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_reuse_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_reuse_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint16_t) - tile_hbytes;
35
36 const uint16_t* i0 = input;
37 uint16_t* o = (uint16_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 uint16x8_t v3_0 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
47 const uint16x8_t v3_1 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
48 const uint16x8_t v3_2 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
49 const uint16x8_t v3_3 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
50 const uint16x8_t v3_4 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
51 const uint16x8_t v3_5 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
52 const uint16x8_t v3_6 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
53 const uint16x8_t v3_7 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
54
55 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
56 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
57 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
58 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
59
60 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
61 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
62 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
63 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
64 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
65 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
66 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
67 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
68
69 o = (uint16_t*) ((uintptr_t) o + oN_offset);
70 vst1q_u16(o, v0_3.val[1]);
71 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
72 if XNN_UNPREDICTABLE(block_width > 7) {
73 o = oN;
74 }
75 vst1q_u16(o, v0_3.val[0]);
76 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
77 if XNN_UNPREDICTABLE(block_width >= 7) {
78 o = oN;
79 }
80 vst1q_u16(o, v0_2.val[1]);
81 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
82 if XNN_UNPREDICTABLE(block_width > 5) {
83 o = oN;
84 }
85 vst1q_u16(o, v0_2.val[0]);
86 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
87 if XNN_UNPREDICTABLE(block_width >= 5) {
88 o = oN;
89 }
90 vst1q_u16(o, v0_1.val[1]);
91 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
92 if XNN_UNPREDICTABLE(block_width > 3) {
93 o = oN;
94 }
95 vst1q_u16(o, v0_1.val[0]);
96 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
97 if XNN_UNPREDICTABLE(block_width >= 3) {
98 o = oN;
99 }
100 vst1q_u16(o, v0_0.val[1]);
101 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
102 if XNN_UNPREDICTABLE(block_width > 1) {
103 o = oN;
104 }
105 vst1q_u16(o, v0_0.val[0]);
106 }
107 o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
108
109 if (bh != 0) {
110 const uint16x8_t v3_0 = vld1q_u16(i0);
111 const uint16_t *i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
112 if XNN_UNPREDICTABLE(bh < 2) {
113 i1 = i0;
114 }
115 const uint16x8_t v3_1 = vld1q_u16(i1);
116 const uint16_t *i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
117 if XNN_UNPREDICTABLE(bh <= 2) {
118 i2 = i1;
119 }
120 const uint16x8_t v3_2 = vld1q_u16(i2);
121 const uint16_t *i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
122 if XNN_UNPREDICTABLE(bh < 4) {
123 i3 = i2;
124 }
125 const uint16x8_t v3_3 = vld1q_u16(i3);
126 const uint16_t *i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
127 if XNN_UNPREDICTABLE(bh <= 4) {
128 i4 = i3;
129 }
130 const uint16x8_t v3_4 = vld1q_u16(i4);
131 const uint16_t *i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
132 if XNN_UNPREDICTABLE(bh < 6) {
133 i5 = i4;
134 }
135 const uint16x8_t v3_5 = vld1q_u16(i5);
136 const uint16_t *i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
137 if XNN_UNPREDICTABLE(bh <= 6) {
138 i6 = i5;
139 }
140 const uint16x8_t v3_6 = vld1q_u16(i6);
141 const uint16x8_t v3_7 = vmovq_n_u16(0);
142
143 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
144 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
145 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
146 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
147
148 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
149 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
150 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
151 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
152 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
153 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
154 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
155 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
156
157 uint16x4_t v0_low = vget_low_u16(v0_0.val[0]);
158 uint16x4_t v1_low = vget_low_u16(v0_0.val[1]);
159 uint16x4_t v2_low = vget_low_u16(v0_1.val[0]);
160 uint16x4_t v3_low = vget_low_u16(v0_1.val[1]);
