1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vrnd/neonfp16arith.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f16_vrndu_ukernel__neonfp16arith_x16(size_t n,const void * input,void * output,const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_vrndu_ukernel__neonfp16arith_x16(
20 size_t n,
21 const void* input,
22 void* output,
23 const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(__fp16) == 0);
27
28 const __fp16* i = (const __fp16*) input;
29 __fp16* o = (__fp16*) output;
30 for (; n >= 16 * sizeof(__fp16); n -= 16 * sizeof(__fp16)) {
31 float16x8_t vacc0 = vld1q_f16(i); i += 8;
32 float16x8_t vacc1 = vld1q_f16(i); i += 8;
33
34 vacc0 = vrndpq_f16(vacc0);
35 vacc1 = vrndpq_f16(vacc1);
36
37 vst1q_f16(o, vacc0); o += 8;
38 vst1q_f16(o, vacc1); o += 8;
39 }
40 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) {
41 float16x8_t vacc = vld1q_f16(i); i += 8;
42 vacc = vrndpq_f16(vacc);
43 vst1q_f16(o, vacc); o += 8;
44 }
45 if XNN_UNLIKELY(n != 0) {
46 float16x8_t vacc = vld1q_f16(i);
47 vacc = vrndpq_f16(vacc);
48 float16x4_t vacc_lo = vget_low_f16(vacc);
49 if (n & (4 * sizeof(__fp16))) {
50 vst1_f16(o, vacc_lo); o += 4;
51 vacc_lo = vget_high_f16(vacc);
52 }
53 if (n & (2 * sizeof(__fp16))) {
54 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2;
55 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
56 }
57 if (n & (1 * sizeof(__fp16))) {
58 vst1_lane_f16(o, vacc_lo, 0);
59 }
60 }
61 }
62