1// Copyright 2022 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$SIMD_TILE = BATCH_TILE // 8 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f16_velu_ukernel__neonfp16arith_rr1_p3_x${BATCH_TILE}( 18 size_t n, 19 const void* input, 20 void* output, 21 const union xnn_f16_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22{ 23 assert(n != 0); 24 assert(n % sizeof(__fp16) == 0); 25 26 const float16x8_t vprescale = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.prescale)); 27 const float16x8_t vsat_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.sat_cutoff)); 28 const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.magic_bias)); 29 const float16x8_t vlog2e = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.log2e)); 30 const float16x8_t vminus_ln2 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.minus_ln2)); 31 const float16x8_t vc3 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.c3)); 32 const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.c2)); 33 const float16x8_t vminus_alpha = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.minus_alpha)); 34 const float16x8_t vbeta = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr1_p3.beta)); 35 36 const __fp16* i = (const __fp16*) input; 37 __fp16* o = (__fp16*) output; 38 $if BATCH_TILE > 8: 39 for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) { 40 $for N in range(SIMD_TILE): 41 float16x8_t vx${N} = vld1q_f16(i); i += 8; 42 43 $for N in range(SIMD_TILE): 44 float16x8_t vz${N} = vmulq_f16(vx${N}, vprescale); 45 46 $for N in range(SIMD_TILE): 47 vz${N} = vmaxq_f16(vz${N}, vsat_cutoff); 48 49 $for N in range(SIMD_TILE): 50 float16x8_t vn${N} = vfmaq_f16(vmagic_bias, vz${N}, vlog2e); 51 52 $for N in range(SIMD_TILE): 53 float16x8_t vs${N} = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn${N}), 10)); 54 vn${N} = vsubq_f16(vn${N}, vmagic_bias); 55 56 $for N in range(SIMD_TILE): 57 float16x8_t vt${N} = vfmaq_f16(vz${N}, vn${N}, vminus_ln2); 58 59 $for N in range(SIMD_TILE): 60 float16x8_t vp${N} = vfmaq_f16(vc2, vc3, vt${N}); 61 vp${N} = vmulq_f16(vp${N}, vt${N}); 62 63 $for N in range(SIMD_TILE): 64 vt${N} = vmulq_f16(vt${N}, vs${N}); 65 vs${N} = vfmsq_f16(vminus_alpha, vs${N}, vminus_alpha); 66 67 $for N in range(SIMD_TILE): 68 vp${N} = vfmaq_f16(vt${N}, vp${N}, vt${N}); 69 70 $for N in range(SIMD_TILE): 71 float16x8_t ve${N} = vfmsq_f16(vs${N}, vp${N}, vminus_alpha); 72 const uint16x8_t vm${N} = vcltq_s16(vreinterpretq_s16_f16(vx${N}), vmovq_n_s16(0)); 73 74 $for N in range(SIMD_TILE): 75 vx${N} = vmulq_f16(vx${N}, vbeta); 76 77 $for N in range(SIMD_TILE): 78 const float16x8_t vy${N} = vbslq_f16(vm${N}, ve${N}, vx${N}); 79 80 $for N in range(SIMD_TILE): 81 vst1q_f16(o, vy${N}); o += 8; 82 } 83 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) { 84 float16x8_t vx = vld1q_f16(i); i += 8; 85 float16x8_t vz = vmulq_f16(vx, vprescale); 86 vz = vmaxq_f16(vz, vsat_cutoff); 87 88 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vlog2e); 89 float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10)); 90 vn = vsubq_f16(vn, vmagic_bias); 91 float16x8_t vt = vfmaq_f16(vz, vn, vminus_ln2); 92 93 float16x8_t vp = vfmaq_f16(vc2, vc3, vt); 94 vp = vmulq_f16(vp, vt); 95 vt = vmulq_f16(vt, vs); 96 vs = vfmsq_f16(vminus_alpha, vs, vminus_alpha); 97 vp = vfmaq_f16(vt, vp, vt); 98 float16x8_t ve = vfmsq_f16(vs, vp, vminus_alpha); 99 100 const uint16x8_t vm = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0)); 101 vx = vmulq_f16(vx, vbeta); 102 const float16x8_t vy = vbslq_f16(vm, ve, vx); 103 vst1q_f16(o, vy); o += 8; 104 } 105 if XNN_UNLIKELY(n != 0) { 106 float16x8_t vx = vld1q_f16(i); i += 8; 107 float16x8_t vz = vmulq_f16(vx, vprescale); 108 vz = vmaxq_f16(vz, vsat_cutoff); 109 110 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vlog2e); 111 float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10)); 112 vn = vsubq_f16(vn, vmagic_bias); 113 float16x8_t vt = vfmaq_f16(vz, vn, vminus_ln2); 114 115 float16x8_t vp = vfmaq_f16(vc2, vc3, vt); 116 vp = vmulq_f16(vp, vt); 117 vt = vmulq_f16(vt, vs); 118 vs = vfmsq_f16(vminus_alpha, vs, vminus_alpha); 119 vp = vfmaq_f16(vt, vp, vt); 120 float16x8_t ve = vfmsq_f16(vs, vp, vminus_alpha); 121 122 const uint16x8_t vm = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0)); 123 vx = vmulq_f16(vx, vbeta); 124 float16x8_t vy = vbslq_f16(vm, ve, vx); 125 float16x4_t vy_lo = vget_low_f16(vy); 126 if (n & (4 * sizeof(__fp16))) { 127 vst1_f16(o, vy_lo); o += 4; 128 vy_lo = vget_high_f16(vy); 129 } 130 if (n & (2 * sizeof(__fp16))) { 131 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vy_lo), 0); o += 2; 132 vy_lo = vext_f16(vy_lo, vy_lo, 2); 133 } 134 if (n & (1 * sizeof(__fp16))) { 135 vst1_lane_f16(o, vy_lo, 0); 136 } 137 } 138} 139