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$assert RR_STEPS in [1, 2] 9$assert DIV_ALGO in ["div", "nr1fma", "nr1fma1adj"] 10$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 11$SIMD_TILE = BATCH_TILE // 16 12#include <assert.h> 13 14#include <immintrin.h> 15 16#include <xnnpack/common.h> 17#include <xnnpack/intrinsics-polyfill.h> 18#include <xnnpack/vunary.h> 19 20 21void xnn_f32_sigmoid_ukernel__avx512f_rr${RR_STEPS}_lut16_p3_perm_scalef_${DIV_ALGO}_x${BATCH_TILE}( 22 size_t n, 23 const float* x, 24 float* y, 25 const void* params) 26{ 27 assert(n % sizeof(float) == 0); 28 29 const __m512i vsign_mask = _mm512_set1_epi32(0x80000000); 30 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); 31 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); 32 const __m512 vtable = _mm512_set_ps( 33 0x1.EA4AFAp+0f, 0x1.D5818Ep+0f, 0x1.C199BEp+0f, 0x1.AE89FAp+0f, 34 0x1.9C4918p+0f, 0x1.8ACE54p+0f, 0x1.7A1148p+0f, 0x1.6A09E6p+0f, 35 0x1.5AB07Ep+0f, 0x1.4BFDAEp+0f, 0x1.3DEA64p+0f, 0x1.306FE0p+0f, 36 0x1.2387A6p+0f, 0x1.172B84p+0f, 0x1.0B5586p+0f, 0x1.000000p+0f); 37 $if RR_STEPS == 1: 38 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); 39 $else: 40 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); 41 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); 42 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); 43 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); 44 const __m512 vone = _mm512_set1_ps(1.0f); 45 46 $if BATCH_TILE > 16: 47 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 48 const __m512 vx${ABC[0]} = _mm512_loadu_ps(x); 49 $for N in range(1, SIMD_TILE): 50 const __m512 vx${ABC[N]} = _mm512_loadu_ps(x + ${N * 16}); 51 x += ${BATCH_TILE}; 52 53 $for N in range(SIMD_TILE): 54 const __m512 vz${ABC[N]} = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx${ABC[N]}), vsign_mask)); 55 56 $for N in range(SIMD_TILE): 57 __m512 vn${ABC[N]} = _mm512_fmadd_ps(vz${ABC[N]}, vlog2e, vmagic_bias); 58 59 $for N in range(SIMD_TILE): 60 const __m512 vl${ABC[N]} = _mm512_permutexvar_ps(_mm512_castps_si512(vn${ABC[N]}), vtable); 61 62 $for N in range(SIMD_TILE): 63 vn${ABC[N]} = _mm512_sub_ps(vn${ABC[N]}, vmagic_bias); 64 65 $if RR_STEPS == 1: 66 $for N in range(SIMD_TILE): 67 __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2, vz${ABC[N]}); 68 $else: 69 $for N in range(SIMD_TILE): 70 __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2_hi, vz${ABC[N]}); 71 72 $for N in range(SIMD_TILE): 73 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2_lo, vt${ABC[N]}); 74 75 $for N in range(SIMD_TILE): 76 __m512 vp${ABC[N]} = _mm512_fmadd_ps(vt${ABC[N]}, vc3, vc2); 77 78 $for N in range(SIMD_TILE): 79 vp${ABC[N]} = _mm512_mul_ps(vp${ABC[N]}, vt${ABC[N]}); 80 81 $for N in range(SIMD_TILE): 82 vp${ABC[N]} = _mm512_fmadd_ps(vt${ABC[N]}, vp${ABC[N]}, vt${ABC[N]}); 83 84 $for N in range(SIMD_TILE): 85 vp${ABC[N]} = _mm512_fmadd_ps(vl${ABC[N]}, vp${ABC[N]}, vl${ABC[N]}); 86 87 $for N in range(SIMD_TILE): 88 const __m512 ve${ABC[N]} = _mm512_scalef_ps(vp${ABC[N]}, vn${ABC[N]}); 89 90 $for N in range(SIMD_TILE): 91 const __m512 vd${ABC[N]} = _mm512_add_ps(ve${ABC[N]}, vone); 92 93 $if DIV_ALGO == "div": 94 $for N in range(SIMD_TILE): 95 __m512 vf${ABC[N]} = _mm512_div_ps(ve${ABC[N]}, vd${ABC[N]}); 96 $else: 97 $for N in range(SIMD_TILE): 98 __m512 vr${ABC[N]} = _mm512_rcp14_ps(vd${ABC[N]}); 99 100 $for N in range(SIMD_TILE): 101 