@cpp-linter/clang-tidy-win32-x64 19.1.0 → 23.1.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +2 -2
- package/bin/clang-tidy.exe +2 -2
- package/lib/clang/23/include/__clang_cuda_complex_builtins.h +236 -0
- package/lib/clang/{19 → 23}/include/__clang_cuda_device_functions.h +15 -1
- package/lib/clang/{19 → 23}/include/__clang_cuda_intrinsics.h +288 -20
- package/lib/clang/{19 → 23}/include/__clang_cuda_math.h +5 -0
- package/lib/clang/{19 → 23}/include/__clang_cuda_math_forward_declares.h +5 -0
- package/lib/clang/{19 → 23}/include/__clang_cuda_runtime_wrapper.h +28 -5
- package/lib/clang/{19 → 23}/include/__clang_cuda_texture_intrinsics.h +437 -2
- package/lib/clang/23/include/__clang_gpu_builtin_vars.h +71 -0
- package/lib/clang/23/include/__clang_gpu_device_functions.h +411 -0
- package/lib/clang/23/include/__clang_gpu_intrinsics.h +242 -0
- package/lib/clang/{19 → 23}/include/__clang_hip_cmath.h +15 -9
- package/lib/clang/{19 → 23}/include/__clang_hip_libdevice_declares.h +24 -139
- package/lib/clang/{19 → 23}/include/__clang_hip_math.h +115 -34
- package/lib/clang/{19 → 23}/include/__clang_hip_runtime_wrapper.h +12 -1
- package/lib/clang/23/include/__clang_spirv_builtins.h +217 -0
- package/lib/clang/23/include/__clang_spirv_libdevice_declares.h +170 -0
- package/lib/clang/23/include/__clang_spirv_math.h +719 -0
- package/lib/clang/{19/include/float.h → 23/include/__float_float.h} +14 -34
- package/lib/clang/23/include/__float_header_macro.h +12 -0
- package/lib/clang/23/include/__float_infinity_nan.h +20 -0
- package/lib/clang/{19 → 23}/include/__stdarg_va_arg.h +2 -2
- package/lib/clang/{19 → 23}/include/adcintrin.h +5 -0
- package/lib/clang/{19 → 23}/include/adxintrin.h +5 -0
- package/lib/clang/{19 → 23}/include/altivec.h +169 -96
- package/lib/clang/23/include/amdgpuintrin.h +164 -0
- package/lib/clang/23/include/amdhsa_abi.h +83 -0
- package/lib/clang/23/include/amo.h +220 -0
- package/lib/clang/23/include/amxavx512intrin.h +596 -0
- package/lib/clang/{19 → 23}/include/amxcomplexintrin.h +4 -6
- package/lib/clang/{19 → 23}/include/amxfp16intrin.h +35 -0
- package/lib/clang/23/include/amxfp8intrin.h +230 -0
- package/lib/clang/{19 → 23}/include/amxintrin.h +3 -35
- package/lib/clang/23/include/amxmovrsintrin.h +48 -0
- package/lib/clang/23/include/andes_vector.h +21 -0
- package/lib/clang/23/include/arm64_neon.h +21 -0
- package/lib/clang/{19 → 23}/include/arm64intr.h +10 -0
- package/lib/clang/{19 → 23}/include/arm_acle.h +67 -89
- package/lib/clang/{19 → 23}/include/arm_fp16.h +89 -89
- package/lib/clang/{19 → 23}/include/arm_neon.h +43614 -38901
- package/lib/clang/{19 → 23}/include/arm_sme.h +1028 -94
- package/lib/clang/{19 → 23}/include/arm_sve.h +7331 -5170
- package/lib/clang/{19 → 23}/include/arm_vector_types.h +87 -0
- package/lib/clang/23/include/avx10_2_512bf16intrin.h +568 -0
- package/lib/clang/23/include/avx10_2_512convertintrin.h +322 -0
- package/lib/clang/23/include/avx10_2_512minmaxintrin.h +127 -0
- package/lib/clang/23/include/avx10_2_512niintrin.h +314 -0
- package/lib/clang/23/include/avx10_2_512satcvtdsintrin.h +307 -0
- package/lib/clang/23/include/avx10_2_512satcvtintrin.h +301 -0
- package/lib/clang/23/include/avx10_2bf16intrin.h +1094 -0
- package/lib/clang/23/include/avx10_2convertintrin.h +3257 -0
- package/lib/clang/23/include/avx10_2copyintrin.h +66 -0
- package/lib/clang/23/include/avx10_2minmaxintrin.h +232 -0
- package/lib/clang/23/include/avx10_2niintrin.h +409 -0
- package/lib/clang/23/include/avx10_2satcvtdsintrin.h +376 -0
- package/lib/clang/23/include/avx10_2satcvtintrin.h +312 -0
- package/lib/clang/{19 → 23}/include/avx2intrin.h +284 -378
- package/lib/clang/{19 → 23}/include/avx512bf16intrin.h +25 -16
- package/lib/clang/{19 → 23}/include/avx512bitalgintrin.h +21 -32
- package/lib/clang/23/include/avx512bmmintrin.h +174 -0
- package/lib/clang/23/include/avx512bmmvlintrin.h +245 -0
- package/lib/clang/{19 → 23}/include/avx512bwintrin.h +405 -534
- package/lib/clang/23/include/avx512cdintrin.h +112 -0
- package/lib/clang/{19 → 23}/include/avx512dqintrin.h +132 -161
- package/lib/clang/{19 → 23}/include/avx512fintrin.h +1068 -1496
- package/lib/clang/{19 → 23}/include/avx512fp16intrin.h +126 -129
- package/lib/clang/23/include/avx512ifmaintrin.h +68 -0
- package/lib/clang/23/include/avx512ifmavlintrin.h +138 -0
- package/lib/clang/{19 → 23}/include/avx512vbmi2intrin.h +55 -48
- package/lib/clang/{19 → 23}/include/avx512vbmiintrin.h +19 -27
- package/lib/clang/{19 → 23}/include/avx512vbmivlintrin.h +35 -49
- package/lib/clang/{19 → 23}/include/avx512vlbf16intrin.h +32 -22
- package/lib/clang/{19 → 23}/include/avx512vlbitalgintrin.h +41 -57
- package/lib/clang/{19 → 23}/include/avx512vlbwintrin.h +494 -609
- package/lib/clang/23/include/avx512vlcdintrin.h +202 -0
- package/lib/clang/{19 → 23}/include/avx512vldqintrin.h +110 -127
- package/lib/clang/{19 → 23}/include/avx512vlfp16intrin.h +130 -111
- package/lib/clang/{19 → 23}/include/avx512vlintrin.h +1025 -1395
- package/lib/clang/{19 → 23}/include/avx512vlvbmi2intrin.h +102 -95
- package/lib/clang/{19 → 23}/include/avx512vlvnniintrin.h +50 -52
- package/lib/clang/{19 → 23}/include/avx512vlvp2intersectintrin.h +2 -2
- package/lib/clang/{19 → 23}/include/avx512vnniintrin.h +42 -45
- package/lib/clang/{19 → 23}/include/avx512vp2intersectintrin.h +2 -3
- package/lib/clang/{19 → 23}/include/avx512vpopcntdqintrin.h +10 -4
- package/lib/clang/{19 → 23}/include/avx512vpopcntdqvlintrin.h +20 -10
- package/lib/clang/{19 → 23}/include/avxifmaintrin.h +16 -0
- package/lib/clang/{19 → 23}/include/avxintrin.h +186 -248
- package/lib/clang/{19 → 23}/include/avxvnniint16intrin.h +130 -152
- package/lib/clang/{19 → 23}/include/avxvnniint8intrin.h +68 -85
- package/lib/clang/{19 → 23}/include/avxvnniintrin.h +39 -26
- package/lib/clang/{19 → 23}/include/bmi2intrin.h +15 -17
- package/lib/clang/{19 → 23}/include/bmiintrin.h +31 -41
- package/lib/clang/{19 → 23}/include/cmpccxaddintrin.h +1 -1
- package/lib/clang/{19 → 23}/include/cpuid.h +137 -37
- package/lib/clang/{19 → 23}/include/crc32intrin.h +10 -8
- package/lib/clang/23/include/cuda_wrappers/__utility/declval.h +28 -0
- package/lib/clang/23/include/cuda_wrappers/bits/c++config.h +61 -0
- package/lib/clang/23/include/cuda_wrappers/cmath +44 -0
- package/lib/clang/{19 → 23}/include/cuda_wrappers/new +2 -0
- package/lib/clang/{19 → 23}/include/emmintrin.h +306 -272
- package/lib/clang/23/include/endian.h +91 -0
- package/lib/clang/{19 → 23}/include/f16cintrin.h +23 -9
- package/lib/clang/23/include/float.h +48 -0
- package/lib/clang/23/include/fma4intrin.h +220 -0
- package/lib/clang/{19 → 23}/include/fmaintrin.h +96 -66
- package/lib/clang/{19 → 23}/include/gfniintrin.h +39 -12
- package/lib/clang/23/include/gpuintrin.h +348 -0
- package/lib/clang/{19 → 23}/include/hexagon_types.h +31 -24
- package/lib/clang/23/include/hlsl/hlsl_alias_intrinsics.h +222 -0
- package/lib/clang/23/include/hlsl/hlsl_alias_intrinsics_gen.inc +3869 -0
- package/lib/clang/23/include/hlsl/hlsl_basic_types.h +370 -0
- package/lib/clang/23/include/hlsl/hlsl_compat_overloads.h +559 -0
- package/lib/clang/23/include/hlsl/hlsl_detail.h +56 -0
- package/lib/clang/23/include/hlsl/hlsl_inline_intrinsics_gen.inc +546 -0
- package/lib/clang/23/include/hlsl/hlsl_intrinsic_helpers.h +170 -0
- package/lib/clang/23/include/hlsl/hlsl_intrinsics.h +331 -0
- package/lib/clang/23/include/hlsl/hlsl_spirv.h +28 -0
- package/lib/clang/23/include/hlsl.h +38 -0
- package/lib/clang/{19 → 23}/include/hvx_hexagon_protos.h +1041 -803
- package/lib/clang/{19 → 23}/include/immintrin.h +35 -185
- package/lib/clang/{19 → 23}/include/intrin.h +62 -21
- package/lib/clang/{19 → 23}/include/intrin0.h +10 -1
- package/lib/clang/{19 → 23}/include/keylockerintrin.h +0 -9
- package/lib/clang/{19 → 23}/include/larchintrin.h +24 -8
- package/lib/clang/{19 → 23}/include/lasxintrin.h +140 -25
- package/lib/clang/{19 → 23}/include/limits.h +7 -4
- package/lib/clang/{19 → 23}/include/llvm_libc_wrappers/assert.h +3 -5
- package/lib/clang/23/include/llvm_libc_wrappers/ctype.h +29 -0
- package/lib/clang/{19 → 23}/include/llvm_libc_wrappers/inttypes.h +3 -5
- package/lib/clang/23/include/llvm_libc_wrappers/stdio.h +52 -0
- package/lib/clang/{19 → 23}/include/llvm_libc_wrappers/stdlib.h +3 -16
- package/lib/clang/23/include/llvm_libc_wrappers/string.h +29 -0
- package/lib/clang/{19 → 23}/include/llvm_libc_wrappers/time.h +4 -10
- package/lib/clang/23/include/llvm_offload_wrappers/__llvm_offload.h +31 -0
- package/lib/clang/23/include/llvm_offload_wrappers/__llvm_offload_device.h +10 -0
- package/lib/clang/23/include/llvm_offload_wrappers/__llvm_offload_host.h +15 -0
- package/lib/clang/{19 → 23}/include/lsxintrin.h +27 -25
- package/lib/clang/{19 → 23}/include/lzcntintrin.h +13 -8
- package/lib/clang/{19 → 23}/include/mmintrin.h +241 -256
- package/lib/clang/{19 → 23}/include/module.modulemap +54 -9
- package/lib/clang/23/include/movrs_avx10_2_512intrin.h +98 -0
- package/lib/clang/23/include/movrs_avx10_2intrin.h +174 -0
- package/lib/clang/23/include/movrsintrin.h +59 -0
- package/lib/clang/23/include/nvptxintrin.h +208 -0
- package/lib/clang/{19 → 23}/include/opencl-c-base.h +28 -92
- package/lib/clang/{19 → 23}/include/opencl-c.h +506 -140
- package/lib/clang/{19 → 23}/include/openmp_wrappers/__clang_openmp_device_functions.h +18 -5
- package/lib/clang/{19 → 23}/include/openmp_wrappers/complex +7 -1
- package/lib/clang/{19 → 23}/include/openmp_wrappers/complex.h +6 -0
- package/lib/clang/{19 → 23}/include/openmp_wrappers/complex_cmath.h +7 -2
- package/lib/clang/{19 → 23}/include/openmp_wrappers/math.h +10 -0
- package/lib/clang/{19 → 23}/include/pmmintrin.h +24 -23
- package/lib/clang/{19 → 23}/include/popcntintrin.h +7 -7
- package/lib/clang/{19 → 23}/include/prfchwintrin.h +10 -13
- package/lib/clang/{19 → 23}/include/ptrauth.h +169 -14
- package/lib/clang/{19 → 23}/include/riscv_bitmanip.h +7 -0
- package/lib/clang/23/include/riscv_corev_alu.h +130 -0
- package/lib/clang/{19 → 23}/include/riscv_crypto.h +8 -0
- package/lib/clang/23/include/riscv_mips.h +36 -0
- package/lib/clang/23/include/riscv_nds.h +92 -0
- package/lib/clang/{19 → 23}/include/riscv_ntlh.h +2 -0
- package/lib/clang/23/include/riscv_packed_simd.h +644 -0
- package/lib/clang/{19 → 23}/include/riscv_vector.h +43 -5
- package/lib/clang/{19 → 23}/include/shaintrin.h +3 -2
- package/lib/clang/{19 → 23}/include/sifive_vector.h +76 -2
- package/lib/clang/23/include/sm4evexintrin.h +32 -0
- package/lib/clang/{19 → 23}/include/smmintrin.h +80 -56
- package/lib/clang/23/include/spirvintrin.h +167 -0
- package/lib/clang/{19 → 23}/include/stdalign.h +0 -5
- package/lib/clang/23/include/stdcountof.h +15 -0
- package/lib/clang/23/include/stddefer.h +19 -0
- package/lib/clang/{19 → 23}/include/stdint.h +23 -143
- package/lib/clang/{19 → 23}/include/tbmintrin.h +25 -37
- package/lib/clang/{19 → 23}/include/tmmintrin.h +125 -130
- package/lib/clang/{19 → 23}/include/vaesintrin.h +1 -2
- package/lib/clang/{19 → 23}/include/vecintrin.h +1783 -60
- package/lib/clang/{19 → 23}/include/wasm_simd128.h +184 -21
- package/lib/clang/{19 → 23}/include/x86gprintrin.h +0 -14
- package/lib/clang/{19 → 23}/include/x86intrin.h +0 -18
- package/lib/clang/{19 → 23}/include/xmmintrin.h +190 -221
- package/lib/clang/{19 → 23}/include/xopintrin.h +20 -10
- package/lib/clang/23/include/zos_wrappers/grp.h +19 -0
- package/lib/clang/23/include/zos_wrappers/locale.h +19 -0
- package/lib/clang/23/include/zos_wrappers/math.h +31 -0
- package/lib/clang/23/include/zos_wrappers/poll.h +19 -0
- package/lib/clang/23/include/zos_wrappers/stdlib.h +20 -0
- package/lib/clang/23/include/zos_wrappers/string.h +19 -0
- package/lib/clang/23/include/zos_wrappers/time.h +19 -0
- package/lib/clang/23/include/zos_wrappers/variant.h +19 -0
- package/package.json +6 -6
- package/lib/clang/19/include/__clang_cuda_complex_builtins.h +0 -285
- package/lib/clang/19/include/avx512cdintrin.h +0 -125
- package/lib/clang/19/include/avx512ifmaintrin.h +0 -70
- package/lib/clang/19/include/avx512ifmavlintrin.h +0 -111
- package/lib/clang/19/include/avx512vlcdintrin.h +0 -230
- package/lib/clang/19/include/cuda_wrappers/cmath +0 -90
- package/lib/clang/19/include/fma4intrin.h +0 -218
- package/lib/clang/19/include/hlsl/hlsl_basic_types.h +0 -74
- package/lib/clang/19/include/hlsl/hlsl_intrinsics.h +0 -1697
- package/lib/clang/19/include/hlsl.h +0 -15
- package/lib/clang/19/include/llvm_libc_wrappers/ctype.h +0 -102
- package/lib/clang/19/include/llvm_libc_wrappers/stdio.h +0 -80
- package/lib/clang/19/include/llvm_libc_wrappers/string.h +0 -96
- /package/lib/clang/{19 → 23}/include/__clang_cuda_builtin_vars.h +0 -0
- /package/lib/clang/{19 → 23}/include/__clang_cuda_cmath.h +0 -0
- /package/lib/clang/{19 → 23}/include/__clang_cuda_libdevice_declares.h +0 -0
- /package/lib/clang/{19 → 23}/include/__clang_hip_stdlib.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stdarg___gnuc_va_list.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stdarg___va_copy.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stdarg_header_macro.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stdarg_va_copy.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stdarg_va_list.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_header_macro.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_max_align_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_null.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_nullptr_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_offsetof.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_ptrdiff_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_rsize_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_size_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_unreachable.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_wchar_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__stddef_wint_t.h +0 -0
- /package/lib/clang/{19 → 23}/include/__wmmintrin_aes.h +0 -0
- /package/lib/clang/{19 → 23}/include/__wmmintrin_pclmul.h +0 -0
- /package/lib/clang/{19 → 23}/include/ammintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/arm_bf16.h +0 -0
- /package/lib/clang/{19 → 23}/include/arm_cde.h +0 -0
- /package/lib/clang/{19 → 23}/include/arm_cmse.h +0 -0
- /package/lib/clang/{19 → 23}/include/arm_mve.h +0 -0
- /package/lib/clang/{19 → 23}/include/arm_neon_sve_bridge.h +0 -0
- /package/lib/clang/{19 → 23}/include/armintr.h +0 -0
- /package/lib/clang/{19 → 23}/include/avxneconvertintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/builtins.h +0 -0
- /package/lib/clang/{19 → 23}/include/cet.h +0 -0
- /package/lib/clang/{19 → 23}/include/cetintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/cldemoteintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/clflushoptintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/clwbintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/clzerointrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/cuda_wrappers/algorithm +0 -0
- /package/lib/clang/{19 → 23}/include/cuda_wrappers/bits/basic_string.h +0 -0
- /package/lib/clang/{19 → 23}/include/cuda_wrappers/bits/basic_string.tcc +0 -0
- /package/lib/clang/{19 → 23}/include/cuda_wrappers/bits/shared_ptr_base.h +0 -0
- /package/lib/clang/{19 → 23}/include/cuda_wrappers/complex +0 -0
- /package/lib/clang/{19 → 23}/include/enqcmdintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/fxsrintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/hexagon_circ_brev_intrinsics.h +0 -0
- /package/lib/clang/{19 → 23}/include/hexagon_protos.h +0 -0
- /package/lib/clang/{19 → 23}/include/hresetintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/htmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/htmxlintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ia32intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/inttypes.h +0 -0
- /package/lib/clang/{19 → 23}/include/invpcidintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/iso646.h +0 -0
- /package/lib/clang/{19 → 23}/include/lwpintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/mm3dnow.h +0 -0
- /package/lib/clang/{19 → 23}/include/mm_malloc.h +0 -0
- /package/lib/clang/{19 → 23}/include/movdirintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/msa.h +0 -0
- /package/lib/clang/{19 → 23}/include/mwaitxintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/nmmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/openmp_wrappers/cmath +0 -0
- /package/lib/clang/{19 → 23}/include/openmp_wrappers/new +0 -0
- /package/lib/clang/{19 → 23}/include/pconfigintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/pkuintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/bmi2intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/bmiintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/emmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/immintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/mm_malloc.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/mmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/nmmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/pmmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/smmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/tmmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/x86gprintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/x86intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ppc_wrappers/xmmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/prfchiintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/ptwriteintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/raointintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/rdpruintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/rdseedintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/rtmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/s390intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/serializeintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/sgxintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/sha512intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/sm3intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/sm4intrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/stdarg.h +0 -0
- /package/lib/clang/{19 → 23}/include/stdatomic.h +0 -0
- /package/lib/clang/{19 → 23}/include/stdbool.h +0 -0
- /package/lib/clang/{19 → 23}/include/stdckdint.h +0 -0
- /package/lib/clang/{19 → 23}/include/stddef.h +0 -0
- /package/lib/clang/{19 → 23}/include/stdnoreturn.h +0 -0
