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_x8_transposec_ukernel__8x8_reuse_mov_zip_neon(const uint8_t * input,uint8_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)18 void xnn_x8_transposec_ukernel__8x8_reuse_mov_zip_neon(
19 const uint8_t* input,
20 uint8_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(uint8_t));
27 assert(input_stride >= block_width * sizeof(uint8_t));
28
29 const size_t tile_height = 8;
30 const size_t tile_width = 8;
31 const size_t tile_hbytes = tile_height * sizeof(uint8_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint8_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_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint8_t) - tile_hbytes;
35
36 const uint8_t* i0 = input;
37 uint8_t* o = (uint8_t*) ((uintptr_t) output - tile_hbytes);
38 const size_t minus_output_stride = -output_stride;
39
40 do {
41 const size_t rem = min(block_width - 1, 7);
42 const size_t oN_stride = rem * output_stride;
43 const size_t oN_offset = oN_stride + tile_hbytes;
44 size_t bh = block_height;
45 for (; bh >= 8; bh -= 8) {
46 const uint8x8_t v3_0 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
47 const uint8x8_t v3_1 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
48 const uint8x8_t v3_2 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
49 const uint8x8_t v3_3 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
50 const uint8x8_t v3_4 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
51 const uint8x8_t v3_5 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
52 const uint8x8_t v3_6 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
53 const uint8x8_t v3_7 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
54
55 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
56 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
57 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
58 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
59
60 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
61 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
62 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
63 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
64 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
65 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
66 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
67 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
68
69 o = (uint8_t*) ((uintptr_t) o + oN_offset);
70 vst1_u8(o, v0_3.val[1]);
71 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
72 if XNN_UNPREDICTABLE(block_width > 7) {
73 o = oN;
74 }
75 vst1_u8(o, v0_3.val[0]);
76 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
77 if XNN_UNPREDICTABLE(block_width >= 7) {
78 o = oN;
79 }
80 vst1_u8(o, v0_2.val[1]);
81 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
82 if XNN_UNPREDICTABLE(block_width > 5) {
83 o = oN;
84 }
85 vst1_u8(o, v0_2.val[0]);
86 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
87 if XNN_UNPREDICTABLE(block_width >= 5) {
88 o = oN;
89 }
90 vst1_u8(o, v0_1.val[1]);
91 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
92 if XNN_UNPREDICTABLE(block_width > 3) {
93 o = oN;
94 }
95 vst1_u8(o, v0_1.val[0]);
96 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
97 if XNN_UNPREDICTABLE(block_width >= 3) {
98 o = oN;
99 }
100 vst1_u8(o, v0_0.val[1]);
101 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
102 if XNN_UNPREDICTABLE(block_width > 1) {
103 o = oN;
104 }
105 vst1_u8(o, v0_0.val[0]);
106 }
107 o = (uint8_t*) ((uintptr_t) o + tile_hbytes);
108
109 if (bh != 0) {
110 const uint8x8_t v3_0 = vld1_u8(i0);
111 const uint8_t *i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
112 if XNN_UNPREDICTABLE(bh < 2) {
113 i1 = i0;
114 }
115 const uint8x8_t v3_1 = vld1_u8(i1);
116 const uint8_t *i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
117 if XNN_UNPREDICTABLE(bh <= 2) {
118 i2 = i1;
119 }
120 const uint8x8_t v3_2 = vld1_u8(i2);
121 const uint8_t *i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
122 if XNN_UNPREDICTABLE(bh < 4) {
123 i3 = i2;
124 }
125 const uint8x8_t v3_3 = vld1_u8(i3);
126 const uint8_t *i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
127 if XNN_UNPREDICTABLE(bh <= 4) {
128 i4 = i3;
129 }
130 const uint8x8_t v3_4 = vld1_u8(i4);
131 const uint8_t *i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
132 if XNN_UNPREDICTABLE(bh < 6) {
133 i5 = i4;
134 }
135 const uint8x8_t v3_5 = vld1_u8(i5);
136 const uint8_t *i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
137 if XNN_UNPREDICTABLE(bh <= 6) {
138 i6 = i5;
139 }
140 const uint8x8_t v3_6 = vld1_u8(i6);
141 const uint8x8_t v3_7 = vmov_n_u8(0);
142
143 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
144 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
145 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
146 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
147
148 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
149 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
150 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
151 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
152 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
153 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
154 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
