chris@163: /**
chris@163:  * Copyright (c) 2014, 2015, Enzien Audio Ltd.
chris@163:  *
chris@163:  * Permission to use, copy, modify, and/or distribute this software for any
chris@163:  * purpose with or without fee is hereby granted, provided that the above
chris@163:  * copyright notice and this permission notice appear in all copies.
chris@163:  *
chris@163:  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
chris@163:  * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
chris@163:  * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
chris@163:  * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
chris@163:  * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
chris@163:  * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
chris@163:  * PERFORMANCE OF THIS SOFTWARE.
chris@163:  */
chris@163: 
chris@163: #ifndef _HEAVY_MATH_H_
chris@163: #define _HEAVY_MATH_H_
chris@163: 
chris@163: #include "Utils.h"
chris@163: 
chris@163: // https://software.intel.com/sites/landingpage/IntrinsicsGuide/
chris@163: // https://gcc.gnu.org/onlinedocs/gcc-4.8.1/gcc/ARM-NEON-Intrinsics.html
chris@163: // http://codesuppository.blogspot.co.uk/2015/02/sse2neonh-porting-guide-and-header-file.html
chris@163: 
chris@163: static inline void __hv_zero_f(hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_setzero_ps();
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_setzero_ps();
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vdupq_n_f32(0.0f);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_load_f(float *bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_load_ps(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_load_ps(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vld1q_f32(bIn);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = *bIn;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_store_f(float *bOut, hv_bInf_t bIn) {
chris@163: #if HV_SIMD_AVX
chris@163:   _mm256_store_ps(bOut, bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   _mm_store_ps(bOut, bIn);
chris@163: #elif HV_SIMD_NEON
chris@163:   vst1q_f32(bOut, bIn);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = bIn;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_log_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_log_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_log_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_log_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn > 0.0f) ? hv_log_f(bIn) : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_log10_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_log10_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_log10_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_log10_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn > 0.0f) ? hv_log10_f(bIn) : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_log2_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_log2_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_log2_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_log2_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn > 0.0f) ? hv_log2_f(bIn) : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // NOTE(mhroth): this is a pretty ghetto implementation
chris@163: static inline void __hv_cos_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_set_ps(
chris@163:       hv_cos_f(bIn[7]), hv_cos_f(bIn[6]), hv_cos_f(bIn[5]), hv_cos_f(bIn[4]),
chris@163:       hv_cos_f(bIn[3]), hv_cos_f(bIn[2]), hv_cos_f(bIn[1]), hv_cos_f(bIn[0]));
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_set_ps(hv_cos_f(bIn[3]), hv_cos_f(bIn[2]), hv_cos_f(bIn[1]), hv_cos_f(bIn[0]));
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = (float32x4_t) {hv_cos_f(bIn[0]), hv_cos_f(bIn[1]), hv_cos_f(bIn[2]), hv_cos_f(bIn[3])};
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_cos_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_acos_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_acos_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_acos_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_acos_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_acos_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_cosh_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_cosh_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_cosh_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_cosh_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_cosh_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_acosh_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_acosh_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_acosh_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_acosh_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_acosh_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_sin_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_sin_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_sin_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_sin_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_sin_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_asin_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_asin_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_asin_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_asin_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_asin_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_sinh_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_sinh_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_sinh_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_sinh_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_sinh_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_asinh_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_asinh_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_asinh_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_asinh_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_asinh_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_tan_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_tan_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_tan_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_tan_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_tan_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_atan_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_atan_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_atan_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_atan_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_atan_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_atan2_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_atan2_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_atan2_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_atan2_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_atan2_f(bIn0, bIn1);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_tanh_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_tanh_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_tanh_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_tanh_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_tanh_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_atanh_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_atanh_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_atanh_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_atanh_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_atanh_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // NOTE(mhroth): use of sqrt is absolute and total MURDER. Make do with recipocal sqrt if possible!!
chris@163: static inline void __hv_sqrt_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_sqrt_ps(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_sqrt_ps(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_sqrt_f() numerical results may be inexact
chris@163:   *bOut = vrecpeq_f32(vrsqrteq_f32(bIn));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_sqrt_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_rsqrt_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_rsqrt_ps(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_rsqrt_ps(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_rsqrt_f() numerical results may be inexact
chris@163:   *bOut = vrsqrteq_f32(bIn);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = 1.0f/hv_sqrt_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_abs_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_andnot_ps(_mm_set1_ps(-0.0f), bIn); // == 1 << 31
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vabsq_f32(bIn);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_abs_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_exp_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_exp_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_exp_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_exp_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_exp_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_ceil_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_ceil_ps(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_ceil_ps(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163: #if __ARM_ARCH >= 8
chris@163:   *bOut = vrndpq_f32(bIn);
chris@163: #else
chris@163: #warning A slow NEON implementation of __hv_ceil_f() is being used because the necessary intrinsic cannot be found. It is only available in ARMv8.