vr${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr${ABC[N]}, vd${ABC[N]}, vone), vr${ABC[N]}, vr${ABC[N]}); 102 103 $for N in range(SIMD_TILE): 104 __m512 vf${ABC[N]} = _mm512_mul_ps(ve${ABC[N]}, vr${ABC[N]}); 105 106 $if DIV_ALGO == "nr1fma1adj": 107 $for N in range(SIMD_TILE): 108 vf${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf${ABC[N]}, vd${ABC[N]}, ve${ABC[N]}), vr${ABC[N]}, vf${ABC[N]}); 109 110 $for N in range(SIMD_TILE): 111 vf${ABC[N]} = _mm512_mask_sub_ps(vf${ABC[N]}, _mm512_testn_epi32_mask(_mm512_castps_si512(vx${ABC[N]}), vsign_mask), vone, vf${ABC[N]}); 112 113 _mm512_storeu_ps(y, vf${ABC[0]}); 114 $for N in range(1, SIMD_TILE): 115 _mm512_storeu_ps(y + ${N * 16}, vf${ABC[N]}); 116 y += ${BATCH_TILE}; 117 } 118 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { 119 const __m512 vx = _mm512_loadu_ps(x); 120 x += 16; 121 122 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); 123 124 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); 125 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); 126 vn = _mm512_sub_ps(vn, vmagic_bias); 127 128 $if RR_STEPS == 1: 129 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 130 $else: 131 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz); 132 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); 133 134 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); 135 vp = _mm512_mul_ps(vp, vt); 136 vp = _mm512_fmadd_ps(vt, vp, vt); 137 vp = _mm512_fmadd_ps(vl, vp, vl); 138 139 const __m512 ve = _mm512_scalef_ps(vp, vn); 140 const __m512 vd = _mm512_add_ps(ve, vone); 141 142 $if DIV_ALGO == "div": 143 __m512 vf = _mm512_div_ps(ve, vd); 144 $else: 145 __m512 vr = _mm512_rcp14_ps(vd); 146 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); 147 148 __m512 vf = _mm512_mul_ps(ve, vr); 149 $if DIV_ALGO == "nr1fma1adj": 150 vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf); 151 152 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf); 153 154 _mm512_storeu_ps(y, vf); 155 y += 16; 156 } 157 if XNN_UNLIKELY(n != 0) { 158 assert(n >= 1 * sizeof(float)); 159 assert(n <= 15 * sizeof(float)); 160 161 // Prepare mask for valid 32-bit elements (depends on n). 162 n >>= 2 /* log2(sizeof(float)) */; 163 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); 164 165 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); 166 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); 167 168 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); 169 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); 170 vn = _mm512_sub_ps(vn, vmagic_bias); 171 172 $if RR_STEPS == 1: 173 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 174 $else: 175 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz); 176 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); 177 178 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); 179 vp = _mm512_mul_ps(vp, vt); 180 vp = _mm512_fmadd_ps(vt, vp, vt); 181 vp = _mm512_fmadd_ps(vl, vp, vl); 182 183 const __m512 ve = _mm512_scalef_ps(vp, vn); 184 const __m512 vd = _mm512_add_ps(ve, vone); 185 186 $if DIV_ALGO == "div": 187 __m512 vf = _mm512_div_ps(ve, vd); 188 $else: 189 __m512 vr = _mm512_rcp14_ps(vd); 190 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); 191 192 __m512 vf = _mm512_mul_ps(ve, vr); 193 $if DIV_ALGO == "nr1fma1adj": 194 vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf); 195 196 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf); 197 198 _mm512_mask_storeu_ps(y, vmask, vf); 199 } 200} 201