Added AVX10_2 and AVX10_2_512 targets

[johnplatts]

Added preliminary support for the HWY_AVX10_2 (AVX10.2 with 256-bit vectors) and HWY_AVX10_2_512 (AVX10.2 with 512-bit vectors) targets.

Added CPUID detection for AVX10.1 and AVX10.2 support in hwy/targets.cc.

Also added a new hwy/ops/x86_avx3-inl.h header, and moved some of the AVX3/AVX10-specific ops that have dependencies on hwy/ops/x86_512-inl.h if HWY_MAX_BYTES == 64 into the new hwy/ops/x86_avx3-inl.h.

Also moved some of the AVX3/AVX3_DL-specific ops that operate on 256-bit or smaller vectors into hwy/ops/x86_128-inl.h and hwy/ops/x86_256-inl.h to support AVX10.2 targets that do not support 512-bit vectors.

Also refactored some of the AVX3 target macros in hwy/ops/set_macros-inl.h as follows:

• Added a new HWY_TARGET_STR_AVX3_VL512 macro that expands to ",evex512" for GCC/Clang versions that support the -mevex512 option, and is otherwise defined as an empty macro
• Refactored the HWY_TARGET_STR_AVX3 macro to HWY_TARGET_STR_AVX3_256 (which does not include ",evex512" or ",no-evex512") and HWY_TARGET_STR_AVX3 (which includes ",evex512" if compiling with a GCC/Clang version that supports the "-mevex512" option)
• Refactored the HWY_TARGET_STR_AVX3_DL macro to HWY_TARGET_STR_AVX3_DL_256 (which does not include ",evex512" or ",no-evex512") and HWY_TARGET_STR_AVX3_DL (which includes ",evex512" if compiling with a GCC/Clang version that supports the -mevex512 option)
• Refactored the HWY_TARGET_STR_AVX3_ZEN4 macro to HWY_TARGET_STR_AVX3_ZEN4_256 (which does not include ",evex512" or ",no-evex512") and HWY_TARGET_STR_AVX3_ZEN4 (which includes ",evex512" if compiling with a GCC/Clang version that supports the -mevex512 option)
• Refactored the HWY_TARGET_STR_AVX3_SPR macro to HWY_TARGET_STR_AVX3_SPR_256 (which does not include ",evex512" or ",no-evex512") and HWY_TARGET_STR_AVX3_SPR (which includes ",evex512" if compiling with a GCC/Clang version that supports the -mevex512 option)

To compile and run the Google Highway unit tests for the HWY_AVX10_2 and HWY_AVX10_2_512 targets, Clang 19 or later and Intel SDE 9.44 or later are needed.

There are some compilation issues with compiling the HWY_AVX10_2 and HWY_AVX10_2_512 with Clang 18 and GCC 14, even with both Clang 18 and GCC 14 supporting the -mno-evex512 option, including a compiler crash when compiling the matvec_test.cc for the HWY_AVX10_2 target with Clang 18 and compiler warnings that are emitted by GCC 14 when casting a int16_t to __bf16 when compiling with -march=sapphirerapids.

The HWY_AVX10_2 and HWY_AVX10_2_512 are not included in HWY_ATTAINABLE_TARGETS_X86 by default due to compiler issues with GCC or Clang 18.

[johnplatts]

Here is the CMake command line to configure to build only the HWY_AVX10_2 target (but not any other targets) with Clang 19 (where <path_to_hwy_src> is substituted with the path to the Google Highway sources <path_to_sde_directory> is substituted with the directory containing the sde64 executable):
CC=clang-19 CXX=clang++-19 cmake <path_to_hwy_src> -DCMAKE_CXX_FLAGS='-march=sapphirerapids -mno-evex512 -DHWY_BASELINE_TARGETS=0x20 -DHWY_ATTAINABLE_TARGETS_X86=0x20 -DHWY_DISABLED_TARGETS=0x6000000000007FDF -DHWY_COMPILE_ONLY_STATIC=1' -DHWY_ENABLE_CONTRIB=OFF -DCMAKE_CXX_STANDARD=17 -DHWY_WARNINGS_ARE_ERRORS=ON -DCMAKE_CROSSCOMPILING_EMULATOR='<path_to_sde_directory>/sde64;-future;--'

Here is the CMake command line to configure to build only the HWY_AVX10_2_512 target (but not any other targets) with Clang 19 (where <path_to_hwy_src> is substituted with the path to the Google Highway sources <path_to_sde_directory> is substituted with the directory containing the sde64 executable):
CC=clang-19 CXX=clang++-19 cmake <path_to_hwy_src> -DCMAKE_CXX_FLAGS='-march=sapphirerapids -mevex512 -DHWY_BASELINE_TARGETS=0x08 -DHWY_ATTAINABLE_TARGETS_X86=0x08 -DHWY_DISABLED_TARGETS=0x6000000000007FF7 -DHWY_COMPILE_ONLY_STATIC=1' -DHWY_ENABLE_CONTRIB=OFF -DCMAKE_CXX_STANDARD=17 -DHWY_WARNINGS_ARE_ERRORS=ON -DCMAKE_CROSSCOMPILING_EMULATOR='<path_to_sde_directory>/sde64;-future;--'

Here is the CMake command line to configure to build the HWY_AVX10_2 and HWY_AVX10_2_512 targets (but not any other targets) with Clang 19 (where <path_to_hwy_src> is substituted with the path to the Google Highway sources <path_to_sde_directory> is substituted with the directory containing the sde64 executable):
CC=clang-19 CXX=clang++-19 cmake <path_to_hwy_src> -DCMAKE_CXX_FLAGS='-march=sapphirerapids -mno-evex512 -DHWY_BASELINE_TARGETS=0x20 -DHWY_ATTAINABLE_TARGETS_X86=0x28 -DHWY_DISABLED_TARGETS=0x6000000000007FD7' -DHWY_ENABLE_CONTRIB=OFF -DCMAKE_CXX_STANDARD=17 -DHWY_WARNINGS_ARE_ERRORS=ON -DCMAKE_CROSSCOMPILING_EMULATOR='<path_to_sde_directory>/sde64;-future;--'

[jan-wassenberg]

Very nice! Thanks for adding this already. Just one small typo, but we'll go ahead and try to import/run CI in case other changes are required.

[jan-wassenberg]

byte -> bit :)