PATH:
lib64
/
llvm17
/
lib
/
clang
/
17
/
include
/*===---- velintrin_approx.h - VEL intrinsics helper for VE ----------------=== * * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. * See https://llvm.org/LICENSE.txt for license information. * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception * *===-----------------------------------------------------------------------=== */ #ifndef __VEL_INTRIN_APPROX_H__ #define __VEL_INTRIN_APPROX_H__ static inline __vr _vel_approx_vfdivs_vvvl(__vr v0, __vr v1, int l) { float s0; __vr v2, v3, v4, v5; v5 = _vel_vrcps_vvl(v1, l); s0 = 1.0; v4 = _vel_vfnmsbs_vsvvl(s0, v1, v5, l); v3 = _vel_vfmads_vvvvl(v5, v5, v4, l); v2 = _vel_vfmuls_vvvl(v0, v3, l); v4 = _vel_vfnmsbs_vvvvl(v0, v2, v1, l); v2 = _vel_vfmads_vvvvl(v2, v5, v4, l); v0 = _vel_vfnmsbs_vvvvl(v0, v2, v1, l); v0 = _vel_vfmads_vvvvl(v2, v3, v0, l); return v0; } static inline __vr _vel_approx_pvfdiv_vvvl(__vr v0, __vr v1, int l) { float s0; __vr v2, v3, v4, v5; v5 = _vel_pvrcp_vvl(v1, l); s0 = 1.0; v4 = _vel_pvfnmsb_vsvvl(s0, v1, v5, l); v3 = _vel_pvfmad_vvvvl(v5, v5, v4, l); v2 = _vel_pvfmul_vvvl(v0, v3, l); v4 = _vel_pvfnmsb_vvvvl(v0, v2, v1, l); v2 = _vel_pvfmad_vvvvl(v2, v5, v4, l); v0 = _vel_pvfnmsb_vvvvl(v0, v2, v1, l); v0 = _vel_pvfmad_vvvvl(v2, v3, v0, l); return v0; } static inline __vr _vel_approx_vfdivs_vsvl(float s0, __vr v0, int l) { float s1; __vr v1, v2, v3, v4; v4 = _vel_vrcps_vvl(v0, l); s1 = 1.0; v2 = _vel_vfnmsbs_vsvvl(s1, v0, v4, l); v2 = _vel_vfmads_vvvvl(v4, v4, v2, l); v1 = _vel_vfmuls_vsvl(s0, v2, l); v3 = _vel_vfnmsbs_vsvvl(s0, v1, v0, l); v1 = _vel_vfmads_vvvvl(v1, v4, v3, l); v3 = _vel_vfnmsbs_vsvvl(s0, v1, v0, l); v0 = _vel_vfmads_vvvvl(v1, v2, v3, l); return v0; } static inline __vr _vel_approx_vfdivs_vvsl(__vr v0, float s0, int l) { float s1; __vr v1, v2; s1 = 1.0f / s0; v1 = _vel_vfmuls_vsvl(s1, v0, l); v2 = _vel_vfnmsbs_vvsvl(v0, s0, v1, l); v0 = _vel_vfmads_vvsvl(v1, s1, v2, l); return v0; } static inline __vr _vel_approx_vfdivd_vsvl(double s0, __vr v0, int l) { __vr v1, v2, v3; v2 = _vel_vrcpd_vvl(v0, l); double s1 = 1.0; v3 = _vel_vfnmsbd_vsvvl(s1, v0, v2, l); v2 = _vel_vfmadd_vvvvl(v2, v2, v3, l); v1 = _vel_vfnmsbd_vsvvl(s1, v0, v2, l); v1 = _vel_vfmadd_vvvvl(v2, v2, v1, l); v1 = _vel_vaddul_vsvl(1, v1, l); v3 = _vel_vfnmsbd_vsvvl(s1, v0, v1, l); v3 = _vel_vfmadd_vvvvl(v1, v1, v3, l); v1 = _vel_vfmuld_vsvl(s0, v3, l); v0 = _vel_vfnmsbd_vsvvl(s0, v1, v0, l); v0 = _vel_vfmadd_vvvvl(v1, v3, v0, l); return v0; } static inline __vr _vel_approx_vfsqrtd_vvl(__vr v0, int l) { double s0, s1; __vr v1, v2, v3; v2 = _vel_vrsqrtdnex_vvl(v0, l); v1 = _vel_vfmuld_vvvl(v0, v2, l); s0 = 1.0; s1 = 0.5; v3 = _vel_vfnmsbd_vsvvl(s0, v1, v2, l); v3 = _vel_vfmuld_vsvl(s1, v3, l); v2 = _vel_vfmadd_vvvvl(v2, v2, v3, l); v1 = _vel_vfmuld_vvvl(v0, v2, l); v3 = _vel_vfnmsbd_vsvvl(s0, v1, v2, l); v3 = _vel_vfmuld_vsvl(s1, v3, l); v0 = _vel_vfmadd_vvvvl(v1, v1, v3, l); return v0; } static inline __vr _vel_approx_vfsqrts_vvl(__vr v0, int l) { float s0, s1; __vr v1, v2, v3; v0 = _vel_vcvtds_vvl(v0, l); v2 = _vel_vrsqrtdnex_vvl(v0, l); v1 = _vel_vfmuld_vvvl(v0, v2, l); s0 = 1.0; s1 = 0.5; v3 = _vel_vfnmsbd_vsvvl(s0, v1, v2, l); v3 = _vel_vfmuld_vsvl(s1, v3, l); v2 = _vel_vfmadd_vvvvl(v2, v2, v3, l); v1 = _vel_vfmuld_vvvl(v0, v2, l); v3 = _vel_vfnmsbd_vsvvl(s0, v1, v2, l); v3 = _vel_vfmuld_vsvl(s1, v3, l); v0 = _vel_vfmadd_vvvvl(v1, v1, v3, l); v0 = _vel_vcvtsd_vvl(v0, l); return v0; } #endif
[+]
..
