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$assert BATCH_TILE % 4 == 0 7$assert BATCH_TILE >= 4 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/vunary.h> 14#include <xnnpack/common.h> 15 16 17void xnn_f32_relu_ukernel__neon_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 assert(x != NULL); 26 assert(y != NULL); 27 28 const float32x4_t vzero = vmovq_n_f32(0.0f); 29 30 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 31 $for N in range(0, BATCH_TILE, 4): 32 float32x4_t vacc${ABC[N:N+4]} = vld1q_f32(x); x += 4; 33 34 $for N in range(0, BATCH_TILE, 4): 35 vacc${ABC[N:N+4]} = vmaxq_f32(vacc${ABC[N:N+4]}, vzero); 36 37 $for N in range(0, BATCH_TILE, 4): 38 vst1q_f32(y, vacc${ABC[N:N+4]}); y += 4; 39 } 40 $if BATCH_TILE > 4: 41 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 42 float32x4_t vacc = vld1q_f32(x); x += 4; 43 vacc = vmaxq_f32(vacc, vzero); 44 vst1q_f32(y, vacc); y += 4; 45 } 46 if XNN_UNLIKELY(n != 0) { 47 float32x4_t vacc = vld1q_f32(x); 48 vacc = vmaxq_f32(vacc, vzero); 49 50 float32x2_t vacc_lo = vget_low_f32(vacc); 51 if (n & (2 * sizeof(float))) { 52 vst1_f32(y, vacc_lo); y += 2; 53 vacc_lo = vget_high_f32(vacc); 54 } 55 if (n & (1 * sizeof(float))) { 56 vst1_lane_f32(y, vacc_lo, 0); 57 } 58 } 59} 60