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_c4(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_c4(
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 >= 4 * sizeof(uint8_t); c -= 4 * 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 const int32_t vtl2 = (int32_t) i0[2];
52 const int32_t vtr2 = (int32_t) i1[2];
53 const int32_t vbl2 = (int32_t) i2[2];
54 const int32_t vbr2 = (int32_t) i3[2];
55 const int32_t vtl3 = (int32_t) i0[3];
56 const int32_t vtr3 = (int32_t) i1[3];
57 const int32_t vbl3 = (int32_t) i2[3];
58 const int32_t vbr3 = (int32_t) i3[3];
59 i0 += 4;
60 i1 += 4;
61 i2 += 4;
62 i3 += 4;
63
64 const int32_t vtd0 = vtr0 - vtl0;
65 const int32_t vbd0 = vbr0 - vbl0;
66 const int32_t vtd1 = vtr1 - vtl1;
67 const int32_t vbd1 = vbr1 - vbl1;
68 const int32_t vtd2 = vtr2 - vtl2;
69 const int32_t vbd2 = vbr2 - vbl2;
70 const int32_t vtd3 = vtr3 - vtl3;
71 const int32_t vbd3 = vbr3 - vbl3;
72
73 const int32_t vt0 = (int32_t) ((uint32_t) vtl0 << 11) + vtd0 * valphah;
74 const int32_t vb0 = (int32_t) ((uint32_t) vbl0 << 11) + vbd0 * valphah;
75 const int32_t vt1 = (int32_t) ((uint32_t) vtl1 << 11) + vtd1 * valphah;
76 const int32_t vb1 = (int32_t) ((uint32_t) vbl1 << 11) + vbd1 * valphah;
77 const int32_t vt2 = (int32_t) ((uint32_t) vtl2 << 11) + vtd2 * valphah;
78 const int32_t vb2 = (int32_t) ((uint32_t) vbl2 << 11) + vbd2 * valphah;
79 const int32_t vt3 = (int32_t) ((uint32_t) vtl3 << 11) + vtd3 * valphah;
80 const int32_t vb3 = (int32_t) ((uint32_t) vbl3 << 11) + vbd3 * valphah;
81
82 const int32_t vd0 = vb0 - vt0;
83 const int32_t vd1 = vb1 - vt1;
84 const int32_t vd2 = vb2 - vt2;
85 const int32_t vd3 = vb3 - vt3;
86
87 const int32_t vacc0 = (int32_t) ((uint32_t) vt0 << 11) + vd0 * valphav;
88 const int32_t vacc1 = (int32_t) ((uint32_t) vt1 << 11) + vd1 * valphav;
89 const int32_t vacc2 = (int32_t) ((uint32_t) vt2 << 11) + vd2 * valphav;
90 const int32_t vacc3 = (int32_t) ((uint32_t) vt3 << 11) + vd3 * valphav;
91
92 const int32_t vo0 = math_asr_s32(vacc0 + vrounding, 22);
93 const int32_t vo1 = math_asr_s32(vacc1 + vrounding, 22);
94 const int32_t vo2 = math_asr_s32(vacc2 + vrounding, 22);
95 const int32_t vo3 = math_asr_s32(vacc3 + vrounding, 22);
96
97 output[0] = (uint8_t) vo0;
98 output[1] = (uint8_t) vo1;
99 output[2] = (uint8_t) vo2;
100 output[3] = (uint8_t) vo3;
101 output += 4;
102 }
103 for (; c >= sizeof(uint8_t); c -= sizeof(uint8_t)) {
104 const int32_t vtl = (int32_t) *i0++;
105 const int32_t vtr = (int32_t) *i1++;
106 const int32_t vbl = (int32_t) *i2++;
107 const int32_t vbr = (int32_t) *i3++;
108
109 const int32_t vtd = vtr - vtl;
110 const int32_t vbd = vbr - vbl;
111
112 const int32_t vt = (int32_t) ((uint32_t) vtl << 11) + vtd * valphah;
113 const int32_t vb = (int32_t) ((uint32_t) vbl << 11) + vbd * valphah;
114
115 const int32_t vd = vb - vt;
116
117 const int32_t vacc = (int32_t) ((uint32_t) vt << 11) + vd * valphav;
118
119 const int32_t vo = math_asr_s32(vacc + vrounding, 22);
120
121 *output++ = vo;
122 }
123
124 output = (uint8_t*) ((uintptr_t) output + output_increment);
125 } while (--output_pixels != 0);
126 }
127