chris@163:   *bOut = (float32x4_t) {hv_ceil_f(bIn[0]), hv_ceil_f(bIn[1]), hv_ceil_f(bIn[2]), hv_ceil_f(bIn[3])};
chris@163: #endif // vrndpq_f32
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_ceil_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_floor_f(hv_bInf_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_floor_ps(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_floor_ps(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163: #if __ARM_ARCH >= 8
chris@163:   *bOut = vrndmq_f32(bIn);
chris@163: #else
chris@163: #warning A slow NEON implementation of __hv_floor_f() is being used because the necessary intrinsic cannot be found. It is only available in ARMv8.
chris@163:   *bOut = (float32x4_t) {hv_floor_f(bIn[0]), hv_floor_f(bIn[1]), hv_floor_f(bIn[2]), hv_floor_f(bIn[3])};
chris@163: #endif // vrndmq_f32
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_floor_f(bIn);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __add~f
chris@163: static inline void __hv_add_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_add_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_add_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vaddq_f32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = bIn0 + bIn1;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __add~i
chris@163: static inline void __hv_add_i(hv_bIni_t bIn0, hv_bIni_t bIn1, hv_bOuti_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   __m128i x = _mm_add_epi32(_mm256_castsi256_si128(bIn0), _mm256_castsi256_si128(bIn1));
chris@163:   __m128i y = _mm_add_epi32(_mm256_extractf128_si256(bIn0, 1), _mm256_extractf128_si256(bIn1, 1));
chris@163:   *bOut = _mm256_insertf128_si256(_mm256_castsi128_si256(x), y, 1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_add_epi32(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vaddq_s32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = bIn0 + bIn1;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __sub~f
chris@163: static inline void __hv_sub_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_sub_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_sub_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vsubq_f32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = bIn0 - bIn1;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __mul~f
chris@163: static inline void __hv_mul_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_mul_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_mul_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vmulq_f32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = bIn0 * bIn1;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __*~i
chris@163: static inline void __hv_mul_i(hv_bIni_t bIn0, hv_bIni_t bIn1, hv_bOuti_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   __m128i x = _mm_mullo_epi32(_mm256_castsi256_si128(bIn0), _mm256_castsi256_si128(bIn1));
chris@163:   __m128i y = _mm_mullo_epi32(_mm256_extractf128_si256(bIn0, 1), _mm256_extractf128_si256(bIn1, 1));
chris@163:   *bOut = _mm256_insertf128_si256(_mm256_castsi128_si256(x), y, 1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_mullo_epi32(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vmulq_s32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = bIn0 * bIn1;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __cast~if
chris@163: static inline void __hv_cast_if(hv_bIni_t bIn, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cvtepi32_ps(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cvtepi32_ps(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vcvtq_f32_s32(bIn);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (float) bIn;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // __cast~fi
chris@163: static inline void __hv_cast_fi(hv_bInf_t bIn, hv_bOuti_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cvtps_epi32(bIn);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cvtps_epi32(bIn);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vcvtq_s32_f32(bIn);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (int) bIn;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_div_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_div_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_div_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_div_f() numerical results may be inexact
chris@163:   *bOut = vmulq_f32(bIn0, vrecpeq_f32(bIn1));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn1 != 0.0f) ? (bIn0 / bIn1) : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_min_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_min_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_min_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vminq_f32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_min_f(bIn0, bIn1);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_min_i(hv_bIni_t bIn0, hv_bIni_t bIn1, hv_bOuti_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   __m128i x = _mm_min_epi32(_mm256_castsi256_si128(bIn0), _mm256_castsi256_si128(bIn1));
chris@163:   __m128i y = _mm_min_epi32(_mm256_extractf128_si256(bIn0, 1), _mm256_extractf128_si256(bIn1, 1));
chris@163:   *bOut = _mm256_insertf128_si256(_mm256_castsi128_si256(x), y, 1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_min_epi32(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vminq_s32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_min_i(bIn0, bIn1);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_max_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_max_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_max_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vmaxq_f32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_max_f(bIn0, bIn1);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_max_i(hv_bIni_t bIn0, hv_bIni_t bIn1, hv_bOuti_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   __m128i x = _mm_max_epi32(_mm256_castsi256_si128(bIn0), _mm256_castsi256_si128(bIn1));
