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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f16_vlrelu_ukernel__neonfp16arith_x${BATCH_TILE}( 18 size_t batch, 19 const void* input, 20 void* output, 21 const union xnn_f16_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22{ 23 assert(batch != 0); 24 assert(batch % sizeof(__fp16) == 0); 25 26 const float16x8_t vslope = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.slope)); 27 const __fp16* i = (const __fp16*) input; 28 __fp16* o = (__fp16*) output; 29 $if BATCH_TILE > 4: 30 for (; batch >= ${BATCH_TILE} * sizeof(__fp16); batch -= ${BATCH_TILE} * sizeof(__fp16)) { 31 $for N in range(0, BATCH_TILE, 8): 32 const float16x8_t vx${ABC[N:N+8]} = vld1q_f16(i); i += 8; 33 34 $for N in range(0, BATCH_TILE, 8): 35 float16x8_t vacc${ABC[N:N+8]} = vmulq_f16(vx${ABC[N:N+8]}, vslope); 36 const uint16x8_t vmask${ABC[N:N+8]} = vcltq_s16(vreinterpretq_s16_f16(vx${ABC[N:N+8]}), vmovq_n_s16(0)); 37 38 $for N in range(0, BATCH_TILE, 8): 39 vacc${ABC[N:N+8]} = vbslq_f16(vmask${ABC[N:N+8]}, vacc${ABC[N:N+8]}, vx${ABC[N:N+8]}); 40 41 $for N in range(0, BATCH_TILE, 8): 42 vst1q_f16(o, vacc${ABC[N:N+8]}); o += 8; 43 } 44 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) { 45 const float16x8_t vx = vld1q_f16(i); i += 8; 46 float16x8_t vacc = vmulq_f16(vx, vslope); 47 const uint16x8_t vmask = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0)); 48 vacc = vbslq_f16(vmask, vacc, vx); 49 vst1q_f16(o, vacc); o += 8; 50 } 51 if XNN_UNLIKELY(batch != 0) { 52 const float16x8_t vx = vld1q_f16(i); 53 float16x8_t vacc = vmulq_f16(vx, vslope); 54 const uint16x8_t vmask = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0)); 55 vacc = vbslq_f16(vmask, vacc, vx); 56 57 float16x4_t vacc_lo = vget_low_f16(vacc); 58 if (batch & (4 * sizeof(__fp16))) { 59 vst1_f16(o, vacc_lo); o += 4; 60 vacc_lo = vget_high_f16(vacc); 61 } 62 if (batch & (2 * sizeof(__fp16))) { 63 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2; 64 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2); 65 } 66 if (batch & (1 * sizeof(__fp16))) { 67 vst1_lane_f16(o, vacc_lo, 0); 68 } 69 } 70} 71