161 uint16x4_t v4_low = vget_low_u16(v0_2.val[0]);
162 uint16x4_t v5_low = vget_low_u16(v0_2.val[1]);
163 uint16x4_t v6_low = vget_low_u16(v0_3.val[0]);
164 uint16x4_t v7_low = vget_low_u16(v0_3.val[1]);
165
166 if (bh & 4) {
167 o = (uint16_t*) ((uintptr_t) o + oN_stride);
168 vst1_u16(o, v7_low);
169 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
170 if XNN_UNPREDICTABLE(block_width > 7) {
171 o = oN;
172 }
173 vst1_u16(o, v6_low);
174 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
175 if XNN_UNPREDICTABLE(block_width >= 7) {
176 o = oN;
177 }
178 vst1_u16(o, v5_low);
179 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
180 if XNN_UNPREDICTABLE(block_width > 5) {
181 o = oN;
182 }
183 vst1_u16(o, v4_low);
184 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
185 if XNN_UNPREDICTABLE(block_width >= 5) {
186 o = oN;
187 }
188 vst1_u16(o, v3_low);
189 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
190 if XNN_UNPREDICTABLE(block_width > 3) {
191 o = oN;
192 }
193 vst1_u16(o, v2_low);
194 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
195 if XNN_UNPREDICTABLE(block_width >= 3) {
196 o = oN;
197 }
198 vst1_u16(o, v1_low);
199 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
200 if XNN_UNPREDICTABLE(block_width > 1) {
201 o = oN;
202 }
203 vst1_u16(o, v0_low); o += 4;
204 v0_low = vget_high_u16(v0_0.val[0]);
205 v1_low = vget_high_u16(v0_0.val[1]);
206 v2_low = vget_high_u16(v0_1.val[0]);
207 v3_low = vget_high_u16(v0_1.val[1]);
208 v4_low = vget_high_u16(v0_2.val[0]);
209 v5_low = vget_high_u16(v0_2.val[1]);
210 v6_low = vget_high_u16(v0_3.val[0]);
211 v7_low = vget_high_u16(v0_3.val[1]);
212 }
213
214 if (bh & 2) {
215 o = (uint16_t*) ((uintptr_t) o + oN_stride);
216 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v7_low), 0);
217 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
218 if XNN_UNPREDICTABLE(block_width > 7) {
219 o = oN;
220 }
221 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v6_low), 0);
222 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
223 if XNN_UNPREDICTABLE(block_width >= 7) {
224 o = oN;
225 }
226 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v5_low), 0);
227 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
228 if XNN_UNPREDICTABLE(block_width > 5) {
229 o = oN;
230 }
231 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v4_low), 0);
232 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
233 if XNN_UNPREDICTABLE(block_width >= 5) {
234 o = oN;
235 }
236 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v3_low), 0);
237 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
238 if XNN_UNPREDICTABLE(block_width > 3) {
239 o = oN;
240 }
241 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v2_low), 0);
242 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
243 if XNN_UNPREDICTABLE(block_width >= 3) {
244 o = oN;
245 }
246 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v1_low), 0);
247 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
248 if XNN_UNPREDICTABLE(block_width > 1) {
249 o = oN;
250 }
251 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
252 v0_low = vext_u16(v0_low, v0_low, 2);
253 v1_low = vext_u16(v1_low, v1_low, 2);
254 v2_low = vext_u16(v2_low, v2_low, 2);
255 v3_low = vext_u16(v3_low, v3_low, 2);
256 v4_low = vext_u16(v4_low, v4_low, 2);
257 v5_low = vext_u16(v5_low, v5_low, 2);
258 v6_low = vext_u16(v6_low, v6_low, 2);
259 v7_low = vext_u16(v7_low, v7_low, 2);
260 }
261 if (bh & 1) {
262 o = (uint16_t*) ((uintptr_t) o + oN_stride);
263 vst1_lane_u16(o, v7_low, 0);
264 uint16_t *oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
265 if XNN_UNPREDICTABLE(block_width > 7) {
266 o = oN;
267 }
268 vst1_lane_u16(o, v6_low, 0);
269 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
270 if XNN_UNPREDICTABLE(block_width >= 7) {
271 o = oN;
272 }
273 vst1_lane_u16(o, v5_low, 0);
274 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
275 if XNN_UNPREDICTABLE(block_width > 5) {
276 o = oN;
277 }
278 vst1_lane_u16(o, v4_low, 0);
279 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
280 if XNN_UNPREDICTABLE(block_width >= 5) {
281 o = oN;
282 }
283 vst1_lane_u16(o, v3_low, 0);
284 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
285 if XNN_UNPREDICTABLE(block_width > 3) {
286 o = oN;
287 }
288 vst1_lane_u16(o, v2_low, 0);
289 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
290 if XNN_UNPREDICTABLE(block_width >= 3) {
291 o = oN;
292 }
293 vst1_lane_u16(o, v1_low, 0);
294 oN = (uint16_t*) ((uintptr_t) o + minus_output_stride);
295 if XNN_UNPREDICTABLE(block_width > 1) {
296 o = oN;
297 }
298 vst1_lane_u16(o, v0_low, 0);
299 }
300 }
301
302 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
303 o = (uint16_t*) ((uintptr_t) o + output_reset);
304 block_width = doz(block_width, tile_width);
305 } while (block_width != 0);
306 }
307