xref: /aosp_15_r20/external/XNNPACK/src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vbinary/vop-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_vsub_minmax_ukernel__neonfp16arith_x16(size_t n,const void * restrict a_ptr,const void * restrict b_ptr,void * restrict y_ptr,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vsub_minmax_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 union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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   const float16x8_t vy_min = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
36   const float16x8_t vy_max = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
37 
38   for (; n >= 16 * sizeof(__fp16); n -= 16 * sizeof(__fp16)) {
39     const float16x8_t va01234567 = vld1q_f16(a); a += 8;
40     const float16x8_t vb01234567 = vld1q_f16(b); b += 8;
41     const float16x8_t va456789AB = vld1q_f16(a); a += 8;
42     const float16x8_t vb456789AB = vld1q_f16(b); b += 8;
43 
44     float16x8_t vy01234567 = vsubq_f16(va01234567, vb01234567);
45     float16x8_t vy456789AB = vsubq_f16(va456789AB, vb456789AB);
46 
47 
48     vy01234567 = vmaxq_f16(vy01234567, vy_min);
49     vy456789AB = vmaxq_f16(vy456789AB, vy_min);
50 
51     vy01234567 = vminq_f16(vy01234567, vy_max);
52     vy456789AB = vminq_f16(vy456789AB, vy_max);
53 
54     vst1q_f16(y, vy01234567); y += 8;
55     vst1q_f16(y, vy456789AB); y += 8;
56   }
57   for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) {
58     const float16x8_t va01234567 = vld1q_f16(a); a += 8;
59     const float16x8_t vb01234567 = vld1q_f16(b); b += 8;
60 
61     float16x8_t vy01234567 = vsubq_f16(va01234567, vb01234567);
62     vy01234567 = vmaxq_f16(vy01234567, vy_min);
63     vy01234567 = vminq_f16(vy01234567, vy_max);
64     vst1q_f16(y, vy01234567); y += 8;
65   }
66   if XNN_UNLIKELY(n != 0) {
67     const float16x8_t va01234567 = vld1q_f16(a);
68     const float16x8_t vb01234567 = vld1q_f16(b);
69 
70     float16x8_t vy01234567 = vsubq_f16(va01234567, vb01234567);
71     vy01234567 = vmaxq_f16(vy01234567, vy_min);
72     vy01234567 = vminq_f16(vy01234567, vy_max);
73 
74     float16x4_t vy0123 = vget_low_f16(vy01234567);
75     if (n & (4 * sizeof(__fp16))) {
76       vst1_f16(y, vy0123); y += 4;
77       vy0123 = vget_high_f16(vy01234567);
78     }
79 
80     if (n & (2 * sizeof(__fp16))) {
81       vst1_lane_u32((void*) y, vreinterpret_u32_f16(vy0123), 0); y += 2;
82       vy0123 = vext_f16(vy0123, vy0123, 2);
83     }
84 
85     if (n & (1 * sizeof(__fp16))) {
86       vst1_lane_f16(y, vy0123, 0);
87     }
88   }
89 }
90