1 // Copyright 2021 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[16] = {0, 1, 2, 3, 16, 17, 18, 19, 32, 33, 34, 35, 48, 49, 50, 51};
15 static const uint8_t pos1[16] = {4, 5, 6, 7, 20, 21, 22, 23, 36, 37, 38, 39, 52, 53, 54, 55};
16 static const uint8_t pos2[16] = {8, 9, 10, 11, 24, 25, 26, 27, 40, 41, 42, 43, 56, 57, 58, 59};
17 static const uint8_t pos3[16] = {12, 13, 14, 15, 28, 29, 30, 31, 44, 45, 46, 47, 60, 61, 62, 63};
18
xnn_x32_transposec_ukernel__4x4_aarch64_neon_tbl(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)19 void xnn_x32_transposec_ukernel__4x4_aarch64_neon_tbl(
20 const uint32_t* input,
21 uint32_t* output,
22 size_t input_stride,
23 size_t output_stride,
24 size_t block_width,
25 size_t block_height) XNN_OOB_READS
26 {
27 assert(output_stride >= block_height * sizeof(uint32_t));
28 assert(input_stride >= block_width * sizeof(uint32_t));
29
30 const size_t tile_height = 4;
31 const size_t tile_width = 4;
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_height * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);
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 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
40 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
41
42 uint8_t* o0 = (uint8_t*) output;
43 uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
44 uint8_t* o2 = (uint8_t*) ((uintptr_t) o1 + output_stride);
45 uint8_t* o3 = (uint8_t*) ((uintptr_t) o2 + output_stride);
46
47 const uint8x16_t vperm0 = vld1q_u8(pos0);
48 const uint8x16_t vperm1 = vld1q_u8(pos1);
49 const uint8x16_t vperm2 = vld1q_u8(pos2);
50 const uint8x16_t vperm3 = vld1q_u8(pos3);
51 do {
52 if XNN_UNPREDICTABLE(block_width < 2) {
53 o1 = o0;
54 }
55 if XNN_UNPREDICTABLE(block_width <= 2) {
56 o2 = o0;
57 }
58 if XNN_UNPREDICTABLE(block_width < 4) {
59 o3 = o0;
60 }
61 size_t bh = block_height;
62 for (; bh >= 4; bh -= 4) {
63 uint8x16x4_t v;
64 v.val[0] = vld1q_u8(i0); i0 = (const uint8_t*) ((uintptr_t) i0 + tile_stride);
65 v.val[1] = vld1q_u8(i1); i1 = (const uint8_t*) ((uintptr_t) i1 + tile_stride);
66 v.val[2] = vld1q_u8(i2); i2 = (const uint8_t*) ((uintptr_t) i2 + tile_stride);
67 v.val[3] = vld1q_u8(i3); i3 = (const uint8_t*) ((uintptr_t) i3 + tile_stride);
68
69 uint8x16_t vres0 = vqtbl4q_u8(v, vperm0);
70 uint8x16_t vres1 = vqtbl4q_u8(v, vperm1);
71 uint8x16_t vres2 = vqtbl4q_u8(v, vperm2);
72 uint8x16_t vres3 = vqtbl4q_u8(v, vperm3);
73
74 vst1q_u8(o3, vres3); o3 = (uint8_t*) ((uintptr_t) o3 + tile_wbytes);
75 vst1q_u8(o2, vres2); o2 = (uint8_t*) ((uintptr_t) o2 + tile_wbytes);
76 vst1q_u8(o1, vres1); o1 = (uint8_t*) ((uintptr_t) o1 + tile_wbytes);
77 vst1q_u8(o0, vres0); o0 = (uint8_t*) ((uintptr_t) o0 + tile_wbytes);
78 }
79
80 if (bh != 0) {
81 if XNN_UNPREDICTABLE(bh <= 2) {
82 i2 = i0;
83 }
84 if XNN_UNPREDICTABLE(bh < 2) {
85 i1 = i0;
86 }
87 uint8x16x4_t v;
88 v.val[0] = vld1q_u8(i0);
89 v.val[1] = vld1q_u8(i1);
90 v.val[2] = vld1q_u8(i2);
91
92 uint8x16_t vres0 = vqtbl4q_u8(v, vperm0);
93 uint8x16_t vres1 = vqtbl4q_u8(v, vperm1);
94 uint8x16_t vres2 = vqtbl4q_u8(v, vperm2);
95 uint8x16_t vres3 = vqtbl4q_u8(v, vperm3);
96
97 uint8x8_t vres0_low = vget_low_u8(vres0);
98 uint8x8_t vres1_low = vget_low_u8(vres1);
99 uint8x8_t vres2_low = vget_low_u8(vres2);
100 uint8x8_t vres3_low = vget_low_u8(vres3);
101
102 if (bh & 2) {
103 vst1_u8(o3, vres3_low); o3 += 8;
104 vst1_u8(o2, vres2_low); o2 += 8;
105 vst1_u8(o1, vres1_low); o1 += 8;
106 vst1_u8(o0, vres0_low); o0 += 8;
107 vres0_low = vget_high_u8(vres0);
108 vres1_low = vget_high_u8(vres1);
109 vres2_low = vget_high_u8(vres2);
110 vres3_low = vget_high_u8(vres3);
111 }
112 if (bh & 1) {
113 vst1_lane_u32((void*) o3, vreinterpret_u32_u8(vres3_low), 0);
114 vst1_lane_u32((void*) o2, vreinterpret_u32_u8(vres2_low), 0);
115 vst1_lane_u32((void*) o1, vreinterpret_u32_u8(vres1_low), 0);
116 vst1_lane_u32((void*) o0, vreinterpret_u32_u8(vres0_low), 0);
117 }
118 }
119 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
120 i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
121 i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
122 i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
123 o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
124 o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
125 o2 = (uint8_t*) ((uintptr_t) o2 + output_reset);
126 o3 = (uint8_t*) ((uintptr_t) o3 + output_reset);
127 block_width = doz(block_width, tile_width);
128 } while (block_width != 0);
129 }
130