- /package/lib/clang/{19 → 23}/include/tgmath.h +0 -0
- /package/lib/clang/{19 → 23}/include/tsxldtrkintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/uintrintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/unwind.h +0 -0
- /package/lib/clang/{19 → 23}/include/usermsrintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/vadefs.h +0 -0
- /package/lib/clang/{19 → 23}/include/varargs.h +0 -0
- /package/lib/clang/{19 → 23}/include/velintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/velintrin_approx.h +0 -0
- /package/lib/clang/{19 → 23}/include/velintrin_gen.h +0 -0
- /package/lib/clang/{19 → 23}/include/vpclmulqdqintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/waitpkgintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/wbnoinvdintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/wmmintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/xsavecintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/xsaveintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/xsaveoptintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/xsavesintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/xtestintrin.h +0 -0
- /package/lib/clang/{19 → 23}/include/yvals_core.h +0 -0
- /package/lib/clang/{19 → 23}/include/zos_wrappers/builtins.h +0 -0
|
@@ -17,12 +17,18 @@
|
|
|
17
17
|
/* Define the default attributes for the functions in this file. */
|
|
18
18
|
#define __DEFAULT_FN_ATTRS128 \
|
|
19
19
|
__attribute__((__always_inline__, __nodebug__, \
|
|
20
|
-
__target__("avx512vl,avx512bw
|
|
21
|
-
__min_vector_width__(128)))
|
|
20
|
+
__target__("avx512vl,avx512bw"), __min_vector_width__(128)))
|
|
22
21
|
#define __DEFAULT_FN_ATTRS256 \
|
|
23
22
|
__attribute__((__always_inline__, __nodebug__, \
|
|
24
|
-
__target__("avx512vl,avx512bw
|
|
25
|
-
|
|
23
|
+
__target__("avx512vl,avx512bw"), __min_vector_width__(256)))
|
|
24
|
+
|
|
25
|
+
#if defined(__cplusplus) && (__cplusplus >= 201103L)
|
|
26
|
+
#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 constexpr
|
|
27
|
+
#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 constexpr
|
|
28
|
+
#else
|
|
29
|
+
#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128
|
|
30
|
+
#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256
|
|
31
|
+
#endif
|
|
26
32
|
|
|
27
33
|
/* Integer compare */
|
|
28
34
|
|
|
@@ -306,250 +312,238 @@
|
|
|
306
312
|
#define _mm256_mask_cmpneq_epu16_mask(k, A, B) \
|
|
307
313
|
_mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE)
|
|
308
314
|
|
|
309
|
-
static __inline__ __m256i
|
|
310
|
-
_mm256_mask_add_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B){
|
|
315
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
316
|
+
_mm256_mask_add_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
|
|
311
317
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
312
318
|
(__v32qi)_mm256_add_epi8(__A, __B),
|
|
313
319
|
(__v32qi)__W);
|
|
314
320
|
}
|
|
315
321
|
|
|
316
|
-
static __inline__ __m256i
|
|
322
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
317
323
|
_mm256_maskz_add_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
318
324
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
319
325
|
(__v32qi)_mm256_add_epi8(__A, __B),
|
|
320
326
|
(__v32qi)_mm256_setzero_si256());
|
|
321
327
|
}
|
|
322
328
|
|
|
323
|
-
static __inline__ __m256i
|
|
329
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
324
330
|
_mm256_mask_add_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
325
331
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
326
332
|
(__v16hi)_mm256_add_epi16(__A, __B),
|
|
327
333
|
(__v16hi)__W);
|
|
328
334
|
}
|
|
329
335
|
|
|
330
|
-
static __inline__ __m256i
|
|
336
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
331
337
|
_mm256_maskz_add_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
332
338
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
333
339
|
(__v16hi)_mm256_add_epi16(__A, __B),
|
|
334
340
|
(__v16hi)_mm256_setzero_si256());
|
|
335
341
|
}
|
|
336
342
|
|
|
337
|
-
static __inline__ __m256i
|
|
343
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
338
344
|
_mm256_mask_sub_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
|
|
339
345
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
340
346
|
(__v32qi)_mm256_sub_epi8(__A, __B),
|
|
341
347
|
(__v32qi)__W);
|
|
342
348
|
}
|
|
343
349
|
|
|
344
|
-
static __inline__ __m256i
|
|
350
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
345
351
|
_mm256_maskz_sub_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
346
352
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
347
353
|
(__v32qi)_mm256_sub_epi8(__A, __B),
|
|
348
354
|
(__v32qi)_mm256_setzero_si256());
|
|
349
355
|
}
|
|
350
356
|
|
|
351
|
-
static __inline__ __m256i
|
|
357
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
352
358
|
_mm256_mask_sub_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
353
359
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
354
360
|
(__v16hi)_mm256_sub_epi16(__A, __B),
|
|
355
361
|
(__v16hi)__W);
|
|
356
362
|
}
|
|
357
363
|
|
|
358
|
-
static __inline__ __m256i
|
|
364
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
359
365
|
_mm256_maskz_sub_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
360
366
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
361
367
|
(__v16hi)_mm256_sub_epi16(__A, __B),
|
|
362
368
|
(__v16hi)_mm256_setzero_si256());
|
|
363
369
|
}
|
|
364
370
|
|
|
365
|
-
static __inline__ __m128i
|
|
371
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
366
372
|
_mm_mask_add_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
|
|
367
373
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
368
374
|
(__v16qi)_mm_add_epi8(__A, __B),
|
|
369
375
|
(__v16qi)__W);
|
|
370
376
|
}
|
|
371
377
|
|
|
372
|
-
static __inline__ __m128i
|
|
378
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
373
379
|
_mm_maskz_add_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
374
380
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
375
381
|
(__v16qi)_mm_add_epi8(__A, __B),
|
|
376
382
|
(__v16qi)_mm_setzero_si128());
|
|
377
383
|
}
|
|
378
384
|
|
|
379
|
-
static __inline__ __m128i
|
|
385
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
380
386
|
_mm_mask_add_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
381
387
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
382
388
|
(__v8hi)_mm_add_epi16(__A, __B),
|
|
383
389
|
(__v8hi)__W);
|
|
384
390
|
}
|
|
385
391
|
|
|
386
|
-
static __inline__ __m128i
|
|
392
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
387
393
|
_mm_maskz_add_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
388
394
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
389
395
|
(__v8hi)_mm_add_epi16(__A, __B),
|
|
390
396
|
(__v8hi)_mm_setzero_si128());
|
|
391
397
|
}
|
|
392
398
|
|
|
393
|
-
static __inline__ __m128i
|
|
399
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
394
400
|
_mm_mask_sub_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
|
|
395
401
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
396
402
|
(__v16qi)_mm_sub_epi8(__A, __B),
|
|
397
403
|
(__v16qi)__W);
|
|
398
404
|
}
|
|
399
405
|
|
|
400
|
-
static __inline__ __m128i
|
|
406
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
401
407
|
_mm_maskz_sub_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
402
408
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
403
409
|
(__v16qi)_mm_sub_epi8(__A, __B),
|
|
404
410
|
(__v16qi)_mm_setzero_si128());
|
|
405
411
|
}
|
|
406
412
|
|
|
407
|
-
static __inline__ __m128i
|
|
413
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
408
414
|
_mm_mask_sub_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
409
415
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
410
416
|
(__v8hi)_mm_sub_epi16(__A, __B),
|
|
411
417
|
(__v8hi)__W);
|
|
412
418
|
}
|
|
413
419
|
|
|
414
|
-
static __inline__ __m128i
|
|
420
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
415
421
|
_mm_maskz_sub_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
416
422
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
417
423
|
(__v8hi)_mm_sub_epi16(__A, __B),
|
|
418
424
|
(__v8hi)_mm_setzero_si128());
|
|
419
425
|
}
|
|
420
426
|
|
|
421
|
-
static __inline__ __m256i
|
|
427
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
422
428
|
_mm256_mask_mullo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
423
429
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
424
430
|
(__v16hi)_mm256_mullo_epi16(__A, __B),
|
|
425
431
|
(__v16hi)__W);
|
|
426
432
|
}
|
|
427
433
|
|
|
428
|
-
static __inline__ __m256i
|
|
434
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
429
435
|
_mm256_maskz_mullo_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
430
436
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
431
437
|
(__v16hi)_mm256_mullo_epi16(__A, __B),
|
|
432
438
|
(__v16hi)_mm256_setzero_si256());
|
|
433
439
|
}
|
|
434
440
|
|
|
435
|
-
static __inline__ __m128i
|
|
441
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
436
442
|
_mm_mask_mullo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
437
443
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
438
444
|
(__v8hi)_mm_mullo_epi16(__A, __B),
|
|
439
445
|
(__v8hi)__W);
|
|
440
446
|
}
|
|
441
447
|
|
|
442
|
-
static __inline__ __m128i
|
|
448
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
443
449
|
_mm_maskz_mullo_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
444
450
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
445
451
|
(__v8hi)_mm_mullo_epi16(__A, __B),
|
|
446
452
|
(__v8hi)_mm_setzero_si128());
|
|
447
453
|
}
|
|
448
454
|
|
|
449
|
-
static __inline__ __m128i
|
|
450
|
-
_mm_mask_blend_epi8
|
|
451
|
-
{
|
|
455
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
456
|
+
_mm_mask_blend_epi8(__mmask16 __U, __m128i __A, __m128i __W) {
|
|
452
457
|
return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U,
|
|
453
458
|
(__v16qi) __W,
|
|
454
459
|
(__v16qi) __A);
|
|
455
460
|
}
|
|
456
461
|
|
|
457
|
-
static __inline__ __m256i
|
|
458
|
-
_mm256_mask_blend_epi8
|
|
459
|
-
{
|
|
462
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
463
|
+
_mm256_mask_blend_epi8(__mmask32 __U, __m256i __A, __m256i __W) {
|
|
460
464
|
return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U,
|
|
461
465
|
(__v32qi) __W,
|
|
462
466
|
(__v32qi) __A);
|
|
463
467
|
}
|
|
464
468
|
|
|
465
|
-
static __inline__ __m128i
|
|
466
|
-
_mm_mask_blend_epi16
|
|
467
|
-
{
|
|
469
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
470
|
+
_mm_mask_blend_epi16(__mmask8 __U, __m128i __A, __m128i __W) {
|
|
468
471
|
return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U,
|
|
469
472
|
(__v8hi) __W,
|
|
470
473
|
(__v8hi) __A);
|
|
471
474
|
}
|
|
472
475
|
|
|
473
|
-
static __inline__ __m256i
|
|
474
|
-
_mm256_mask_blend_epi16
|
|
475
|
-
{
|
|
476
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
477
|
+
_mm256_mask_blend_epi16(__mmask16 __U, __m256i __A, __m256i __W) {
|
|
476
478
|
return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U,
|
|
477
479
|
(__v16hi) __W,
|
|
478
480
|
(__v16hi) __A);
|
|
479
481
|
}
|
|
480
482
|
|
|
481
|
-
static __inline__ __m128i
|
|
482
|
-
_mm_mask_abs_epi8(__m128i __W, __mmask16 __U, __m128i __A)
|
|
483
|
-
{
|
|
483
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
484
|
+
_mm_mask_abs_epi8(__m128i __W, __mmask16 __U, __m128i __A) {
|
|
484
485
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
485
486
|
(__v16qi)_mm_abs_epi8(__A),
|
|
486
487
|
(__v16qi)__W);
|
|
487
488
|
}
|
|
488
489
|
|
|
489
|
-
static __inline__ __m128i
|
|
490
|
-
_mm_maskz_abs_epi8(__mmask16 __U, __m128i __A)
|
|
491
|
-
{
|
|
490
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
491
|
+
_mm_maskz_abs_epi8(__mmask16 __U, __m128i __A) {
|
|
492
492
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
493
493
|
(__v16qi)_mm_abs_epi8(__A),
|
|
494
494
|
(__v16qi)_mm_setzero_si128());
|
|
495
495
|
}
|
|
496
496
|
|
|
497
|
-
static __inline__ __m256i
|
|
498
|
-
_mm256_mask_abs_epi8(__m256i __W, __mmask32 __U, __m256i __A)
|
|
499
|
-
{
|
|
497
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
498
|
+
_mm256_mask_abs_epi8(__m256i __W, __mmask32 __U, __m256i __A) {
|
|
500
499
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
501
500
|
(__v32qi)_mm256_abs_epi8(__A),
|
|
502
501
|
(__v32qi)__W);
|
|
503
502
|
}
|
|
504
503
|
|
|
505
|
-
static __inline__ __m256i
|
|
506
|
-
_mm256_maskz_abs_epi8
|
|
507
|
-
{
|
|
504
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
505
|
+
_mm256_maskz_abs_epi8(__mmask32 __U, __m256i __A) {
|
|
508
506
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
509
507
|
(__v32qi)_mm256_abs_epi8(__A),
|
|
510
508
|
(__v32qi)_mm256_setzero_si256());
|
|
511
509
|
}
|
|
512
510
|
|
|
513
|
-
static __inline__ __m128i
|
|
514
|
-
_mm_mask_abs_epi16(__m128i __W, __mmask8 __U, __m128i __A)
|
|
515
|
-
{
|
|
511
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
512
|
+
_mm_mask_abs_epi16(__m128i __W, __mmask8 __U, __m128i __A) {
|
|
516
513
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
517
514
|
(__v8hi)_mm_abs_epi16(__A),
|
|
518
515
|
(__v8hi)__W);
|
|
519
516
|
}
|
|
520
517
|
|
|
521
|
-
static __inline__ __m128i
|
|
522
|
-
_mm_maskz_abs_epi16(__mmask8 __U, __m128i __A)
|
|
523
|
-
{
|
|
518
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
519
|
+
_mm_maskz_abs_epi16(__mmask8 __U, __m128i __A) {
|
|
524
520
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
525
521
|
(__v8hi)_mm_abs_epi16(__A),
|
|
526
522
|
(__v8hi)_mm_setzero_si128());
|
|
527
523
|
}
|
|
528
524
|
|
|
529
|
-
static __inline__ __m256i
|
|
530
|
-
_mm256_mask_abs_epi16(__m256i __W, __mmask16 __U, __m256i __A)
|
|
531
|
-
{
|
|
525
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
526
|
+
_mm256_mask_abs_epi16(__m256i __W, __mmask16 __U, __m256i __A) {
|
|
532
527
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
533
528
|
(__v16hi)_mm256_abs_epi16(__A),
|
|
534
529
|
(__v16hi)__W);
|
|
535
530
|
}
|
|
536
531
|
|
|
537
|
-
static __inline__ __m256i
|
|
538
|
-
_mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A)
|
|
539
|
-
{
|
|
532
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
533
|
+
_mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A) {
|
|
540
534
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
541
535
|
(__v16hi)_mm256_abs_epi16(__A),
|
|
542
536
|
(__v16hi)_mm256_setzero_si256());
|
|
543
537
|
}
|
|
544
538
|
|
|
545
|
-
static __inline__ __m128i
|
|
539
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
546
540
|
_mm_maskz_packs_epi32(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
547
541
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
548
542
|
(__v8hi)_mm_packs_epi32(__A, __B),
|
|
549
543
|
(__v8hi)_mm_setzero_si128());
|
|
550
544
|
}
|
|
551
545
|
|
|
552
|
-
static __inline__ __m128i
|
|
546
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
553
547
|
_mm_mask_packs_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
554
548
|
{
|
|
555
549
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
@@ -557,7 +551,7 @@ _mm_mask_packs_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
|
557
551
|
(__v8hi)__W);
|
|
558
552
|
}
|
|
559
553
|
|
|
560
|
-
static __inline__ __m256i
|
|
554
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
561
555
|
_mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B)
|
|
562
556
|
{
|
|
563
557
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
@@ -565,7 +559,7 @@ _mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B)
|
|
|
565
559
|
(__v16hi)_mm256_setzero_si256());
|
|
566
560
|
}
|
|
567
561
|
|
|
568
|
-
static __inline__ __m256i
|
|
562
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
569
563
|
_mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
570
564
|
{
|
|
571
565
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
@@ -573,7 +567,7 @@ _mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
|
573
567
|
(__v16hi)__W);
|
|
574
568
|
}
|
|
575
569
|
|
|
576
|
-
static __inline__ __m128i
|
|
570
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
577
571
|
_mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B)
|
|
578
572
|
{
|
|
579
573
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
@@ -581,7 +575,7 @@ _mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B)
|
|
|
581
575
|
(__v16qi)_mm_setzero_si128());
|
|
582
576
|
}
|
|
583
577
|
|
|
584
|
-
static __inline__ __m128i
|
|
578
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
585
579
|
_mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
586
580
|
{
|
|
587
581
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
@@ -589,7 +583,7 @@ _mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
|
589
583
|
(__v16qi)__W);
|
|
590
584
|
}
|
|
591
585
|
|
|
592
|
-
static __inline__ __m256i
|
|
586
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
593
587
|
_mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B)
|
|
594
588
|
{
|
|
595
589
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
@@ -597,7 +591,7 @@ _mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B)
|
|
|
597
591
|
(__v32qi)_mm256_setzero_si256());
|
|
598
592
|
}
|
|
599
593
|
|
|
600
|
-
static __inline__ __m256i
|
|
594
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
601
595
|
_mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
602
596
|
{
|
|
603
597
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
@@ -605,7 +599,7 @@ _mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
|
605
599
|
(__v32qi)__W);
|
|
606
600
|
}
|
|
607
601
|
|
|
608
|
-
static __inline__ __m128i
|
|
602
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
609
603
|
_mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B)
|
|
610
604
|
{
|
|
611
605
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
@@ -613,7 +607,7 @@ _mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B)
|
|
|
613
607
|
(__v8hi)_mm_setzero_si128());
|
|
614
608
|
}
|
|
615
609
|
|
|
616
|
-
static __inline__ __m128i
|
|
610
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
617
611
|
_mm_mask_packus_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
618
612
|
{
|
|
619
613
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
@@ -621,7 +615,7 @@ _mm_mask_packus_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
|
621
615
|
(__v8hi)__W);
|
|
622
616
|
}
|
|
623
617
|
|
|
624
|
-
static __inline__ __m256i
|
|
618
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