chris@163:   __m128i y = _mm_max_epi32(_mm256_extractf128_si256(bIn0, 1), _mm256_extractf128_si256(bIn1, 1));
chris@163:   *bOut = _mm256_insertf128_si256(_mm256_castsi128_si256(x), y, 1);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_max_epi32(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vmaxq_s32(bIn0, bIn1);
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_max_i(bIn0, bIn1);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_pow_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_set_ps(
chris@163:       hv_pow_f(bIn0[7], bIn1[7]),
chris@163:       hv_pow_f(bIn0[6], bIn1[6]),
chris@163:       hv_pow_f(bIn0[5], bIn1[5]),
chris@163:       hv_pow_f(bIn0[4], bIn1[4]),
chris@163:       hv_pow_f(bIn0[3], bIn1[3]),
chris@163:       hv_pow_f(bIn0[2], bIn1[2]),
chris@163:       hv_pow_f(bIn0[1], bIn1[1]),
chris@163:       hv_pow_f(bIn0[0], bIn1[0]));
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_set_ps(
chris@163:       hv_pow_f(bIn0[3], bIn1[3]),
chris@163:       hv_pow_f(bIn0[2], bIn1[2]),
chris@163:       hv_pow_f(bIn0[1], bIn1[1]),
chris@163:       hv_pow_f(bIn0[0], bIn1[0]));
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = (float32x4_t) {
chris@163:       hv_pow_f(bIn0[0], bIn1[0]),
chris@163:       hv_pow_f(bIn0[1], bIn1[1]),
chris@163:       hv_pow_f(bIn0[2], bIn1[2]),
chris@163:       hv_pow_f(bIn0[3], bIn1[3])};
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_pow_f(bIn0, bIn1);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_gt_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cmp_ps(bIn0, bIn1, _CMP_GT_OQ);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cmpgt_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vreinterpretq_f32_u32(vcgtq_f32(bIn0, bIn1));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn0 > bIn1) ? 1.0f : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_gte_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cmp_ps(bIn0, bIn1, _CMP_GE_OQ);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cmpge_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vreinterpretq_f32_u32(vcgeq_f32(bIn0, bIn1));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn0 >= bIn1) ? 1.0f : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_lt_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cmp_ps(bIn0, bIn1, _CMP_LT_OQ);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cmplt_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vreinterpretq_f32_u32(vcltq_f32(bIn0, bIn1));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn0 < bIn1) ? 1.0f : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_lte_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cmp_ps(bIn0, bIn1, _CMP_LE_OQ);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cmple_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vreinterpretq_f32_u32(vcleq_f32(bIn0, bIn1));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn0 <= bIn1) ? 1.0f : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_neq_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_cmp_ps(bIn0, bIn1, _CMP_NEQ_OQ);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_cmpneq_ps(bIn0, bIn1);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vreinterpretq_f32_u32(vmvnq_u32(vceqq_f32(bIn0, bIn1)));
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (bIn0 != bIn1) ? 1.0f : 0.0f;
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_xor_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #warning __hv_xor_f() not implemented
chris@163: #elif HV_SIMD_SSE
chris@163: #warning __hv_xor_f() not implemented
chris@163: #elif HV_SIMD_NEON
chris@163: #warning __hv_xor_f() not implemented
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = (float) (((int) bIn0) ^ ((int) bIn1));
chris@163: #endif
chris@163: }
chris@163: 
chris@163: static inline void __hv_and_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163:   *bOut = _mm256_and_ps(bIn1, bIn0);
chris@163: #elif HV_SIMD_SSE
chris@163:   *bOut = _mm_and_ps(bIn1, bIn0);
chris@163: #elif HV_SIMD_NEON
chris@163:   *bOut = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(bIn1), vreinterpretq_u32_f32(bIn0)));
chris@163: #else // HV_SIMD_NONE
chris@163:   if (bIn0 == 0.0f || bIn1 == 0.0f) *bOut = 0.0f;
chris@163:   else if (bIn0 == 1.0f) *bOut = bIn1;
chris@163:   else if (bIn1 == 1.0f) *bOut = bIn0;
chris@163:   else hv_assert(0); // NOTE(mhroth): floating point & is pretty much a bad idea, only used for if~
chris@163: #endif
chris@163: }
chris@163: 
chris@163: // bOut = (bIn0 * bIn1) + bIn2
chris@163: static inline void __hv_fma_f(hv_bInf_t bIn0, hv_bInf_t bIn1, hv_bInf_t bIn2, hv_bOutf_t bOut) {
chris@163: #if HV_SIMD_AVX
chris@163: #if HV_SIMD_FMA
chris@163:   *bOut = _mm256_fmadd_ps(bIn0, bIn1, bIn2);
chris@163: #else
chris@163:   *bOut = _mm256_add_ps(_mm256_mul_ps(bIn0, bIn1), bIn2);
chris@163: #endif // HV_SIMD_FMA
chris@163: #elif HV_SIMD_SSE
chris@163: #if HV_SIMD_FMA
chris@163:   *bOut = _mm_fmadd_ps(bIn0, bIn1, bIn2);
chris@163: #else
chris@163:   *bOut = _mm_add_ps(_mm_mul_ps(bIn0, bIn1), bIn2);
chris@163: #endif // HV_SIMD_FMA
chris@163: #elif HV_SIMD_NEON
chris@163: #if __ARM_ARCH >= 8
chris@163:   *bOut = vfmaq_f32(bIn2, bIn0, bIn1);
chris@163: #else
chris@163:   // NOTE(mhroth): it turns out, fma SUUUUCKS on lesser ARM architectures
chris@163:   // But in fact ideally fma would be disabled in ir2c for ARM architectures.
chris@163:   // LLVM does a much better job handling fma than we do.
chris@163:   *bOut = vaddq_f32(vmulq_f32(bIn0, bIn1), bIn2);
chris@163: #endif
chris@163: #else // HV_SIMD_NONE
chris@163:   *bOut = hv_fma_f(bIn0, bIn1, bIn2);
chris@163: #endif
chris@163: }
chris@163: 
chris@163: #endif // _HEAVY_MATH_H_