[-] riscv_ntlh.h
[edit]
[-] avx512vlbitalgintrin.h
[edit]
[-] avxifmaintrin.h
[edit]
[-] mm_malloc.h
[edit]
[-] avx512vlbf16intrin.h
[edit]
[-] mwaitxintrin.h
[edit]
[-] adxintrin.h
[edit]
[-] arm_cmse.h
[edit]
[-] __clang_cuda_intrinsics.h
[edit]
[-] module.modulemap
[edit]
[-] crc32intrin.h
[edit]
[+]
cuda_wrappers
[-] msa.h
[edit]
[-] __clang_hip_stdlib.h
[edit]
[-] arm_acle.h
[edit]
[-] arm_neon_sve_bridge.h
[edit]
[-] xsaveintrin.h
[edit]
[-] velintrin_gen.h
[edit]
[-] stdatomic.h
[edit]
[-] hvx_hexagon_protos.h
[edit]
[-] uintrintrin.h
[edit]
[-] fxsrintrin.h
[edit]
[-] sha512intrin.h
[edit]
[-] __clang_cuda_texture_intrinsics.h
[edit]
[-] iso646.h
[edit]
[-] unwind.h
[edit]
[-] avx512vlbwintrin.h
[edit]
[-] avxvnniint8intrin.h
[edit]
[-] avx512erintrin.h
[edit]
[-] avxvnniintrin.h
[edit]
[-] ia32intrin.h
[edit]
[-] rdseedintrin.h
[edit]
[-] prfchiintrin.h
[edit]
[-] amxcomplexintrin.h
[edit]
[-] clflushoptintrin.h
[edit]
[-] htmxlintrin.h
[edit]
[-] avx512fintrin.h
[edit]
[-] gfniintrin.h
[edit]
[-] arm_cde.h
[edit]
[+]
ppc_wrappers
[-] amxfp16intrin.h
[edit]
[-] velintrin_approx.h
[edit]
[-] avx512pfintrin.h
[edit]
[-] stdarg.h
[edit]
[-] cmpccxaddintrin.h
[edit]
[-] avx512vlintrin.h
[edit]
[-] __clang_hip_math.h
[edit]
[-] vecintrin.h
[edit]
[-] xtestintrin.h
[edit]
[-] __wmmintrin_aes.h
[edit]
[-] arm_neon.h
[edit]
[-] immintrin.h
[edit]
[-] ammintrin.h
[edit]
[-] waitpkgintrin.h
[edit]
[-] vpclmulqdqintrin.h
[edit]
[-] fmaintrin.h
[edit]
[-] tsxldtrkintrin.h
[edit]
[-] prfchwintrin.h
[edit]
[-] avx512bitalgintrin.h
[edit]
[-] bmiintrin.h
[edit]
[-] __wmmintrin_pclmul.h
[edit]
[-] htmintrin.h
[edit]
[-] mm3dnow.h
[edit]
[-] __clang_cuda_builtin_vars.h
[edit]
[-] __clang_hip_runtime_wrapper.h
[edit]
[-] stdbool.h
[edit]
[-] altivec.h
[edit]
[-] wbnoinvdintrin.h
[edit]
[-] keylockerintrin.h
[edit]
[-] tgmath.h
[edit]
[-] hexagon_circ_brev_intrinsics.h
[edit]
[-] x86intrin.h
[edit]
[-] pkuintrin.h
[edit]
[-] avx512vbmivlintrin.h
[edit]
[-] avxneconvertintrin.h
[edit]
[-] __clang_hip_cmath.h
[edit]
[-] sgxintrin.h
[edit]
[-] f16cintrin.h
[edit]
[-] opencl-c-base.h
[edit]
[-] cpuid.h
[edit]
[-] raointintrin.h
[edit]
[-] builtins.h
[edit]
[-] emmintrin.h
[edit]
[-] smmintrin.h
[edit]
[-] vaesintrin.h
[edit]
[-] larchintrin.h
[edit]
[-] avx512ifmaintrin.h
[edit]
[-] intrin.h
[edit]
[-] avx512vlvp2intersectintrin.h
[edit]
[-] fma4intrin.h
[edit]