625
619
|
_mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B)
|
|
626
620
|
{
|
|
627
621
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
@@ -629,7 +623,7 @@ _mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B)
|
|
|
629
623
|
(__v16hi)_mm256_setzero_si256());
|
|
630
624
|
}
|
|
631
625
|
|
|
632
|
-
static __inline__ __m256i
|
|
626
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
633
627
|
_mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
634
628
|
{
|
|
635
629
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
@@ -637,7 +631,7 @@ _mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
|
637
631
|
(__v16hi)__W);
|
|
638
632
|
}
|
|
639
633
|
|
|
640
|
-
static __inline__ __m128i
|
|
634
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
641
635
|
_mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B)
|
|
642
636
|
{
|
|
643
637
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
@@ -645,7 +639,7 @@ _mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B)
|
|
|
645
639
|
(__v16qi)_mm_setzero_si128());
|
|
646
640
|
}
|
|
647
641
|
|
|
648
|
-
static __inline__ __m128i
|
|
642
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
649
643
|
_mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
650
644
|
{
|
|
651
645
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
@@ -653,7 +647,7 @@ _mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
|
653
647
|
(__v16qi)__W);
|
|
654
648
|
}
|
|
655
649
|
|
|
656
|
-
static __inline__ __m256i
|
|
650
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
657
651
|
_mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B)
|
|
658
652
|
{
|
|
659
653
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
@@ -661,7 +655,7 @@ _mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B)
|
|
|
661
655
|
(__v32qi)_mm256_setzero_si256());
|
|
662
656
|
}
|
|
663
657
|
|
|
664
|
-
static __inline__ __m256i
|
|
658
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
665
659
|
_mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
666
660
|
{
|
|
667
661
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
@@ -669,7 +663,7 @@ _mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
|
669
663
|
(__v32qi)__W);
|
|
670
664
|
}
|
|
671
665
|
|
|
672
|
-
static __inline__ __m128i
|
|
666
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
673
667
|
_mm_mask_adds_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
674
668
|
{
|
|
675
669
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -677,7 +671,7 @@ _mm_mask_adds_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
|
677
671
|
(__v16qi)__W);
|
|
678
672
|
}
|
|
679
673
|
|
|
680
|
-
static __inline__ __m128i
|
|
674
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
681
675
|
_mm_maskz_adds_epi8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
682
676
|
{
|
|
683
677
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -685,7 +679,7 @@ _mm_maskz_adds_epi8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
|
685
679
|
(__v16qi)_mm_setzero_si128());
|
|
686
680
|
}
|
|
687
681
|
|
|
688
|
-
static __inline__ __m256i
|
|
682
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
689
683
|
_mm256_mask_adds_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
690
684
|
{
|
|
691
685
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -693,7 +687,7 @@ _mm256_mask_adds_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
|
693
687
|
(__v32qi)__W);
|
|
694
688
|
}
|
|
695
689
|
|
|
696
|
-
static __inline__ __m256i
|
|
690
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
697
691
|
_mm256_maskz_adds_epi8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
698
692
|
{
|
|
699
693
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -701,7 +695,7 @@ _mm256_maskz_adds_epi8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
|
701
695
|
(__v32qi)_mm256_setzero_si256());
|
|
702
696
|
}
|
|
703
697
|
|
|
704
|
-
static __inline__ __m128i
|
|
698
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
705
699
|
_mm_mask_adds_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
706
700
|
{
|
|
707
701
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -709,7 +703,7 @@ _mm_mask_adds_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
709
703
|
(__v8hi)__W);
|
|
710
704
|
}
|
|
711
705
|
|
|
712
|
-
static __inline__ __m128i
|
|
706
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
713
707
|
_mm_maskz_adds_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
714
708
|
{
|
|
715
709
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -717,7 +711,7 @@ _mm_maskz_adds_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
717
711
|
(__v8hi)_mm_setzero_si128());
|
|
718
712
|
}
|
|
719
713
|
|
|
720
|
-
static __inline__ __m256i
|
|
714
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
721
715
|
_mm256_mask_adds_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
722
716
|
{
|
|
723
717
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -725,7 +719,7 @@ _mm256_mask_adds_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
|
725
719
|
(__v16hi)__W);
|
|
726
720
|
}
|
|
727
721
|
|
|
728
|
-
static __inline__ __m256i
|
|
722
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
729
723
|
_mm256_maskz_adds_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
730
724
|
{
|
|
731
725
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -733,7 +727,7 @@ _mm256_maskz_adds_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
733
727
|
(__v16hi)_mm256_setzero_si256());
|
|
734
728
|
}
|
|
735
729
|
|
|
736
|
-
static __inline__ __m128i
|
|
730
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
737
731
|
_mm_mask_adds_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
738
732
|
{
|
|
739
733
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -741,7 +735,7 @@ _mm_mask_adds_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
|
741
735
|
(__v16qi)__W);
|
|
742
736
|
}
|
|
743
737
|
|
|
744
|
-
static __inline__ __m128i
|
|
738
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
745
739
|
_mm_maskz_adds_epu8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
746
740
|
{
|
|
747
741
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -749,7 +743,7 @@ _mm_maskz_adds_epu8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
|
749
743
|
(__v16qi)_mm_setzero_si128());
|
|
750
744
|
}
|
|
751
745
|
|
|
752
|
-
static __inline__ __m256i
|
|
746
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
753
747
|
_mm256_mask_adds_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
754
748
|
{
|
|
755
749
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -757,7 +751,7 @@ _mm256_mask_adds_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
|
757
751
|
(__v32qi)__W);
|
|
758
752
|
}
|
|
759
753
|
|
|
760
|
-
static __inline__ __m256i
|
|
754
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
761
755
|
_mm256_maskz_adds_epu8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
762
756
|
{
|
|
763
757
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -765,7 +759,7 @@ _mm256_maskz_adds_epu8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
|
765
759
|
(__v32qi)_mm256_setzero_si256());
|
|
766
760
|
}
|
|
767
761
|
|
|
768
|
-
static __inline__ __m128i
|
|
762
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
769
763
|
_mm_mask_adds_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
770
764
|
{
|
|
771
765
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -773,7 +767,7 @@ _mm_mask_adds_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
773
767
|
(__v8hi)__W);
|
|
774
768
|
}
|
|
775
769
|
|
|
776
|
-
static __inline__ __m128i
|
|
770
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
777
771
|
_mm_maskz_adds_epu16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
778
772
|
{
|
|
779
773
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -781,7 +775,7 @@ _mm_maskz_adds_epu16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
781
775
|
(__v8hi)_mm_setzero_si128());
|
|
782
776
|
}
|
|
783
777
|
|
|
784
|
-
static __inline__ __m256i
|
|
778
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
785
779
|
_mm256_mask_adds_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
786
780
|
{
|
|
787
781
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -789,7 +783,7 @@ _mm256_mask_adds_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
|
789
783
|
(__v16hi)__W);
|
|
790
784
|
}
|
|
791
785
|
|
|
792
|
-
static __inline__ __m256i
|
|
786
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
793
787
|
_mm256_maskz_adds_epu16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
794
788
|
{
|
|
795
789
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -797,359 +791,311 @@ _mm256_maskz_adds_epu16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
797
791
|
(__v16hi)_mm256_setzero_si256());
|
|
798
792
|
}
|
|
799
793
|
|
|
800
|
-
static __inline__ __m128i
|
|
801
|
-
_mm_mask_avg_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
802
|
-
|
|
803
|
-
|
|
804
|
-
(__v16qi)_mm_avg_epu8(__A, __B),
|
|
805
|
-
(__v16qi)__W);
|
|
794
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
795
|
+
_mm_mask_avg_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
|
|
796
|
+
return (__m128i)__builtin_ia32_selectb_128(
|
|
797
|
+
(__mmask16)__U, (__v16qi)_mm_avg_epu8(__A, __B), (__v16qi)__W);
|
|
806
798
|
}
|
|
807
799
|
|
|
808
|
-
static __inline__ __m128i
|
|
809
|
-
_mm_maskz_avg_epu8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
810
|
-
{
|
|
800
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
801
|
+
_mm_maskz_avg_epu8(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
811
802
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
812
803
|
(__v16qi)_mm_avg_epu8(__A, __B),
|
|
813
804
|
(__v16qi)_mm_setzero_si128());
|
|
814
805
|
}
|
|
815
806
|
|
|
816
|
-
static __inline__ __m256i
|
|
817
|
-
_mm256_mask_avg_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
818
|
-
|
|
819
|
-
|
|
820
|
-
(__v32qi)_mm256_avg_epu8(__A, __B),
|
|
821
|
-
(__v32qi)__W);
|
|
807
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
808
|
+
_mm256_mask_avg_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
|
|
809
|
+
return (__m256i)__builtin_ia32_selectb_256(
|
|
810
|
+
(__mmask32)__U, (__v32qi)_mm256_avg_epu8(__A, __B), (__v32qi)__W);
|
|
822
811
|
}
|
|
823
812
|
|
|
824
|
-
static __inline__ __m256i
|
|
825
|
-
_mm256_maskz_avg_epu8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
826
|
-
{
|
|
813
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
814
|
+
_mm256_maskz_avg_epu8(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
827
815
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
828
816
|
(__v32qi)_mm256_avg_epu8(__A, __B),
|
|
829
817
|
(__v32qi)_mm256_setzero_si256());
|
|
830
818
|
}
|
|
831
819
|
|
|
832
|
-
static __inline__ __m128i
|
|
833
|
-
_mm_mask_avg_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
834
|
-
|
|
835
|
-
|
|
836
|
-
(__v8hi)_mm_avg_epu16(__A, __B),
|
|
837
|
-
(__v8hi)__W);
|
|
820
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
821
|
+
_mm_mask_avg_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
822
|
+
return (__m128i)__builtin_ia32_selectw_128(
|
|
823
|
+
(__mmask8)__U, (__v8hi)_mm_avg_epu16(__A, __B), (__v8hi)__W);
|
|
838
824
|
}
|
|
839
825
|
|
|
840
|
-
static __inline__ __m128i
|
|
841
|
-
_mm_maskz_avg_epu16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
842
|
-
{
|
|
826
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
827
|
+
_mm_maskz_avg_epu16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
843
828
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
844
829
|
(__v8hi)_mm_avg_epu16(__A, __B),
|
|
845
830
|
(__v8hi)_mm_setzero_si128());
|
|
846
831
|
}
|
|
847
832
|
|
|
848
|
-
static __inline__ __m256i
|
|
849
|
-
_mm256_mask_avg_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
850
|
-
|
|
851
|
-
|
|
852
|
-
(__v16hi)_mm256_avg_epu16(__A, __B),
|
|
853
|
-
(__v16hi)__W);
|
|
833
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
834
|
+
_mm256_mask_avg_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
835
|
+
return (__m256i)__builtin_ia32_selectw_256(
|
|
836
|
+
(__mmask16)__U, (__v16hi)_mm256_avg_epu16(__A, __B), (__v16hi)__W);
|
|
854
837
|
}
|
|
855
838
|
|
|
856
|
-
static __inline__ __m256i
|
|
857
|
-
_mm256_maskz_avg_epu16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
858
|
-
|
|
859
|
-
|
|
860
|
-
|
|
861
|
-
(__v16hi)_mm256_setzero_si256());
|
|
839
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
840
|
+
_mm256_maskz_avg_epu16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
841
|
+
return (__m256i)__builtin_ia32_selectw_256(
|
|
842
|
+
(__mmask16)__U, (__v16hi)_mm256_avg_epu16(__A, __B),
|
|
843
|
+
(__v16hi)_mm256_setzero_si256());
|
|
862
844
|
}
|
|
863
845
|
|
|
864
|
-
static __inline__ __m128i
|
|
865
|
-
_mm_maskz_max_epi8(__mmask16 __M, __m128i __A, __m128i __B)
|
|
866
|
-
{
|
|
846
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
847
|
+
_mm_maskz_max_epi8(__mmask16 __M, __m128i __A, __m128i __B) {
|
|
867
848
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
868
849
|
(__v16qi)_mm_max_epi8(__A, __B),
|
|
869
850
|
(__v16qi)_mm_setzero_si128());
|
|
870
851
|
}
|
|
871
852
|
|
|
872
|
-
static __inline__ __m128i
|
|
873
|
-
_mm_mask_max_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
874
|
-
{
|
|
853
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
854
|
+
_mm_mask_max_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) {
|
|
875
855
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
876
856
|
(__v16qi)_mm_max_epi8(__A, __B),
|
|
877
857
|
(__v16qi)__W);
|
|
878
858
|
}
|
|
879
859
|
|
|
880
|
-
static __inline__ __m256i
|
|
881
|
-
_mm256_maskz_max_epi8(__mmask32 __M, __m256i __A, __m256i __B)
|
|
882
|
-
{
|
|
860
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
861
|
+
_mm256_maskz_max_epi8(__mmask32 __M, __m256i __A, __m256i __B) {
|
|
883
862
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
884
863
|
(__v32qi)_mm256_max_epi8(__A, __B),
|
|
885
864
|
(__v32qi)_mm256_setzero_si256());
|
|
886
865
|
}
|
|
887
866
|
|
|
888
|
-
static __inline__ __m256i
|
|
889
|
-
_mm256_mask_max_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
890
|
-
{
|
|
867
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
868
|
+
_mm256_mask_max_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) {
|
|
891
869
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
892
870
|
(__v32qi)_mm256_max_epi8(__A, __B),
|
|
893
871
|
(__v32qi)__W);
|
|
894
872
|
}
|
|
895
873
|
|
|
896
|
-
static __inline__ __m128i
|
|
897
|
-
_mm_maskz_max_epi16(__mmask8 __M, __m128i __A, __m128i __B)
|
|
898
|
-
{
|
|
874
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
875
|
+
_mm_maskz_max_epi16(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
899
876
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
900
877
|
(__v8hi)_mm_max_epi16(__A, __B),
|
|
901
878
|
(__v8hi)_mm_setzero_si128());
|
|
902
879
|
}
|
|
903
880
|
|
|
904
|
-
static __inline__ __m128i
|
|
905
|
-
_mm_mask_max_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
906
|
-
{
|
|
881
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
882
|
+
_mm_mask_max_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
907
883
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
908
884
|
(__v8hi)_mm_max_epi16(__A, __B),
|
|
909
885
|
(__v8hi)__W);
|
|
910
886
|
}
|
|
911
887
|
|
|
912
|
-
static __inline__ __m256i
|
|
913
|
-
_mm256_maskz_max_epi16(__mmask16 __M, __m256i __A, __m256i __B)
|
|
914
|
-
{
|
|
888
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
889
|
+
_mm256_maskz_max_epi16(__mmask16 __M, __m256i __A, __m256i __B) {
|
|
915
890
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
916
891
|
(__v16hi)_mm256_max_epi16(__A, __B),
|
|
917
892
|
(__v16hi)_mm256_setzero_si256());
|
|
918
893
|
}
|
|
919
894
|
|
|
920
|
-
static __inline__ __m256i
|
|
921
|
-
_mm256_mask_max_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
922
|
-
{
|
|
895
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
896
|
+
_mm256_mask_max_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) {
|
|
923
897
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
924
898
|
(__v16hi)_mm256_max_epi16(__A, __B),
|
|
925
899
|
(__v16hi)__W);
|
|
926
900
|
}
|
|
927
901
|
|
|
928
|
-
static __inline__ __m128i
|
|
929
|
-
_mm_maskz_max_epu8(__mmask16 __M, __m128i __A, __m128i __B)
|
|
930
|
-
{
|
|
902
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
903
|
+
_mm_maskz_max_epu8(__mmask16 __M, __m128i __A, __m128i __B) {
|
|
931
904
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
932
905
|
(__v16qi)_mm_max_epu8(__A, __B),
|
|
933
906
|
(__v16qi)_mm_setzero_si128());
|
|
934
907
|
}
|
|
935
908
|
|
|
936
|
-
static __inline__ __m128i
|
|
937
|
-
_mm_mask_max_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
938
|
-
{
|
|
909
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
910
|
+
_mm_mask_max_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) {
|
|
939
911
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
940
912
|
(__v16qi)_mm_max_epu8(__A, __B),
|
|
941
913
|
(__v16qi)__W);
|
|
942
914
|
}
|
|
943
915
|
|
|
944
|
-
static __inline__ __m256i
|
|
945
|
-
_mm256_maskz_max_epu8
|
|
946
|
-
{
|
|
916
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
917
|
+
_mm256_maskz_max_epu8(__mmask32 __M, __m256i __A, __m256i __B) {
|
|
947
918
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
948
919
|
(__v32qi)_mm256_max_epu8(__A, __B),
|
|
949
920
|
(__v32qi)_mm256_setzero_si256());
|
|
950
921
|
}
|
|
951
922
|
|
|
952
|
-
static __inline__ __m256i
|
|
953
|
-
_mm256_mask_max_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
954
|
-
{
|
|
923
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
924
|
+
_mm256_mask_max_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) {
|
|
955
925
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
956
926
|
(__v32qi)_mm256_max_epu8(__A, __B),
|
|
957
927
|
(__v32qi)__W);
|
|
958
928
|
}
|
|
959
929
|
|
|
960
|
-
static __inline__ __m128i
|
|
961
|
-
_mm_maskz_max_epu16(__mmask8 __M, __m128i __A, __m128i __B)
|
|
962
|
-
{
|
|
930
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
931
|
+
_mm_maskz_max_epu16(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
963
932
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
964
933
|
(__v8hi)_mm_max_epu16(__A, __B),
|
|
965
934
|
(__v8hi)_mm_setzero_si128());
|
|
966
935
|
}
|
|
967
936
|
|
|
968
|
-
static __inline__ __m128i
|
|
969
|
-
