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__4x4_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__4x4_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 = 4;
30 const size_t tile_width = 4;
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 const uint32_t* i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
40 const uint32_t* i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
41 uint32_t* o0 = (uint32_t*) output;
42 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
43 uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
44 uint32_t* o3 = (uint32_t*) ((uintptr_t) o2 + output_stride);
45
46 do {
47 if XNN_UNPREDICTABLE(block_width < 2) {
48 o1 = o0;
49 }
50 if XNN_UNPREDICTABLE(block_width <= 2) {
51 o2 = o0;
52 }
53 if XNN_UNPREDICTABLE(block_width < 4) {
54 o3 = o0;
55 }
56 size_t bh = block_height;
57 for (; bh >= 4; bh -= 4) {
58 const uint32x4_t v2_0 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
59 const uint32x4_t v2_1 = vld1q_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
60 const uint32x4_t v2_2 = vld1q_u32(i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
61 const uint32x4_t v2_3 = vld1q_u32(i3); i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
62
63 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
64 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
65
66 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
67 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
68
69 vst1q_u32(o3, v0_1.val[1]); o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
70 vst1q_u32(o2, v0_1.val[0]); o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
71 vst1q_u32(o1, v0_0.val[1]); o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
72 vst1q_u32(o0, v0_0.val[0]); o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
73 }
74
75 if (bh != 0) {
76 const uint32x4_t v2_0 = vld1q_u32(i0);
77 if XNN_UNPREDICTABLE(bh < 2) {
78 i1 = i0;
79 }
80 const uint32x4_t v2_1 = vld1q_u32(i1);
81 if XNN_UNPREDICTABLE(bh <= 2) {
82 i2 = i0;
83 }
84 const uint32x4_t v2_2 = vld1q_u32(i2);
85 const uint32x4_t v2_3 = vmovq_n_u32(0);
86
87 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
88 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
89
90 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
91 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
92
93 uint32x2_t v0_low = vget_low_u32(v0_0.val[0]);
94 uint32x2_t v1_low = vget_low_u32(v0_0.val[1]);
95 uint32x2_t v2_low = vget_low_u32(v0_1.val[0]);
96 uint32x2_t v3_low = vget_low_u32(v0_1.val[1]);
97
98 if (bh & 2) {
99 vst1_u32(o3, v3_low); o3 += 2;
100 vst1_u32(o2, v2_low); o2 += 2;
101 vst1_u32(o1, v1_low); o1 += 2;
102 vst1_u32(o0, v0_low); o0 += 2;
103 v0_low = vget_high_u32(v0_0.val[0]);
104 v1_low = vget_high_u32(v0_0.val[1]);
105 v2_low = vget_high_u32(v0_1.val[0]);
106 v3_low = vget_high_u32(v0_1.val[1]);
107 }
108
109 if (bh & 1) {
110 vst1_lane_u32(o3, v3_low, 0);
111 vst1_lane_u32(o2, v2_low, 0);
112 vst1_lane_u32(o1, v1_low, 0);
113 vst1_lane_u32(o0, v0_low, 0);
114 }
115 }
116
117 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
118 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
119 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
120 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
121 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
122 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
123 o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
124 o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
125 block_width = doz(block_width, tile_width);
126 } while (block_width != 0);
127 }
128