1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/x32-transposec/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 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_x32_transposec_ukernel__2x2_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*4bdc9457SAndroid Build Coastguard Worker void xnn_x32_transposec_ukernel__2x2_scalar_int(
18*4bdc9457SAndroid Build Coastguard Worker const uint32_t *input,
19*4bdc9457SAndroid Build Coastguard Worker uint32_t * 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) XNN_OOB_READS
24*4bdc9457SAndroid Build Coastguard Worker {
25*4bdc9457SAndroid Build Coastguard Worker assert(output_stride >= block_height * sizeof(int));
26*4bdc9457SAndroid Build Coastguard Worker assert(input_stride >= block_width * sizeof(int));
27*4bdc9457SAndroid Build Coastguard Worker
28*4bdc9457SAndroid Build Coastguard Worker const size_t tile_height = 2;
29*4bdc9457SAndroid Build Coastguard Worker const size_t tile_width = 2;
30*4bdc9457SAndroid Build Coastguard Worker const size_t tile_wbytes = tile_width * sizeof(int);
31*4bdc9457SAndroid Build Coastguard Worker const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
32*4bdc9457SAndroid Build Coastguard Worker const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(int);
33*4bdc9457SAndroid Build Coastguard Worker const size_t input_offset = tile_height * input_stride;
34*4bdc9457SAndroid Build Coastguard Worker
35*4bdc9457SAndroid Build Coastguard Worker const int* i0 = (const int*) input;
36*4bdc9457SAndroid Build Coastguard Worker const int* i1 = (const int*) ((uintptr_t) i0 + input_stride);
37*4bdc9457SAndroid Build Coastguard Worker
38*4bdc9457SAndroid Build Coastguard Worker int* o0 = (int*) output;
39*4bdc9457SAndroid Build Coastguard Worker int* o1 = (int*) ((uintptr_t) o0 + output_stride);
40*4bdc9457SAndroid Build Coastguard Worker
41*4bdc9457SAndroid Build Coastguard Worker do {
42*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(block_width < 2) {
43*4bdc9457SAndroid Build Coastguard Worker o1 = o0;
44*4bdc9457SAndroid Build Coastguard Worker }
45*4bdc9457SAndroid Build Coastguard Worker size_t bh = block_height;
46*4bdc9457SAndroid Build Coastguard Worker for (; bh >= 2; bh -= 2) {
47*4bdc9457SAndroid Build Coastguard Worker *o1++ = i0[1];
48*4bdc9457SAndroid Build Coastguard Worker *o1++ = i1[1];
49*4bdc9457SAndroid Build Coastguard Worker *o0++ = i0[0];
50*4bdc9457SAndroid Build Coastguard Worker *o0++ = i1[0];
51*4bdc9457SAndroid Build Coastguard Worker i0 = (const int*) ((uintptr_t) i0 + input_offset);
52*4bdc9457SAndroid Build Coastguard Worker i1 = (const int*) ((uintptr_t) i1 + input_offset);
53*4bdc9457SAndroid Build Coastguard Worker }
54*4bdc9457SAndroid Build Coastguard Worker if (bh & 1) {
55*4bdc9457SAndroid Build Coastguard Worker o1[0] = i0[1];
56*4bdc9457SAndroid Build Coastguard Worker o0[0] = i0[0];
57*4bdc9457SAndroid Build Coastguard Worker }
58*4bdc9457SAndroid Build Coastguard Worker
59*4bdc9457SAndroid Build Coastguard Worker i0 = (const int*) ((uintptr_t) i0 + input_reset);
60*4bdc9457SAndroid Build Coastguard Worker i1 = (const int*) ((uintptr_t) i0 + input_stride);
61*4bdc9457SAndroid Build Coastguard Worker o0 = (int*) ((uintptr_t) o0 + output_reset);
62*4bdc9457SAndroid Build Coastguard Worker o1 = (int*) ((uintptr_t) o1 + output_reset);
63*4bdc9457SAndroid Build Coastguard Worker block_width = doz(block_width, tile_width);
64*4bdc9457SAndroid Build Coastguard Worker } while (block_width != 0);
65*4bdc9457SAndroid Build Coastguard Worker }
66