_mm_mask_max_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
970
|
-
{
|
|
937
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
938
|
+
_mm_mask_max_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
971
939
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
972
940
|
(__v8hi)_mm_max_epu16(__A, __B),
|
|
973
941
|
(__v8hi)__W);
|
|
974
942
|
}
|
|
975
943
|
|
|
976
|
-
static __inline__ __m256i
|
|
977
|
-
_mm256_maskz_max_epu16(__mmask16 __M, __m256i __A, __m256i __B)
|
|
978
|
-
{
|
|
944
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
945
|
+
_mm256_maskz_max_epu16(__mmask16 __M, __m256i __A, __m256i __B) {
|
|
979
946
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
980
947
|
(__v16hi)_mm256_max_epu16(__A, __B),
|
|
981
948
|
(__v16hi)_mm256_setzero_si256());
|
|
982
949
|
}
|
|
983
950
|
|
|
984
|
-
static __inline__ __m256i
|
|
985
|
-
_mm256_mask_max_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
986
|
-
{
|
|
951
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
952
|
+
_mm256_mask_max_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) {
|
|
987
953
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
988
954
|
(__v16hi)_mm256_max_epu16(__A, __B),
|
|
989
955
|
(__v16hi)__W);
|
|
990
956
|
}
|
|
991
957
|
|
|
992
|
-
static __inline__ __m128i
|
|
993
|
-
_mm_maskz_min_epi8(__mmask16 __M, __m128i __A, __m128i __B)
|
|
994
|
-
{
|
|
958
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
959
|
+
_mm_maskz_min_epi8(__mmask16 __M, __m128i __A, __m128i __B) {
|
|
995
960
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
996
961
|
(__v16qi)_mm_min_epi8(__A, __B),
|
|
997
962
|
(__v16qi)_mm_setzero_si128());
|
|
998
963
|
}
|
|
999
964
|
|
|
1000
|
-
static __inline__ __m128i
|
|
1001
|
-
_mm_mask_min_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
1002
|
-
{
|
|
965
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
966
|
+
_mm_mask_min_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) {
|
|
1003
967
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
1004
968
|
(__v16qi)_mm_min_epi8(__A, __B),
|
|
1005
969
|
(__v16qi)__W);
|
|
1006
970
|
}
|
|
1007
971
|
|
|
1008
|
-
static __inline__ __m256i
|
|
1009
|
-
_mm256_maskz_min_epi8(__mmask32 __M, __m256i __A, __m256i __B)
|
|
1010
|
-
{
|
|
972
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
973
|
+
_mm256_maskz_min_epi8(__mmask32 __M, __m256i __A, __m256i __B) {
|
|
1011
974
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
1012
975
|
(__v32qi)_mm256_min_epi8(__A, __B),
|
|
1013
976
|
(__v32qi)_mm256_setzero_si256());
|
|
1014
977
|
}
|
|
1015
978
|
|
|
1016
|
-
static __inline__ __m256i
|
|
1017
|
-
_mm256_mask_min_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
1018
|
-
{
|
|
979
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
980
|
+
_mm256_mask_min_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) {
|
|
1019
981
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
1020
982
|
(__v32qi)_mm256_min_epi8(__A, __B),
|
|
1021
983
|
(__v32qi)__W);
|
|
1022
984
|
}
|
|
1023
985
|
|
|
1024
|
-
static __inline__ __m128i
|
|
1025
|
-
_mm_maskz_min_epi16(__mmask8 __M, __m128i __A, __m128i __B)
|
|
1026
|
-
{
|
|
986
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
987
|
+
_mm_maskz_min_epi16(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
1027
988
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
1028
989
|
(__v8hi)_mm_min_epi16(__A, __B),
|
|
1029
990
|
(__v8hi)_mm_setzero_si128());
|
|
1030
991
|
}
|
|
1031
992
|
|
|
1032
|
-
static __inline__ __m128i
|
|
1033
|
-
_mm_mask_min_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
1034
|
-
{
|
|
993
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
994
|
+
_mm_mask_min_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
1035
995
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
1036
996
|
(__v8hi)_mm_min_epi16(__A, __B),
|
|
1037
997
|
(__v8hi)__W);
|
|
1038
998
|
}
|
|
1039
999
|
|
|
1040
|
-
static __inline__ __m256i
|
|
1041
|
-
_mm256_maskz_min_epi16(__mmask16 __M, __m256i __A, __m256i __B)
|
|
1042
|
-
{
|
|
1000
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1001
|
+
_mm256_maskz_min_epi16(__mmask16 __M, __m256i __A, __m256i __B) {
|
|
1043
1002
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
1044
1003
|
(__v16hi)_mm256_min_epi16(__A, __B),
|
|
1045
1004
|
(__v16hi)_mm256_setzero_si256());
|
|
1046
1005
|
}
|
|
1047
1006
|
|
|
1048
|
-
static __inline__ __m256i
|
|
1049
|
-
_mm256_mask_min_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
1050
|
-
{
|
|
1007
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1008
|
+
_mm256_mask_min_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) {
|
|
1051
1009
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
1052
1010
|
(__v16hi)_mm256_min_epi16(__A, __B),
|
|
1053
1011
|
(__v16hi)__W);
|
|
1054
1012
|
}
|
|
1055
1013
|
|
|
1056
|
-
static __inline__ __m128i
|
|
1057
|
-
_mm_maskz_min_epu8(__mmask16 __M, __m128i __A, __m128i __B)
|
|
1058
|
-
{
|
|
1014
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1015
|
+
_mm_maskz_min_epu8(__mmask16 __M, __m128i __A, __m128i __B) {
|
|
1059
1016
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
1060
1017
|
(__v16qi)_mm_min_epu8(__A, __B),
|
|
1061
1018
|
(__v16qi)_mm_setzero_si128());
|
|
1062
1019
|
}
|
|
1063
1020
|
|
|
1064
|
-
static __inline__ __m128i
|
|
1065
|
-
_mm_mask_min_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
|
|
1066
|
-
{
|
|
1021
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1022
|
+
_mm_mask_min_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) {
|
|
1067
1023
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
1068
1024
|
(__v16qi)_mm_min_epu8(__A, __B),
|
|
1069
1025
|
(__v16qi)__W);
|
|
1070
1026
|
}
|
|
1071
1027
|
|
|
1072
|
-
static __inline__ __m256i
|
|
1073
|
-
_mm256_maskz_min_epu8
|
|
1074
|
-
{
|
|
1028
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1029
|
+
_mm256_maskz_min_epu8(__mmask32 __M, __m256i __A, __m256i __B) {
|
|
1075
1030
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
1076
1031
|
(__v32qi)_mm256_min_epu8(__A, __B),
|
|
1077
1032
|
(__v32qi)_mm256_setzero_si256());
|
|
1078
1033
|
}
|
|
1079
1034
|
|
|
1080
|
-
static __inline__ __m256i
|
|
1081
|
-
_mm256_mask_min_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
|
|
1082
|
-
{
|
|
1035
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1036
|
+
_mm256_mask_min_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) {
|
|
1083
1037
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
|
|
1084
1038
|
(__v32qi)_mm256_min_epu8(__A, __B),
|
|
1085
1039
|
(__v32qi)__W);
|
|
1086
1040
|
}
|
|
1087
1041
|
|
|
1088
|
-
static __inline__ __m128i
|
|
1089
|
-
_mm_maskz_min_epu16(__mmask8 __M, __m128i __A, __m128i __B)
|
|
1090
|
-
{
|
|
1042
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1043
|
+
_mm_maskz_min_epu16(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
1091
1044
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
1092
1045
|
(__v8hi)_mm_min_epu16(__A, __B),
|
|
1093
1046
|
(__v8hi)_mm_setzero_si128());
|
|
1094
1047
|
}
|
|
1095
1048
|
|
|
1096
|
-
static __inline__ __m128i
|
|
1097
|
-
_mm_mask_min_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
1098
|
-
{
|
|
1049
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1050
|
+
_mm_mask_min_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
1099
1051
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
1100
1052
|
(__v8hi)_mm_min_epu16(__A, __B),
|
|
1101
1053
|
(__v8hi)__W);
|
|
1102
1054
|
}
|
|
1103
1055
|
|
|
1104
|
-
static __inline__ __m256i
|
|
1105
|
-
_mm256_maskz_min_epu16(__mmask16 __M, __m256i __A, __m256i __B)
|
|
1106
|
-
{
|
|
1056
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1057
|
+
_mm256_maskz_min_epu16(__mmask16 __M, __m256i __A, __m256i __B) {
|
|
1107
1058
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
1108
1059
|
(__v16hi)_mm256_min_epu16(__A, __B),
|
|
1109
1060
|
(__v16hi)_mm256_setzero_si256());
|
|
1110
1061
|
}
|
|
1111
1062
|
|
|
1112
|
-
static __inline__ __m256i
|
|
1113
|
-
_mm256_mask_min_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
|
|
1114
|
-
{
|
|
1063
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1064
|
+
_mm256_mask_min_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) {
|
|
1115
1065
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
1116
1066
|
(__v16hi)_mm256_min_epu16(__A, __B),
|
|
1117
1067
|
(__v16hi)__W);
|
|
1118
1068
|
}
|
|
1119
1069
|
|
|
1120
|
-
static __inline__ __m128i
|
|
1121
|
-
_mm_mask_shuffle_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
1122
|
-
{
|
|
1070
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1071
|
+
_mm_mask_shuffle_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
|
|
1123
1072
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
1124
1073
|
(__v16qi)_mm_shuffle_epi8(__A, __B),
|
|
1125
1074
|
(__v16qi)__W);
|
|
1126
1075
|
}
|
|
1127
1076
|
|
|
1128
|
-
static __inline__ __m128i
|
|
1129
|
-
_mm_maskz_shuffle_epi8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
1130
|
-
{
|
|
1077
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1078
|
+
_mm_maskz_shuffle_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
1131
1079
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
1132
1080
|
(__v16qi)_mm_shuffle_epi8(__A, __B),
|
|
1133
1081
|
(__v16qi)_mm_setzero_si128());
|
|
1134
1082
|
}
|
|
1135
1083
|
|
|
1136
|
-
static __inline__ __m256i
|
|
1137
|
-
_mm256_mask_shuffle_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
1138
|
-
{
|
|
1084
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1085
|
+
_mm256_mask_shuffle_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
|
|
1139
1086
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
1140
1087
|
(__v32qi)_mm256_shuffle_epi8(__A, __B),
|
|
1141
1088
|
(__v32qi)__W);
|
|
1142
1089
|
}
|
|
1143
1090
|
|
|
1144
|
-
static __inline__ __m256i
|
|
1145
|
-
_mm256_maskz_shuffle_epi8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
1146
|
-
{
|
|
1091
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1092
|
+
_mm256_maskz_shuffle_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
1147
1093
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
1148
1094
|
(__v32qi)_mm256_shuffle_epi8(__A, __B),
|
|
1149
1095
|
(__v32qi)_mm256_setzero_si256());
|
|
1150
1096
|
}
|
|
1151
1097
|
|
|
1152
|
-
static __inline__ __m128i
|
|
1098
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1153
1099
|
_mm_mask_subs_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
1154
1100
|
{
|
|
1155
1101
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -1157,7 +1103,7 @@ _mm_mask_subs_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
|
1157
1103
|
(__v16qi)__W);
|
|
1158
1104
|
}
|
|
1159
1105
|
|
|
1160
|
-
static __inline__ __m128i
|
|
1106
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1161
1107
|
_mm_maskz_subs_epi8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
1162
1108
|
{
|
|
1163
1109
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -1165,7 +1111,7 @@ _mm_maskz_subs_epi8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
|
1165
1111
|
(__v16qi)_mm_setzero_si128());
|
|
1166
1112
|
}
|
|
1167
1113
|
|
|
1168
|
-
static __inline__ __m256i
|
|
1114
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1169
1115
|
_mm256_mask_subs_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
1170
1116
|
{
|
|
1171
1117
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -1173,7 +1119,7 @@ _mm256_mask_subs_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
|
1173
1119
|
(__v32qi)__W);
|
|
1174
1120
|
}
|
|
1175
1121
|
|
|
1176
|
-
static __inline__ __m256i
|
|
1122
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1177
1123
|
_mm256_maskz_subs_epi8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
1178
1124
|
{
|
|
1179
1125
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -1181,7 +1127,7 @@ _mm256_maskz_subs_epi8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
|
1181
1127
|
(__v32qi)_mm256_setzero_si256());
|
|
1182
1128
|
}
|
|
1183
1129
|
|
|
1184
|
-
static __inline__ __m128i
|
|
1130
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1185
1131
|
_mm_mask_subs_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
1186
1132
|
{
|
|
1187
1133
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1189,7 +1135,7 @@ _mm_mask_subs_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
1189
1135
|
(__v8hi)__W);
|
|
1190
1136
|
}
|
|
1191
1137
|
|
|
1192
|
-
static __inline__ __m128i
|
|
1138
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1193
1139
|
_mm_maskz_subs_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
1194
1140
|
{
|
|
1195
1141
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1197,7 +1143,7 @@ _mm_maskz_subs_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
1197
1143
|
(__v8hi)_mm_setzero_si128());
|
|
1198
1144
|
}
|
|
1199
1145
|
|
|
1200
|
-
static __inline__ __m256i
|
|
1146
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1201
1147
|
_mm256_mask_subs_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
1202
1148
|
{
|
|
1203
1149
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1205,7 +1151,7 @@ _mm256_mask_subs_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1205
1151
|
(__v16hi)__W);
|
|
1206
1152
|
}
|
|
1207
1153
|
|
|
1208
|
-
static __inline__ __m256i
|
|
1154
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1209
1155
|
_mm256_maskz_subs_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
1210
1156
|
{
|
|
1211
1157
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1213,7 +1159,7 @@ _mm256_maskz_subs_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1213
1159
|
(__v16hi)_mm256_setzero_si256());
|
|
1214
1160
|
}
|
|
1215
1161
|
|
|
1216
|
-
static __inline__ __m128i
|
|
1162
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1217
1163
|
_mm_mask_subs_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
1218
1164
|
{
|
|
1219
1165
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -1221,7 +1167,7 @@ _mm_mask_subs_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
|
|
|
1221
1167
|
(__v16qi)__W);
|
|
1222
1168
|
}
|
|
1223
1169
|
|
|
1224
|
-
static __inline__ __m128i
|
|
1170
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1225
1171
|
_mm_maskz_subs_epu8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
1226
1172
|
{
|
|
1227
1173
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
@@ -1229,7 +1175,7 @@ _mm_maskz_subs_epu8(__mmask16 __U, __m128i __A, __m128i __B)
|
|
|
1229
1175
|
(__v16qi)_mm_setzero_si128());
|
|
1230
1176
|
}
|
|
1231
1177
|
|
|
1232
|
-
static __inline__ __m256i
|
|
1178
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1233
1179
|
_mm256_mask_subs_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
1234
1180
|
{
|
|
1235
1181
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -1237,7 +1183,7 @@ _mm256_mask_subs_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
|
|
|
1237
1183
|
(__v32qi)__W);
|
|
1238
1184
|
}
|
|
1239
1185
|
|
|
1240
|
-
static __inline__ __m256i
|
|
1186
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1241
1187
|
_mm256_maskz_subs_epu8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
1242
1188
|
{
|
|
1243
1189
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
@@ -1245,7 +1191,7 @@ _mm256_maskz_subs_epu8(__mmask32 __U, __m256i __A, __m256i __B)
|
|
|
1245
1191
|
(__v32qi)_mm256_setzero_si256());
|
|
1246
1192
|
}
|
|
1247
1193
|
|
|
1248
|
-
static __inline__ __m128i
|
|
1194
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1249
1195
|
_mm_mask_subs_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
1250
1196
|
{
|
|
1251
1197
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1253,7 +1199,7 @@ _mm_mask_subs_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
1253
1199
|
(__v8hi)__W);
|
|
1254
1200
|
}
|
|
1255
1201
|
|
|
1256
|
-
static __inline__ __m128i
|
|
1202
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1257
1203
|
_mm_maskz_subs_epu16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
1258
1204
|
{
|
|
1259
1205
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1261,7 +1207,7 @@ _mm_maskz_subs_epu16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
1261
1207
|
(__v8hi)_mm_setzero_si128());
|
|
1262
1208
|
}
|
|
1263
1209
|
|
|
1264
|
-
static __inline__ __m256i
|
|
1210
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1265
1211
|
_mm256_mask_subs_epu16(__m256i __W, __mmask16 __U, __m256i __A,
|
|
1266
1212
|
__m256i __B) {
|
|
1267
1213
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1269,7 +1215,7 @@ _mm256_mask_subs_epu16(__m256i __W, __mmask16 __U, __m256i __A,
|
|
|
1269
1215
|
(__v16hi)__W);
|
|
1270
1216
|
}
|
|
1271
1217
|
|
|
1272
|
-
static __inline__ __m256i
|
|
1218
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1273
1219
|
_mm256_maskz_subs_epu16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
1274
1220
|
{
|
|
1275
1221
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1277,89 +1223,81 @@ _mm256_maskz_subs_epu16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1277
1223
|
(__v16hi)_mm256_setzero_si256());
|
|
1278
1224
|
}
|
|
1279
1225
|
|
|
1280
|
-
static __inline__ __m128i
|
|
1281
|
-
_mm_permutex2var_epi16(__m128i __A, __m128i __I, __m128i __B)
|
|
1282
|
-
{
|
|
1226
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1227
|
+
_mm_permutex2var_epi16(__m128i __A, __m128i __I, __m128i __B) {
|
|
1283
1228
|
return (__m128i)__builtin_ia32_vpermi2varhi128((__v8hi)__A, (__v8hi)__I,
|
|
1284
1229
|
(__v8hi) __B);
|
|
1285
1230
|
}
|
|
1286
1231
|
|
|
1287
|
-
static __inline__ __m128i
|
|
1232
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1288
1233
|
_mm_mask_permutex2var_epi16(__m128i __A, __mmask8 __U, __m128i __I,
|
|
1289
|
-
__m128i __B)
|
|
1290
|
-
{
|
|
1234
|
+
__m128i __B) {
|
|
1291
1235
|
return (__m128i)__builtin_ia32_selectw_128(__U,
|
|
1292
1236
|
(__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
|
|
1293
1237
|
(__v8hi)__A);
|
|
1294
1238
|
}
|
|
1295
1239
|
|
|
1296
|
-
static __inline__ __m128i
|
|
1240
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1297
1241
|
_mm_mask2_permutex2var_epi16(__m128i __A, __m128i __I, __mmask8 __U,
|
|
1298
|
-
__m128i __B)
|
|
1299
|
-
{
|
|
1242
|
+
__m128i __B) {
|
|
1300
1243
|
return (__m128i)__builtin_ia32_selectw_128(__U,
|
|
1301
1244
|
(__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
|
|
1302
1245
|
(__v8hi)__I);
|
|
1303
1246
|
}
|
|
1304
1247
|
|
|
1305
|
-
static __inline__ __m128i
|
|
1306