[-] pmmintrin.h
[edit]
[-] __clang_hip_libdevice_declares.h
[edit]
[-] limits.h
[edit]
[-] clwbintrin.h
[edit]
[-] rtmintrin.h
[edit]
[-] mmintrin.h
[edit]
[-] stddef.h
[edit]
[-] invpcidintrin.h
[edit]
[-] avx512vp2intersectintrin.h
[edit]
[-] cet.h
[edit]
[-] xopintrin.h
[edit]
[-] avx512vlvnniintrin.h
[edit]
[-] avx512vlfp16intrin.h
[edit]
[-] stdint.h
[edit]
[-] arm64intr.h
[edit]
[-] sm4intrin.h
[edit]
[-] avx512vnniintrin.h
[edit]
[-] avx2intrin.h
[edit]
[-] movdirintrin.h
[edit]
[-] tbmintrin.h
[edit]
[-] arm_mve.h
[edit]
[-] avx512ifmavlintrin.h
[edit]
[-] amxintrin.h
[edit]
[-] opencl-c.h
[edit]
[-] stdalign.h
[edit]
[-] __clang_cuda_device_functions.h
[edit]
[-] pconfigintrin.h
[edit]
[-] avx512fp16intrin.h
[edit]
[-] inttypes.h
[edit]
[-] arm_bf16.h
[edit]
[-] __clang_cuda_math_forward_declares.h
[edit]
[-] vadefs.h
[edit]
[-] shaintrin.h
[edit]
[-] hexagon_protos.h
[edit]
[-] ptwriteintrin.h
[edit]
[-] xsaveoptintrin.h
[edit]
[-] enqcmdintrin.h
[edit]
[-] x86gprintrin.h
[edit]
[-] tmmintrin.h
[edit]
[-] stdnoreturn.h
[edit]
[-] avx512bf16intrin.h
[edit]
[-] varargs.h
[edit]
[-] s390intrin.h
[edit]
[-] avx512vbmiintrin.h
[edit]
[+]
openmp_wrappers
[-] wmmintrin.h
[edit]
[-] __clang_cuda_cmath.h
[edit]
[-] clzerointrin.h
[edit]
[-] xsavesintrin.h
[edit]
[-] __clang_cuda_libdevice_declares.h
[edit]
[-] nmmintrin.h
[edit]
[-] wasm_simd128.h
[edit]
[-] xsavecintrin.h
[edit]
[-] avx512dqintrin.h
[edit]
[-] lwpintrin.h
[edit]
[-] serializeintrin.h
[edit]
[-] arm_sme_draft_spec_subject_to_change.h
[edit]
[-] avxintrin.h
[edit]
[-] __stddef_max_align_t.h
[edit]
[-] sm3intrin.h
[edit]
[-] velintrin.h
[edit]
[-] __clang_cuda_complex_builtins.h
[edit]
[-] armintr.h
[edit]
[-] avx512cdintrin.h
[edit]
[-] float.h
[edit]
[-] avx512vbmi2intrin.h
[edit]
[-] lzcntintrin.h
[edit]
[-] sifive_vector.h
[edit]
[-] rdpruintrin.h
[edit]
[-] arm_sve.h
[edit]
[-] avx512vpopcntdqintrin.h
[edit]
[-] xmmintrin.h
[edit]
[-] hresetintrin.h
[edit]
[-] bmi2intrin.h
[edit]
[-] hexagon_types.h
[edit]
[-] avx512bwintrin.h
[edit]
[-] cetintrin.h
[edit]
[-] __clang_cuda_math.h
[edit]
[-] avx512vlvbmi2intrin.h
[edit]
[-] arm_fp16.h
[edit]
[-] avx512vpopcntdqvlintrin.h
[edit]
[-] avxvnniint16intrin.h
[edit]
[+]
llvm_libc_wrappers
[-] __clang_cuda_runtime_wrapper.h
[edit]
[-] cldemoteintrin.h
[edit]
[-] avx512vldqintrin.h
[edit]
[-] avx512vlcdintrin.h
[edit]
[-] popcntintrin.h
[edit]