xref: /aosp_15_r20/external/XNNPACK/src/x24-transposec/gen/2x2-scalar.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/x24-transposec/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2022 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker 
11*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/transpose.h>
16*4bdc9457SAndroid Build Coastguard Worker 
xnn_x24_transposec_ukernel__2x2_scalar(const void * input,void * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)17*4bdc9457SAndroid Build Coastguard Worker void xnn_x24_transposec_ukernel__2x2_scalar(
18*4bdc9457SAndroid Build Coastguard Worker     const void *input,
19*4bdc9457SAndroid Build Coastguard Worker     void * output,
20*4bdc9457SAndroid Build Coastguard Worker     size_t input_stride,
21*4bdc9457SAndroid Build Coastguard Worker     size_t output_stride,
22*4bdc9457SAndroid Build Coastguard Worker     size_t block_width,
23*4bdc9457SAndroid Build Coastguard Worker     size_t block_height)
24*4bdc9457SAndroid Build Coastguard Worker {
25*4bdc9457SAndroid Build Coastguard Worker   assert(output_stride >= block_height * 3);
26*4bdc9457SAndroid Build Coastguard Worker   assert(input_stride >= block_width * 3);
27*4bdc9457SAndroid Build Coastguard Worker 
28*4bdc9457SAndroid Build Coastguard Worker   const size_t input_reset = 6 - round_down_po2(block_height, 2) * input_stride;
29*4bdc9457SAndroid Build Coastguard Worker   const size_t output_reset = 2 * output_stride - block_height * 3;
30*4bdc9457SAndroid Build Coastguard Worker   const size_t input_offset = 2 * input_stride;
31*4bdc9457SAndroid Build Coastguard Worker 
32*4bdc9457SAndroid Build Coastguard Worker   const uint8_t* i0 = (const uint8_t*) input;
33*4bdc9457SAndroid Build Coastguard Worker   const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker   uint8_t* o0 = (uint8_t*) output;
36*4bdc9457SAndroid Build Coastguard Worker   uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker   do {
39*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(block_width < 2) {
40*4bdc9457SAndroid Build Coastguard Worker       o1 = o0;
41*4bdc9457SAndroid Build Coastguard Worker     }
42*4bdc9457SAndroid Build Coastguard Worker     size_t bh = block_height;
43*4bdc9457SAndroid Build Coastguard Worker     for (; bh >= 2; bh -= 2) {
44*4bdc9457SAndroid Build Coastguard Worker       o1[0] = i0[3];
45*4bdc9457SAndroid Build Coastguard Worker       o1[1] = i0[4];
46*4bdc9457SAndroid Build Coastguard Worker       o1[2] = i0[5];
47*4bdc9457SAndroid Build Coastguard Worker       o1[3] = i1[3];
48*4bdc9457SAndroid Build Coastguard Worker       o1[4] = i1[4];
49*4bdc9457SAndroid Build Coastguard Worker       o1[5] = i1[5];
50*4bdc9457SAndroid Build Coastguard Worker       o1 += 6;
51*4bdc9457SAndroid Build Coastguard Worker       o0[0] = i0[0];
52*4bdc9457SAndroid Build Coastguard Worker       o0[1] = i0[1];
53*4bdc9457SAndroid Build Coastguard Worker       o0[2] = i0[2];
54*4bdc9457SAndroid Build Coastguard Worker       o0[3] = i1[0];
55*4bdc9457SAndroid Build Coastguard Worker       o0[4] = i1[1];
56*4bdc9457SAndroid Build Coastguard Worker       o0[5] = i1[2];
57*4bdc9457SAndroid Build Coastguard Worker       o0 += 6;
58*4bdc9457SAndroid Build Coastguard Worker       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
59*4bdc9457SAndroid Build Coastguard Worker       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
60*4bdc9457SAndroid Build Coastguard Worker     }
61*4bdc9457SAndroid Build Coastguard Worker     const uint8_t* i = i0;
62*4bdc9457SAndroid Build Coastguard Worker     if (bh & 1) {
63*4bdc9457SAndroid Build Coastguard Worker       o1[0] = i[3];
64*4bdc9457SAndroid Build Coastguard Worker       o1[1] = i[4];
65*4bdc9457SAndroid Build Coastguard Worker       o1[2] = i[5];
66*4bdc9457SAndroid Build Coastguard Worker       o1 += 3;
67*4bdc9457SAndroid Build Coastguard Worker       o0[0] = i[0];
68*4bdc9457SAndroid Build Coastguard Worker       o0[1] = i[1];
69*4bdc9457SAndroid Build Coastguard Worker       o0[2] = i[2];
70*4bdc9457SAndroid Build Coastguard Worker       o0 += 3;
71*4bdc9457SAndroid Build Coastguard Worker     }
72*4bdc9457SAndroid Build Coastguard Worker 
73*4bdc9457SAndroid Build Coastguard Worker     i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
74*4bdc9457SAndroid Build Coastguard Worker     i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
75*4bdc9457SAndroid Build Coastguard Worker     o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
76*4bdc9457SAndroid Build Coastguard Worker     o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
77*4bdc9457SAndroid Build Coastguard Worker     block_width = doz(block_width, 2);
78*4bdc9457SAndroid Build Coastguard Worker   } while (block_width != 0);
79*4bdc9457SAndroid Build Coastguard Worker }
80