|
-
_mm_maskz_permutex2var_epi16
|
|
1307
|
-
|
|
1308
|
-
{
|
|
1248
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1249
|
+
_mm_maskz_permutex2var_epi16(__mmask8 __U, __m128i __A, __m128i __I,
|
|
1250
|
+
__m128i __B) {
|
|
1309
1251
|
return (__m128i)__builtin_ia32_selectw_128(__U,
|
|
1310
1252
|
(__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
|
|
1311
1253
|
(__v8hi)_mm_setzero_si128());
|
|
1312
1254
|
}
|
|
1313
1255
|
|
|
1314
|
-
static __inline__ __m256i
|
|
1315
|
-
_mm256_permutex2var_epi16(__m256i __A, __m256i __I, __m256i __B)
|
|
1316
|
-
{
|
|
1256
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1257
|
+
_mm256_permutex2var_epi16(__m256i __A, __m256i __I, __m256i __B) {
|
|
1317
1258
|
return (__m256i)__builtin_ia32_vpermi2varhi256((__v16hi)__A, (__v16hi)__I,
|
|
1318
1259
|
(__v16hi)__B);
|
|
1319
1260
|
}
|
|
1320
1261
|
|
|
1321
|
-
static __inline__ __m256i
|
|
1262
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1322
1263
|
_mm256_mask_permutex2var_epi16(__m256i __A, __mmask16 __U, __m256i __I,
|
|
1323
|
-
__m256i __B)
|
|
1324
|
-
{
|
|
1264
|
+
__m256i __B) {
|
|
1325
1265
|
return (__m256i)__builtin_ia32_selectw_256(__U,
|
|
1326
1266
|
(__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
|
|
1327
1267
|
(__v16hi)__A);
|
|
1328
1268
|
}
|
|
1329
1269
|
|
|
1330
|
-
static __inline__ __m256i
|
|
1270
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1331
1271
|
_mm256_mask2_permutex2var_epi16(__m256i __A, __m256i __I, __mmask16 __U,
|
|
1332
|
-
__m256i __B)
|
|
1333
|
-
{
|
|
1272
|
+
__m256i __B) {
|
|
1334
1273
|
return (__m256i)__builtin_ia32_selectw_256(__U,
|
|
1335
1274
|
(__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
|
|
1336
1275
|
(__v16hi)__I);
|
|
1337
1276
|
}
|
|
1338
1277
|
|
|
1339
|
-
static __inline__ __m256i
|
|
1340
|
-
_mm256_maskz_permutex2var_epi16
|
|
1341
|
-
|
|
1342
|
-
{
|
|
1278
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1279
|
+
_mm256_maskz_permutex2var_epi16(__mmask16 __U, __m256i __A, __m256i __I,
|
|
1280
|
+
__m256i __B) {
|
|
1343
1281
|
return (__m256i)__builtin_ia32_selectw_256(__U,
|
|
1344
1282
|
(__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
|
|
1345
1283
|
(__v16hi)_mm256_setzero_si256());
|
|
1346
1284
|
}
|
|
1347
1285
|
|
|
1348
|
-
static __inline__ __m128i
|
|
1286
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1349
1287
|
_mm_mask_maddubs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) {
|
|
1350
1288
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1351
1289
|
(__v8hi)_mm_maddubs_epi16(__X, __Y),
|
|
1352
1290
|
(__v8hi)__W);
|
|
1353
1291
|
}
|
|
1354
1292
|
|
|
1355
|
-
static __inline__ __m128i
|
|
1293
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1356
1294
|
_mm_maskz_maddubs_epi16(__mmask8 __U, __m128i __X, __m128i __Y) {
|
|
1357
1295
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1358
1296
|
(__v8hi)_mm_maddubs_epi16(__X, __Y),
|
|
1359
1297
|
(__v8hi)_mm_setzero_si128());
|
|
1360
1298
|
}
|
|
1361
1299
|
|
|
1362
|
-
static __inline__ __m256i
|
|
1300
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1363
1301
|
_mm256_mask_maddubs_epi16(__m256i __W, __mmask16 __U, __m256i __X,
|
|
1364
1302
|
__m256i __Y) {
|
|
1365
1303
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1367,35 +1305,35 @@ _mm256_mask_maddubs_epi16(__m256i __W, __mmask16 __U, __m256i __X,
|
|
|
1367
1305
|
(__v16hi)__W);
|
|
1368
1306
|
}
|
|
1369
1307
|
|
|
1370
|
-
static __inline__ __m256i
|
|
1308
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1371
1309
|
_mm256_maskz_maddubs_epi16(__mmask16 __U, __m256i __X, __m256i __Y) {
|
|
1372
1310
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1373
1311
|
(__v16hi)_mm256_maddubs_epi16(__X, __Y),
|
|
1374
1312
|
(__v16hi)_mm256_setzero_si256());
|
|
1375
1313
|
}
|
|
1376
1314
|
|
|
1377
|
-
static __inline__ __m128i
|
|
1315
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1378
1316
|
_mm_mask_madd_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
1379
1317
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
1380
1318
|
(__v4si)_mm_madd_epi16(__A, __B),
|
|
1381
1319
|
(__v4si)__W);
|
|
1382
1320
|
}
|
|
1383
1321
|
|
|
1384
|
-
static __inline__ __m128i
|
|
1322
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1385
1323
|
_mm_maskz_madd_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
1386
1324
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
1387
1325
|
(__v4si)_mm_madd_epi16(__A, __B),
|
|
1388
1326
|
(__v4si)_mm_setzero_si128());
|
|
1389
1327
|
}
|
|
1390
1328
|
|
|
1391
|
-
static __inline__ __m256i
|
|
1329
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1392
1330
|
_mm256_mask_madd_epi16(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
|
|
1393
1331
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
1394
1332
|
(__v8si)_mm256_madd_epi16(__A, __B),
|
|
1395
1333
|
(__v8si)__W);
|
|
1396
1334
|
}
|
|
1397
1335
|
|
|
1398
|
-
static __inline__ __m256i
|
|
1336
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1399
1337
|
_mm256_maskz_madd_epi16(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
1400
1338
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
1401
1339
|
(__v8si)_mm256_madd_epi16(__A, __B),
|
|
@@ -1486,8 +1424,8 @@ _mm256_maskz_cvtusepi16_epi8 (__mmask16 __M, __m256i __A) {
|
|
|
1486
1424
|
__M);
|
|
1487
1425
|
}
|
|
1488
1426
|
|
|
1489
|
-
static __inline__ __m128i
|
|
1490
|
-
_mm_cvtepi16_epi8
|
|
1427
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1428
|
+
_mm_cvtepi16_epi8(__m128i __A) {
|
|
1491
1429
|
return (__m128i)__builtin_shufflevector(
|
|
1492
1430
|
__builtin_convertvector((__v8hi)__A, __v8qi),
|
|
1493
1431
|
(__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
|
|
@@ -1527,20 +1465,20 @@ _mm_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
|
|
|
1527
1465
|
__builtin_ia32_pmovuswb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M);
|
|
1528
1466
|
}
|
|
1529
1467
|
|
|
1530
|
-
static __inline__ __m128i
|
|
1531
|
-
_mm256_cvtepi16_epi8
|
|
1468
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1469
|
+
_mm256_cvtepi16_epi8(__m256i __A) {
|
|
1532
1470
|
return (__m128i)__builtin_convertvector((__v16hi) __A, __v16qi);
|
|
1533
1471
|
}
|
|
1534
1472
|
|
|
1535
|
-
static __inline__ __m128i
|
|
1536
|
-
_mm256_mask_cvtepi16_epi8
|
|
1473
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1474
|
+
_mm256_mask_cvtepi16_epi8(__m128i __O, __mmask16 __M, __m256i __A) {
|
|
1537
1475
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
1538
1476
|
(__v16qi)_mm256_cvtepi16_epi8(__A),
|
|
1539
1477
|
(__v16qi)__O);
|
|
1540
1478
|
}
|
|
1541
1479
|
|
|
1542
|
-
static __inline__ __m128i
|
|
1543
|
-
_mm256_maskz_cvtepi16_epi8
|
|
1480
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1481
|
+
_mm256_maskz_cvtepi16_epi8(__mmask16 __M, __m256i __A) {
|
|
1544
1482
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
|
|
1545
1483
|
(__v16qi)_mm256_cvtepi16_epi8(__A),
|
|
1546
1484
|
(__v16qi)_mm_setzero_si128());
|
|
@@ -1564,203 +1502,198 @@ _mm256_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A)
|
|
|
1564
1502
|
__builtin_ia32_pmovuswb256mem_mask ((__v16qi*) __P, (__v16hi) __A, __M);
|
|
1565
1503
|
}
|
|
1566
1504
|
|
|
1567
|
-
static __inline__ __m128i
|
|
1505
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1568
1506
|
_mm_mask_mulhrs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) {
|
|
1569
1507
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1570
1508
|
(__v8hi)_mm_mulhrs_epi16(__X, __Y),
|
|
1571
1509
|
(__v8hi)__W);
|
|
1572
1510
|
}
|
|
1573
1511
|
|
|
1574
|
-
static __inline__ __m128i
|
|
1512
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1575
1513
|
_mm_maskz_mulhrs_epi16(__mmask8 __U, __m128i __X, __m128i __Y) {
|
|
1576
1514
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1577
1515
|
(__v8hi)_mm_mulhrs_epi16(__X, __Y),
|
|
1578
1516
|
(__v8hi)_mm_setzero_si128());
|
|
1579
1517
|
}
|
|
1580
1518
|
|
|
1581
|
-
static __inline__ __m256i
|
|
1519
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1582
1520
|
_mm256_mask_mulhrs_epi16(__m256i __W, __mmask16 __U, __m256i __X, __m256i __Y) {
|
|
1583
1521
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1584
1522
|
(__v16hi)_mm256_mulhrs_epi16(__X, __Y),
|
|
1585
1523
|
(__v16hi)__W);
|
|
1586
1524
|
}
|
|
1587
1525
|
|
|
1588
|
-
static __inline__ __m256i
|
|
1526
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1589
1527
|
_mm256_maskz_mulhrs_epi16(__mmask16 __U, __m256i __X, __m256i __Y) {
|
|
1590
1528
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1591
1529
|
(__v16hi)_mm256_mulhrs_epi16(__X, __Y),
|
|
1592
1530
|
(__v16hi)_mm256_setzero_si256());
|
|
1593
1531
|
}
|
|
1594
1532
|
|
|
1595
|
-
static __inline__ __m128i
|
|
1533
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1596
1534
|
_mm_mask_mulhi_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
1597
|
-
return (__m128i)__builtin_ia32_selectw_128(
|
|
1598
|
-
|
|
1599
|
-
(__v8hi)__W);
|
|
1535
|
+
return (__m128i)__builtin_ia32_selectw_128(
|
|
1536
|
+
(__mmask8)__U, (__v8hi)_mm_mulhi_epu16(__A, __B), (__v8hi)__W);
|
|
1600
1537
|
}
|
|
1601
1538
|
|
|
1602
|
-
static __inline__ __m128i
|
|
1539
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1603
1540
|
_mm_maskz_mulhi_epu16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
1604
1541
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1605
1542
|
(__v8hi)_mm_mulhi_epu16(__A, __B),
|
|
1606
1543
|
(__v8hi)_mm_setzero_si128());
|
|
1607
1544
|
}
|
|
1608
1545
|
|
|
1609
|
-
static __inline__ __m256i
|
|
1546
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1610
1547
|
_mm256_mask_mulhi_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
1611
|
-
return (__m256i)__builtin_ia32_selectw_256(
|
|
1612
|
-
|
|
1613
|
-
(__v16hi)__W);
|
|
1548
|
+
return (__m256i)__builtin_ia32_selectw_256(
|
|
1549
|
+
(__mmask16)__U, (__v16hi)_mm256_mulhi_epu16(__A, __B), (__v16hi)__W);
|
|
1614
1550
|
}
|
|
1615
1551
|
|
|
1616
|
-
static __inline__ __m256i
|
|
1552
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1617
1553
|
_mm256_maskz_mulhi_epu16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
1618
|
-
return (__m256i)__builtin_ia32_selectw_256(
|
|
1619
|
-
|
|
1620
|
-
|
|
1554
|
+
return (__m256i)__builtin_ia32_selectw_256(
|
|
1555
|
+
(__mmask16)__U, (__v16hi)_mm256_mulhi_epu16(__A, __B),
|
|
1556
|
+
(__v16hi)_mm256_setzero_si256());
|
|
1621
1557
|
}
|
|
1622
1558
|
|
|
1623
|
-
static __inline__ __m128i
|
|
1559
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1624
1560
|
_mm_mask_mulhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
1625
|
-
return (__m128i)__builtin_ia32_selectw_128(
|
|
1626
|
-
|
|
1627
|
-
(__v8hi)__W);
|
|
1561
|
+
return (__m128i)__builtin_ia32_selectw_128(
|
|
1562
|
+
(__mmask8)__U, (__v8hi)_mm_mulhi_epi16(__A, __B), (__v8hi)__W);
|
|
1628
1563
|
}
|
|
1629
1564
|
|
|
1630
|
-
static __inline__ __m128i
|
|
1565
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1631
1566
|
_mm_maskz_mulhi_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
1632
1567
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1633
1568
|
(__v8hi)_mm_mulhi_epi16(__A, __B),
|
|
1634
1569
|
(__v8hi)_mm_setzero_si128());
|
|
1635
1570
|
}
|
|
1636
1571
|
|
|
1637
|
-
static __inline__ __m256i
|
|
1572
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1638
1573
|
_mm256_mask_mulhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
1639
|
-
return (__m256i)__builtin_ia32_selectw_256(
|
|
1640
|
-
|
|
1641
|
-
(__v16hi)__W);
|
|
1574
|
+
return (__m256i)__builtin_ia32_selectw_256(
|
|
1575
|
+
(__mmask16)__U, (__v16hi)_mm256_mulhi_epi16(__A, __B), (__v16hi)__W);
|
|
1642
1576
|
}
|
|
1643
1577
|
|
|
1644
|
-
static __inline__ __m256i
|
|
1578
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1645
1579
|
_mm256_maskz_mulhi_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
1646
|
-
return (__m256i)__builtin_ia32_selectw_256(
|
|
1647
|
-
|
|
1648
|
-
|
|
1580
|
+
return (__m256i)__builtin_ia32_selectw_256(
|
|
1581
|
+
(__mmask16)__U, (__v16hi)_mm256_mulhi_epi16(__A, __B),
|
|
1582
|
+
(__v16hi)_mm256_setzero_si256());
|
|
1649
1583
|
}
|
|
1650
1584
|
|
|
1651
|
-
static __inline__ __m128i
|
|
1585
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1652
1586
|
_mm_mask_unpackhi_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
|
|
1653
|
-
return (__m128i)__builtin_ia32_selectb_128(
|
|
1654
|
-
|
|
1655
|
-
(__v16qi)__W);
|
|
1587
|
+
return (__m128i)__builtin_ia32_selectb_128(
|
|
1588
|
+
(__mmask16)__U, (__v16qi)_mm_unpackhi_epi8(__A, __B), (__v16qi)__W);
|
|
1656
1589
|
}
|
|
1657
1590
|
|
|
1658
|
-
static __inline__ __m128i
|
|
1591
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1659
1592
|
_mm_maskz_unpackhi_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
1660
1593
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
1661
1594
|
(__v16qi)_mm_unpackhi_epi8(__A, __B),
|
|
1662
1595
|
(__v16qi)_mm_setzero_si128());
|
|
1663
1596
|
}
|
|
1664
1597
|
|
|
1665
|
-
static __inline__ __m256i
|
|
1598
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1666
1599
|
_mm256_mask_unpackhi_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
|
|
1667
1600
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
1668
1601
|
(__v32qi)_mm256_unpackhi_epi8(__A, __B),
|
|
1669
1602
|
(__v32qi)__W);
|
|
1670
1603
|
}
|
|
1671
1604
|
|
|
1672
|
-
static __inline__ __m256i
|
|
1605
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1673
1606
|
_mm256_maskz_unpackhi_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
1674
1607
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
1675
1608
|
(__v32qi)_mm256_unpackhi_epi8(__A, __B),
|
|
1676
1609
|
(__v32qi)_mm256_setzero_si256());
|
|
1677
1610
|
}
|
|
1678
1611
|
|
|
1679
|
-
static __inline__ __m128i
|
|
1612
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1680
1613
|
_mm_mask_unpackhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
1681
1614
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1682
1615
|
(__v8hi)_mm_unpackhi_epi16(__A, __B),
|
|
1683
1616
|
(__v8hi)__W);
|
|
1684
1617
|
}
|
|
1685
1618
|
|
|
1686
|
-
static __inline__ __m128i
|
|
1619
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1687
1620
|
_mm_maskz_unpackhi_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
1688
1621
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1689
1622
|
(__v8hi)_mm_unpackhi_epi16(__A, __B),
|
|
1690
1623
|
(__v8hi) _mm_setzero_si128());
|
|
1691
1624
|
}
|
|
1692
1625
|
|
|
1693
|
-
static __inline__ __m256i
|
|
1626
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1694
1627
|
_mm256_mask_unpackhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
1695
1628
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1696
1629
|
(__v16hi)_mm256_unpackhi_epi16(__A, __B),
|
|
1697
1630
|
(__v16hi)__W);
|
|
1698
1631
|
}
|
|
1699
1632
|
|
|
1700
|
-
static __inline__ __m256i
|
|
1633
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1701
1634
|
_mm256_maskz_unpackhi_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
1702
1635
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1703
1636
|
(__v16hi)_mm256_unpackhi_epi16(__A, __B),
|
|
1704
1637
|
(__v16hi)_mm256_setzero_si256());
|
|
1705
1638
|
}
|
|
1706
1639
|
|
|
1707
|
-
static __inline__ __m128i
|
|
1640
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1708
1641
|
_mm_mask_unpacklo_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
|
|
1709
1642
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
1710
1643
|
(__v16qi)_mm_unpacklo_epi8(__A, __B),
|
|
1711
1644
|
(__v16qi)__W);
|
|
1712
1645
|
}
|
|
1713
1646
|
|
|
1714
|
-
static __inline__ __m128i
|
|
1647
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1715
1648
|
_mm_maskz_unpacklo_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
1716
1649
|
return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
|
|
1717
1650
|
(__v16qi)_mm_unpacklo_epi8(__A, __B),
|
|
1718
1651
|
(__v16qi)_mm_setzero_si128());
|
|
1719
1652
|
}
|
|
1720
1653
|
|
|
1721
|
-
static __inline__ __m256i
|
|
1654
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1722
1655
|
_mm256_mask_unpacklo_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
|
|
1723
1656
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
1724
1657
|
(__v32qi)_mm256_unpacklo_epi8(__A, __B),
|
|
1725
1658
|
(__v32qi)__W);
|
|
1726
1659
|
}
|
|
1727
1660
|
|
|
1728
|
-
static __inline__ __m256i
|
|
1661
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1729
1662
|
_mm256_maskz_unpacklo_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
1730
1663
|
return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
|
|
1731
1664
|
(__v32qi)_mm256_unpacklo_epi8(__A, __B),
|
|
1732
1665
|
(__v32qi)_mm256_setzero_si256());
|
|
1733
1666
|
}
|
|
1734
1667
|
|
|
1735
|
-
static __inline__ __m128i
|
|
1668
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1736
1669
|
_mm_mask_unpacklo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
1737
1670
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1738
1671
|
(__v8hi)_mm_unpacklo_epi16(__A, __B),
|
|
1739
1672
|
(__v8hi)__W);
|
|
1740
1673
|
}
|
|
1741
1674
|
|
|
1742
|
-
static __inline__ __m128i
|
|
1675
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1743
1676
|
_mm_maskz_unpacklo_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
1744
1677
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
1745
1678
|
(__v8hi)_mm_unpacklo_epi16(__A, __B),
|
|
1746
1679
|
(__v8hi) _mm_setzero_si128());
|
|
1747
1680
|
}
|
|
1748
1681
|
|
|
1749
|
-
static __inline__ __m256i
|
|
1682
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1750
1683
|
_mm256_mask_unpacklo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
|
|
1751
1684
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1752
1685
|
(__v16hi)_mm256_unpacklo_epi16(__A, __B),
|
|
1753
1686
|
(__v16hi)__W);
|
|
1754
1687
|
}
|
|
1755
1688
|
|
|
1756
|
-
static __inline__ __m256i
|
|
1689
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1757
1690
|
_mm256_maskz_unpacklo_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
1758
1691
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1759
1692
|
(__v16hi)_mm256_unpacklo_epi16(__A, __B),
|
|
1760
1693
|
(__v16hi)_mm256_setzero_si256());
|
|
1761
1694
|
}
|
|
1762
1695
|
|
|
1763
|
-
static __inline__ __m128i
|
|
1696
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1764
1697
|
_mm_mask_cvtepi8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
|
|
1765
1698
|
{
|
