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$SIMD_TILE = BATCH_TILE // 8 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) 22{ 23 assert(n % sizeof(float) == 0); 24 25 const __m256 vprescale = _mm256_load_ps(params->avx2_rr1_lut8_p4.prescale); 26 const __m256 valpha = _mm256_load_ps(params->avx2_rr1_lut8_p4.alpha); 27 const __m256 vbeta = _mm256_load_ps(params->avx2_rr1_lut8_p4.beta); 28 const __m256 vsat_cutoff = _mm256_load_ps(params->avx2_rr1_lut8_p4.sat_cutoff); 29 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_lut8_p4.magic_bias); 30 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_lut8_p4.log2e); 31 const __m256i vtable = _mm256_load_si256((const __m256i*) params->avx2_rr1_lut8_p4.table); 32 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_lut8_p4.minus_ln2); 33 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_lut8_p4.c4); 34 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_lut8_p4.c3); 35 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_lut8_p4.c2); 36 37 $if BATCH_TILE > 8: 38 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 39 __m256 vx0 = _mm256_loadu_ps(x); 40 $for N in range(1, SIMD_TILE): 41 __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8}); 42 x += ${BATCH_TILE}; 43 44 $for N in range(SIMD_TILE): 45 const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale)); 46 47 $for N in range(SIMD_TILE): 48 __m256 vn${N} = _mm256_fmadd_ps(vz${N}, vlog2e, vmagic_bias); 49 50 $for N in range(SIMD_TILE): 51 const __m256i ven${N} = _mm256_slli_epi32(_mm256_castps_si256(vn${N}), 20); 52 const __m256i vl${N} = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn${N})); 53 vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias); 54 55 $for N in range(SIMD_TILE): 56 __m256 vs${N} = _mm256_castsi256_ps(_mm256_add_epi32(vl${N}, ven${N})); 57 __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vz${N}); 58 59 $for N in range(SIMD_TILE): 60 __m256 vp${N} = _mm256_fmadd_ps(vc4, vt${N}, vc3); 61 62 $for N in range(SIMD_TILE): 63 vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc2); 64 65 $for N in range(SIMD_TILE): 66 vp${N} = _mm256_mul_ps(vp${N}, vt${N}); 67 vt${N} = _mm256_mul_ps(vt${N}, vs${N}); 68 69 $for N in range(SIMD_TILE): 70 vs${N} = _mm256_fmsub_ps(vs${N}, valpha, valpha); 71 vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vt${N}); 72 73 $for N in range(SIMD_TILE): 74 const __m256 ve${N} = _mm256_fmadd_ps(vp${N}, valpha, vs${N}); 75 vx${N} = _mm256_mul_ps(vx${N}, vbeta); 76 77 $for N in range(SIMD_TILE): 78 const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N}); 79 80 _mm256_storeu_ps(y, vy0); 81 $for N in range(1, SIMD_TILE): 82 _mm256_storeu_ps(y + ${N * 8}, vy${N}); 83 y += ${BATCH_TILE}; 84 } 85 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 86 __m256 vx = _mm256_loadu_ps(x); 87 x += 8; 88 89 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale)); 90 91 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias); 92 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20); 93 const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn)); 94 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven)); 95 vn = _mm256_sub_ps(vn, vmagic_bias); 96 97 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz); 98 99 __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3); 100 vp = _mm256_fmadd_ps(vp, vt, vc2); 101 vp = _mm256_mul_ps(vp, vt); 102 103 vt = _mm256_mul_ps(vt, vs); 104 vs = _mm256_fmsub_ps(vs, valpha, valpha); 105 vp = _mm256_fmadd_ps(vp, vt, vt); 106 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs); 107 108 vx = _mm256_mul_ps(vx, vbeta); 109 const __m256 vy = _mm256_blendv_ps(vx, ve, vx); 110 111 _mm256_storeu_ps(y, vy); 112 y += 8; 113 } 114 if XNN_UNLIKELY(n != 0) { 115 assert(n >= 1 * sizeof(float)); 116 assert(n <= 7 * sizeof(float)); 117 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_lut8_p4.mask_table[7] - n)); 118 119 __m256 vx = _mm256_maskload_ps(x, vmask); 120 121 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale)); 122 123 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias); 124 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20); 125 const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn)); 126 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven)); 127 vn = _mm256_sub_ps(vn, vmagic_bias); 128 129 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz); 130 131 __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3); 132 vp = _mm256_fmadd_ps(vp, vt, vc2); 133 vp = _mm256_mul_ps(vp, vt); 134 135 vt = _mm256_mul_ps(vt, vs); 136 vs = _mm256_fmsub_ps(vs, valpha, valpha); 137 vp = _mm256_fmadd_ps(vp, vt, vt); 138 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs); 139 140 vx = _mm256_mul_ps(vx, vbeta); 141 const __m256 vy = _mm256_blendv_ps(vx, ve, vx); 142 143 __m128 vy_lo = _mm256_castps256_ps128(vy); 144 if (n & (4 * sizeof(float))) { 145 _mm_storeu_ps(y, vy_lo); 146 vy_lo = _mm256_extractf128_ps(vy, 1); 147 y += 4; 148 } 149 if (n & (2 * sizeof(float))) { 150 _mm_storel_pi((__m64*) y, vy_lo); 151 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 152 y += 2; 153 } 154 if (n & (1 * sizeof(float))) { 155 _mm_store_ss(y, vy_lo); 156 } 157 } 158} 159