1 // Auto-generated file. Do not edit!
2 // Template: src/x24-transposec/scalar.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
11 #include <assert.h>
12
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 #include <xnnpack/transpose.h>
16
xnn_x24_transposec_ukernel__2x4_scalar(const void * input,void * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)17 void xnn_x24_transposec_ukernel__2x4_scalar(
18 const void *input,
19 void * output,
20 size_t input_stride,
21 size_t output_stride,
22 size_t block_width,
23 size_t block_height)
24 {
25 assert(output_stride >= block_height * 3);
26 assert(input_stride >= block_width * 3);
27
28 const size_t input_reset = 12 - round_down_po2(block_height, 2) * input_stride;
29 const size_t output_reset = 4 * output_stride - block_height * 3;
30 const size_t input_offset = 2 * input_stride;
31
32 const uint8_t* i0 = (const uint8_t*) input;
33 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
34
35 uint8_t* o0 = (uint8_t*) output;
36 uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
37 uint8_t* o2 = (uint8_t*) ((uintptr_t) o1 + output_stride);
38 uint8_t* o3 = (uint8_t*) ((uintptr_t) o2 + output_stride);
39
40 do {
41 if XNN_UNPREDICTABLE(block_width < 2) {
42 o1 = o0;
43 }
44 if XNN_UNPREDICTABLE(block_width <= 2) {
45 o2 = o0;
46 }
47 if XNN_UNPREDICTABLE(block_width < 4) {
48 o3 = o0;
49 }
50 size_t bh = block_height;
51 for (; bh >= 2; bh -= 2) {
52 o3[0] = i0[9];
53 o3[1] = i0[10];
54 o3[2] = i0[11];
55 o3[3] = i1[9];
56 o3[4] = i1[10];
57 o3[5] = i1[11];
58 o3 += 6;
59 o2[0] = i0[6];
60 o2[1] = i0[7];
61 o2[2] = i0[8];
62 o2[3] = i1[6];
63 o2[4] = i1[7];
64 o2[5] = i1[8];
65 o2 += 6;
66 o1[0] = i0[3];
67 o1[1] = i0[4];
68 o1[2] = i0[5];
69 o1[3] = i1[3];
70 o1[4] = i1[4];
71 o1[5] = i1[5];
72 o1 += 6;
73 o0[0] = i0[0];
74 o0[1] = i0[1];
75 o0[2] = i0[2];
76 o0[3] = i1[0];
77 o0[4] = i1[1];
78 o0[5] = i1[2];
79 o0 += 6;
80 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
81 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
82 }
83 const uint8_t* i = i0;
84 if (bh & 1) {
85 o3[0] = i[9];
86 o3[1] = i[10];
87 o3[2] = i[11];
88 o3 += 3;
89 o2[0] = i[6];
90 o2[1] = i[7];
91 o2[2] = i[8];
92 o2 += 3;
93 o1[0] = i[3];
94 o1[1] = i[4];
95 o1[2] = i[5];
96 o1 += 3;
97 o0[0] = i[0];
98 o0[1] = i[1];
99 o0[2] = i[2];
100 o0 += 3;
101 }
102
103 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
104 i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
105 o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
106 o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
107 o2 = (uint8_t*) ((uintptr_t) o2 + output_reset);
108 o3 = (uint8_t*) ((uintptr_t) o3 + output_reset);
109 block_width = doz(block_width, 4);
110 } while (block_width != 0);
111 }
112