1// Copyright 2019 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$assert ELEMENTS_TILE % 16 == 0 7$assert ELEMENTS_TILE >= 16 8$SIMD_TILE = ELEMENTS_TILE // 16 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/intrinsics-polyfill.h> 16#include <xnnpack/vscaleextexp.h> 17 18 19void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x${ELEMENTS_TILE}( 20 size_t elements, 21 const float* x, 22 float* y, 23 float scale_value, 24 float scale_exp) 25{ 26 assert(elements % sizeof(float) == 0); 27 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); 29 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); 30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); 31 32 const __m512 vc0 = _mm512_set1_ps(1.0f); 33 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); 34 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); 35 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); 36 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); 37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); 38 39 const __m512 vscalev = _mm512_set1_ps(scale_value); 40 const __m512 vscalee = _mm512_set1_ps(scale_exp); 41 42 for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) { 43 // Load ${ELEMENTS_TILE} (${SIMD_TILE}x16) inputs at a time. 44 const __m512 vx0 = _mm512_loadu_ps(x); 45 $for N in range(1, SIMD_TILE): 46 const __m512 vx${N} = _mm512_loadu_ps(x + ${N * 16}); 47 x += ${ELEMENTS_TILE}; 48 49 // Compute reduced argument elements := round(x / log(2)). 50 $for N in range(SIMD_TILE): 51 const __m512 vn${N} = _mm512_roundscale_ps(_mm512_mul_ps(vx${N}, vlog2e), 0); 52 53 // Compute reduced argument t := x - elements * log(2). 54 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. 55 $for N in range(SIMD_TILE): 56 __m512 vt${N} = _mm512_fmadd_ps(vn${N}, vminus_ln2_hi, vx${N}); 57 58 $for N in range(SIMD_TILE): 59 vt${N} = _mm512_fmadd_ps(vn${N}, vminus_ln2_lo, vt${N}); 60 61 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. 62 $for N in range(SIMD_TILE): 63 __m512 vp${N} = _mm512_fmadd_ps(vc5, vt${N}, vc4); 64 65 $for N in range(SIMD_TILE): 66 vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vc3); 67 68 $for N in range(SIMD_TILE): 69 vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vc2); 70 71 $for N in range(SIMD_TILE): 72 vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vc1); 73 74 $for N in range(SIMD_TILE): 75 vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vc0); 76 77 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where 78 // - vnX is "exponent" 79 // - vpX is "mantissa" 80 // 81 // exp2(ae) * av * exp2(be) * bv = 82 // = exp2(ae + be) * (av * bv) 83 $for N in range(SIMD_TILE): 84 __m512 vf${N} = _mm512_mul_ps(vp${N}, vscalev); 85 86 $for N in range(SIMD_TILE): 87 const __m512 ve${N} = _mm512_add_ps(vn${N}, vscalee); 88 89 // Multiply "mantissa" by the exp2("exponent"). 90 $for N in range(SIMD_TILE): 91 vf${N} = _mm512_scalef_ps(vf${N}, ve${N}); 92 93 // Store 128 (8x16) results at a time. 94 _mm512_storeu_ps(y, vf0); 95 $for N in range(SIMD_TILE): 96 _mm512_storeu_ps(y + ${N * 16}, vf${N}); 97 y += ${ELEMENTS_TILE}; 98 } 99 100 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) { 101 // Load 16 inputs at a time. 102 const __m512 vx = _mm512_loadu_ps(x); 103 x += 16; 104 105 // Compute reduced argument elements := round(x / log(2)). 106 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); 107 108 // Compute reduced argument t := x - elements * log(2). 109 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. 110 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); 111 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); 112 113 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. 114 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); 115 vp = _mm512_fmadd_ps(vp, vt, vc3); 116 vp = _mm512_fmadd_ps(vp, vt, vc2); 117 vp = _mm512_fmadd_ps(vp, vt, vc1); 118 vp = _mm512_fmadd_ps(vp, vt, vc0); 119 120 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation. 121 __m512 vf = _mm512_mul_ps(vp, vscalev); 122 const __m512 ve = _mm512_add_ps(vn, vscalee); 123 124 // Multiply "mantissa" by the exp2("exponent"). 125 vf = _mm512_scalef_ps(vf, ve); 126 127 // Store 16 results at a time. 128 _mm512_storeu_ps(y, vf); 129 y += 16; 130 } 131 if XNN_UNLIKELY(elements != 0) { 132 // Prepare mask for valid 32-bit elements (depends on elements). 133 elements >>= 2 /* log2(sizeof(float)) */; 134 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1))); 135 136 // Load up to 15 inputs at a time. 137 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); 138 139 // Compute reduced argument elements := round(x / log(2)). 140 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); 141 142 // Compute reduced argument t := x - elements * log(2). 143 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. 144 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); 145 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); 146 147 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. 148 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); 149 vp = _mm512_fmadd_ps(vp, vt, vc3); 150 vp = _mm512_fmadd_ps(vp, vt, vc2); 151 vp = _mm512_fmadd_ps(vp, vt, vc1); 152 vp = _mm512_fmadd_ps(vp, vt, vc0); 153 154 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation. 155 __m512 vf = _mm512_mul_ps(vp, vscalev); 156 const __m512 ve = _mm512_add_ps(vn, vscalee); 157 158 // Multiply "mantissa" by the exp2("exponent"). 159 vf = _mm512_scalef_ps(vf, ve); 160 161 // Store up to 15 results at a time. 162 _mm512_mask_storeu_ps(y, vmask, vf); 163 } 164 _mm256_zeroupper(); 165} 166