1 // Copyright 2022 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <arm_neon.h>
9
10 #include <xnnpack/common.h>
11 #include <xnnpack/math.h>
12 #include <xnnpack/transpose.h>
13
14 static const uint8_t pos0[8] = {0, 1, 2, 8, 9, 10, 0, 0};
15 static const uint8_t pos1[8] = {3, 4, 5, 11, 12, 13, 0, 0};
16
xnn_x24_transposec_ukernel__2x2_neon_tbl(const void * input,void * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)17 void xnn_x24_transposec_ukernel__2x2_neon_tbl(
18 const void* input,
19 void* 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 * 3);
26 assert(input_stride >= block_width * 3);
27
28 const size_t tile_height = 2;
29 const size_t tile_width = 2;
30 const size_t tile_wbytes = tile_width * 3;
31 const size_t tile_wbytes_minus_4 = tile_wbytes - 4;
32 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
33 const size_t output_reset = tile_height * output_stride - block_height * 3;
34
35 const size_t tile_stride = tile_height * input_stride;
36
37 const uint8_t* i0 = (const uint8_t*) input;
38 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
39
40 uint8_t* o0 = (uint8_t*) output;
41 uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
42
43 const uint8x8_t vperm0 = vld1_u8(pos0);
44 const uint8x8_t vperm1 = vld1_u8(pos1);
45 do {
46 if XNN_UNPREDICTABLE(block_width < 2) {
47 o1 = o0;
48 }
49 size_t bh = block_height;
50 for (; bh >= 2; bh -= 2) {
51 uint8x8x2_t v;
52 v.val[0] = vld1_u8(i0); i0 = (const uint8_t*) ((uintptr_t) i0 + tile_stride);
53 v.val[1] = vld1_u8(i1); i1 = (const uint8_t*) ((uintptr_t) i1 + tile_stride);
54
55 const uint8x8_t vres0 = vtbl2_u8(v, vperm0);
56 const uint8x8_t vres1 = vtbl2_u8(v, vperm1);
57
58 vst1_lane_u32((void*) o1, vreinterpret_u32_u8(vres1), 0); o1 = (uint8_t*) ((uintptr_t) o1 + 4);
59 vst1_lane_u32((void*) o0, vreinterpret_u32_u8(vres0), 0); o0 = (uint8_t*) ((uintptr_t) o0 + 4);
60 vst1_lane_u16((void*) o1, vreinterpret_u16_u8(vres1), 2); o1 = (uint8_t*) ((uintptr_t) o1 + tile_wbytes_minus_4);
61 vst1_lane_u16((void*) o0, vreinterpret_u16_u8(vres0), 2); o0 = (uint8_t*) ((uintptr_t) o0 + tile_wbytes_minus_4);
62 }
63
64 if (bh != 0) {
65 if XNN_UNPREDICTABLE(bh < 2) {
66 i1 = i0;
67 }
68 uint8x8_t v = vld1_u8(i0);
69
70 const uint8x8_t vres0 = vtbl1_u8(v, vperm0);
71 const uint8x8_t vres1 = vtbl1_u8(v, vperm1);
72
73 if (bh & 1) {
74 vst1_lane_u16((void*) o1, vreinterpret_u16_u8(vres1), 0); o1 += 2;
75 vst1_lane_u16((void*) o0, vreinterpret_u16_u8(vres0), 0); o0 += 2;
76 vst1_lane_u8(o1, vres1, 2); o1 += 1;
77 vst1_lane_u8(o0, vres0, 2); o0 += 1;
78 }
79 }
80 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
81 i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
82 o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
83 o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
84 block_width = doz(block_width, tile_width);
85 } while (block_width != 0);
86 }
87