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_mov_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_mov_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) - tile_hbytes;
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* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
42 const size_t minus_output_stride = -output_stride;
43
44 do {
45 const size_t rem = min(block_width - 1, 3);
46 const size_t oN_stride = rem * output_stride;
47 const size_t oN_offset = oN_stride + tile_hbytes;
48 size_t bh = block_height;
49 for (; bh >= 4; bh -= 4) {
50 const uint32x4_t v2_0 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
51 const uint32x4_t v2_1 = vld1q_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
52 const uint32x4_t v2_2 = vld1q_u32(i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
53 const uint32x4_t v2_3 = vld1q_u32(i3); i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
54
55 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
56 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
57
58 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
59 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
60
61 o = (uint32_t*) ((uintptr_t) o + oN_offset);
62 vst1q_u32(o, v0_1.val[1]);
63 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
64 if XNN_UNPREDICTABLE(block_width > 3) {
65 o = oN;
66 }
67 vst1q_u32(o, v0_1.val[0]);
68 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
69 if XNN_UNPREDICTABLE(block_width >= 3) {
70 o = oN;
71 }
72 vst1q_u32(o, v0_0.val[1]);
73 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
74 if XNN_UNPREDICTABLE(block_width > 1) {
75 o = oN;
76 }
77 vst1q_u32(o, v0_0.val[0]);
78 }
79 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
80
81 if (bh != 0) {
82 const uint32x4_t v2_0 = vld1q_u32(i0);
83 if XNN_UNPREDICTABLE(bh < 2) {
84 i1 = i0;
85 }
86 const uint32x4_t v2_1 = vld1q_u32(i1);
87 if XNN_UNPREDICTABLE(bh <= 2) {
88 i2 = i0;
89 }
90 const uint32x4_t v2_2 = vld1q_u32(i2);
91 const uint32x4_t v2_3 = vmovq_n_u32(0);
92
93 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
94 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
95
96 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
97 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
98
99 uint32x2_t v0_low = vget_low_u32(v0_0.val[0]);
100 uint32x2_t v1_low = vget_low_u32(v0_0.val[1]);
101 uint32x2_t v2_low = vget_low_u32(v0_1.val[0]);
102 uint32x2_t v3_low = vget_low_u32(v0_1.val[1]);
103
104 if (bh & 2) {
105 o = (uint32_t*) ((uintptr_t) o + oN_stride);
106 vst1_u32(o, v3_low);
107 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
108 if XNN_UNPREDICTABLE(block_width > 3) {
109 o = oN;
110 }
111 vst1_u32(o, v2_low);
112 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
113 if XNN_UNPREDICTABLE(block_width >= 3) {
114 o = oN;
115 }
116 vst1_u32(o, v1_low);
117 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
118 if XNN_UNPREDICTABLE(block_width > 1) {
119 o = oN;
120 }
121 vst1_u32(o, v0_low); o += 2;
122 v0_low = vget_high_u32(v0_0.val[0]);
123 v1_low = vget_high_u32(v0_0.val[1]);
124 v2_low = vget_high_u32(v0_1.val[0]);
125 v3_low = vget_high_u32(v0_1.val[1]);
126 }
127
128 if (bh & 1) {
129 o = (uint32_t*) ((uintptr_t) o + oN_stride);
130 vst1_lane_u32(o, v3_low, 0);
131 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
132 if XNN_UNPREDICTABLE(block_width > 3) {
133 o = oN;
134 }
135 vst1_lane_u32(o, v2_low, 0);
136 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
137 if XNN_UNPREDICTABLE(block_width >= 3) {
138 o = oN;
139 }
140 vst1_lane_u32(o, v1_low, 0);
141 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
142 if XNN_UNPREDICTABLE(block_width > 1) {
143 o = oN;
144 }
145 vst1_lane_u32(o, v0_low, 0);
146 }
147 }
148
149 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
150 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
151 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
152 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
153 o = (uint32_t*) ((uintptr_t) o + output_reset);
154 block_width = doz(block_width, tile_width);
155 } while (block_width != 0);
156 }
157