1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transposec/neon-zip.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 <arm_neon.h>
11
12 #include <assert.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/transpose.h>
17
xnn_x32_transposec_ukernel__2x2_multi_multi_zip_neon(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)18 void xnn_x32_transposec_ukernel__2x2_multi_multi_zip_neon(
19 const uint32_t* input,
20 uint32_t* output,
21 size_t input_stride,
22 size_t output_stride,
23 size_t block_width,
24 size_t block_height) XNN_OOB_READS
25 {
26 assert(output_stride >= block_height * sizeof(uint32_t));
27 assert(input_stride >= block_width * sizeof(uint32_t));
28
29 const size_t tile_height = 2;
30 const size_t tile_width = 2;
31 const size_t tile_hbytes = tile_height * sizeof(uint32_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint32_t);
33 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
34 const size_t input_offset = tile_height * input_stride;
35 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);
36
37 const uint32_t* i0 = input;
38 const uint32_t* i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
39 uint32_t* o0 = (uint32_t*) output;
40 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
41
42 do {
43 if XNN_UNPREDICTABLE(block_width < 2) {
44 o1 = o0;
45 }
46 size_t bh = block_height;
47 for (; bh >= 2; bh -= 2) {
48 const uint32x2_t v1_0 = vld1_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
49 const uint32x2_t v1_1 = vld1_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
50
51 const uint32x2x2_t v0_0 = vzip_u32(v1_0, v1_1);
52
53
54 vst1_u32(o1, v0_0.val[1]); o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
55 vst1_u32(o0, v0_0.val[0]); o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
56 }
57
58 if (bh != 0) {
59 const uint32x2_t v1_0 = vld1_u32(i0);
60 const uint32x2_t v1_1 = vmov_n_u32(0);
61
62 const uint32x2x2_t v0_0 = vzip_u32(v1_0, v1_1);
63
64
65 uint32x2_t v0_low = v0_0.val[0];
66 uint32x2_t v1_low = v0_0.val[1];
67
68 if (bh & 1) {
69 vst1_lane_u32(o1, v1_low, 0);
70 vst1_lane_u32(o0, v0_low, 0);
71 }
72 }
73
74 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
75 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
76 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
77 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
78 block_width = doz(block_width, tile_width);
79 } while (block_width != 0);
80 }
81