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_multi_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_multi_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 input_offset = tile_height * input_stride;
35 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint8_t) - tile_hbytes;
36
37 const uint8_t* i0 = 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 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
42 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
43 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
44 const uint8_t* i7 = (const uint8_t*) ((uintptr_t) i6 + input_stride);
45 uint8_t* o = (uint8_t*) ((uintptr_t) output - tile_hbytes);
46 const size_t minus_output_stride = -output_stride;
47
48 do {
49 const size_t rem = min(block_width - 1, 7);
50 const size_t oN_stride = rem * output_stride;
51 const size_t oN_offset = oN_stride + tile_hbytes;
52 size_t bh = block_height;
53 for (; bh >= 8; bh -= 8) {
54 const uint8x8_t v3_0 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_offset);
55 const uint8x8_t v3_1 = vld1_u8(i1); i1 = (uint8_t*) ((uintptr_t) i1 + input_offset);
56 const uint8x8_t v3_2 = vld1_u8(i2); i2 = (uint8_t*) ((uintptr_t) i2 + input_offset);
57 const uint8x8_t v3_3 = vld1_u8(i3); i3 = (uint8_t*) ((uintptr_t) i3 + input_offset);
58 const uint8x8_t v3_4 = vld1_u8(i4); i4 = (uint8_t*) ((uintptr_t) i4 + input_offset);
59 const uint8x8_t v3_5 = vld1_u8(i5); i5 = (uint8_t*) ((uintptr_t) i5 + input_offset);
60 const uint8x8_t v3_6 = vld1_u8(i6); i6 = (uint8_t*) ((uintptr_t) i6 + input_offset);
61 const uint8x8_t v3_7 = vld1_u8(i7); i7 = (uint8_t*) ((uintptr_t) i7 + input_offset);
62
63 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
64 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
65 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
66 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
67
68 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
69 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
70 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
71 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
72 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
73 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
74 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
75 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
76
77 o = (uint8_t*) ((uintptr_t) o + oN_offset);
78 vst1_u8(o, v0_3.val[1]);
79 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
80 if XNN_UNPREDICTABLE(block_width > 7) {
81 o = oN;
82 }
83 vst1_u8(o, v0_3.val[0]);
84 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
85 if XNN_UNPREDICTABLE(block_width >= 7) {
86 o = oN;
87 }
88 vst1_u8(o, v0_2.val[1]);
89 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
90 if XNN_UNPREDICTABLE(block_width > 5) {
91 o = oN;
92 }
93 vst1_u8(o, v0_2.val[0]);
94 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
95 if XNN_UNPREDICTABLE(block_width >= 5) {
96 o = oN;
97 }
98 vst1_u8(o, v0_1.val[1]);
99 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
100 if XNN_UNPREDICTABLE(block_width > 3) {
101 o = oN;
102 }
103 vst1_u8(o, v0_1.val[0]);
104 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
105 if XNN_UNPREDICTABLE(block_width >= 3) {
106 o = oN;
107 }
108 vst1_u8(o, v0_0.val[1]);
109 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
110 if XNN_UNPREDICTABLE(block_width > 1) {
111 o = oN;
112 }
113 vst1_u8(o, v0_0.val[0]);
114 }
115 o = (uint8_t*) ((uintptr_t) o + tile_hbytes);
116
117 if (bh != 0) {
118 const uint8x8_t v3_0 = vld1_u8(i0);
119 if XNN_UNPREDICTABLE(bh < 2) {
120 i1 = i0;
121 }
122 const uint8x8_t v3_1 = vld1_u8(i1);
123 if XNN_UNPREDICTABLE(bh <= 2) {
124 i2 = i0;
125 }
126 const uint8x8_t v3_2 = vld1_u8(i2);
127 if XNN_UNPREDICTABLE(bh < 4) {
128 i3 = i0;
129 }
130 const uint8x8_t v3_3 = vld1_u8(i3);
131 if XNN_UNPREDICTABLE(bh <= 4) {
132 i4 = i0;
133 }
134 const uint8x8_t v3_4 = vld1_u8(i4);
135 if XNN_UNPREDICTABLE(bh < 6) {
136 i5 = i0;
137 }
138 const uint8x8_t v3_5 = vld1_u8(i5);
139 if XNN_UNPREDICTABLE(bh <= 6) {
140 i6 = i0;
141 }
142 const uint8x8_t v3_6 = vld1_u8(i6);
143 const uint8x8_t v3_7 = vmov_n_u8(0);
144
145 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
146 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
147 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
148 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
149
150 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
151 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
152 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
153 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
154 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
155 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
156 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
157 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
158
159 uint8x8_t v0_low = v0_0.val[0];
160 uint8x8_t v1_low = v0_0.val[1];
161 uint8x8_t v2_low = v0_1.val[0];
162 uint8x8_t v3_low = v0_1.val[1];
163 uint8x8_t v4_low = v0_2.val[0];
164 uint8x8_t v5_low = v0_2.val[1];
165 uint8x8_t v6_low = v0_3.val[0];
166 uint8x8_t v7_low = v0_3.val[1];
167
168 if (bh & 4) {
169 o = (uint8_t*) ((uintptr_t) o + oN_stride);
170 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v7_low), 0);
