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$SIMD_TILE = BATCH_TILE // 8 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <arm_neon.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/vunary.h> 16 17 18void xnn_f16_vsqrt_ukernel__neonfp16arith_sqrt_x${BATCH_TILE}( 19 size_t n, 20 const void* input, 21 void* output, 22 const union xnn_f16_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 23{ 24 assert(n != 0); 25 assert(n % sizeof(__fp16) == 0); 26 assert(input != NULL); 27 assert(output != NULL); 28 29 const __fp16* i = (const __fp16*) input; 30 __fp16* o = (__fp16*) output; 31 $if SIMD_TILE > 8: 32 for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) { 33 $for N in range(SIMD_TILE): 34 float16x8_t vacc${ABC[N]} = vld1q_f16(i); i += 8; 35 36 $for N in range(SIMD_TILE): 37 vacc${ABC[N]} = vsqrtq_f16(vacc${ABC[N]}); 38 39 $for N in range(SIMD_TILE): 40 vst1q_f16(o, vacc${ABC[N]}); o += 8; 41 } 42 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) { 43 float16x8_t vacc = vld1q_f16(i); i += 8; 44 vacc = vsqrtq_f16(vacc); 45 vst1q_f16(o, vacc); o += 8; 46 } 47 if XNN_UNLIKELY(n != 0) { 48 const float16x8_t vacc = vld1q_f16(i); 49 float16x4_t vacc_lo = vsqrt_f16(vget_low_f16(vacc)); 50 if (n & (4 * sizeof(__fp16))) { 51 vst1_f16(o, vacc_lo); o += 4; 52 vacc_lo = vsqrt_f16(vget_high_f16(vacc)); 53 } 54 if (n & (2 * sizeof(__fp16))) { 55 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2; 56 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2); 57 } 58 if (n & (1 * sizeof(__fp16))) { 59 vst1_lane_f16(o, vacc_lo, 0); 60 } 61 } 62} 63