_AVX512 OFF)
+  SET(EMBREE_ISA_AVX10_1 OFF)
   MESSAGE(STATUS "Detecting default ISA...")
   INCLUDE(check_isa_default)
   CHECK_ISA_DEFAULT(EMBREE_ISA_DEFAULT)
@@ -427,6 +432,7 @@ ELSE()
   UNSET(EMBREE_ISA_AVX CACHE)
   UNSET(EMBREE_ISA_AVX2 CACHE)
   UNSET(EMBREE_ISA_AVX512 CACHE)
+  UNSET(EMBREE_ISA_AVX10_1 CACHE)
 
   IF(EMBREE_MAX_ISA STREQUAL "NEON")
     SET(ISA  1)
@@ -444,6 +450,8 @@ ELSE()
     SET(ISA 5)
   ELSEIF(EMBREE_MAX_ISA STREQUAL "AVX512SKX")  # just for compatibility
     SET(ISA 5)
+  ELSEIF(EMBREE_MAX_ISA STREQUAL "AVX10.1")
+    SET(ISA 6)
   ELSE()
     MESSAGE(FATAL_ERROR "Unsupported ISA specified: " ${EMBREE_MAX_ISA})
   ENDIF()
@@ -455,6 +463,7 @@ ELSE()
   SET(EMBREE_ISA_AVX OFF)
   SET(EMBREE_ISA_AVX2 OFF)
   SET(EMBREE_ISA_AVX512 OFF)
+  SET(EMBREE_ISA_AVX10_1 OFF)
 
   IF (EMBREE_ARM)
     IF (ISA GREATER 0)
@@ -479,6 +488,9 @@ ELSE()
     IF (ISA GREATER 4)
       SET(EMBREE_ISA_AVX512  ON)
     ENDIF ()
+    IF (ISA GREATER 5)
+      SET(EMBREE_ISA_AVX10_1  ON)
+    ENDIF ()
   ENDIF()
 ENDIF()
 
@@ -511,6 +523,9 @@ IF (APPLE AND EMBREE_STATIC_LIB)
   IF (EMBREE_ISA_AVX512)
     MATH(EXPR NUMISA "${NUMISA}+1")
   ENDIF()
+  IF (EMBREE_ISA_AVX10_1)
+    MATH(EXPR NUMISA "${NUMISA}+1")
+  ENDIF()
 
   IF (NUMISA GREATER 1)
     IF (${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
@@ -530,6 +545,7 @@ SET(SSE42 1)
 SET(AVX 2)
 SET(AVX2 3)
 SET(AVX512 4)
+SET(AVX10_1 5)
 
 UNSET(FLAGS_LOWEST)
 SET(ISA_LOWEST -1)
@@ -612,6 +628,19 @@ IF (EMBREE_ISA_AVX512)
   ENDIF()
 ENDIF ()
 
+IF (EMBREE_ISA_AVX10_1)
+  ADD_DEFINITIONS(-DEMBREE_TARGET_AVX10_1)
+  IF (NOT EMBREE_ARM)
+    # AVX 10.1 support in ISPC?
+    # LIST(APPEND ISPC_TARGETS "")
+  ENDIF()
+  IF(NOT FLAGS_LOWEST)
+    SET(ISA_LOWEST ${AVX10_1})
+    SET(ISA_LOWEST_AVX ${AVX10_1})
+    SET(FLAGS_LOWEST ${FLAGS_AVX10_1})
+  ENDIF()
+ENDIF ()
+
 IF (ISA_LOWEST EQUAL -1)
   MESSAGE(FATAL_ERROR "You have to enable at least one ISA!")
 ENDIF()

common/cmake/msvc.cmake

@@ -6,6 +6,7 @@ SET(FLAGS_SSE42 "${FLAGS_SSE2} /D__SSE3__ /D__SSSE3__ /D__SSE4_1__ /D__SSE4_2__"
 SET(FLAGS_AVX   "${FLAGS_SSE42} /arch:AVX")
 SET(FLAGS_AVX2  "${FLAGS_SSE42} /arch:AVX2")
 SET(FLAGS_AVX512  "${FLAGS_AVX2} /arch:AVX512")
+SET(FLAGS_AVX10_1  "${FLAGS_AVX2} /arch:AVX10.1")
 
