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$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"] 11#include <assert.h> 12 13#include <arm_neon.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/math.h> 17#include <xnnpack/vunary.h> 18 19 20$VRNDQ_F16 = { 21$ "RNDNE": "vrndnq_f16", 22$ "RNDZ": "vrndq_f16", 23$ "RNDU": "vrndpq_f16", 24$ "RNDD": "vrndmq_f16", 25$}[OP] 26void xnn_f16_v${OP.lower()}_ukernel__neonfp16arith_x${BATCH_TILE}( 27 size_t n, 28 const void* input, 29 void* output, 30 const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 31{ 32 assert(n != 0); 33 assert(n % sizeof(__fp16) == 0); 34 35 const __fp16* i = (const __fp16*) input; 36 __fp16* o = (__fp16*) output; 37 $if BATCH_TILE > 8: 38 for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) { 39 $for N in range(SIMD_TILE): 40 float16x8_t vacc${ABC[N]} = vld1q_f16(i); i += 8; 41 42 $for N in range(SIMD_TILE): 43 vacc${ABC[N]} = ${VRNDQ_F16}(vacc${ABC[N]}); 44 45 $for N in range(SIMD_TILE): 46 vst1q_f16(o, vacc${ABC[N]}); o += 8; 47 } 48 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) { 49 float16x8_t vacc = vld1q_f16(i); i += 8; 50 vacc = ${VRNDQ_F16}(vacc); 51 vst1q_f16(o, vacc); o += 8; 52 } 53 if XNN_UNLIKELY(n != 0) { 54 float16x8_t vacc = vld1q_f16(i); 55 vacc = ${VRNDQ_F16}(vacc); 56 float16x4_t vacc_lo = vget_low_f16(vacc); 57 if (n & (4 * sizeof(__fp16))) { 58 vst1_f16(o, vacc_lo); o += 4; 59 vacc_lo = vget_high_f16(vacc); 60 } 61 if (n & (2 * sizeof(__fp16))) { 62 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2; 63 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2); 64 } 65 if (n & (1 * sizeof(__fp16))) { 66 vst1_lane_f16(o, vacc_lo, 0); 67 } 68 } 69} 70