1 // Auto-generated file. Do not edit!
2 // Template: src/f16-clamp/neonfp16arith.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/clamp.h>
15 #include <xnnpack/common.h>
16
17
xnn_f16_clamp_ukernel__neonfp16arith_x16(size_t n,const void * restrict x_ptr,void * restrict y_ptr,const struct xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_clamp_ukernel__neonfp16arith_x16(
19 size_t n,
20 const void* restrict x_ptr,
21 void* restrict y_ptr,
22 const struct xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(__fp16) == 0);
26 assert(x_ptr != NULL);
27 assert(y_ptr != NULL);
28
29 const __fp16* x = (const __fp16*) x_ptr;
30 __fp16* y = (__fp16*) y_ptr;
31
32 const float16x8_t vy_min = vld1q_dup_f16(¶ms->min);
33 const float16x8_t vy_max = vld1q_dup_f16(¶ms->max);
34
35 for (; n >= 16 * sizeof(__fp16); n -= 16 * sizeof(__fp16)) {
36 float16x8_t vacc01234567 = vld1q_f16(x); x += 8;
37 float16x8_t vacc89ABCDEF = vld1q_f16(x); x += 8;
38
39 vacc01234567 = vmaxq_f16(vacc01234567, vy_min);
40 vacc89ABCDEF = vmaxq_f16(vacc89ABCDEF, vy_min);
41
42 vacc01234567 = vminq_f16(vacc01234567, vy_max);
43 vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vy_max);
44
45 vst1q_f16(y, vacc01234567); y += 8;
46 vst1q_f16(y, vacc89ABCDEF); y += 8;
47 }
48 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) {
49 float16x8_t vacc = vld1q_f16(x); x += 8;
50 vacc = vmaxq_f16(vacc, vy_min);
51 vacc = vminq_f16(vacc, vy_max);
52 vst1q_f16(y, vacc); y += 8;
53 }
54 if XNN_UNLIKELY(n != 0) {
55 float16x8_t vacc = vld1q_f16(x);
56 vacc = vmaxq_f16(vacc, vy_min);
57 vacc = vminq_f16(vacc, vy_max);
58
59 float16x4_t vacc_lo = vget_low_f16(vacc);
60 if (n & (4 * sizeof(__fp16))) {
61 vst1_f16(y, vacc_lo); y += 4;
62 vacc_lo = vget_high_f16(vacc);
63 }
64 if (n & (2 * sizeof(__fp16))) {
65 vst1_lane_u32(__builtin_assume_aligned(y, 1), vreinterpret_u32_f16(vacc_lo), 0); y += 2;
66 vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
67 }
68 if (n & (1 * sizeof(__fp16))) {
69 vst1_lane_f16(y, vacc_lo, 0);
70 }
71 }
72 }
73