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_x16_transposec_ukernel__4x4_multi_mov_zip_neon(const uint16_t * input,uint16_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)18 void xnn_x16_transposec_ukernel__4x4_multi_mov_zip_neon(
19 const uint16_t* input,
20 uint16_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(uint16_t));
27 assert(input_stride >= block_width * sizeof(uint16_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(uint16_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint16_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(uint16_t) - tile_hbytes;
36
37 const uint16_t* i0 = input;
38 const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
39 const uint16_t* i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
40 const uint16_t* i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
41 uint16_t* o = (uint16_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 uint16x4_t v2_0 = vld1_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_offset);
51 const uint16x4_t v2_1 = vld1_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset);
52 const uint16x4_t v2_2 = vld1_u16(i2); i2 = (uint16_t*) ((uintptr_t) i2 + input_offset);
53 const uint16x4_t v2_3 = vld1_u16(i3); i3 = (uint16_t*) ((uintptr_t) i3 + input_offset);
54
55 const uint16x4x2_t v1_0 = vzip_u16(v2_0, v2_2);
56 const uint16x4x2_t v1_1 = vzip_u16(v2_1, v2_3);
57
58 const uint16x4x2_t v0_0 = vzip_u16(v1_0.val[0], v1_1.val[0]);
59 const uint16x4x2_t v0_1 = vzip_u16(v1_0.val[1], v1_1.val[1]);
60
61 o = (uint16_t*) ((uintptr_t) o + oN_offset);
62 vst1_u16(o, v0_1.val[1]);
63 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
64 if XNN_UNPREDICTABLE(block_width > 3) {
65 o = oN;
66 }
67 vst1_u16(o, v0_1.val[0]);
68 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
69 if XNN_UNPREDICTABLE(block_width >= 3) {
70 o = oN;
71 }
72 vst1_u16(o, v0_0.val[1]);
73 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
74 if XNN_UNPREDICTABLE(block_width > 1) {
75 o = oN;
76 }
77 vst1_u16(o, v0_0.val[0]);
78 }
79 o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
80
81 if (bh != 0) {
82 const uint16x4_t v2_0 = vld1_u16(i0);
83 if XNN_UNPREDICTABLE(bh < 2) {
84 i1 = i0;
85 }
86 const uint16x4_t v2_1 = vld1_u16(i1);
87 if XNN_UNPREDICTABLE(bh <= 2) {
88 i2 = i0;
89 }
90 const uint16x4_t v2_2 = vld1_u16(i2);
91 const uint16x4_t v2_3 = vmov_n_u16(0);
92
93 const uint16x4x2_t v1_0 = vzip_u16(v2_0, v2_2);
94 const uint16x4x2_t v1_1 = vzip_u16(v2_1, v2_3);
95
96 const uint16x4x2_t v0_0 = vzip_u16(v1_0.val[0], v1_1.val[0]);
97 const uint16x4x2_t v0_1 = vzip_u16(v1_0.val[1], v1_1.val[1]);
98
99 uint16x4_t v0_low = v0_0.val[0];
100 uint16x4_t v1_low = v0_0.val[1];
101 uint16x4_t v2_low = v0_1.val[0];
102 uint16x4_t v3_low = v0_1.val[1];
103
104 if (bh & 2) {
105 o = (uint16_t*) ((uintptr_t) o + oN_stride);
106 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v3_low), 0);
107 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
108 if XNN_UNPREDICTABLE(block_width > 3) {
109 o = oN;
110 }
111 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v2_low), 0);
112 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
113 if XNN_UNPREDICTABLE(block_width >= 3) {
114 o = oN;
115 }
116 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v1_low), 0);
117 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
118 if XNN_UNPREDICTABLE(block_width > 1) {
119 o = oN;
120 }
121 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
122 v0_low = vext_u16(v0_low, v0_low, 2);
123 v1_low = vext_u16(v1_low, v1_low, 2);
124 v2_low = vext_u16(v2_low, v2_low, 2);
125 v3_low = vext_u16(v3_low, v3_low, 2);
126 }
127 if (bh & 1) {
128 o = (uint16_t*) ((uintptr_t) o + oN_stride);
129 vst1_lane_u16(o, v3_low, 0);
130 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
131 if XNN_UNPREDICTABLE(block_width > 3) {
132 o = oN;
133 }
134 vst1_lane_u16(o, v2_low, 0);
135 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
136 if XNN_UNPREDICTABLE(block_width >= 3) {
137 o = oN;
138 }
139 vst1_lane_u16(o, v1_low, 0);
140 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
141 if XNN_UNPREDICTABLE(block_width > 1) {
142 o = oN;
143 }
144 vst1_lane_u16(o, v0_low, 0);
145 }
146 }
147
148 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
149 i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
150 i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
151 i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
152 o = (uint16_t*) ((uintptr_t) o + output_reset);
153 block_width = doz(block_width, tile_width);
154 } while (block_width != 0);
155 }
156