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$assert BATCH_TILE % 8 == 0 7$assert BATCH_TILE >= 8 8$SIMD_TILE = BATCH_TILE // 8 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0}; 18 19void xnn_f32_velu_ukernel__avx_rr2_p6_x${BATCH_TILE}( 20 size_t n, 21 const float* x, 22 float* y, 23 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) 24{ 25 assert(n % sizeof(float) == 0); 26 27 const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale); 28 const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha); 29 const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta); 30 31 const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f); 32 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f); 33 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f); 34 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E440p-1f); 35 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.0105C6p-21f); 36 const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f); 37 const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f); 38 const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f); 39 const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f); 40 const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f); 41 const __m256 vone = _mm256_set1_ps(1.0f); 42 43 $if BATCH_TILE > 8: 44 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 45 __m256 vx0 = _mm256_loadu_ps(x); 46 $for N in range(1, SIMD_TILE): 47 __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8}); 48 x += ${BATCH_TILE}; 49 50 $for N in range(SIMD_TILE): 51 const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale)); 52 53 $for N in range(SIMD_TILE): 54 __m256 vn${N} = _mm256_add_ps(_mm256_mul_ps(vz${N}, vlog2e), vmagic_bias); 55 56 $for N in range(SIMD_TILE): 57 const __m128 vs${N}_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn${N})), 23)); 58 const __m128 vs${N}_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn${N}, 1)), 23)); 59 vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias); 60 61 $for N in range(SIMD_TILE): 62 __m256 vt${N} = _mm256_add_ps(_mm256_mul_ps(vn${N}, vminus_ln2_hi), vz${N}); 63 __m256 vs${N} = _mm256_insertf128_ps(_mm256_castps128_ps256(vs${N}_lo), vs${N}_hi, 1); 64 65 $for N in range(SIMD_TILE): 66 vt${N} = _mm256_add_ps(_mm256_mul_ps(vn${N}, vminus_ln2_lo), vt${N}); 67 68 $for N in range(SIMD_TILE): 69 __m256 vp${N} = _mm256_add_ps(_mm256_mul_ps(vc6, vt${N}), vc5); 70 71 $for N in range(SIMD_TILE): 72 vp${N} = _mm256_add_ps(_mm256_mul_ps(vp${N}, vt${N}), vc4); 73 74 $for N in range(SIMD_TILE): 75 vp${N} = _mm256_add_ps(_mm256_mul_ps(vp${N}, vt${N}), vc3); 76 77 $for N in range(SIMD_TILE): 78 vp${N} = _mm256_add_ps(_mm256_mul_ps(vp${N}, vt${N}), vc2); 79 80 $for N in range(SIMD_TILE): 81 vp${N} = _mm256_mul_ps(vp${N}, vt${N}); 82 83 $for N in range(SIMD_TILE): 84 vt${N} = _mm256_mul_ps(vt${N}, vs${N}); 85 vs${N} = _mm256_sub_ps(vs${N}, vone); 86 87 $for N in range(SIMD_TILE): 88 vp${N} = _mm256_add_ps(_mm256_mul_ps(vp${N}, vt${N}), vt${N}); 89 90 $for N in range(SIMD_TILE): 91 const __m256 ve${N} = _mm256_mul_ps(_mm256_add_ps(vp${N}, vs${N}), valpha); 92 vx${N} = _mm256_mul_ps(vx${N}, vbeta); 93 94 $for N in range(SIMD_TILE): 95 const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N}); 96 97 _mm256_storeu_ps(y, vy0); 98 $for N in range(1, SIMD_TILE): 99 _mm256_storeu_ps(y + ${N * 8}, vy${N}); 100 y += ${BATCH_TILE}; 101 } 102 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 103 __m256 vx = _mm256_loadu_ps(x); 104 x += 8; 105 106 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale)); 107 108 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias); 109 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23)); 110 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23)); 111 vn = _mm256_sub_ps(vn, vmagic_bias); 112 113 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz); 114 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1); 115 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt); 116 117 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5); 118 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4); 119 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3); 120 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2); 121 vp = _mm256_mul_ps(vp, vt); 122 123 vt = _mm256_mul_ps(vt, vs); 124 vs = _mm256_sub_ps(vs, vone); 125 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt); 126 127 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha); 128 vx = _mm256_mul_ps(vx, vbeta); 129 const __m256 vy = _mm256_blendv_ps(vx, ve, vx); 130 131 _mm256_storeu_ps(y, vy); 132 y += 8; 133 } 134 if XNN_UNLIKELY(n != 0) { 135 assert(n >= 1 * sizeof(float)); 136 assert(n <= 7 * sizeof(float)); 137 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n)); 138 139 __m256 vx = _mm256_maskload_ps(x, vmask); 140 141 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale)); 142 143 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias); 144 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23)); 145 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23)); 146 vn = _mm256_sub_ps(vn, vmagic_bias); 147 148 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz); 149 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1); 150 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt); 151 152 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5); 153 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4); 154 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3); 155 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2); 156 vp = _mm256_mul_ps(vp, vt); 157 158 vt = _mm256_mul_ps(vt, vs); 159 vs = _mm256_sub_ps(vs, vone); 160 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt); 161 162 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha); 163 vx = _mm256_mul_ps(vx, vbeta); 164 const __m256 vy = _mm256_blendv_ps(vx, ve, vx); 165 166 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug). 167 __m128 vy_lo = _mm256_castps256_ps128(vy); 168 if (n & (4 * sizeof(float))) { 169 _mm_storeu_ps(y, vy_lo); 170 vy_lo = _mm256_extractf128_ps(vy, 1); 171 y += 4; 172 } 173 if (n & (2 * sizeof(float))) { 174 _mm_storel_pi((__m64*) y, vy_lo); 175 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 176 y += 2; 177 } 178 if (n & (1 * sizeof(float))) { 179 _mm_store_ss(y, vy_lo); 180 } 181 } 182} 183