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_dec_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_dec_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 if XNN_UNPREDICTABLE(block_width > 7) {
80 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
81 }
82 vst1_u8(o, v0_3.val[0]);
83 if XNN_UNPREDICTABLE(block_width >= 7) {
84 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
85 }
86 vst1_u8(o, v0_2.val[1]);
87 if XNN_UNPREDICTABLE(block_width > 5) {
88 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
89 }
90 vst1_u8(o, v0_2.val[0]);
91 if XNN_UNPREDICTABLE(block_width >= 5) {
92 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
93 }
94 vst1_u8(o, v0_1.val[1]);
95 if XNN_UNPREDICTABLE(block_width > 3) {
96 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
97 }
98 vst1_u8(o, v0_1.val[0]);
99 if XNN_UNPREDICTABLE(block_width >= 3) {
100 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
101 }
102 vst1_u8(o, v0_0.val[1]);
103 if XNN_UNPREDICTABLE(block_width > 1) {
104 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
105 }
106 vst1_u8(o, v0_0.val[0]);
107 }
108 o = (uint8_t*) ((uintptr_t) o + tile_hbytes);
109
110 if (bh != 0) {
111 const uint8x8_t v3_0 = vld1_u8(i0);
112 if XNN_UNPREDICTABLE(bh < 2) {
113 i1 = i0;
114 }
115 const uint8x8_t v3_1 = vld1_u8(i1);
116 if XNN_UNPREDICTABLE(bh <= 2) {
117 i2 = i0;
118 }
119 const uint8x8_t v3_2 = vld1_u8(i2);
120 if XNN_UNPREDICTABLE(bh < 4) {
121 i3 = i0;
122 }
123 const uint8x8_t v3_3 = vld1_u8(i3);
124 if XNN_UNPREDICTABLE(bh <= 4) {
125 i4 = i0;
126 }
127 const uint8x8_t v3_4 = vld1_u8(i4);
128 if XNN_UNPREDICTABLE(bh < 6) {
129 i5 = i0;
130 }
131 const uint8x8_t v3_5 = vld1_u8(i5);
132 if XNN_UNPREDICTABLE(bh <= 6) {
133 i6 = i0;
134 }
135 const uint8x8_t v3_6 = vld1_u8(i6);
136 const uint8x8_t v3_7 = vmov_n_u8(0);
137
138 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
139 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
140 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
141 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
142
143 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
144 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
145 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
146 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
147 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
148 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
149 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
150 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
151
152 uint8x8_t v0_low = v0_0.val[0];
153 uint8x8_t v1_low = v0_0.val[1];
154 uint8x8_t v2_low = v0_1.val[0];
155 uint8x8_t v3_low = v0_1.val[1];
156 uint8x8_t v4_low = v0_2.val[0];
157 uint8x8_t v5_low = v0_2.val[1];
158 uint8x8_t v6_low = v0_3.val[0];
159 uint8x8_t v7_low = v0_3.val[1];
160
161 if (bh & 4) {
162 o = (uint8_t*) ((uintptr_t) o + oN_stride);
163 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v7_low), 0);
164 if XNN_UNPREDICTABLE(block_width > 7) {
165 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
166 }
167 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v6_low), 0);
168 if XNN_UNPREDICTABLE(block_width >= 7) {
169 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
170 }
171 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v5_low), 0);
172 if XNN_UNPREDICTABLE(block_width > 5) {
173 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
174 }
175 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v4_low), 0);
176 if XNN_UNPREDICTABLE(block_width >= 5) {
177 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
178 }
179 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v3_low), 0);
180 if XNN_UNPREDICTABLE(block_width > 3) {
181 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
182 }
183 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v2_low), 0);
184 if XNN_UNPREDICTABLE(block_width >= 3) {
185 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
186 }
187 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v1_low), 0);
188 if XNN_UNPREDICTABLE(block_width > 1) {
189 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
190 }
191 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v0_low), 0); o += 4;
192 v0_low = vext_u8(v0_low, v0_low, 4);
193 v1_low = vext_u8(v1_low, v1_low, 4);
194 v2_low = vext_u8(v2_low, v2_low, 4);
195 v3_low = vext_u8(v3_low, v3_low, 4);
196 v4_low = vext_u8(v4_low, v4_low, 4);
197 v5_low = vext_u8(v5_low, v5_low, 4);
198 v6_low = vext_u8(v6_low, v6_low, 4);
199 v7_low = vext_u8(v7_low, v7_low, 4);
200 }
201 if (bh & 2) {
202 o = (uint8_t*) ((uintptr_t) o + oN_stride);
203 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v7_low), 0);
204 if XNN_UNPREDICTABLE(block_width > 7) {
205 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
206 }
207 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v6_low), 0);
208 if XNN_UNPREDICTABLE(block_width >= 7) {
209 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
210 }
211 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v5_low), 0);
212 if XNN_UNPREDICTABLE(block_width > 5) {
213 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
214 }
215 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v4_low), 0);
216 if XNN_UNPREDICTABLE(block_width >= 5) {
217 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
218 }
219 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v3_low), 0);
220 if XNN_UNPREDICTABLE(block_width > 3) {
221 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
222 }
223 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v2_low), 0);
224 if XNN_UNPREDICTABLE(block_width >= 3) {
225 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
226 }
227 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v1_low), 0);
228 if XNN_UNPREDICTABLE(block_width > 1) {
229 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
230 }
231 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v0_low), 0); o += 2;
232 v0_low = vext_u8(v0_low, v0_low, 2);
233 v1_low = vext_u8(v1_low, v1_low, 2);
234 v2_low = vext_u8(v2_low, v2_low, 2);
235 v3_low = vext_u8(v3_low, v3_low, 2);
236 v4_low = vext_u8(v4_low, v4_low, 2);
237 v5_low = vext_u8(v5_low, v5_low, 2);
238 v6_low = vext_u8(v6_low, v6_low, 2);
239 v7_low = vext_u8(v7_low, v7_low, 2);
240 }
241 if (bh & 1) {
242 o = (uint8_t*) ((uintptr_t) o + oN_stride);
243 vst1_lane_u8(o, v7_low, 0);
244 if XNN_UNPREDICTABLE(block_width > 7) {
245 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
246 }
247 vst1_lane_u8(o, v6_low, 0);
248 if XNN_UNPREDICTABLE(block_width >= 7) {
249 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
250 }
251 vst1_lane_u8(o, v5_low, 0);
252 if XNN_UNPREDICTABLE(block_width > 5) {
253 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
254 }
255 vst1_lane_u8(o, v4_low, 0);
256 if XNN_UNPREDICTABLE(block_width >= 5) {
257 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
258 }
259 vst1_lane_u8(o, v3_low, 0);
260 if XNN_UNPREDICTABLE(block_width > 3) {
261 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
262 }
263 vst1_lane_u8(o, v2_low, 0);
264 if XNN_UNPREDICTABLE(block_width >= 3) {
265 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
266 }
267 vst1_lane_u8(o, v1_low, 0);
268 if XNN_UNPREDICTABLE(block_width > 1) {
269 o = (uint8_t*) ((uintptr_t) o + minus_output_stride);
270 }
271 vst1_lane_u8(o, v0_low, 0);
272 }
273 }
274
275 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
276 i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
277 i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
278 i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
279 i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
280 i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
281 i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
282 i7 = (const uint8_t*) ((uintptr_t) i6 + input_stride);
283 o = (uint8_t*) ((uintptr_t) o + output_reset);
284 block_width = doz(block_width, tile_width);
285 } while (block_width != 0);
286 }
287