|
1766
1699
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1768,7 +1701,7 @@ _mm_mask_cvtepi8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
|
|
|
1768
1701
|
(__v8hi)__W);
|
|
1769
1702
|
}
|
|
1770
1703
|
|
|
1771
|
-
static __inline__ __m128i
|
|
1704
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1772
1705
|
_mm_maskz_cvtepi8_epi16(__mmask8 __U, __m128i __A)
|
|
1773
1706
|
{
|
|
1774
1707
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1776,7 +1709,7 @@ _mm_maskz_cvtepi8_epi16(__mmask8 __U, __m128i __A)
|
|
|
1776
1709
|
(__v8hi)_mm_setzero_si128());
|
|
1777
1710
|
}
|
|
1778
1711
|
|
|
1779
|
-
static __inline__ __m256i
|
|
1712
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1780
1713
|
_mm256_mask_cvtepi8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
|
|
1781
1714
|
{
|
|
1782
1715
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1784,7 +1717,7 @@ _mm256_mask_cvtepi8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
|
|
|
1784
1717
|
(__v16hi)__W);
|
|
1785
1718
|
}
|
|
1786
1719
|
|
|
1787
|
-
static __inline__ __m256i
|
|
1720
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1788
1721
|
_mm256_maskz_cvtepi8_epi16(__mmask16 __U, __m128i __A)
|
|
1789
1722
|
{
|
|
1790
1723
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1793,7 +1726,7 @@ _mm256_maskz_cvtepi8_epi16(__mmask16 __U, __m128i __A)
|
|
|
1793
1726
|
}
|
|
1794
1727
|
|
|
1795
1728
|
|
|
1796
|
-
static __inline__ __m128i
|
|
1729
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1797
1730
|
_mm_mask_cvtepu8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
|
|
1798
1731
|
{
|
|
1799
1732
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1801,7 +1734,7 @@ _mm_mask_cvtepu8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
|
|
|
1801
1734
|
(__v8hi)__W);
|
|
1802
1735
|
}
|
|
1803
1736
|
|
|
1804
|
-
static __inline__ __m128i
|
|
1737
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1805
1738
|
_mm_maskz_cvtepu8_epi16(__mmask8 __U, __m128i __A)
|
|
1806
1739
|
{
|
|
1807
1740
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1809,7 +1742,7 @@ _mm_maskz_cvtepu8_epi16(__mmask8 __U, __m128i __A)
|
|
|
1809
1742
|
(__v8hi)_mm_setzero_si128());
|
|
1810
1743
|
}
|
|
1811
1744
|
|
|
1812
|
-
static __inline__ __m256i
|
|
1745
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1813
1746
|
_mm256_mask_cvtepu8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
|
|
1814
1747
|
{
|
|
1815
1748
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1817,7 +1750,7 @@ _mm256_mask_cvtepu8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
|
|
|
1817
1750
|
(__v16hi)__W);
|
|
1818
1751
|
}
|
|
1819
1752
|
|
|
1820
|
-
static __inline__ __m256i
|
|
1753
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1821
1754
|
_mm256_maskz_cvtepu8_epi16 (__mmask16 __U, __m128i __A)
|
|
1822
1755
|
{
|
|
1823
1756
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1868,13 +1801,13 @@ _mm256_maskz_cvtepu8_epi16 (__mmask16 __U, __m128i __A)
|
|
|
1868
1801
|
(imm)), \
|
|
1869
1802
|
(__v16hi)_mm256_setzero_si256()))
|
|
1870
1803
|
|
|
1871
|
-
static __inline__ __m256i
|
|
1804
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1872
1805
|
_mm256_sllv_epi16(__m256i __A, __m256i __B)
|
|
1873
1806
|
{
|
|
1874
1807
|
return (__m256i)__builtin_ia32_psllv16hi((__v16hi)__A, (__v16hi)__B);
|
|
1875
1808
|
}
|
|
1876
1809
|
|
|
1877
|
-
static __inline__ __m256i
|
|
1810
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1878
1811
|
_mm256_mask_sllv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
1879
1812
|
{
|
|
1880
1813
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1882,7 +1815,7 @@ _mm256_mask_sllv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1882
1815
|
(__v16hi)__W);
|
|
1883
1816
|
}
|
|
1884
1817
|
|
|
1885
|
-
static __inline__ __m256i
|
|
1818
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1886
1819
|
_mm256_maskz_sllv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
1887
1820
|
{
|
|
1888
1821
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1890,13 +1823,13 @@ _mm256_maskz_sllv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1890
1823
|
(__v16hi)_mm256_setzero_si256());
|
|
1891
1824
|
}
|
|
1892
1825
|
|
|
1893
|
-
static __inline__ __m128i
|
|
1826
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1894
1827
|
_mm_sllv_epi16(__m128i __A, __m128i __B)
|
|
1895
1828
|
{
|
|
1896
1829
|
return (__m128i)__builtin_ia32_psllv8hi((__v8hi)__A, (__v8hi)__B);
|
|
1897
1830
|
}
|
|
1898
1831
|
|
|
1899
|
-
static __inline__ __m128i
|
|
1832
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1900
1833
|
_mm_mask_sllv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
1901
1834
|
{
|
|
1902
1835
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1904,7 +1837,7 @@ _mm_mask_sllv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
1904
1837
|
(__v8hi)__W);
|
|
1905
1838
|
}
|
|
1906
1839
|
|
|
1907
|
-
static __inline__ __m128i
|
|
1840
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1908
1841
|
_mm_maskz_sllv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
1909
1842
|
{
|
|
1910
1843
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1944,7 +1877,7 @@ _mm256_maskz_sll_epi16(__mmask16 __U, __m256i __A, __m128i __B)
|
|
|
1944
1877
|
(__v16hi)_mm256_setzero_si256());
|
|
1945
1878
|
}
|
|
1946
1879
|
|
|
1947
|
-
static __inline__ __m128i
|
|
1880
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1948
1881
|
_mm_mask_slli_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
|
|
1949
1882
|
{
|
|
1950
1883
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1952,7 +1885,7 @@ _mm_mask_slli_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
|
|
|
1952
1885
|
(__v8hi)__W);
|
|
1953
1886
|
}
|
|
1954
1887
|
|
|
1955
|
-
static __inline__ __m128i
|
|
1888
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1956
1889
|
_mm_maskz_slli_epi16 (__mmask8 __U, __m128i __A, unsigned int __B)
|
|
1957
1890
|
{
|
|
1958
1891
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -1960,30 +1893,28 @@ _mm_maskz_slli_epi16 (__mmask8 __U, __m128i __A, unsigned int __B)
|
|
|
1960
1893
|
(__v8hi)_mm_setzero_si128());
|
|
1961
1894
|
}
|
|
1962
1895
|
|
|
1963
|
-
static __inline__ __m256i
|
|
1896
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1964
1897
|
_mm256_mask_slli_epi16(__m256i __W, __mmask16 __U, __m256i __A,
|
|
1965
|
-
unsigned int __B)
|
|
1966
|
-
{
|
|
1898
|
+
unsigned int __B) {
|
|
1967
1899
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1968
1900
|
(__v16hi)_mm256_slli_epi16(__A, (int)__B),
|
|
1969
1901
|
(__v16hi)__W);
|
|
1970
1902
|
}
|
|
1971
1903
|
|
|
1972
|
-
static __inline__ __m256i
|
|
1973
|
-
_mm256_maskz_slli_epi16(__mmask16 __U, __m256i __A, unsigned int __B)
|
|
1974
|
-
{
|
|
1904
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1905
|
+
_mm256_maskz_slli_epi16(__mmask16 __U, __m256i __A, unsigned int __B) {
|
|
1975
1906
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
1976
1907
|
(__v16hi)_mm256_slli_epi16(__A, (int)__B),
|
|
1977
1908
|
(__v16hi)_mm256_setzero_si256());
|
|
1978
1909
|
}
|
|
1979
1910
|
|
|
1980
|
-
static __inline__ __m256i
|
|
1911
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1981
1912
|
_mm256_srlv_epi16(__m256i __A, __m256i __B)
|
|
1982
1913
|
{
|
|
1983
1914
|
return (__m256i)__builtin_ia32_psrlv16hi((__v16hi)__A, (__v16hi)__B);
|
|
1984
1915
|
}
|
|
1985
1916
|
|
|
1986
|
-
static __inline__ __m256i
|
|
1917
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1987
1918
|
_mm256_mask_srlv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
1988
1919
|
{
|
|
1989
1920
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1991,7 +1922,7 @@ _mm256_mask_srlv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1991
1922
|
(__v16hi)__W);
|
|
1992
1923
|
}
|
|
1993
1924
|
|
|
1994
|
-
static __inline__ __m256i
|
|
1925
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1995
1926
|
_mm256_maskz_srlv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
1996
1927
|
{
|
|
1997
1928
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -1999,13 +1930,13 @@ _mm256_maskz_srlv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
1999
1930
|
(__v16hi)_mm256_setzero_si256());
|
|
2000
1931
|
}
|
|
2001
1932
|
|
|
2002
|
-
static __inline__ __m128i
|
|
1933
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2003
1934
|
_mm_srlv_epi16(__m128i __A, __m128i __B)
|
|
2004
1935
|
{
|
|
2005
1936
|
return (__m128i)__builtin_ia32_psrlv8hi((__v8hi)__A, (__v8hi)__B);
|
|
2006
1937
|
}
|
|
2007
1938
|
|
|
2008
|
-
static __inline__ __m128i
|
|
1939
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2009
1940
|
_mm_mask_srlv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
2010
1941
|
{
|
|
2011
1942
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -2013,7 +1944,7 @@ _mm_mask_srlv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
2013
1944
|
(__v8hi)__W);
|
|
2014
1945
|
}
|
|
2015
1946
|
|
|
2016
|
-
static __inline__ __m128i
|
|
1947
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2017
1948
|
_mm_maskz_srlv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
2018
1949
|
{
|
|
2019
1950
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -2021,13 +1952,13 @@ _mm_maskz_srlv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
2021
1952
|
(__v8hi)_mm_setzero_si128());
|
|
2022
1953
|
}
|
|
2023
1954
|
|
|
2024
|
-
static __inline__ __m256i
|
|
1955
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2025
1956
|
_mm256_srav_epi16(__m256i __A, __m256i __B)
|
|
2026
1957
|
{
|
|
2027
1958
|
return (__m256i)__builtin_ia32_psrav16hi((__v16hi)__A, (__v16hi)__B);
|
|
2028
1959
|
}
|
|
2029
1960
|
|
|
2030
|
-
static __inline__ __m256i
|
|
1961
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2031
1962
|
_mm256_mask_srav_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
2032
1963
|
{
|
|
2033
1964
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -2035,7 +1966,7 @@ _mm256_mask_srav_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
|
|
|
2035
1966
|
(__v16hi)__W);
|
|
2036
1967
|
}
|
|
2037
1968
|
|
|
2038
|
-
static __inline__ __m256i
|
|
1969
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2039
1970
|
_mm256_maskz_srav_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
2040
1971
|
{
|
|
2041
1972
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
@@ -2043,13 +1974,13 @@ _mm256_maskz_srav_epi16(__mmask16 __U, __m256i __A, __m256i __B)
|
|
|
2043
1974
|
(__v16hi)_mm256_setzero_si256());
|
|
2044
1975
|
}
|
|
2045
1976
|
|
|
2046
|
-
static __inline__ __m128i
|
|
1977
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2047
1978
|
_mm_srav_epi16(__m128i __A, __m128i __B)
|
|
2048
1979
|
{
|
|
2049
1980
|
return (__m128i)__builtin_ia32_psrav8hi((__v8hi)__A, (__v8hi)__B);
|
|
2050
1981
|
}
|
|
2051
1982
|
|
|
2052
|
-
static __inline__ __m128i
|
|
1983
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2053
1984
|
_mm_mask_srav_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
2054
1985
|
{
|
|
2055
1986
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -2057,7 +1988,7 @@ _mm_mask_srav_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
2057
1988
|
(__v8hi)__W);
|
|
2058
1989
|
}
|
|
2059
1990
|
|
|
2060
|
-
static __inline__ __m128i
|
|
1991
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2061
1992
|
_mm_maskz_srav_epi16(__mmask8 __U, __m128i __A, __m128i __B)
|
|
2062
1993
|
{
|
|
2063
1994
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
@@ -2097,34 +2028,30 @@ _mm256_maskz_sra_epi16(__mmask16 __U, __m256i __A, __m128i __B)
|
|
|
2097
2028
|
(__v16hi)_mm256_setzero_si256());
|
|
2098
2029
|
}
|
|
2099
2030
|
|
|
2100
|
-
static __inline__ __m128i
|
|
2101
|
-
_mm_mask_srai_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
|
|
2102
|
-
{
|
|
2031
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2032
|
+
_mm_mask_srai_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B) {
|
|
2103
2033
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
2104
2034
|
(__v8hi)_mm_srai_epi16(__A, (int)__B),
|
|
2105
2035
|
(__v8hi)__W);
|
|
2106
2036
|
}
|
|
2107
2037
|
|
|
2108
|
-
static __inline__ __m128i
|
|
2109
|
-
_mm_maskz_srai_epi16(__mmask8 __U, __m128i __A, unsigned int __B)
|
|
2110
|
-
{
|
|
2038
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2039
|
+
_mm_maskz_srai_epi16(__mmask8 __U, __m128i __A, unsigned int __B) {
|
|
2111
2040
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
2112
2041
|
(__v8hi)_mm_srai_epi16(__A, (int)__B),
|
|
2113
2042
|
(__v8hi)_mm_setzero_si128());
|
|
2114
2043
|
}
|
|
2115
2044
|
|
|
2116
|
-
static __inline__ __m256i
|
|
2045
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2117
2046
|
_mm256_mask_srai_epi16(__m256i __W, __mmask16 __U, __m256i __A,
|
|
2118
|
-
unsigned int __B)
|
|
2119
|
-
{
|
|
2047
|
+
unsigned int __B) {
|
|
2120
2048
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
2121
2049
|
(__v16hi)_mm256_srai_epi16(__A, (int)__B),
|
|
2122
2050
|
(__v16hi)__W);
|
|
2123
2051
|
}
|
|
2124
2052
|
|
|
2125
|
-
static __inline__ __m256i
|
|
2126
|
-
_mm256_maskz_srai_epi16(__mmask16 __U, __m256i __A, unsigned int __B)
|
|
2127
|
-
{
|
|
2053
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2054
|
+
_mm256_maskz_srai_epi16(__mmask16 __U, __m256i __A, unsigned int __B) {
|
|
2128
2055
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
2129
2056
|
(__v16hi)_mm256_srai_epi16(__A, (int)__B),
|
|
2130
2057
|
(__v16hi)_mm256_setzero_si256());
|
|
@@ -2162,104 +2089,91 @@ _mm256_maskz_srl_epi16(__mmask16 __U, __m256i __A, __m128i __B)
|
|
|
2162
2089
|
(__v16hi)_mm256_setzero_si256());
|
|
2163
2090
|
}
|
|
2164
2091
|
|
|
2165
|
-
static __inline__ __m128i
|
|
2166
|
-
_mm_mask_srli_epi16(__m128i __W, __mmask8 __U, __m128i __A, int __B)
|
|
2167
|
-
{
|
|
2092
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2093
|
+
_mm_mask_srli_epi16(__m128i __W, __mmask8 __U, __m128i __A, int __B) {
|
|
2168
2094
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
2169
2095
|
(__v8hi)_mm_srli_epi16(__A, __B),
|
|
2170
2096
|
(__v8hi)__W);
|
|
2171
2097
|
}
|
|
2172
2098
|
|
|
2173
|
-
static __inline__ __m128i
|
|
2174
|
-
_mm_maskz_srli_epi16
|
|
2175
|
-
{
|
|
2099
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2100
|
+
_mm_maskz_srli_epi16(__mmask8 __U, __m128i __A, int __B) {
|
|
2176
2101
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
|
|
2177
2102
|
(__v8hi)_mm_srli_epi16(__A, __B),
|
|
2178
2103
|
(__v8hi)_mm_setzero_si128());
|
|
2179
2104
|
}
|
|
2180
2105
|
|
|
2181
|
-
static __inline__ __m256i
|
|
2182
|
-
_mm256_mask_srli_epi16(__m256i __W, __mmask16 __U, __m256i __A, int __B)
|
|
2183
|
-
{
|
|
2106
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2107
|
+
_mm256_mask_srli_epi16(__m256i __W, __mmask16 __U, __m256i __A, int __B) {
|
|
2184
2108
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
2185
2109
|
(__v16hi)_mm256_srli_epi16(__A, __B),
|
|
2186
2110
|
(__v16hi)__W);
|
|
2187
2111
|
}
|
|
2188
2112
|
|
|
2189
|
-
static __inline__ __m256i
|
|
2190
|
-
_mm256_maskz_srli_epi16(__mmask16 __U, __m256i __A, int __B)
|
|
2191
|
-
{
|
|
2113
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2114
|
+
_mm256_maskz_srli_epi16(__mmask16 __U, __m256i __A, int __B) {
|
|
2192
2115
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
|
|
2193
2116
|
(__v16hi)_mm256_srli_epi16(__A, __B),
|
|
2194
2117
|
(__v16hi)_mm256_setzero_si256());
|
|
2195
2118
|
}
|
|
2196
2119
|
|
|
2197
|
-
static __inline__ __m128i
|
|
2198
|
-
_mm_mask_mov_epi16
|
|
2199
|
-
{
|
|
2120
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2121
|
+
_mm_mask_mov_epi16(__m128i __W, __mmask8 __U, __m128i __A) {
|
|
2200
2122
|
return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U,
|
|
2201
2123
|
(__v8hi) __A,
|
|
2202
2124
|
(__v8hi) __W);
|
|
2203
2125
|
}
|
|
2204
2126
|
|
|
2205
|
-
static __inline__ __m128i
|
|
2206
|
-
_mm_maskz_mov_epi16
|
|
2207
|
-
{
|
|
2127
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2128
|
+
_mm_maskz_mov_epi16(__mmask8 __U, __m128i __A) {
|
|
2208
2129
|
return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U,
|
|
2209
2130
|
(__v8hi) __A,
|
|
2210
2131
|
(__v8hi) _mm_setzero_si128 ());
|
|
2211
2132
|
}
|
|
2212
2133
|
|
|
2213
|
-
static __inline__ __m256i
|
|
2214
|
-
_mm256_mask_mov_epi16
|
|
2215
|
-
{
|
|
2134
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2135
|
+
_mm256_mask_mov_epi16(__m256i __W, __mmask16 __U, __m256i __A) {
|
|
2216
2136
|
return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U,
|
|
2217
2137
|
(__v16hi) __A,
|
|
2218
2138
|
(__v16hi) __W);
|
|
2219
2139
|
}
|
|
2220
2140
|
|
|
2221
|
-
static __inline__ __m256i
|
|
2222
|
-
_mm256_maskz_mov_epi16
|
|
2223
|
-
{
|
|
2141
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2142
|
+
_mm256_maskz_mov_epi16(__mmask16 __U, __m256i __A) {
|
|
2224
2143
|
return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U,
|
|
2225
2144
|
(__v16hi) __A,
|
|
2226
2145
|
(__v16hi) _mm256_setzero_si256 ());
|
|
2227
2146
|
}
|
|
2228
2147
|
|
|
2229
|
-
static __inline__ __m128i
|
|
2230
|
-
_mm_mask_mov_epi8
|
|
2231
|
-
{
|
|
2148
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2149
|
+
_mm_mask_mov_epi8(__m128i __W, __mmask16 __U, __m128i __A) {
|
|
2232
2150
|
return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U,
|
|
2233
2151
|
(__v16qi) __A,
|
|
2234
2152
|
(__v16qi) __W);
|
|
2235
2153
|
}
|
|
2236
2154
|
|
|
2237
|
-
static __inline__ __m128i
|
|
2238
|
-
_mm_maskz_mov_epi8
|
|
2239
|
-
{
|
|
2155
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2156
|
+
_mm_maskz_mov_epi8(__mmask16 __U, __m128i __A) {
|
|
2240
2157
|
return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U,
|
|
2241
2158
|
(__v16qi) __A,
|
|
2242
2159
|
(__v16qi) _mm_setzero_si128 ());
|
|
2243
2160
|
}
|
|
2244
2161
|
|
|
2245
|
-
static __inline__ __m256i
|
|
2246
|
-
_mm256_mask_mov_epi8
|
|
2247
|
-
{
|
|
2162
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2163
|
+
_mm256_mask_mov_epi8(__m256i __W, __mmask32 __U, __m256i __A) {
|
|
2248
2164
|
return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U,
|
|
2249
2165
|
(__v32qi) __A,
|
|
2250
2166
|
(__v32qi) __W);
|
|
2251
2167
|
}
|
|
2252
2168
|
|
|
2253
|
-
static __inline__ __m256i
|
|
2254
|
-
_mm256_maskz_mov_epi8
|
|
2255
|
-
{
|
|
2169
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2170
|
+
_mm256_maskz_mov_epi8(__mmask32 __U, __m256i __A) {
|
|
2256
2171
|
return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U,