155 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
156
157 uint8x8_t v0_low = v0_0.val[0];
158 uint8x8_t v1_low = v0_0.val[1];
159 uint8x8_t v2_low = v0_1.val[0];
160 uint8x8_t v3_low = v0_1.val[1];
161 uint8x8_t v4_low = v0_2.val[0];
162 uint8x8_t v5_low = v0_2.val[1];
163 uint8x8_t v6_low = v0_3.val[0];
164 uint8x8_t v7_low = v0_3.val[1];
165
166 if (bh & 4) {
167 o = (uint8_t*) ((uintptr_t) o + oN_stride);
168 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v7_low), 0);
169 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
170 if XNN_UNPREDICTABLE(block_width > 7) {
171 o = oN;
172 }
173 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v6_low), 0);
174 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
175 if XNN_UNPREDICTABLE(block_width >= 7) {
176 o = oN;
177 }
178 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v5_low), 0);
179 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
180 if XNN_UNPREDICTABLE(block_width > 5) {
181 o = oN;
182 }
183 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v4_low), 0);
184 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
185 if XNN_UNPREDICTABLE(block_width >= 5) {
186 o = oN;
187 }
188 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v3_low), 0);
189 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
190 if XNN_UNPREDICTABLE(block_width > 3) {
191 o = oN;
192 }
193 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v2_low), 0);
194 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
195 if XNN_UNPREDICTABLE(block_width >= 3) {
196 o = oN;
197 }
198 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v1_low), 0);
199 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
200 if XNN_UNPREDICTABLE(block_width > 1) {
201 o = oN;
202 }
203 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v0_low), 0); o += 4;
204 v0_low = vext_u8(v0_low, v0_low, 4);
205 v1_low = vext_u8(v1_low, v1_low, 4);
206 v2_low = vext_u8(v2_low, v2_low, 4);
207 v3_low = vext_u8(v3_low, v3_low, 4);
208 v4_low = vext_u8(v4_low, v4_low, 4);
209 v5_low = vext_u8(v5_low, v5_low, 4);
210 v6_low = vext_u8(v6_low, v6_low, 4);
211 v7_low = vext_u8(v7_low, v7_low, 4);
212 }
213 if (bh & 2) {
214 o = (uint8_t*) ((uintptr_t) o + oN_stride);
215 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v7_low), 0);
216 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
217 if XNN_UNPREDICTABLE(block_width > 7) {
218 o = oN;
219 }
220 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v6_low), 0);
221 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
222 if XNN_UNPREDICTABLE(block_width >= 7) {
223 o = oN;
224 }
225 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v5_low), 0);
226 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
227 if XNN_UNPREDICTABLE(block_width > 5) {
228 o = oN;
229 }
230 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v4_low), 0);
231 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
232 if XNN_UNPREDICTABLE(block_width >= 5) {
233 o = oN;
234 }
235 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v3_low), 0);
236 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
237 if XNN_UNPREDICTABLE(block_width > 3) {
238 o = oN;
239 }
240 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v2_low), 0);
241 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
242 if XNN_UNPREDICTABLE(block_width >= 3) {
243 o = oN;
244 }
245 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v1_low), 0);
246 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
247 if XNN_UNPREDICTABLE(block_width > 1) {
248 o = oN;
249 }
250 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v0_low), 0); o += 2;
251 v0_low = vext_u8(v0_low, v0_low, 2);
252 v1_low = vext_u8(v1_low, v1_low, 2);
253 v2_low = vext_u8(v2_low, v2_low, 2);
254 v3_low = vext_u8(v3_low, v3_low, 2);
255 v4_low = vext_u8(v4_low, v4_low, 2);
256 v5_low = vext_u8(v5_low, v5_low, 2);
257 v6_low = vext_u8(v6_low, v6_low, 2);
258 v7_low = vext_u8(v7_low, v7_low, 2);
259 }
260 if (bh & 1) {
261 o = (uint8_t*) ((uintptr_t) o + oN_stride);
262 vst1_lane_u8(o, v7_low, 0);
263 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
264 if XNN_UNPREDICTABLE(block_width > 7) {
265 o = oN;
266 }
267 vst1_lane_u8(o, v6_low, 0);
268 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
269 if XNN_UNPREDICTABLE(block_width >= 7) {
270 o = oN;
271 }
272 vst1_lane_u8(o, v5_low, 0);
273 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
274 if XNN_UNPREDICTABLE(block_width > 5) {
275 o = oN;
276 }
277 vst1_lane_u8(o, v4_low, 0);
278 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
279 if XNN_UNPREDICTABLE(block_width >= 5) {
280 o = oN;
281 }
282 vst1_lane_u8(o, v3_low, 0);
283 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
284 if XNN_UNPREDICTABLE(block_width > 3) {
285 o = oN;
286 }
287 vst1_lane_u8(o, v2_low, 0);
288 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
289 if XNN_UNPREDICTABLE(block_width >= 3) {
290 o = oN;
291 }
292 vst1_lane_u8(o, v1_low, 0);
293 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
294 if XNN_UNPREDICTABLE(block_width > 1) {
295 o = oN;
296 }
297 vst1_lane_u8(o, v0_low, 0);
298 }
299 }
300
301 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
302 o = (uint8_t*) ((uintptr_t) o + output_reset);
303 block_width = doz(block_width, tile_width);
304 } while (block_width != 0);
305 }
306