1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vbinary/vopc-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/common.h>
15 #include <xnnpack/vbinary.h>
16
17
xnn_f16_vsqrdiffc_ukernel__neonfp16arith_x16(size_t n,const void * restrict a_ptr,const void * restrict b_ptr,void * restrict y_ptr,const struct xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vsqrdiffc_ukernel__neonfp16arith_x16(
19 size_t n,
20 const void* restrict a_ptr,
21 const void* restrict b_ptr,
22 void* restrict y_ptr,
23 const struct xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25 assert(n != 0);
26 assert(n % sizeof(__fp16) == 0);
27 assert(a_ptr != NULL);
28 assert(b_ptr != NULL);
29 assert(y_ptr != NULL);
30
31 const __fp16* a = (const __fp16*) a_ptr;
32 const __fp16* b = (const __fp16*) b_ptr;
33 __fp16* y = (__fp16*) y_ptr;
34
35
36 const float16x8_t vb = vld1q_dup_f16(b);
37 for (; n >= 16 * sizeof(__fp16); n -= 16 * sizeof(__fp16)) {
38 const float16x8_t va01234567 = vld1q_f16(a); a += 8;
39 const float16x8_t va456789AB = vld1q_f16(a); a += 8;
40
41 float16x8_t vy01234567 = vsubq_f16(va01234567, vb);
42 float16x8_t vy456789AB = vsubq_f16(va456789AB, vb);
43
44 vy01234567 = vmulq_f16(vy01234567, vy01234567);
45 vy456789AB = vmulq_f16(vy456789AB, vy456789AB);
46
47
48 vst1q_f16(y, vy01234567); y += 8;
49 vst1q_f16(y, vy456789AB); y += 8;
50 }
51 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) {
52 const float16x8_t va01234567 = vld1q_f16(a); a += 8;
53
54 float16x8_t vy01234567 = vsubq_f16(va01234567, vb);
55 vy01234567 = vmulq_f16(vy01234567, vy01234567);
56 vst1q_f16(y, vy01234567); y += 8;
57 }
58 if XNN_UNLIKELY(n != 0) {
59 const float16x8_t va01234567 = vld1q_f16(a);
60
61 float16x8_t vy01234567 = vsubq_f16(va01234567, vb);
62 vy01234567 = vmulq_f16(vy01234567, vy01234567);
63
64 float16x4_t vy0123 = vget_low_f16(vy01234567);
65 if (n & (4 * sizeof(__fp16))) {
66 vst1_f16(y, vy0123); y += 4;
67 vy0123 = vget_high_f16(vy01234567);
68 }
69
70 if (n & (2 * sizeof(__fp16))) {
71 vst1_lane_u32(__builtin_assume_aligned(y, 1), vreinterpret_u32_f16(vy0123), 0); y += 2;
72 vy0123 = vext_f16(vy0123, vy0123, 2);
73 }
74
75 if (n & (1 * sizeof(__fp16))) {
76 vst1_lane_f16(y, vy0123, 0);
77 }
78 }
79 }
80