|
|
2257
2172
|
(__v32qi) __A,
|
|
2258
2173
|
(__v32qi) _mm256_setzero_si256 ());
|
|
2259
2174
|
}
|
|
2260
2175
|
|
|
2261
|
-
|
|
2262
|
-
static __inline__ __m128i __DEFAULT_FN_ATTRS128
|
|
2176
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2263
2177
|
_mm_mask_set1_epi8 (__m128i __O, __mmask16 __M, char __A)
|
|
2264
2178
|
{
|
|
2265
2179
|
return (__m128i) __builtin_ia32_selectb_128(__M,
|
|
@@ -2267,7 +2181,7 @@ _mm_mask_set1_epi8 (__m128i __O, __mmask16 __M, char __A)
|
|
|
2267
2181
|
(__v16qi) __O);
|
|
2268
2182
|
}
|
|
2269
2183
|
|
|
2270
|
-
static __inline__ __m128i
|
|
2184
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2271
2185
|
_mm_maskz_set1_epi8 (__mmask16 __M, char __A)
|
|
2272
2186
|
{
|
|
2273
2187
|
return (__m128i) __builtin_ia32_selectb_128(__M,
|
|
@@ -2275,7 +2189,7 @@ _mm_maskz_set1_epi8 (__mmask16 __M, char __A)
|
|
|
2275
2189
|
(__v16qi) _mm_setzero_si128());
|
|
2276
2190
|
}
|
|
2277
2191
|
|
|
2278
|
-
static __inline__ __m256i
|
|
2192
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2279
2193
|
_mm256_mask_set1_epi8 (__m256i __O, __mmask32 __M, char __A)
|
|
2280
2194
|
{
|
|
2281
2195
|
return (__m256i) __builtin_ia32_selectb_256(__M,
|
|
@@ -2283,7 +2197,7 @@ _mm256_mask_set1_epi8 (__m256i __O, __mmask32 __M, char __A)
|
|
|
2283
2197
|
(__v32qi) __O);
|
|
2284
2198
|
}
|
|
2285
2199
|
|
|
2286
|
-
static __inline__ __m256i
|
|
2200
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2287
2201
|
_mm256_maskz_set1_epi8 (__mmask32 __M, char __A)
|
|
2288
2202
|
{
|
|
2289
2203
|
return (__m256i) __builtin_ia32_selectb_256(__M,
|
|
@@ -2463,162 +2377,138 @@ _mm256_mask_storeu_epi8 (void *__P, __mmask32 __U, __m256i __A)
|
|
|
2463
2377
|
(__mmask32) __U);
|
|
2464
2378
|
}
|
|
2465
2379
|
|
|
2466
|
-
static __inline__ __mmask16
|
|
2467
|
-
_mm_test_epi8_mask
|
|
2468
|
-
{
|
|
2380
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2381
|
+
_mm_test_epi8_mask(__m128i __A, __m128i __B) {
|
|
2469
2382
|
return _mm_cmpneq_epi8_mask (_mm_and_si128(__A, __B), _mm_setzero_si128());
|
|
2470
2383
|
}
|
|
2471
2384
|
|
|
2472
|
-
static __inline__ __mmask16
|
|
2473
|
-
_mm_mask_test_epi8_mask
|
|
2474
|
-
{
|
|
2385
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2386
|
+
_mm_mask_test_epi8_mask(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
2475
2387
|
return _mm_mask_cmpneq_epi8_mask (__U, _mm_and_si128 (__A, __B),
|
|
2476
2388
|
_mm_setzero_si128());
|
|
2477
2389
|
}
|
|
2478
2390
|
|
|
2479
|
-
static __inline__ __mmask32
|
|
2480
|
-
_mm256_test_epi8_mask
|
|
2481
|
-
{
|
|
2391
|
+
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2392
|
+
_mm256_test_epi8_mask(__m256i __A, __m256i __B) {
|
|
2482
2393
|
return _mm256_cmpneq_epi8_mask (_mm256_and_si256(__A, __B),
|
|
2483
2394
|
_mm256_setzero_si256());
|
|
2484
2395
|
}
|
|
2485
2396
|
|
|
2486
|
-
static __inline__ __mmask32
|
|
2487
|
-
_mm256_mask_test_epi8_mask
|
|
2488
|
-
{
|
|
2397
|
+
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2398
|
+
_mm256_mask_test_epi8_mask(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
2489
2399
|
return _mm256_mask_cmpneq_epi8_mask (__U, _mm256_and_si256(__A, __B),
|
|
2490
2400
|
_mm256_setzero_si256());
|
|
2491
2401
|
}
|
|
2492
2402
|
|
|
2493
|
-
static __inline__ __mmask8
|
|
2494
|
-
_mm_test_epi16_mask
|
|
2495
|
-
{
|
|
2403
|
+
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2404
|
+
_mm_test_epi16_mask(__m128i __A, __m128i __B) {
|
|
2496
2405
|
return _mm_cmpneq_epi16_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128());
|
|
2497
2406
|
}
|
|
2498
2407
|
|
|
2499
|
-
static __inline__ __mmask8
|
|
2500
|
-
_mm_mask_test_epi16_mask
|
|
2501
|
-
{
|
|
2408
|
+
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2409
|
+
_mm_mask_test_epi16_mask(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
2502
2410
|
return _mm_mask_cmpneq_epi16_mask (__U, _mm_and_si128 (__A, __B),
|
|
2503
2411
|
_mm_setzero_si128());
|
|
2504
2412
|
}
|
|
2505
2413
|
|
|
2506
|
-
static __inline__ __mmask16
|
|
2507
|
-
_mm256_test_epi16_mask
|
|
2508
|
-
{
|
|
2414
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2415
|
+
_mm256_test_epi16_mask(__m256i __A, __m256i __B) {
|
|
2509
2416
|
return _mm256_cmpneq_epi16_mask (_mm256_and_si256 (__A, __B),
|
|
2510
2417
|
_mm256_setzero_si256 ());
|
|
2511
2418
|
}
|
|
2512
2419
|
|
|
2513
|
-
static __inline__ __mmask16
|
|
2514
|
-
_mm256_mask_test_epi16_mask
|
|
2515
|
-
{
|
|
2420
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2421
|
+
_mm256_mask_test_epi16_mask(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
2516
2422
|
return _mm256_mask_cmpneq_epi16_mask (__U, _mm256_and_si256(__A, __B),
|
|
2517
2423
|
_mm256_setzero_si256());
|
|
2518
2424
|
}
|
|
2519
2425
|
|
|
2520
|
-
static __inline__ __mmask16
|
|
2521
|
-
_mm_testn_epi8_mask
|
|
2522
|
-
{
|
|
2426
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2427
|
+
_mm_testn_epi8_mask(__m128i __A, __m128i __B) {
|
|
2523
2428
|
return _mm_cmpeq_epi8_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128());
|
|
2524
2429
|
}
|
|
2525
2430
|
|
|
2526
|
-
static __inline__ __mmask16
|
|
2527
|
-
_mm_mask_testn_epi8_mask
|
|
2528
|
-
{
|
|
2431
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2432
|
+
_mm_mask_testn_epi8_mask(__mmask16 __U, __m128i __A, __m128i __B) {
|
|
2529
2433
|
return _mm_mask_cmpeq_epi8_mask (__U, _mm_and_si128 (__A, __B),
|
|
2530
2434
|
_mm_setzero_si128());
|
|
2531
2435
|
}
|
|
2532
2436
|
|
|
2533
|
-
static __inline__ __mmask32
|
|
2534
|
-
_mm256_testn_epi8_mask
|
|
2535
|
-
{
|
|
2437
|
+
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2438
|
+
_mm256_testn_epi8_mask(__m256i __A, __m256i __B) {
|
|
2536
2439
|
return _mm256_cmpeq_epi8_mask (_mm256_and_si256 (__A, __B),
|
|
2537
2440
|
_mm256_setzero_si256());
|
|
2538
2441
|
}
|
|
2539
2442
|
|
|
2540
|
-
static __inline__ __mmask32
|
|
2541
|
-
_mm256_mask_testn_epi8_mask
|
|
2542
|
-
{
|
|
2443
|
+
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2444
|
+
_mm256_mask_testn_epi8_mask(__mmask32 __U, __m256i __A, __m256i __B) {
|
|
2543
2445
|
return _mm256_mask_cmpeq_epi8_mask (__U, _mm256_and_si256 (__A, __B),
|
|
2544
2446
|
_mm256_setzero_si256());
|
|
2545
2447
|
}
|
|
2546
2448
|
|
|
2547
|
-
static __inline__ __mmask8
|
|
2548
|
-
_mm_testn_epi16_mask
|
|
2549
|
-
{
|
|
2449
|
+
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2450
|
+
_mm_testn_epi16_mask(__m128i __A, __m128i __B) {
|
|
2550
2451
|
return _mm_cmpeq_epi16_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128());
|
|
2551
2452
|
}
|
|
2552
2453
|
|
|
2553
|
-
static __inline__ __mmask8
|
|
2554
|
-
_mm_mask_testn_epi16_mask
|
|
2555
|
-
{
|
|
2454
|
+
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2455
|
+
_mm_mask_testn_epi16_mask(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
2556
2456
|
return _mm_mask_cmpeq_epi16_mask (__U, _mm_and_si128(__A, __B), _mm_setzero_si128());
|
|
2557
2457
|
}
|
|
2558
2458
|
|
|
2559
|
-
static __inline__ __mmask16
|
|
2560
|
-
_mm256_testn_epi16_mask
|
|
2561
|
-
{
|
|
2459
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2460
|
+
_mm256_testn_epi16_mask(__m256i __A, __m256i __B) {
|
|
2562
2461
|
return _mm256_cmpeq_epi16_mask (_mm256_and_si256(__A, __B),
|
|
2563
2462
|
_mm256_setzero_si256());
|
|
2564
2463
|
}
|
|
2565
2464
|
|
|
2566
|
-
static __inline__ __mmask16
|
|
2567
|
-
_mm256_mask_testn_epi16_mask
|
|
2568
|
-
{
|
|
2465
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2466
|
+
_mm256_mask_testn_epi16_mask(__mmask16 __U, __m256i __A, __m256i __B) {
|
|
2569
2467
|
return _mm256_mask_cmpeq_epi16_mask (__U, _mm256_and_si256 (__A, __B),
|
|
2570
2468
|
_mm256_setzero_si256());
|
|
2571
2469
|
}
|
|
2572
2470
|
|
|
2573
|
-
static __inline__ __mmask16
|
|
2574
|
-
_mm_movepi8_mask
|
|
2575
|
-
{
|
|
2471
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2472
|
+
_mm_movepi8_mask(__m128i __A) {
|
|
2576
2473
|
return (__mmask16) __builtin_ia32_cvtb2mask128 ((__v16qi) __A);
|
|
2577
2474
|
}
|
|
2578
2475
|
|
|
2579
|
-
static __inline__ __mmask32
|
|
2580
|
-
_mm256_movepi8_mask
|
|
2581
|
-
{
|
|
2476
|
+
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2477
|
+
_mm256_movepi8_mask(__m256i __A) {
|
|
2582
2478
|
return (__mmask32) __builtin_ia32_cvtb2mask256 ((__v32qi) __A);
|
|
2583
2479
|
}
|
|
2584
2480
|
|
|
2585
|
-
static __inline__ __mmask8
|
|
2586
|
-
_mm_movepi16_mask
|
|
2587
|
-
{
|
|
2481
|
+
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2482
|
+
_mm_movepi16_mask(__m128i __A) {
|
|
2588
2483
|
return (__mmask8) __builtin_ia32_cvtw2mask128 ((__v8hi) __A);
|
|
2589
2484
|
}
|
|
2590
2485
|
|
|
2591
|
-
static __inline__ __mmask16
|
|
2592
|
-
_mm256_movepi16_mask
|
|
2593
|
-
{
|
|
2486
|
+
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2487
|
+
_mm256_movepi16_mask(__m256i __A) {
|
|
2594
2488
|
return (__mmask16) __builtin_ia32_cvtw2mask256 ((__v16hi) __A);
|
|
2595
2489
|
}
|
|
2596
2490
|
|
|
2597
|
-
static __inline__ __m128i
|
|
2598
|
-
_mm_movm_epi8
|
|
2599
|
-
{
|
|
2491
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2492
|
+
_mm_movm_epi8(__mmask16 __A) {
|
|
2600
2493
|
return (__m128i) __builtin_ia32_cvtmask2b128 (__A);
|
|
2601
2494
|
}
|
|
2602
2495
|
|
|
2603
|
-
static __inline__ __m256i
|
|
2604
|
-
_mm256_movm_epi8
|
|
2605
|
-
{
|
|
2496
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2497
|
+
_mm256_movm_epi8(__mmask32 __A) {
|
|
2606
2498
|
return (__m256i) __builtin_ia32_cvtmask2b256 (__A);
|
|
2607
2499
|
}
|
|
2608
2500
|
|
|
2609
|
-
static __inline__ __m128i
|
|
2610
|
-
_mm_movm_epi16
|
|
2611
|
-
{
|
|
2501
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2502
|
+
_mm_movm_epi16(__mmask8 __A) {
|
|
2612
2503
|
return (__m128i) __builtin_ia32_cvtmask2w128 (__A);
|
|
2613
2504
|
}
|
|
2614
2505
|
|
|
2615
|
-
static __inline__ __m256i
|
|
2616
|
-
_mm256_movm_epi16
|
|
2617
|
-
{
|
|
2506
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2507
|
+
_mm256_movm_epi16(__mmask16 __A) {
|
|
2618
2508
|
return (__m256i) __builtin_ia32_cvtmask2w256 (__A);
|
|
2619
2509
|
}
|
|
2620
2510
|
|
|
2621
|
-
static __inline__ __m128i
|
|
2511
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2622
2512
|
_mm_mask_broadcastb_epi8 (__m128i __O, __mmask16 __M, __m128i __A)
|
|
2623
2513
|
{
|
|
2624
2514
|
return (__m128i)__builtin_ia32_selectb_128(__M,
|
|
@@ -2626,7 +2516,7 @@ _mm_mask_broadcastb_epi8 (__m128i __O, __mmask16 __M, __m128i __A)
|
|
|
2626
2516
|
(__v16qi) __O);
|
|
2627
2517
|
}
|
|
2628
2518
|
|
|
2629
|
-
static __inline__ __m128i
|
|
2519
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2630
2520
|
_mm_maskz_broadcastb_epi8 (__mmask16 __M, __m128i __A)
|
|
2631
2521
|
{
|
|
2632
2522
|
return (__m128i)__builtin_ia32_selectb_128(__M,
|
|
@@ -2634,7 +2524,7 @@ _mm_maskz_broadcastb_epi8 (__mmask16 __M, __m128i __A)
|
|
|
2634
2524
|
(__v16qi) _mm_setzero_si128());
|
|
2635
2525
|
}
|
|
2636
2526
|
|
|
2637
|
-
static __inline__ __m256i
|
|
2527
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2638
2528
|
_mm256_mask_broadcastb_epi8 (__m256i __O, __mmask32 __M, __m128i __A)
|
|
2639
2529
|
{
|
|
2640
2530
|
return (__m256i)__builtin_ia32_selectb_256(__M,
|
|
@@ -2642,7 +2532,7 @@ _mm256_mask_broadcastb_epi8 (__m256i __O, __mmask32 __M, __m128i __A)
|
|
|
2642
2532
|
(__v32qi) __O);
|
|
2643
2533
|
}
|
|
2644
2534
|
|
|
2645
|
-
static __inline__ __m256i
|
|
2535
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2646
2536
|
_mm256_maskz_broadcastb_epi8 (__mmask32 __M, __m128i __A)
|
|
2647
2537
|
{
|
|
2648
2538
|
return (__m256i)__builtin_ia32_selectb_256(__M,
|
|
@@ -2650,7 +2540,7 @@ _mm256_maskz_broadcastb_epi8 (__mmask32 __M, __m128i __A)
|
|
|
2650
2540
|
(__v32qi) _mm256_setzero_si256());
|
|
2651
2541
|
}
|
|
2652
2542
|
|
|
2653
|
-
static __inline__ __m128i
|
|
2543
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2654
2544
|
_mm_mask_broadcastw_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
|
|
2655
2545
|
{
|
|
2656
2546
|
return (__m128i)__builtin_ia32_selectw_128(__M,
|
|
@@ -2658,7 +2548,7 @@ _mm_mask_broadcastw_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
|
|
|
2658
2548
|
(__v8hi) __O);
|
|
2659
2549
|
}
|
|
2660
2550
|
|
|
2661
|
-
static __inline__ __m128i
|
|
2551
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2662
2552
|
_mm_maskz_broadcastw_epi16 (__mmask8 __M, __m128i __A)
|
|
2663
2553
|
{
|
|
2664
2554
|
return (__m128i)__builtin_ia32_selectw_128(__M,
|
|
@@ -2666,7 +2556,7 @@ _mm_maskz_broadcastw_epi16 (__mmask8 __M, __m128i __A)
|
|
|
2666
2556
|
(__v8hi) _mm_setzero_si128());
|
|
2667
2557
|
}
|
|
2668
2558
|
|
|
2669
|
-
static __inline__ __m256i
|
|
2559
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2670
2560
|
_mm256_mask_broadcastw_epi16 (__m256i __O, __mmask16 __M, __m128i __A)
|
|
2671
2561
|
{
|
|
2672
2562
|
return (__m256i)__builtin_ia32_selectw_256(__M,
|
|
@@ -2674,7 +2564,7 @@ _mm256_mask_broadcastw_epi16 (__m256i __O, __mmask16 __M, __m128i __A)
|
|
|
2674
2564
|
(__v16hi) __O);
|
|
2675
2565
|
}
|
|
2676
2566
|
|
|
2677
|
-
static __inline__ __m256i
|
|
2567
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2678
2568
|
_mm256_maskz_broadcastw_epi16 (__mmask16 __M, __m128i __A)
|
|
2679
2569
|
{
|
|
2680
2570
|
return (__m256i)__builtin_ia32_selectw_256(__M,
|
|
@@ -2682,7 +2572,7 @@ _mm256_maskz_broadcastw_epi16 (__mmask16 __M, __m128i __A)
|
|
|
2682
2572
|
(__v16hi) _mm256_setzero_si256());
|
|
2683
2573
|
}
|
|
2684
2574
|
|
|
2685
|
-
static __inline__ __m256i
|
|
2575
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2686
2576
|
_mm256_mask_set1_epi16 (__m256i __O, __mmask16 __M, short __A)
|
|
2687
2577
|
{
|
|
2688
2578
|
return (__m256i) __builtin_ia32_selectw_256 (__M,
|
|
@@ -2690,7 +2580,7 @@ _mm256_mask_set1_epi16 (__m256i __O, __mmask16 __M, short __A)
|
|
|
2690
2580
|
(__v16hi) __O);
|
|
2691
2581
|
}
|
|
2692
2582
|
|
|
2693
|
-
static __inline__ __m256i
|
|
2583
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2694
2584
|
_mm256_maskz_set1_epi16 (__mmask16 __M, short __A)
|
|
2695
2585
|
{
|
|
2696
2586
|
return (__m256i) __builtin_ia32_selectw_256(__M,
|
|
@@ -2698,7 +2588,7 @@ _mm256_maskz_set1_epi16 (__mmask16 __M, short __A)
|
|
|
2698
2588
|
(__v16hi) _mm256_setzero_si256());
|
|
2699
2589
|
}
|
|
2700
2590
|
|
|
2701
|
-
static __inline__ __m128i
|
|
2591
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2702
2592
|
_mm_mask_set1_epi16 (__m128i __O, __mmask8 __M, short __A)
|
|
2703
2593
|
{
|
|
2704
2594
|
return (__m128i) __builtin_ia32_selectw_128(__M,
|
|
@@ -2706,7 +2596,7 @@ _mm_mask_set1_epi16 (__m128i __O, __mmask8 __M, short __A)
|
|
|
2706
2596
|
(__v8hi) __O);
|
|
2707
2597
|
}
|
|
2708
2598
|
|
|
2709
|
-
static __inline__ __m128i
|
|
2599
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2710
2600
|
_mm_maskz_set1_epi16 (__mmask8 __M, short __A)
|
|
2711
2601
|
{
|
|
2712
2602
|
return (__m128i) __builtin_ia32_selectw_128(__M,
|
|
@@ -2714,48 +2604,41 @@ _mm_maskz_set1_epi16 (__mmask8 __M, short __A)
|
|
|
2714
2604
|
(__v8hi) _mm_setzero_si128());
|
|
2715
2605
|
}
|
|
2716
2606
|
|
|
2717
|
-
static __inline__ __m128i
|
|
2718
|
-
_mm_permutexvar_epi16
|
|
2719
|
-
{
|
|
2607
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2608
|
+
_mm_permutexvar_epi16(__m128i __A, __m128i __B) {
|
|
2720
2609
|
return (__m128i)__builtin_ia32_permvarhi128((__v8hi) __B, (__v8hi) __A);
|
|
2721
2610
|
}
|
|
2722
2611
|
|
|
2723
|
-
static __inline__ __m128i
|
|
2724
|
-
_mm_maskz_permutexvar_epi16
|
|
2725
|
-
{
|
|
2612
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2613
|
+
_mm_maskz_permutexvar_epi16(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
2726
2614
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
2727
2615
|
(__v8hi)_mm_permutexvar_epi16(__A, __B),
|
|
2728
2616
|
(__v8hi) _mm_setzero_si128());
|
|
2729
2617
|
}
|
|
2730
2618
|
|
|
2731
|
-
static __inline__ __m128i
|
|
2732
|
-
_mm_mask_permutexvar_epi16
|
|
2733
|
-
|
|
2734
|
-
{
|
|
2619
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2620
|
+
_mm_mask_permutexvar_epi16(__m128i __W, __mmask8 __M, __m128i __A,
|
|
2621
|
+
__m128i __B) {
|
|
2735
2622
|
return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
|
|
2736
2623
|
(__v8hi)_mm_permutexvar_epi16(__A, __B),
|
|
2737
2624
|
(__v8hi)__W);
|
|
2738
2625
|
}
|
|
2739
2626
|
|
|
2740
|
-
static __inline__ __m256i
|
|
2741
|
-
_mm256_permutexvar_epi16
|
|
2742
|
-
{
|
|
2627
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2628
|
+
_mm256_permutexvar_epi16(__m256i __A, __m256i __B) {
|
|
2743
2629
|
return (__m256i)__builtin_ia32_permvarhi256((__v16hi) __B, (__v16hi) __A);
|
|
2744
2630
|
}
|
|
2745
2631
|
|
|
2746
|
-
static __inline__ __m256i
|
|
2747
|
-
_mm256_maskz_permutexvar_epi16
|
|
2748
|
-
__m256i __B)
|
|
2749
|
-
{
|
|
2632
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2633
|
+
_mm256_maskz_permutexvar_epi16(__mmask16 __M, __m256i __A, __m256i __B) {
|
|
2750
2634
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
2751
2635
|
(__v16hi)_mm256_permutexvar_epi16(__A, __B),
|
|
2752
2636
|
(__v16hi)_mm256_setzero_si256());
|
|
2753
2637
|
}
|
|
2754
2638
|
|
|
2755
|
-
static __inline__ __m256i
|
|
2756
|
-
_mm256_mask_permutexvar_epi16
|
|
2757
|
-
|
|
2758
|
-
{
|
|
2639
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2640
|
+
_mm256_mask_permutexvar_epi16(__m256i __W, __mmask16 __M, __m256i __A,
|
|
2641
|
+
__m256i __B) {
|
|
2759
2642
|
return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
|
|
2760
2643
|
(__v16hi)_mm256_permutexvar_epi16(__A, __B),
|
|
2761
2644
|
(__v16hi)__W);
|
|
@@ -2809,353 +2692,353 @@ _mm256_mask_permutexvar_epi16 (__m256i __W, __mmask16 __M, __m256i __A,
|
|
|
2809
2692
|
(__v16hi)_mm256_dbsad_epu8((A), (B), (imm)), \
|
|
2810
2693
|
(__v16hi)_mm256_setzero_si256()))
|
|
2811
2694
|
|
|
2812
|
-
static __inline__ short
|
|
2695
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2813
2696
|
_mm_reduce_add_epi16(__m128i __W) {
|
|
2814
2697
|
return __builtin_reduce_add((__v8hi)__W);
|
|
2815
2698
|
}
|
|
2816
2699
|
|
|
2817
|
-
static __inline__ short
|
|
2700
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2818
2701
|
_mm_reduce_mul_epi16(__m128i __W) {
|
|
2819
2702
|
return __builtin_reduce_mul((__v8hi)__W);
|
|
2820
2703
|
}
|
|
2821
2704
|
|
|
2822
|
-
static __inline__ short
|
|
2705
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2823
2706
|
_mm_reduce_and_epi16(__m128i __W) {
