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__8x8_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__8x8_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 = 8;
30 const size_t tile_width = 8;
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 const uint16_t* i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
42 const uint16_t* i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
43 const uint16_t* i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
44 const uint16_t* i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
45 uint16_t* o = (uint16_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 uint16x8_t v3_0 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_offset);
55 const uint16x8_t v3_1 = vld1q_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset);
56 const uint16x8_t v3_2 = vld1q_u16(i2); i2 = (uint16_t*) ((uintptr_t) i2 + input_offset);
57 const uint16x8_t v3_3 = vld1q_u16(i3); i3 = (uint16_t*) ((uintptr_t) i3 + input_offset);
58 const uint16x8_t v3_4 = vld1q_u16(i4); i4 = (uint16_t*) ((uintptr_t) i4 + input_offset);
59 const uint16x8_t v3_5 = vld1q_u16(i5); i5 = (uint16_t*) ((uintptr_t) i5 + input_offset);
60 const uint16x8_t v3_6 = vld1q_u16(i6); i6 = (uint16_t*) ((uintptr_t) i6 + input_offset);
61 const uint16x8_t v3_7 = vld1q_u16(i7); i7 = (uint16_t*) ((uintptr_t) i7 + input_offset);
62
63 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
64 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
65 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
66 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
67
68 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
69 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
70 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
71 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
72 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
73 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
74 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
75 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
76
77 o = (uint16_t*) ((uintptr_t) o + oN_offset);
78 vst1q_u16(o, v0_3.val[1]);
79 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
80 if XNN_UNPREDICTABLE(block_width > 7) {
81 o = oN;
82 }
83 vst1q_u16(o, v0_3.val[0]);
84 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
85 if XNN_UNPREDICTABLE(block_width >= 7) {
86 o = oN;
87 }
88 vst1q_u16(o, v0_2.val[1]);
89 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
90 if XNN_UNPREDICTABLE(block_width > 5) {
91 o = oN;
92 }
93 vst1q_u16(o, v0_2.val[0]);
94 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
95 if XNN_UNPREDICTABLE(block_width >= 5) {
96 o = oN;
97 }
98 vst1q_u16(o, v0_1.val[1]);
99 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
100 if XNN_UNPREDICTABLE(block_width > 3) {
101 o = oN;
102 }
103 vst1q_u16(o, v0_1.val[0]);
104 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
105 if XNN_UNPREDICTABLE(block_width >= 3) {
106 o = oN;
107 }
108 vst1q_u16(o, v0_0.val[1]);
109 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
110 if XNN_UNPREDICTABLE(block_width > 1) {
111 o = oN;
112 }
113 vst1q_u16(o, v0_0.val[0]);
114 }
115 o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
116
117 if (bh != 0) {
118 const uint16x8_t v3_0 = vld1q_u16(i0);
119 if XNN_UNPREDICTABLE(bh < 2) {
120 i1 = i0;
121 }
122 const uint16x8_t v3_1 = vld1q_u16(i1);
123 if XNN_UNPREDICTABLE(bh <= 2) {
124 i2 = i0;
125 }
126 const uint16x8_t v3_2 = vld1q_u16(i2);
127 if XNN_UNPREDICTABLE(bh < 4) {
128 i3 = i0;
129 }
130 const uint16x8_t v3_3 = vld1q_u16(i3);
131 if XNN_UNPREDICTABLE(bh <= 4) {
132 i4 = i0;
133 }
134 const uint16x8_t v3_4 = vld1q_u16(i4);
135 if XNN_UNPREDICTABLE(bh < 6) {
136 i5 = i0;
137 }
138 const uint16x8_t v3_5 = vld1q_u16(i5);
139 if XNN_UNPREDICTABLE(bh <= 6) {
140 i6 = i0;
141 }
142 const uint16x8_t v3_6 = vld1q_u16(i6);
143 const uint16x8_t v3_7 = vmovq_n_u16(0);
144
145 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
146 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
147 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
148 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
149
150 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
151 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
152 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
153 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
154 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
155 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
156 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
157 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
158
159 uint16x4_t v0_low = vget_low_u16(v0_0.val[0]);
160 uint16x4_t v1_low = vget_low_u16(v0_0.val[1]);
161 uint16x4_t v2_low = vget_low_u16(v0_1.val[0]);
162 uint16x4_t v3_low = vget_low_u16(v0_1.val[1]);
163 uint16x4_t v4_low = vget_low_u16(v0_2.val[0]);
164 uint16x4_t v5_low = vget_low_u16(v0_2.val[1]);
165 uint16x4_t v6_low = vget_low_u16(v0_3.val[0]);
166 uint16x4_t v7_low = vget_low_u16(v0_3.val[1]);
167
