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$assert BATCH_TILE % 8 == 0 7$assert BATCH_TILE >= 8 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f16_vhswish_ukernel__neonfp16arith_x${BATCH_TILE}( 18 size_t n, 19 const void* restrict x_ptr, 20 void* restrict y_ptr, 21 const union xnn_f16_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22{ 23 assert(n != 0); 24 assert(n % sizeof(__fp16) == 0); 25 26 const __fp16* x = (const __fp16*) x_ptr; 27 __fp16* y = (__fp16*) y_ptr; 28 29 const float16x8_t vsixth = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.sixth)); 30 const float16x8_t vthree = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.three)); 31 const int16x8_t vsix = vreinterpretq_s16_u16(vld1q_dup_u16(¶ms->neon.six)); 32 const int16x8_t vzero = vdupq_n_s16(0); 33 34 $if BATCH_TILE > 8: 35 for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) { 36 $for N in range(0, BATCH_TILE, 8): 37 float16x8_t vx${ABC[N:N+8]} = vld1q_f16(x); x += 8; 38 39 $for N in range(0, BATCH_TILE, 8): 40 float16x8_t vacc${ABC[N:N+8]} = vaddq_f16(vx${ABC[N:N+8]}, vthree); 41 vx${ABC[N:N+8]} = vmulq_f16(vx${ABC[N:N+8]}, vsixth); 42 43 $for N in range(0, BATCH_TILE, 8): 44 vacc${ABC[N:N+8]} = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc${ABC[N:N+8]}), vzero)); 45 46 $for N in range(0, BATCH_TILE, 8): 47 vacc${ABC[N:N+8]} = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc${ABC[N:N+8]}), vsix)); 48 49 $for N in range(0, BATCH_TILE, 8): 50 vacc${ABC[N:N+8]} = vmulq_f16(vacc${ABC[N:N+8]}, vx${ABC[N:N+8]}); 51 52 $for N in range(0, BATCH_TILE, 8): 53 vst1q_f16(y, vacc${ABC[N:N+8]}); y += 8; 54 } 55 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) { 56 float16x8_t vx = vld1q_f16(x); x += 8; 57 float16x8_t vacc = vaddq_f16(vx, vthree); 58 vx = vmulq_f16(vx, vsixth); 59 vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero)); 60 vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix)); 61 vacc = vmulq_f16(vacc, vx); 62 vst1q_f16(y, vacc); y += 8; 63 } 64 if XNN_UNLIKELY(n != 0) { 65 float16x8_t vx = vld1q_f16(x); 66 float16x8_t vacc = vaddq_f16(vx, vthree); 67 vx = vmulq_f16(vx, vsixth); 68 vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero)); 69 vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix)); 70 vacc = vmulq_f16(vacc, vx); 71 72 float16x4_t vacc_lo = vget_low_f16(vacc); 73 if (n & (4 * sizeof(__fp16))) { 74 vst1_f16(y, vacc_lo); y += 4; 75 vacc_lo = vget_high_f16(vacc); 76 } 77 if (n & (2 * sizeof(__fp16))) { 78 vst1_lane_u32((void*) y, vreinterpret_u32_f16(vacc_lo), 0); y += 2; 79 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2); 80 } 81 if (n & (1 * sizeof(__fp16))) { 82 vst1_lane_f16(y, vacc_lo, 0); 83 } 84 } 85} 86