|
|
2824
2707
|
return __builtin_reduce_and((__v8hi)__W);
|
|
2825
2708
|
}
|
|
2826
2709
|
|
|
2827
|
-
static __inline__ short
|
|
2710
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2828
2711
|
_mm_reduce_or_epi16(__m128i __W) {
|
|
2829
2712
|
return __builtin_reduce_or((__v8hi)__W);
|
|
2830
2713
|
}
|
|
2831
2714
|
|
|
2832
|
-
static __inline__ short
|
|
2833
|
-
_mm_mask_reduce_add_epi16(
|
|
2715
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2716
|
+
_mm_mask_reduce_add_epi16(__mmask8 __M, __m128i __W) {
|
|
2834
2717
|
__W = _mm_maskz_mov_epi16(__M, __W);
|
|
2835
2718
|
return __builtin_reduce_add((__v8hi)__W);
|
|
2836
2719
|
}
|
|
2837
2720
|
|
|
2838
|
-
static __inline__ short
|
|
2839
|
-
_mm_mask_reduce_mul_epi16(
|
|
2721
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2722
|
+
_mm_mask_reduce_mul_epi16(__mmask8 __M, __m128i __W) {
|
|
2840
2723
|
__W = _mm_mask_mov_epi16(_mm_set1_epi16(1), __M, __W);
|
|
2841
2724
|
return __builtin_reduce_mul((__v8hi)__W);
|
|
2842
2725
|
}
|
|
2843
2726
|
|
|
2844
|
-
static __inline__ short
|
|
2845
|
-
_mm_mask_reduce_and_epi16(
|
|
2727
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2728
|
+
_mm_mask_reduce_and_epi16(__mmask8 __M, __m128i __W) {
|
|
2846
2729
|
__W = _mm_mask_mov_epi16(_mm_set1_epi16(-1), __M, __W);
|
|
2847
2730
|
return __builtin_reduce_and((__v8hi)__W);
|
|
2848
2731
|
}
|
|
2849
2732
|
|
|
2850
|
-
static __inline__ short
|
|
2733
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2851
2734
|
_mm_mask_reduce_or_epi16(__mmask8 __M, __m128i __W) {
|
|
2852
2735
|
__W = _mm_maskz_mov_epi16(__M, __W);
|
|
2853
2736
|
return __builtin_reduce_or((__v8hi)__W);
|
|
2854
2737
|
}
|
|
2855
2738
|
|
|
2856
|
-
static __inline__ short
|
|
2739
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2857
2740
|
_mm_reduce_max_epi16(__m128i __V) {
|
|
2858
2741
|
return __builtin_reduce_max((__v8hi)__V);
|
|
2859
2742
|
}
|
|
2860
2743
|
|
|
2861
|
-
static __inline__ unsigned short
|
|
2744
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2862
2745
|
_mm_reduce_max_epu16(__m128i __V) {
|
|
2863
2746
|
return __builtin_reduce_max((__v8hu)__V);
|
|
2864
2747
|
}
|
|
2865
2748
|
|
|
2866
|
-
static __inline__ short
|
|
2749
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2867
2750
|
_mm_reduce_min_epi16(__m128i __V) {
|
|
2868
2751
|
return __builtin_reduce_min((__v8hi)__V);
|
|
2869
2752
|
}
|
|
2870
2753
|
|
|
2871
|
-
static __inline__ unsigned short
|
|
2754
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2872
2755
|
_mm_reduce_min_epu16(__m128i __V) {
|
|
2873
2756
|
return __builtin_reduce_min((__v8hu)__V);
|
|
2874
2757
|
}
|
|
2875
2758
|
|
|
2876
|
-
static __inline__ short
|
|
2759
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2877
2760
|
_mm_mask_reduce_max_epi16(__mmask16 __M, __m128i __V) {
|
|
2878
2761
|
__V = _mm_mask_mov_epi16(_mm_set1_epi16(-32767-1), __M, __V);
|
|
2879
2762
|
return __builtin_reduce_max((__v8hi)__V);
|
|
2880
2763
|
}
|
|
2881
2764
|
|
|
2882
|
-
static __inline__ unsigned short
|
|
2765
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2883
2766
|
_mm_mask_reduce_max_epu16(__mmask16 __M, __m128i __V) {
|
|
2884
2767
|
__V = _mm_maskz_mov_epi16(__M, __V);
|
|
2885
2768
|
return __builtin_reduce_max((__v8hu)__V);
|
|
2886
2769
|
}
|
|
2887
2770
|
|
|
2888
|
-
static __inline__ short
|
|
2771
|
+
static __inline__ short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2889
2772
|
_mm_mask_reduce_min_epi16(__mmask16 __M, __m128i __V) {
|
|
2890
2773
|
__V = _mm_mask_mov_epi16(_mm_set1_epi16(32767), __M, __V);
|
|
2891
2774
|
return __builtin_reduce_min((__v8hi)__V);
|
|
2892
2775
|
}
|
|
2893
2776
|
|
|
2894
|
-
static __inline__ unsigned short
|
|
2777
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2895
2778
|
_mm_mask_reduce_min_epu16(__mmask16 __M, __m128i __V) {
|
|
2896
2779
|
__V = _mm_mask_mov_epi16(_mm_set1_epi16(-1), __M, __V);
|
|
2897
2780
|
return __builtin_reduce_min((__v8hu)__V);
|
|
2898
2781
|
}
|
|
2899
2782
|
|
|
2900
|
-
static __inline__ short
|
|
2783
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2901
2784
|
_mm256_reduce_add_epi16(__m256i __W) {
|
|
2902
2785
|
return __builtin_reduce_add((__v16hi)__W);
|
|
2903
2786
|
}
|
|
2904
2787
|
|
|
2905
|
-
static __inline__ short
|
|
2788
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2906
2789
|
_mm256_reduce_mul_epi16(__m256i __W) {
|
|
2907
2790
|
return __builtin_reduce_mul((__v16hi)__W);
|
|
2908
2791
|
}
|
|
2909
2792
|
|
|
2910
|
-
static __inline__ short
|
|
2793
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2911
2794
|
_mm256_reduce_and_epi16(__m256i __W) {
|
|
2912
2795
|
return __builtin_reduce_and((__v16hi)__W);
|
|
2913
2796
|
}
|
|
2914
2797
|
|
|
2915
|
-
static __inline__ short
|
|
2798
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2916
2799
|
_mm256_reduce_or_epi16(__m256i __W) {
|
|
2917
2800
|
return __builtin_reduce_or((__v16hi)__W);
|
|
2918
2801
|
}
|
|
2919
2802
|
|
|
2920
|
-
static __inline__ short
|
|
2921
|
-
_mm256_mask_reduce_add_epi16(
|
|
2803
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2804
|
+
_mm256_mask_reduce_add_epi16(__mmask16 __M, __m256i __W) {
|
|
2922
2805
|
__W = _mm256_maskz_mov_epi16(__M, __W);
|
|
2923
2806
|
return __builtin_reduce_add((__v16hi)__W);
|
|
2924
2807
|
}
|
|
2925
2808
|
|
|
2926
|
-
static __inline__ short
|
|
2927
|
-
_mm256_mask_reduce_mul_epi16(
|
|
2809
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2810
|
+
_mm256_mask_reduce_mul_epi16(__mmask16 __M, __m256i __W) {
|
|
2928
2811
|
__W = _mm256_mask_mov_epi16(_mm256_set1_epi16(1), __M, __W);
|
|
2929
2812
|
return __builtin_reduce_mul((__v16hi)__W);
|
|
2930
2813
|
}
|
|
2931
2814
|
|
|
2932
|
-
static __inline__ short
|
|
2933
|
-
_mm256_mask_reduce_and_epi16(
|
|
2815
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2816
|
+
_mm256_mask_reduce_and_epi16(__mmask16 __M, __m256i __W) {
|
|
2934
2817
|
__W = _mm256_mask_mov_epi16(_mm256_set1_epi16(-1), __M, __W);
|
|
2935
2818
|
return __builtin_reduce_and((__v16hi)__W);
|
|
2936
2819
|
}
|
|
2937
2820
|
|
|
2938
|
-
static __inline__ short
|
|
2821
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2939
2822
|
_mm256_mask_reduce_or_epi16(__mmask16 __M, __m256i __W) {
|
|
2940
2823
|
__W = _mm256_maskz_mov_epi16(__M, __W);
|
|
2941
2824
|
return __builtin_reduce_or((__v16hi)__W);
|
|
2942
2825
|
}
|
|
2943
2826
|
|
|
2944
|
-
static __inline__ short
|
|
2827
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2945
2828
|
_mm256_reduce_max_epi16(__m256i __V) {
|
|
2946
2829
|
return __builtin_reduce_max((__v16hi)__V);
|
|
2947
2830
|
}
|
|
2948
2831
|
|
|
2949
|
-
static __inline__ unsigned short
|
|
2832
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2950
2833
|
_mm256_reduce_max_epu16(__m256i __V) {
|
|
2951
2834
|
return __builtin_reduce_max((__v16hu)__V);
|
|
2952
2835
|
}
|
|
2953
2836
|
|
|
2954
|
-
static __inline__ short
|
|
2837
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2955
2838
|
_mm256_reduce_min_epi16(__m256i __V) {
|
|
2956
2839
|
return __builtin_reduce_min((__v16hi)__V);
|
|
2957
2840
|
}
|
|
2958
2841
|
|
|
2959
|
-
static __inline__ unsigned short
|
|
2842
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2960
2843
|
_mm256_reduce_min_epu16(__m256i __V) {
|
|
2961
2844
|
return __builtin_reduce_min((__v16hu)__V);
|
|
2962
2845
|
}
|
|
2963
2846
|
|
|
2964
|
-
static __inline__ short
|
|
2847
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2965
2848
|
_mm256_mask_reduce_max_epi16(__mmask16 __M, __m256i __V) {
|
|
2966
2849
|
__V = _mm256_mask_mov_epi16(_mm256_set1_epi16(-32767-1), __M, __V);
|
|
2967
2850
|
return __builtin_reduce_max((__v16hi)__V);
|
|
2968
2851
|
}
|
|
2969
2852
|
|
|
2970
|
-
static __inline__ unsigned short
|
|
2853
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2971
2854
|
_mm256_mask_reduce_max_epu16(__mmask16 __M, __m256i __V) {
|
|
2972
2855
|
__V = _mm256_maskz_mov_epi16(__M, __V);
|
|
2973
2856
|
return __builtin_reduce_max((__v16hu)__V);
|
|
2974
2857
|
}
|
|
2975
2858
|
|
|
2976
|
-
static __inline__ short
|
|
2859
|
+
static __inline__ short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2977
2860
|
_mm256_mask_reduce_min_epi16(__mmask16 __M, __m256i __V) {
|
|
2978
2861
|
__V = _mm256_mask_mov_epi16(_mm256_set1_epi16(32767), __M, __V);
|
|
2979
2862
|
return __builtin_reduce_min((__v16hi)__V);
|
|
2980
2863
|
}
|
|
2981
2864
|
|
|
2982
|
-
static __inline__ unsigned short
|
|
2865
|
+
static __inline__ unsigned short __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2983
2866
|
_mm256_mask_reduce_min_epu16(__mmask16 __M, __m256i __V) {
|
|
2984
2867
|
__V = _mm256_mask_mov_epi16(_mm256_set1_epi16(-1), __M, __V);
|
|
2985
2868
|
return __builtin_reduce_min((__v16hu)__V);
|
|
2986
2869
|
}
|
|
2987
2870
|
|
|
2988
|
-
static __inline__ signed char
|
|
2871
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2989
2872
|
_mm_reduce_add_epi8(__m128i __W) {
|
|
2990
2873
|
return __builtin_reduce_add((__v16qs)__W);
|
|
2991
2874
|
}
|
|
2992
2875
|
|
|
2993
|
-
static __inline__ signed char
|
|
2876
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2994
2877
|
_mm_reduce_mul_epi8(__m128i __W) {
|
|
2995
2878
|
return __builtin_reduce_mul((__v16qs)__W);
|
|
2996
2879
|
}
|
|
2997
2880
|
|
|
2998
|
-
static __inline__ signed char
|
|
2881
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2999
2882
|
_mm_reduce_and_epi8(__m128i __W) {
|
|
3000
2883
|
return __builtin_reduce_and((__v16qs)__W);
|
|
3001
2884
|
}
|
|
3002
2885
|
|
|
3003
|
-
static __inline__ signed char
|
|
2886
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3004
2887
|
_mm_reduce_or_epi8(__m128i __W) {
|
|
3005
2888
|
return __builtin_reduce_or((__v16qs)__W);
|
|
3006
2889
|
}
|
|
3007
2890
|
|
|
3008
|
-
static __inline__ signed char
|
|
2891
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3009
2892
|
_mm_mask_reduce_add_epi8(__mmask16 __M, __m128i __W) {
|
|
3010
2893
|
__W = _mm_maskz_mov_epi8(__M, __W);
|
|
3011
2894
|
return __builtin_reduce_add((__v16qs)__W);
|
|
3012
2895
|
}
|
|
3013
2896
|
|
|
3014
|
-
static __inline__ signed char
|
|
2897
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3015
2898
|
_mm_mask_reduce_mul_epi8(__mmask16 __M, __m128i __W) {
|
|
3016
2899
|
__W = _mm_mask_mov_epi8(_mm_set1_epi8(1), __M, __W);
|
|
3017
2900
|
return __builtin_reduce_mul((__v16qs)__W);
|
|
3018
2901
|
}
|
|
3019
2902
|
|
|
3020
|
-
static __inline__ signed char
|
|
2903
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3021
2904
|
_mm_mask_reduce_and_epi8(__mmask16 __M, __m128i __W) {
|
|
3022
2905
|
__W = _mm_mask_mov_epi8(_mm_set1_epi8(-1), __M, __W);
|
|
3023
2906
|
return __builtin_reduce_and((__v16qs)__W);
|
|
3024
2907
|
}
|
|
3025
2908
|
|
|
3026
|
-
static __inline__ signed char
|
|
2909
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3027
2910
|
_mm_mask_reduce_or_epi8(__mmask16 __M, __m128i __W) {
|
|
3028
2911
|
__W = _mm_maskz_mov_epi8(__M, __W);
|
|
3029
2912
|
return __builtin_reduce_or((__v16qs)__W);
|
|
3030
2913
|
}
|
|
3031
2914
|
|
|
3032
|
-
static __inline__ signed char
|
|
2915
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3033
2916
|
_mm_reduce_max_epi8(__m128i __V) {
|
|
3034
2917
|
return __builtin_reduce_max((__v16qs)__V);
|
|
3035
2918
|
}
|
|
3036
2919
|
|
|
3037
|
-
static __inline__ unsigned char
|
|
2920
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3038
2921
|
_mm_reduce_max_epu8(__m128i __V) {
|
|
3039
2922
|
return __builtin_reduce_max((__v16qu)__V);
|
|
3040
2923
|
}
|
|
3041
2924
|
|
|
3042
|
-
static __inline__ signed char
|
|
2925
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3043
2926
|
_mm_reduce_min_epi8(__m128i __V) {
|
|
3044
2927
|
return __builtin_reduce_min((__v16qs)__V);
|
|
3045
2928
|
}
|
|
3046
2929
|
|
|
3047
|
-
static __inline__ unsigned char
|
|
2930
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3048
2931
|
_mm_reduce_min_epu8(__m128i __V) {
|
|
3049
2932
|
return __builtin_reduce_min((__v16qu)__V);
|
|
3050
2933
|
}
|
|
3051
2934
|
|
|
3052
|
-
static __inline__ signed char
|
|
2935
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3053
2936
|
_mm_mask_reduce_max_epi8(__mmask16 __M, __m128i __V) {
|
|
3054
2937
|
__V = _mm_mask_mov_epi8(_mm_set1_epi8(-127-1), __M, __V);
|
|
3055
2938
|
return __builtin_reduce_max((__v16qs)__V);
|
|
3056
2939
|
}
|
|
3057
2940
|
|
|
3058
|
-
static __inline__ unsigned char
|
|
2941
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3059
2942
|
_mm_mask_reduce_max_epu8(__mmask16 __M, __m128i __V) {
|
|
3060
2943
|
__V = _mm_maskz_mov_epi8(__M, __V);
|
|
3061
2944
|
return __builtin_reduce_max((__v16qu)__V);
|
|
3062
2945
|
}
|
|
3063
2946
|
|
|
3064
|
-
static __inline__ signed char
|
|
2947
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3065
2948
|
_mm_mask_reduce_min_epi8(__mmask16 __M, __m128i __V) {
|
|
3066
2949
|
__V = _mm_mask_mov_epi8(_mm_set1_epi8(127), __M, __V);
|
|
3067
2950
|
return __builtin_reduce_min((__v16qs)__V);
|
|
3068
2951
|
}
|
|
3069
2952
|
|
|
3070
|
-
static __inline__ unsigned char
|
|
2953
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3071
2954
|
_mm_mask_reduce_min_epu8(__mmask16 __M, __m128i __V) {
|
|
3072
2955
|
__V = _mm_mask_mov_epi8(_mm_set1_epi8(-1), __M, __V);
|
|
3073
2956
|
return __builtin_reduce_min((__v16qu)__V);
|
|
3074
2957
|
}
|
|
3075
2958
|
|
|
3076
|
-
static __inline__ signed char
|
|
2959
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3077
2960
|
_mm256_reduce_add_epi8(__m256i __W) {
|
|
3078
2961
|
return __builtin_reduce_add((__v32qs)__W);
|
|
3079
2962
|
}
|
|
3080
2963
|
|
|
3081
|
-
static __inline__ signed char
|
|
2964
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3082
2965
|
_mm256_reduce_mul_epi8(__m256i __W) {
|
|
3083
2966
|
return __builtin_reduce_mul((__v32qs)__W);
|
|
3084
2967
|
}
|
|
3085
2968
|
|
|
3086
|
-
static __inline__ signed char
|
|
2969
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3087
2970
|
_mm256_reduce_and_epi8(__m256i __W) {
|
|
3088
2971
|
return __builtin_reduce_and((__v32qs)__W);
|
|
3089
2972
|
}
|
|
3090
2973
|
|
|
3091
|
-
static __inline__ signed char
|
|
2974
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3092
2975
|
_mm256_reduce_or_epi8(__m256i __W) {
|
|
3093
2976
|
return __builtin_reduce_or((__v32qs)__W);
|
|
3094
2977
|
}
|
|
3095
2978
|
|
|
3096
|
-
static __inline__ signed char
|
|
2979
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3097
2980
|
_mm256_mask_reduce_add_epi8(__mmask32 __M, __m256i __W) {
|
|
3098
2981
|
__W = _mm256_maskz_mov_epi8(__M, __W);
|
|
3099
2982
|
return __builtin_reduce_add((__v32qs)__W);
|
|
3100
2983
|
}
|
|
3101
2984
|
|
|
3102
|
-
static __inline__ signed char
|
|
2985
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3103
2986
|
_mm256_mask_reduce_mul_epi8(__mmask32 __M, __m256i __W) {
|
|
3104
2987
|
__W = _mm256_mask_mov_epi8(_mm256_set1_epi8(1), __M, __W);
|
|
3105
2988
|
return __builtin_reduce_mul((__v32qs)__W);
|
|
3106
2989
|
}
|
|
3107
2990
|
|
|
3108
|
-
static __inline__ signed char
|
|
2991
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3109
2992
|
_mm256_mask_reduce_and_epi8(__mmask32 __M, __m256i __W) {
|
|
3110
2993
|
__W = _mm256_mask_mov_epi8(_mm256_set1_epi8(-1), __M, __W);
|
|
3111
2994
|
return __builtin_reduce_and((__v32qs)__W);
|
|
3112
2995
|
}
|
|
3113
2996
|
|
|
3114
|
-
static __inline__ signed char
|
|
2997
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3115
2998
|
_mm256_mask_reduce_or_epi8(__mmask32 __M, __m256i __W) {
|
|
3116
2999
|
__W = _mm256_maskz_mov_epi8(__M, __W);
|
|
3117
3000
|
return __builtin_reduce_or((__v32qs)__W);
|
|
3118
3001
|
}
|
|
3119
3002
|
|
|
3120
|
-
static __inline__ signed char
|
|
3003
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3121
3004
|
_mm256_reduce_max_epi8(__m256i __V) {
|
|
3122
3005
|
return __builtin_reduce_max((__v32qs)__V);
|
|
3123
3006
|
}
|
|
3124
3007
|
|
|
3125
|
-
static __inline__ unsigned char
|
|
3008
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3126
3009
|
_mm256_reduce_max_epu8(__m256i __V) {
|
|
3127
3010
|
return __builtin_reduce_max((__v32qu)__V);
|
|
3128
3011
|
}
|
|
3129
3012
|
|
|
3130
|
-
static __inline__ signed char
|
|
3013
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3131
3014
|
_mm256_reduce_min_epi8(__m256i __V) {
|
|
3132
3015
|
return __builtin_reduce_min((__v32qs)__V);
|
|
3133
3016
|
}
|
|
3134
3017
|
|
|
3135
|
-
static __inline__ unsigned char
|
|
3018
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3136
3019
|
_mm256_reduce_min_epu8(__m256i __V) {
|
|
3137
3020
|
return __builtin_reduce_min((__v32qu)__V);
|
|
3138
3021
|
}
|
|
3139
3022
|
|
|
3140
|
-
static __inline__ signed char
|
|
3023
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3141
3024
|
_mm256_mask_reduce_max_epi8(__mmask32 __M, __m256i __V) {
|
|
3142
3025
|
__V = _mm256_mask_mov_epi8(_mm256_set1_epi8(-127-1), __M, __V);
|
|
3143
3026
|
return __builtin_reduce_max((__v32qs)__V);
|
|
3144
3027
|
}
|
|
3145
3028
|
|
|
3146
|
-
static __inline__ unsigned char
|
|
3029
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3147
3030
|
_mm256_mask_reduce_max_epu8(__mmask32 __M, __m256i __V) {
|
|
3148
3031
|
__V = _mm256_maskz_mov_epi8(__M, __V);
|
|
3149
3032
|
return __builtin_reduce_max((__v32qu)__V);
|
|
3150
3033
|
}
|
|
3151
3034
|
|
|
3152
|
-
static __inline__ signed char
|
|
3035
|
+
static __inline__ signed char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3153
3036
|
_mm256_mask_reduce_min_epi8(__mmask32 __M, __m256i __V) {
|
|
3154
3037
|
__V = _mm256_mask_mov_epi8(_mm256_set1_epi8(127), __M, __V);
|
|
3155
3038
|
return __builtin_reduce_min((__v32qs)__V);
|
|
3156
3039
|
}
|
|
3157
3040
|
|
|
3158
|
-
static __inline__ unsigned char
|
|
3041
|
+
static __inline__ unsigned char __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3159
3042
|
_mm256_mask_reduce_min_epu8(__mmask32 __M, __m256i __V) {
|
|
3160
3043
|
__V = _mm256_mask_mov_epi8(_mm256_set1_epi8(-1), __M, __V);
|
|
3161
3044
|
return __builtin_reduce_min((__v32qu)__V);
|
|
@@ -3163,5 +3046,7 @@ _mm256_mask_reduce_min_epu8(__mmask32 __M, __m256i __V) {
|
|
|
3163
3046
|
|
|
3164
3047
|
#undef __DEFAULT_FN_ATTRS128
|
|
3165
3048
|
#undef __DEFAULT_FN_ATTRS256
|
|
3049
|
+
#undef __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3050
|
+
#undef __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3166
3051
|
|
|
3167
3052
|
#endif /* __AVX512VLBWINTRIN_H */
|