168 if (bh & 4) {
169 o = (uint16_t*) ((uintptr_t) o + oN_stride);
170 vst1_u16(o, v7_low);
171 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
172 if XNN_UNPREDICTABLE(block_width > 7) {
173 o = oN;
174 }
175 vst1_u16(o, v6_low);
176 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
177 if XNN_UNPREDICTABLE(block_width >= 7) {
178 o = oN;
179 }
180 vst1_u16(o, v5_low);
181 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
182 if XNN_UNPREDICTABLE(block_width > 5) {
183 o = oN;
184 }
185 vst1_u16(o, v4_low);
186 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
187 if XNN_UNPREDICTABLE(block_width >= 5) {
188 o = oN;
189 }
190 vst1_u16(o, v3_low);
191 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
192 if XNN_UNPREDICTABLE(block_width > 3) {
193 o = oN;
194 }
195 vst1_u16(o, v2_low);
196 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
197 if XNN_UNPREDICTABLE(block_width >= 3) {
198 o = oN;
199 }
200 vst1_u16(o, v1_low);
201 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
202 if XNN_UNPREDICTABLE(block_width > 1) {
203 o = oN;
204 }
205 vst1_u16(o, v0_low); o += 4;
206 v0_low = vget_high_u16(v0_0.val[0]);
207 v1_low = vget_high_u16(v0_0.val[1]);
208 v2_low = vget_high_u16(v0_1.val[0]);
209 v3_low = vget_high_u16(v0_1.val[1]);
210 v4_low = vget_high_u16(v0_2.val[0]);
211 v5_low = vget_high_u16(v0_2.val[1]);
212 v6_low = vget_high_u16(v0_3.val[0]);
213 v7_low = vget_high_u16(v0_3.val[1]);
214 }
215
216 if (bh & 2) {
217 o = (uint16_t*) ((uintptr_t) o + oN_stride);
218 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v7_low), 0);
219 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
220 if XNN_UNPREDICTABLE(block_width > 7) {
221 o = oN;
222 }
223 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v6_low), 0);
224 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
225 if XNN_UNPREDICTABLE(block_width >= 7) {
226 o = oN;
227 }
228 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v5_low), 0);
229 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
230 if XNN_UNPREDICTABLE(block_width > 5) {
231 o = oN;
232 }
233 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v4_low), 0);
234 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
235 if XNN_UNPREDICTABLE(block_width >= 5) {
236 o = oN;
237 }
238 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v3_low), 0);
239 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
240 if XNN_UNPREDICTABLE(block_width > 3) {
241 o = oN;
242 }
243 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v2_low), 0);
244 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
245 if XNN_UNPREDICTABLE(block_width >= 3) {
246 o = oN;
247 }
248 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v1_low), 0);
249 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
250 if XNN_UNPREDICTABLE(block_width > 1) {
251 o = oN;
252 }
253 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
254 v0_low = vext_u16(v0_low, v0_low, 2);
255 v1_low = vext_u16(v1_low, v1_low, 2);
256 v2_low = vext_u16(v2_low, v2_low, 2);
257 v3_low = vext_u16(v3_low, v3_low, 2);
258 v4_low = vext_u16(v4_low, v4_low, 2);
259 v5_low = vext_u16(v5_low, v5_low, 2);
260 v6_low = vext_u16(v6_low, v6_low, 2);
261 v7_low = vext_u16(v7_low, v7_low, 2);
262 }
263 if (bh & 1) {
264 o = (uint16_t*) ((uintptr_t) o + oN_stride);
265 vst1_lane_u16(o, v7_low, 0);
266 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
267 if XNN_UNPREDICTABLE(block_width > 7) {
268 o = oN;
269 }
270 vst1_lane_u16(o, v6_low, 0);
271 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
272 if XNN_UNPREDICTABLE(block_width >= 7) {
273 o = oN;
274 }
275 vst1_lane_u16(o, v5_low, 0);
276 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
277 if XNN_UNPREDICTABLE(block_width > 5) {
278 o = oN;
279 }
280 vst1_lane_u16(o, v4_low, 0);
281 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
282 if XNN_UNPREDICTABLE(block_width >= 5) {
283 o = oN;
284 }
285 vst1_lane_u16(o, v3_low, 0);
286 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
287 if XNN_UNPREDICTABLE(block_width > 3) {
288 o = oN;
289 }
290 vst1_lane_u16(o, v2_low, 0);
291 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
292 if XNN_UNPREDICTABLE(block_width >= 3) {
293 o = oN;
294 }
295 vst1_lane_u16(o, v1_low, 0);
296 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
297 if XNN_UNPREDICTABLE(block_width > 1) {
298 o = oN;
299 }
300 vst1_lane_u16(o, v0_low, 0);
301 }
302 }
303
304 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
305 i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
306 i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
307 i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
308 i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
309 i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
310 i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
311 i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
312 o = (uint16_t*) ((uintptr_t) o + output_reset);
313 block_width = doz(block_width, tile_width);
314 } while (block_width != 0);
315 }
316