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${RR_STEPS}_p5_scalef_${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 vlog2e = _mm512_set1_ps(0x1.715476p+0f); 31 $if RR_STEPS == 1: 32 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); 33 $else: 34 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); 35 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); 37 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); 38 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); 39 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); 40 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); 41 const __m512 vone = _mm512_set1_ps(1.0f); 42 43 $if BATCH_TILE > 16: 44 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 45 const __m512 vx${ABC[0]} = _mm512_loadu_ps(x); 46 $for N in range(1, SIMD_TILE): 47 const __m512 vx${ABC[N]} = _mm512_loadu_ps(x + ${N * 16}); 48 x += ${BATCH_TILE}; 49 50 $for N in range(SIMD_TILE): 51 const __m512 vz${ABC[N]} = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx${ABC[N]}), vsign_mask)); 52 53 $for N in range(SIMD_TILE): 54 __m512 vn${ABC[N]} = _mm512_mul_ps(vz${ABC[N]}, vlog2e); 55 56 $for N in range(SIMD_TILE): 57 vn${ABC[N]} = _mm512_roundscale_ps(vn${ABC[N]}, 0); 58 59 $if RR_STEPS == 1: 60 $for N in range(SIMD_TILE): 61 __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2, vz${ABC[N]}); 62 $else: 63 $for N in range(SIMD_TILE): 64 __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2_hi, vz${ABC[N]}); 65 66 $for N in range(SIMD_TILE): 67 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2_lo, vt${ABC[N]}); 68 69 $for N in range(SIMD_TILE): 70 __m512 vp${ABC[N]} = _mm512_fmadd_ps(vc5, vt${ABC[N]}, vc4); 71 72 $for N in range(SIMD_TILE): 73 vp${ABC[N]} = _mm512_fmadd_ps(vp${ABC[N]}, vt${ABC[N]}, vc3); 74 75 $for N in range(SIMD_TILE): 76 vp${ABC[N]} = _mm512_fmadd_ps(vp${ABC[N]}, vt${ABC[N]}, vc2); 77 78 $for N in range(SIMD_TILE): 79 vp${ABC[N]} = _mm512_fmadd_ps(vp${ABC[N]}, vt${ABC[N]}, vc1); 80 81 $for N in range(SIMD_TILE): 82 vp${ABC[N]} = _mm512_fmadd_ps(vp${ABC[N]}, vt${ABC[N]}, vone); 83 84 $for N in range(SIMD_TILE): 85 const __m512 ve${ABC[N]} = _mm512_scalef_ps(vp${ABC[N]}, vn${ABC[N]}); 86 87 $for N in range(SIMD_TILE): 88 const __m512 vd${ABC[N]} = _mm512_add_ps(ve${ABC[N]}, vone); 89 90 $if DIV_ALGO == "div": 91 $for N in range(SIMD_TILE): 92 __m512 vf${ABC[N]} = _mm512_div_ps(ve${ABC[N]}, vd${ABC[N]}); 93 $else: 94 $for N in range(SIMD_TILE): 95 __m512 vr${ABC[N]} = _mm512_rcp14_ps(vd${ABC[N]}); 96 97 $for N in range(SIMD_TILE): 98 vr${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr${ABC[N]}, vd${ABC[N]}, vone), vr${ABC[N]}, vr${ABC[N]}); 99 100 $for N in range(SIMD_TILE): 101 __m512 vf${ABC[N]} = _mm512_mul_ps(ve${ABC[N]}, vr${ABC[N]}); 102 103 $if DIV_ALGO == "nr1fma1adj": 104 $for N in range(SIMD_TILE): 105 vf${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf${ABC[N]}, vd${ABC[N]}, ve${ABC[N]}), vr${ABC[N]}, vf${ABC[N]}); 106 107 $for N in range(SIMD_TILE): 108 vf${ABC[N]} = _mm512_mask_sub_ps(vf${ABC[N]}, _mm512_testn_epi32_mask(_mm512_castps_si512(vx${ABC[N]}), vsign_mask), vone, vf${ABC[N]}); 109 110 _mm512_storeu_ps(y, vf${ABC[0]}); 111 $for N in range(1, SIMD_TILE): 112 _mm512_storeu_ps(y + ${N * 16}, vf${ABC[N]}); 113 y += ${BATCH_TILE}; 114 } 115 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { 116 const __m512 vx = _mm512_loadu_ps(x); 117 x += 16; 118 119 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); 120 121 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0); 122 123 $if RR_STEPS == 1: 124 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 125 $else: 126 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz); 127 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); 128 129 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); 130 vp = _mm512_fmadd_ps(vp, vt, vc3); 131 vp = _mm512_fmadd_ps(vp, vt, vc2); 132 vp = _mm512_fmadd_ps(vp, vt, vc1); 133 vp = _mm512_fmadd_ps(vp, vt, vone); 134 135 const __m512 ve = _mm512_scalef_ps(vp, vn); 136 const __m512 vd = _mm512_add_ps(ve, vone); 137 138 $if DIV_ALGO == "div": 139 __m512 vf = _mm512_div_ps(ve, vd); 140 $else: 141 __m512 vr = _mm512_rcp14_ps(vd); 142 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); 143 144 __m512 vf = _mm512_mul_ps(ve, vr); 145 $if DIV_ALGO == "nr1fma1adj": 146 vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf); 147 148 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf); 149 150 _mm512_storeu_ps(y, vf); 151 y += 16; 152 } 153 if XNN_UNLIKELY(n != 0) { 154 assert(n >= 1 * sizeof(float)); 155 assert(n <= 15 * sizeof(float)); 156 157 // Prepare mask for valid 32-bit elements (depends on n). 158 n >>= 2 /* log2(sizeof(float)) */; 159 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); 160 161 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); 162 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); 163 164 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0); 165 166 $if RR_STEPS == 1: 167 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 168 $else: 169 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz); 170 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); 171 172 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); 173 vp = _mm512_fmadd_ps(vp, vt, vc3); 174 vp = _mm512_fmadd_ps(vp, vt, vc2); 175 vp = _mm512_fmadd_ps(vp, vt, vc1); 176 vp = _mm512_fmadd_ps(vp, vt, vone); 177 178 const __m512 ve = _mm512_scalef_ps(vp, vn); 179 const __m512 vd = _mm512_add_ps(ve, vone); 180 181 $if DIV_ALGO == "div": 182 __m512 vf = _mm512_div_ps(ve, vd); 183 $else: 184 __m512 vr = _mm512_rcp14_ps(vd); 185 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr); 186 187 __m512 vf = _mm512_mul_ps(ve, vr); 188 $if DIV_ALGO == "nr1fma1adj": 189 vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf); 190 191 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf); 192 193 _mm512_mask_storeu_ps(y, vmask, vf); 194 } 195} 196