1 // Auto-generated file. Do not edit!
2 // Template: src/f16-ibilinear/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/ibilinear.h>
16
17
xnn_f16_ibilinear_ukernel__neonfp16arith_c16(size_t output_pixels,size_t channels,const void ** restrict input,size_t input_offset,const void * restrict weights,void * restrict output,size_t output_increment)18 void xnn_f16_ibilinear_ukernel__neonfp16arith_c16(
19 size_t output_pixels,
20 size_t channels,
21 const void**restrict input,
22 size_t input_offset,
23 const void*restrict weights,
24 void*restrict output,
25 size_t output_increment) XNN_OOB_READS
26 {
27 assert(output_pixels != 0);
28 assert(channels != 0);
29 assert(channels % sizeof(__fp16) == 0);
30
31 __fp16* o = (__fp16*) output;
32 do {
33 const __fp16* i0 = (const __fp16*) ((uintptr_t) input[0] + input_offset);
34 const __fp16* i1 = (const __fp16*) ((uintptr_t) input[1] + input_offset);
35 const __fp16* i2 = (const __fp16*) ((uintptr_t) input[2] + input_offset);
36 const __fp16* i3 = (const __fp16*) ((uintptr_t) input[3] + input_offset);
37 input += 4;
38
39 const float16x8_t valphah = vld1q_dup_f16(weights); weights = (const __fp16*) weights + 1;
40 const float16x8_t valphav = vld1q_dup_f16(weights); weights = (const __fp16*) weights + 1;
41
42 size_t c = channels;
43 for (; c >= 16 * sizeof(__fp16); c -= 16 * sizeof(__fp16)) {
44 const float16x8_t vtl456789AB = vld1q_f16(i0); i0 += 8;
45 const float16x8_t vtr456789AB = vld1q_f16(i1); i1 += 8;
46 const float16x8_t vbl456789AB = vld1q_f16(i2); i2 += 8;
47 const float16x8_t vbr456789AB = vld1q_f16(i3); i3 += 8;
48 const float16x8_t vtlCDEFGHIJ = vld1q_f16(i0); i0 += 8;
49 const float16x8_t vtrCDEFGHIJ = vld1q_f16(i1); i1 += 8;
50 const float16x8_t vblCDEFGHIJ = vld1q_f16(i2); i2 += 8;
51 const float16x8_t vbrCDEFGHIJ = vld1q_f16(i3); i3 += 8;
52
53 const float16x8_t vtd456789AB = vsubq_f16(vtr456789AB, vtl456789AB);
54 const float16x8_t vbd456789AB = vsubq_f16(vbr456789AB, vbl456789AB);
55 const float16x8_t vtdCDEFGHIJ = vsubq_f16(vtrCDEFGHIJ, vtlCDEFGHIJ);
56 const float16x8_t vbdCDEFGHIJ = vsubq_f16(vbrCDEFGHIJ, vblCDEFGHIJ);
57
58 const float16x8_t vt456789AB = vfmaq_f16(vtl456789AB, vtd456789AB, valphah);
59 const float16x8_t vb456789AB = vfmaq_f16(vbl456789AB, vbd456789AB, valphah);
60 const float16x8_t vtCDEFGHIJ = vfmaq_f16(vtlCDEFGHIJ, vtdCDEFGHIJ, valphah);
61 const float16x8_t vbCDEFGHIJ = vfmaq_f16(vblCDEFGHIJ, vbdCDEFGHIJ, valphah);
62
63 const float16x8_t vd456789AB = vsubq_f16(vb456789AB, vt456789AB);
64 const float16x8_t vdCDEFGHIJ = vsubq_f16(vbCDEFGHIJ, vtCDEFGHIJ);
65
66 const float16x8_t vo456789AB = vfmaq_f16(vt456789AB, vd456789AB, valphav);
67 const float16x8_t voCDEFGHIJ = vfmaq_f16(vtCDEFGHIJ, vdCDEFGHIJ, valphav);
68
69 vst1q_f16(o, vo456789AB); o += 8;
70 vst1q_f16(o, voCDEFGHIJ); o += 8;
71 }
72 for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
73 const float16x8_t vtl = vld1q_f16(i0); i0 += 8;
74 const float16x8_t vtr = vld1q_f16(i1); i1 += 8;
75 const float16x8_t vbl = vld1q_f16(i2); i2 += 8;
76 const float16x8_t vbr = vld1q_f16(i3); i3 += 8;
77
78 const float16x8_t vtd = vsubq_f16(vtr, vtl);
79 const float16x8_t vbd = vsubq_f16(vbr, vbl);
80
81 const float16x8_t vt = vfmaq_f16(vtl, vtd, valphah);
82 const float16x8_t vb = vfmaq_f16(vbl, vbd, valphah);
83
84 const float16x8_t vd = vsubq_f16(vb, vt);
85
86 const float16x8_t vo = vfmaq_f16(vt, vd, valphav);
87
88 vst1q_f16(o, vo); o += 8;
89 }
90 if XNN_UNLIKELY(c != 0) {
91 const float16x8_t vtl = vld1q_f16(i0);
92 const float16x8_t vtr = vld1q_f16(i1);
93 const float16x8_t vbl = vld1q_f16(i2);
94 const float16x8_t vbr = vld1q_f16(i3);
95
96 const float16x8_t vtd = vsubq_f16(vtr, vtl);
97 const float16x8_t vbd = vsubq_f16(vbr, vbl);
98
99 const float16x8_t vt = vfmaq_f16(vtl, vtd, valphah);
100 const float16x8_t vb = vfmaq_f16(vbl, vbd, valphah);
101
102 const float16x8_t vd = vsubq_f16(vb, vt);
103
104 float16x8_t vo = vfmaq_f16(vt, vd, valphav);
105
106 float16x4_t vo_lo = vget_low_f16(vo);
107 if (c & (4 * sizeof(__fp16))) {
108 vst1_f16(o, vo_lo); o += 4;
109 vo_lo = vget_high_f16(vo);
110 }
111 if (c & (2 * sizeof(__fp16))) {
112 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vo_lo), 0); o += 2;
113 vo_lo = vext_f16(vo_lo, vo_lo, 2);
114 }
115 if (c & (1 * sizeof(__fp16))) {
116 vst1_lane_f16(o, vo_lo, 0); o += 1;
117 }
118 }
119
120 o = (__fp16*) ((uintptr_t) o + output_increment);
121 } while (--output_pixels != 0);
122 }
123