xref: /aosp_15_r20/external/XNNPACK/src/x32-transposec/gen/2x2-reuse-switch-zip-neon.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_reuse_switch_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_reuse_switch_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);
35 
36   const uint32_t* i0 = input;
37   uint32_t* o = (uint32_t*) output;
38 
39   do {
40     const size_t rem = min(block_width - 1, 1);
41     const size_t oN_stride = rem * output_stride;
42     size_t bh = block_height;
43     for (; bh >= 2; bh -= 2) {
44       const uint32x2_t v1_0 = vld1_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
45       const uint32x2_t v1_1 = vld1_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
46 
47       const uint32x2x2_t v0_0 = vzip_u32(v1_0, v1_1);
48 
49 
50       uint32_t *oN = (uint32_t*) ((uintptr_t) o + oN_stride);
51       switch (rem) {
52         case 1:
53           vst1_u32(oN, v0_0.val[1]);
54         case 0:
55           vst1_u32(o, v0_0.val[0]); o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
56           break;
57         default:
58           XNN_UNREACHABLE;
59       }
60     }
61 
62     if (bh != 0) {
63       const uint32x2_t v1_0 = vld1_u32(i0);
64       const uint32x2_t v1_1 = vmov_n_u32(0);
65 
66       const uint32x2x2_t v0_0 = vzip_u32(v1_0, v1_1);
67 
68 
69       uint32x2_t v0_low = v0_0.val[0];
70       uint32x2_t v1_low = v0_0.val[1];
71 
72       if (bh & 1) {
73         uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
74         switch (rem) {
75           case 1:
76             vst1_lane_u32(oN, v1_low, 0);
77           case 0:
78             vst1_lane_u32(o, v0_low, 0);
79             break;
80           default:
81             XNN_UNREACHABLE;
82         }
83       }
84     }
85 
86     i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
87     o = (uint32_t*) ((uintptr_t) o + output_reset);
88     block_width = doz(block_width, tile_width);
89   } while (block_width != 0);
90 }
91