171 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
172 if XNN_UNPREDICTABLE(block_width > 7) {
173 o = oN;
174 }
175 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v6_low), 0);
176 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
177 if XNN_UNPREDICTABLE(block_width >= 7) {
178 o = oN;
179 }
180 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v5_low), 0);
181 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
182 if XNN_UNPREDICTABLE(block_width > 5) {
183 o = oN;
184 }
185 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v4_low), 0);
186 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
187 if XNN_UNPREDICTABLE(block_width >= 5) {
188 o = oN;
189 }
190 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v3_low), 0);
191 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
192 if XNN_UNPREDICTABLE(block_width > 3) {
193 o = oN;
194 }
195 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v2_low), 0);
196 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
197 if XNN_UNPREDICTABLE(block_width >= 3) {
198 o = oN;
199 }
200 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v1_low), 0);
201 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
202 if XNN_UNPREDICTABLE(block_width > 1) {
203 o = oN;
204 }
205 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v0_low), 0); o += 4;
206 v0_low = vext_u8(v0_low, v0_low, 4);
207 v1_low = vext_u8(v1_low, v1_low, 4);
208 v2_low = vext_u8(v2_low, v2_low, 4);
209 v3_low = vext_u8(v3_low, v3_low, 4);
210 v4_low = vext_u8(v4_low, v4_low, 4);
211 v5_low = vext_u8(v5_low, v5_low, 4);
212 v6_low = vext_u8(v6_low, v6_low, 4);
213 v7_low = vext_u8(v7_low, v7_low, 4);
214 }
215 if (bh & 2) {
216 o = (uint8_t*) ((uintptr_t) o + oN_stride);
217 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v7_low), 0);
218 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
219 if XNN_UNPREDICTABLE(block_width > 7) {
220 o = oN;
221 }
222 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v6_low), 0);
223 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
224 if XNN_UNPREDICTABLE(block_width >= 7) {
225 o = oN;
226 }
227 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v5_low), 0);
228 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
229 if XNN_UNPREDICTABLE(block_width > 5) {
230 o = oN;
231 }
232 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v4_low), 0);
233 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
234 if XNN_UNPREDICTABLE(block_width >= 5) {
235 o = oN;
236 }
237 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v3_low), 0);
238 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
239 if XNN_UNPREDICTABLE(block_width > 3) {
240 o = oN;
241 }
242 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v2_low), 0);
243 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
244 if XNN_UNPREDICTABLE(block_width >= 3) {
245 o = oN;
246 }
247 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v1_low), 0);
248 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
249 if XNN_UNPREDICTABLE(block_width > 1) {
250 o = oN;
251 }
252 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v0_low), 0); o += 2;
253 v0_low = vext_u8(v0_low, v0_low, 2);
254 v1_low = vext_u8(v1_low, v1_low, 2);
255 v2_low = vext_u8(v2_low, v2_low, 2);
256 v3_low = vext_u8(v3_low, v3_low, 2);
257 v4_low = vext_u8(v4_low, v4_low, 2);
258 v5_low = vext_u8(v5_low, v5_low, 2);
259 v6_low = vext_u8(v6_low, v6_low, 2);
260 v7_low = vext_u8(v7_low, v7_low, 2);
261 }
262 if (bh & 1) {
263 o = (uint8_t*) ((uintptr_t) o + oN_stride);
264 vst1_lane_u8(o, v7_low, 0);
265 uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
266 if XNN_UNPREDICTABLE(block_width > 7) {
267 o = oN;
268 }
269 vst1_lane_u8(o, v6_low, 0);
270 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
271 if XNN_UNPREDICTABLE(block_width >= 7) {
272 o = oN;
273 }
274 vst1_lane_u8(o, v5_low, 0);
275 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
276 if XNN_UNPREDICTABLE(block_width > 5) {
277 o = oN;
278 }
279 vst1_lane_u8(o, v4_low, 0);
280 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
281 if XNN_UNPREDICTABLE(block_width >= 5) {
282 o = oN;
283 }
284 vst1_lane_u8(o, v3_low, 0);
285 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
286 if XNN_UNPREDICTABLE(block_width > 3) {
287 o = oN;
288 }
289 vst1_lane_u8(o, v2_low, 0);
290 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
291 if XNN_UNPREDICTABLE(block_width >= 3) {
292 o = oN;
293 }
294 vst1_lane_u8(o, v1_low, 0);
295 oN = (uint8_t*) ((uintptr_t) o + minus_output_stride);
296 if XNN_UNPREDICTABLE(block_width > 1) {
297 o = oN;
298 }
299 vst1_lane_u8(o, v0_low, 0);
300 }
301 }
302
303 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
304 i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
305 i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
306 i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
307 i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
308 i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
309 i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
310 i7 = (const uint8_t*) ((uintptr_t) i6 + input_stride);
311 o = (uint8_t*) ((uintptr_t) o + output_reset);
312 block_width = doz(block_width, tile_width);
313 } while (block_width != 0);
314 }
315