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$SIMD_TILE = BATCH_TILE // 16 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vunary.h> 16 17 18void xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x${BATCH_TILE}( 19 size_t n, 20 const float* x, 21 float* y, 22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) 23{ 24 assert(n != 0); 25 assert(n % sizeof(float) == 0); 26 27 const __m512 vprescale = _mm512_set1_ps(params->avx512_rr1_lut16_p3.prescale); 28 const __m512 valpha = _mm512_set1_ps(params->avx512_rr1_lut16_p3.alpha); 29 const __m512 vbeta = _mm512_set1_ps(params->avx512_rr1_lut16_p3.beta); 30 const __m512 vsat_cutoff = _mm512_set1_ps(params->avx512_rr1_lut16_p3.sat_cutoff); 31 const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr1_lut16_p3.magic_bias); 32 const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_lut16_p3.log2e); 33 const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.minus_ln2); 34 const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c3); 35 const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c2); 36 const __m512i vtable = _mm512_load_si512(params->avx512_rr1_lut16_p3.table); 37 38 $if BATCH_TILE > 16: 39 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 40 __m512 vx0 = _mm512_loadu_ps(x); 41 $for N in range(1, SIMD_TILE): 42 __m512 vx${N} = _mm512_loadu_ps(x + ${N * 16}); 43 x += ${BATCH_TILE}; 44 45 $for N in range(SIMD_TILE): 46 const __m512 vz${N} = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx${N}, vprescale)); 47 48 $for N in range(SIMD_TILE): 49 __m512 vn${N} = _mm512_fmadd_ps(vz${N}, vlog2e, vmagic_bias); 50 51 $for N in range(SIMD_TILE): 52 const __m512i ven${N} = _mm512_slli_epi32(_mm512_castps_si512(vn${N}), 19); 53 const __m512i vl${N} = _mm512_permutexvar_epi32(_mm512_castps_si512(vn${N}), vtable); 54 55 $for N in range(SIMD_TILE): 56 __m512 vs${N} = _mm512_castsi512_ps(_mm512_add_epi32(vl${N}, ven${N})); 57 vn${N} = _mm512_sub_ps(vn${N}, vmagic_bias); 58 59 $for N in range(SIMD_TILE): 60 __m512 vt${N} = _mm512_fmadd_ps(vn${N}, vminus_ln2, vz${N}); 61 62 $for N in range(SIMD_TILE): 63 __m512 vp${N} = _mm512_fmadd_ps(vc3, vt${N}, vc2); 64 65 $for N in range(SIMD_TILE): 66 vp${N} = _mm512_mul_ps(vp${N}, vt${N}); 67 vt${N} = _mm512_mul_ps(vt${N}, vs${N}); 68 69 $for N in range(SIMD_TILE): 70 vs${N} = _mm512_fmsub_ps(vs${N}, valpha, valpha); 71 72 $for N in range(SIMD_TILE): 73 vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vt${N}); 74 75 const __m512 vzero = _mm512_setzero_ps(); 76 $for N in range(SIMD_TILE): 77 __m512 vy${N} = _mm512_fmadd_ps(vp${N}, valpha, vs${N}); 78 const __mmask16 vsign${N} = _mm512_cmp_ps_mask(vx${N}, vzero, _CMP_NLT_US); 79 80 $for N in range(SIMD_TILE): 81 vy${N} = _mm512_mask_mul_ps(vy${N}, vsign${N}, vx${N}, vbeta); 82 83 _mm512_storeu_ps(y, vy0); 84 $for N in range(1, SIMD_TILE): 85 _mm512_storeu_ps(y + ${N * 16}, vy${N}); 86 y += ${BATCH_TILE}; 87 } 88 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { 89 __m512 vx = _mm512_loadu_ps(x); 90 x += 16; 91 92 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale)); 93 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US); 94 95 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); 96 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); 97 const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable); 98 __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven)); 99 vn = _mm512_sub_ps(vn, vmagic_bias); 100 101 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 102 103 __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2); 104 vp = _mm512_mul_ps(vp, vt); 105 106 vt = _mm512_mul_ps(vt, vs); 107 vs = _mm512_fmsub_ps(vs, valpha, valpha); 108 vp = _mm512_fmadd_ps(vp, vt, vt); 109 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs); 110 111 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta); 112 113 _mm512_storeu_ps(y, vy); 114 y += 16; 115 } 116 if XNN_UNLIKELY(n != 0) { 117 assert(n >= 1 * sizeof(float)); 118 assert(n <= 15 * sizeof(float)); 119 // Prepare mask for valid 32-bit elements (depends on n). 120 n >>= 2 /* log2(sizeof(float)) */; 121 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); 122 123 __m512 vx = _mm512_maskz_loadu_ps(vmask, x); 124 125 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale)); 126 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US); 127 128 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); 129 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); 130 const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable); 131 __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven)); 132 vn = _mm512_sub_ps(vn, vmagic_bias); 133 134 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); 135 136 __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2); 137 vp = _mm512_mul_ps(vp, vt); 138 139 vt = _mm512_mul_ps(vt, vs); 140 vs = _mm512_fmsub_ps(vs, valpha, valpha); 141 vp = _mm512_fmadd_ps(vp, vt, vt); 142 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs); 143 144 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta); 145 146 _mm512_mask_storeu_ps(y, vmask, vy); 147 } 148} 149