1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transposec/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
11 #include <assert.h>
12
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 #include <xnnpack/transpose.h>
16
xnn_x32_transposec_ukernel__2x4_scalar_int(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)17 void xnn_x32_transposec_ukernel__2x4_scalar_int(
18 const uint32_t *input,
19 uint32_t * output,
20 size_t input_stride,
21 size_t output_stride,
22 size_t block_width,
23 size_t block_height) XNN_OOB_READS
24 {
25 assert(output_stride >= block_height * sizeof(int));
26 assert(input_stride >= block_width * sizeof(int));
27
28 const size_t tile_height = 2;
29 const size_t tile_width = 4;
30 const size_t tile_wbytes = tile_width * sizeof(int);
31 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
32 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(int);
33 const size_t input_offset = tile_height * input_stride;
34
35 const int* i0 = (const int*) input;
36 const int* i1 = (const int*) ((uintptr_t) i0 + input_stride);
37
38 int* o0 = (int*) output;
39 int* o1 = (int*) ((uintptr_t) o0 + output_stride);
40 int* o2 = (int*) ((uintptr_t) o1 + output_stride);
41 int* o3 = (int*) ((uintptr_t) o2 + output_stride);
42
43 do {
44 if XNN_UNPREDICTABLE(block_width < 2) {
45 o1 = o0;
46 }
47 if XNN_UNPREDICTABLE(block_width <= 2) {
48 o2 = o0;
49 }
50 if XNN_UNPREDICTABLE(block_width < 4) {
51 o3 = o0;
52 }
53 size_t bh = block_height;
54 for (; bh >= 2; bh -= 2) {
55 *o3++ = i0[3];
56 *o3++ = i1[3];
57 *o2++ = i0[2];
58 *o2++ = i1[2];
59 *o1++ = i0[1];
60 *o1++ = i1[1];
61 *o0++ = i0[0];
62 *o0++ = i1[0];
63 i0 = (const int*) ((uintptr_t) i0 + input_offset);
64 i1 = (const int*) ((uintptr_t) i1 + input_offset);
65 }
66 if (bh & 1) {
67 o3[0] = i0[3];
68 o2[0] = i0[2];
69 o1[0] = i0[1];
70 o0[0] = i0[0];
71 }
72
73 i0 = (const int*) ((uintptr_t) i0 + input_reset);
74 i1 = (const int*) ((uintptr_t) i0 + input_stride);
75 o0 = (int*) ((uintptr_t) o0 + output_reset);
76 o1 = (int*) ((uintptr_t) o1 + output_reset);
77 o2 = (int*) ((uintptr_t) o2 + output_reset);
78 o3 = (int*) ((uintptr_t) o3 + output_reset);
79 block_width = doz(block_width, tile_width);
80 } while (block_width != 0);
81 }
82