1// Copyright 2020 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$assert BATCH_TILE % 16 == 0 7$assert BATCH_TILE >= 16 8$assert DIV_ALGO in ["div", "nr1fma", "nr1fma1adj"] 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10$SIMD_TILE = BATCH_TILE // 16 11#include <assert.h> 12 13#include <immintrin.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/intrinsics-polyfill.h> 17#include <xnnpack/vunary.h> 18 19 20void xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_${DIV_ALGO}_x${BATCH_TILE}( 21 size_t n, 22 const float* x, 23 float* y, 24 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) 25{ 26 assert(n % sizeof(float) == 0); 27 28 const __m512i vsign_mask = _mm512_set1_epi32((int) params->avx512_rr1_lut16_p3.sign_mask); 29 const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr1_lut16_p3.magic_bias); 30 const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_lut16_p3.log2e); 31 const __m512 vtable = _mm512_load_ps(params->avx512_rr1_lut16_p3.table); 32 const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.minus_ln2); 33 const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c3); 34 const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c2); 35 const __m512 vone = _mm512_set1_ps(params->avx512_rr1_lut16_p3.one); 36 37 $if BATCH_TILE > 16: 38 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 39 const __m512 vx${ABC[0]} = _mm512_loadu_ps(x); 40 $for N in range(1, SIMD_TILE): 41 const __m512 vx${ABC[N]} = _mm512_loadu_ps(x + ${N * 16}); 42 x += ${BATCH_TILE}; 43 44 $for N in range(SIMD_TILE): 45 const __m512 vz${ABC[N]} = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx${ABC[N]}), vsign_mask)); 46 47 $for N in range(SIMD_TILE): 48 __m512 vn${ABC[N]} = _mm512_fmadd_ps(vz${ABC[N]}, vlog2e, vmagic_bias); 49 50 $for N in range(SIMD_TILE): 51 const __m512 vl${ABC[N]} = _mm512_permutexvar_ps(_mm512_castps_si512(vn${ABC[N]}), vtable); 52 53 $for N in range(SIMD_TILE): 54 vn${ABC[N]} = _mm512_sub_ps(vn${ABC[N]}, vmagic_bias); 55 56 $for N in range(SIMD_TILE): 57 __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2, vz${ABC[N]}); 58 59 $for N in range(SIMD_TILE): 60 __m512 vp${ABC[N]} = _mm512_fmadd_ps(vt${ABC[N]}, vc3, vc2); 61 62 $for N in range(SIMD_TILE): 63 vp${ABC[N]} = _mm512_mul_ps(vp${ABC[N]}, vt${ABC[N]}); 64 65 $for N in range(SIMD_TILE): 66 vp${ABC[N]} = _mm512_fmadd_ps(vt${ABC[N]}, vp${ABC[N]}, vt${ABC[N]}); 67 68 $for N in range(SIMD_TILE): 69 vp${ABC[N]} = _mm512_fmadd_ps(vl${ABC[N]}, vp${ABC[N]}, vl${ABC[N]}); 70 71 $for N in range(SIMD_TILE): 72 const __m512 ve${ABC[N]} = _mm512_scalef_ps(vp${ABC[N]}, vn${ABC[N]}); 73 74 $for N in range(SIMD_TILE): 75 const __m512 vd${ABC[N]} = _mm512_add_ps(ve${ABC[N]}, vone); 76 77 $if DIV_ALGO == "div": 78 $for N in range(SIMD_TILE): 79 __m512 vf${ABC[N]} = _mm512_div_ps(ve${ABC[N]}, vd${ABC[N]}); 80 $else: 81 $for N in range(SIMD_TILE): 82 __m512 vr${ABC[N]} = _mm512_rcp14_ps(vd${ABC[N]}); 83 84 $for N in range(SIMD_TILE): 85 vr${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr${ABC[N]}, vd${ABC[N]}, vone), vr${ABC[N]}, vr${ABC[N]}); 86 87 $for N in range(SIMD_TILE): 88 __m512 vf${ABC[N]} = _mm512_mul_ps(ve${ABC[N]}, vr${ABC[N]}); 89 90 $if DIV_ALGO == "nr1fma1adj": 91 $for N in range(SIMD_TILE): 92 vf${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf${ABC[N]}, vd${ABC[N]}, ve${ABC[N]}), vr${ABC[N]}, vf${ABC[N]}); 93 94 $for N in range(SIMD_TILE): 95 vf${ABC[N]} = _mm512_mask_sub_ps(vf${ABC[N]}, _mm512_testn_epi32_mask(_mm512_castps_si512(vx${ABC[N]}), vsign_mask), vone, vf${ABC[N]}); 96 97 _mm512_storeu_ps(y, vf${ABC[0]}); 98 $for N in range(1, SIMD_TILE): 99 _mm512_storeu_ps(y + ${N * 16}, vf${ABC[N]}); 100 y += ${BATCH_TILE}; 101 } 102 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { 103 const __m512 vx = _mm512_loadu_ps(x); 104 x += 16; 105 106 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); 107 108 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); 109 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); 110 vn = _mm512_sub_ps(vn, vmagic_bias); 111 112 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 113 114 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); 115 vp = _mm512_mul_ps(vp, vt); 116 vp = _mm512_fmadd_ps(vt, vp, vt); 117 vp = _mm512_fmadd_ps(vl, vp, vl); 118 119 const __m512 ve = _mm512_scalef_ps(vp, vn); 120 const __m512 vd = _mm512_add_ps(ve, vone); 121 122 $if DIV_ALGO == "div": 123 __m512 vf = _mm512_div_ps(ve, vd); 124 $else: 125 __m512 vr = _mm512_rcp14_ps(vd); 126 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); 127 128 __m512 vf = _mm512_mul_ps(ve, vr); 129 $if DIV_ALGO == "nr1fma1adj": 130 vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf); 131 132 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf); 133 134 _mm512_storeu_ps(y, vf); 135 y += 16; 136 } 137 if XNN_UNLIKELY(n != 0) { 138 assert(n >= 1 * sizeof(float)); 139 assert(n <= 15 * sizeof(float)); 140 141 // Prepare mask for valid 32-bit elements (depends on n). 142 n >>= 2 /* log2(sizeof(float)) */; 143 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); 144 145 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); 146 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); 147 148 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); 149 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); 150 vn = _mm512_sub_ps(vn, vmagic_bias); 151 152 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 153 154 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); 155 vp = _mm512_mul_ps(vp, vt); 156 vp = _mm512_fmadd_ps(vt, vp, vt); 157 vp = _mm512_fmadd_ps(vl, vp, vl); 158 159 const __m512 ve = _mm512_scalef_ps(vp, vn); 160 const __m512 vd = _mm512_add_ps(ve, vone); 161 162 $if DIV_ALGO == "div": 163 __m512 vf = _mm512_div_ps(ve, vd); 164 $else: 165 __m512 vr = _mm512_rcp14_ps(vd); 166 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); 167 168 __m512 vf = _mm512_mul_ps(ve, vr); 169 $if DIV_ALGO == "nr1fma1adj": 170 vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf); 171 172 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf); 173 174 _mm512_mask_storeu_ps(y, vmask, vf); 175 } 176} 177