 SET(COMMON_CXX_FLAGS "")
 SET(COMMON_CXX_FLAGS "${COMMON_CXX_FLAGS} /EHsc")        # catch C++ exceptions only and extern "C" functions never throw a C++ exception

kernels/CMakeLists.txt

@@ -132,7 +132,7 @@ MACRO(embree_files TARGET ISA)
     LIST(APPEND ${TARGET} geometry/primitive8.cpp)
   ENDIF() 
 
-  IF (${ISA} EQUAL ${SSE2} OR ${ISA} EQUAL ${AVX} OR ${ISA} EQUAL ${AVX2} OR ${ISA} EQUAL ${AVX512} OR ${ISA_LOWEST} EQUAL ${ISA})
+  IF (${ISA} EQUAL ${SSE2} OR ${ISA} EQUAL ${AVX} OR ${ISA} EQUAL ${AVX2} OR ${ISA} EQUAL ${AVX512} OR ${ISA} EQUAL ${AVX10_1} OR ${ISA_LOWEST} EQUAL ${ISA})
 
     LIST(APPEND ${TARGET}
       common/scene_user_geometry.cpp
@@ -209,6 +209,7 @@ embree_files(EMBREE_LIBRARY_FILES_SSE42 ${SSE42})
 embree_files(EMBREE_LIBRARY_FILES_AVX ${AVX})
 embree_files(EMBREE_LIBRARY_FILES_AVX2 ${AVX2})
 embree_files(EMBREE_LIBRARY_FILES_AVX512 ${AVX512})
+embree_files(EMBREE_LIBRARY_FILES_AVX10_1 ${AVX10_1})
 
 #message("SSE2: ${EMBREE_LIBRARY_FILES_SSE2}")
 #message("SSE42: ${EMBREE_LIBRARY_FILES_SSE42}")
@@ -237,6 +238,7 @@ CreateISADummyFiles(EMBREE_LIBRARY_FILES_SSE42     sse42     ${EMBREE_LIBRARY_FI
 CreateISADummyFiles(EMBREE_LIBRARY_FILES_AVX       avx       ${EMBREE_LIBRARY_FILES_AVX})
 CreateISADummyFiles(EMBREE_LIBRARY_FILES_AVX2      avx2      ${EMBREE_LIBRARY_FILES_AVX2})
 CreateISADummyFiles(EMBREE_LIBRARY_FILES_AVX512    avx512    ${EMBREE_LIBRARY_FILES_AVX512})
+CreateISADummyFiles(EMBREE_LIBRARY_FILES_AVX10_1   avx10_1   ${EMBREE_LIBRARY_FILES_AVX10_1})
 
 MACRO (CheckGlobals library)
   IF (NOT WIN32 AND NOT APPLE)
@@ -346,6 +348,20 @@ IF (EMBREE_ISA_AVX512 AND EMBREE_LIBRARY_FILES_AVX512)
   ENDIF()
 ENDIF()
 
+IF (EMBREE_ISA_AVX10_1 AND EMBREE_LIBRARY_FILES_AVX10_1)
+  DISABLE_STACK_PROTECTOR_FOR_INTERSECTORS(${EMBREE_LIBRARY_FILES_AVX10_1})
+  ADD_LIBRARY(embree_avx10_1 STATIC ${EMBREE_LIBRARY_FILES_AVX10_1})
+  TARGET_LINK_LIBRARIES(embree_avx10_1 PRIVATE tasking)
+  SET_TARGET_PROPERTIES(embree_avx10_1 PROPERTIES COMPILE_FLAGS "${FLAGS_AVX10_1}")
+  SET_PROPERTY(TARGET embree_avx10_1 PROPERTY FOLDER kernels)
+  SET(EMBREE_LIBRARIES ${EMBREE_LIBRARIES} embree_avx10_1)
+  CheckGlobals(embree_avx10_1)
+  IF (EMBREE_STATIC_LIB)
+    INSTALL(TARGETS embree_avx10_1 EXPORT embree_avx10_1-targets ARCHIVE DESTINATION "${CMAKE_INSTALL_LIBDIR}" COMPONENT devel)
+    INSTALL(EXPORT embree_avx10_1-targets DESTINATION "${EMBREE_CMAKEEXPORT_DIR}" COMPONENT devel)
+  ENDIF()
+ENDIF()
+
 TARGET_LINK_LIBRARIES(embree PRIVATE ${EMBREE_LIBRARIES} sys math simd lexers tasking)
 IF (EMBREE_SYCL_SUPPORT)
   TARGET_LINK_LIBRARIES(embree PRIVATE ${SYCL_LIB_NAME} ze_wrapper PUBLIC embree_sycl)