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$assert OP in ["ABS", "NEG", "SQR"] 11#include <assert.h> 12 13#include <arm_neon.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/vunary.h> 17 18 19$VOPQ_F16 = { 20$ "ABS": lambda x: "vabsq_f16(%s)" % x, 21$ "NEG": lambda x: "vnegq_f16(%s)" % x, 22$ "SQR": lambda x: "vmulq_f16(%s, %s)" % (x, x), 23$}[OP] 24$PARAMS = { 25$ "ABS": "xnn_f16_abs_params", 26$ "NEG": "xnn_f16_neg_params", 27$ "SQR": "xnn_f16_default_params", 28$}[OP] 29void xnn_f16_v${OP.lower()}_ukernel__neonfp16arith_x${BATCH_TILE}( 30 size_t n, 31 const void* input, 32 void* output, 33 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 34{ 35 assert(n != 0); 36 assert(n % sizeof(__fp16) == 0); 37 assert(input != NULL); 38 assert(output != NULL); 39 40 const __fp16* i = (const __fp16*) input; 41 __fp16* o = (__fp16*) output; 42 $if SIMD_TILE > 8: 43 for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) { 44 $for N in range(SIMD_TILE): 45 float16x8_t vacc${ABC[N]} = vld1q_f16(i); i += 8; 46 47 $for N in range(SIMD_TILE): 48 vacc${ABC[N]} = ${VOPQ_F16("vacc" + ABC[N])}; 49 50 $for N in range(SIMD_TILE): 51 vst1q_f16(o, vacc${ABC[N]}); o += 8; 52 } 53 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) { 54 float16x8_t vacc = vld1q_f16(i); i += 8; 55 vacc = ${VOPQ_F16("vacc")}; 56 vst1q_f16(o, vacc); o += 8; 57 } 58 if XNN_UNLIKELY(n != 0) { 59 float16x8_t vacc = vld1q_f16(i); 60 vacc = ${VOPQ_F16("vacc")}; 61 float16x4_t vacc_lo = vget_low_f16(vacc); 62 if (n & (4 * sizeof(__fp16))) { 63 vst1_f16(o, vacc_lo); o += 4; 64 vacc_lo = vget_high_f16(vacc); 65 } 66 if (n & (2 * sizeof(__fp16))) { 67 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2; 68 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2); 69 } 70 if (n & (1 * sizeof(__fp16))) { 71 vst1_lane_f16(o, vacc_lo, 0); 72 } 73 } 74} 75