1 // Auto-generated file. Do not edit!
2 // Template: src/s8-ibilinear/scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <xnnpack/ibilinear.h>
13 #include <xnnpack/math.h>
14
15
xnn_u8_ibilinear_ukernel__scalar_c2(size_t output_pixels,size_t channels,const uint8_t ** restrict input,size_t input_offset,const int16_t * restrict weights,uint8_t * restrict output,size_t output_increment)16 void xnn_u8_ibilinear_ukernel__scalar_c2(
17 size_t output_pixels,
18 size_t channels,
19 const uint8_t**restrict input,
20 size_t input_offset,
21 const int16_t*restrict weights,
22 uint8_t*restrict output,
23 size_t output_increment)
24 {
25 assert(output_pixels != 0);
26 assert(channels != 0);
27
28 do {
29 const uint8_t* i0 = (const uint8_t*) ((uintptr_t) input[0] + input_offset);
30 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) input[1] + input_offset);
31 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) input[2] + input_offset);
32 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) input[3] + input_offset);
33 input += 4;
34
35 const int32_t valphah = (int32_t) (uint32_t) (uint16_t) weights[0];
36 const int32_t valphav = (int32_t) (uint32_t) (uint16_t) weights[1];
37 weights += 2;
38
39 const int32_t vrounding = INT32_C(0x00200000);
40
41 size_t c = channels;
42 for (; c >= 2 * sizeof(uint8_t); c -= 2 * sizeof(uint8_t)) {
43 const int32_t vtl0 = (int32_t) i0[0];
44 const int32_t vtr0 = (int32_t) i1[0];
45 const int32_t vbl0 = (int32_t) i2[0];
46 const int32_t vbr0 = (int32_t) i3[0];
47 const int32_t vtl1 = (int32_t) i0[1];
48 const int32_t vtr1 = (int32_t) i1[1];
49 const int32_t vbl1 = (int32_t) i2[1];
50 const int32_t vbr1 = (int32_t) i3[1];
51 i0 += 2;
52 i1 += 2;
53 i2 += 2;
54 i3 += 2;
55
56 const int32_t vtd0 = vtr0 - vtl0;
57 const int32_t vbd0 = vbr0 - vbl0;
58 const int32_t vtd1 = vtr1 - vtl1;
59 const int32_t vbd1 = vbr1 - vbl1;
60
61 const int32_t vt0 = (int32_t) ((uint32_t) vtl0 << 11) + vtd0 * valphah;
62 const int32_t vb0 = (int32_t) ((uint32_t) vbl0 << 11) + vbd0 * valphah;
63 const int32_t vt1 = (int32_t) ((uint32_t) vtl1 << 11) + vtd1 * valphah;
64 const int32_t vb1 = (int32_t) ((uint32_t) vbl1 << 11) + vbd1 * valphah;
65
66 const int32_t vd0 = vb0 - vt0;
67 const int32_t vd1 = vb1 - vt1;
68
69 const int32_t vacc0 = (int32_t) ((uint32_t) vt0 << 11) + vd0 * valphav;
70 const int32_t vacc1 = (int32_t) ((uint32_t) vt1 << 11) + vd1 * valphav;
71
72 const int32_t vo0 = math_asr_s32(vacc0 + vrounding, 22);
73 const int32_t vo1 = math_asr_s32(vacc1 + vrounding, 22);
74
75 output[0] = (uint8_t) vo0;
76 output[1] = (uint8_t) vo1;
77 output += 2;
78 }
79 for (; c >= sizeof(uint8_t); c -= sizeof(uint8_t)) {
80 const int32_t vtl = (int32_t) *i0++;
81 const int32_t vtr = (int32_t) *i1++;
82 const int32_t vbl = (int32_t) *i2++;
83 const int32_t vbr = (int32_t) *i3++;
84
85 const int32_t vtd = vtr - vtl;
86 const int32_t vbd = vbr - vbl;
87
88 const int32_t vt = (int32_t) ((uint32_t) vtl << 11) + vtd * valphah;
89 const int32_t vb = (int32_t) ((uint32_t) vbl << 11) + vbd * valphah;
90
91 const int32_t vd = vb - vt;
92
93 const int32_t vacc = (int32_t) ((uint32_t) vt << 11) + vd * valphav;
94
95 const int32_t vo = math_asr_s32(vacc + vrounding, 22);
96
97 *output++ = vo;
98 }
99
100 output = (uint8_t*) ((uintptr_t) output + output_increment);
101 } while (--output_pixels != 0);
102 }
103