xref: /aosp_15_r20/external/XNNPACK/src/x16-transposec/gen/8x8-multi-dec-zip-neon.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_dec_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_dec_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       if XNN_UNPREDICTABLE(block_width > 7) {
80         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
81       }
82       vst1q_u16(o, v0_3.val[0]);
83       if XNN_UNPREDICTABLE(block_width >= 7) {
84         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
85       }
86       vst1q_u16(o, v0_2.val[1]);
87       if XNN_UNPREDICTABLE(block_width > 5) {
88         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
89       }
90       vst1q_u16(o, v0_2.val[0]);
91       if XNN_UNPREDICTABLE(block_width >= 5) {
92         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
93       }
94       vst1q_u16(o, v0_1.val[1]);
95       if XNN_UNPREDICTABLE(block_width > 3) {
96         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
97       }
98       vst1q_u16(o, v0_1.val[0]);
99       if XNN_UNPREDICTABLE(block_width >= 3) {
100         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
101       }
102       vst1q_u16(o, v0_0.val[1]);
103       if XNN_UNPREDICTABLE(block_width > 1) {
104         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
105       }
106       vst1q_u16(o, v0_0.val[0]);
107     }
108     o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
109 
110     if (bh != 0) {
111       const uint16x8_t v3_0 = vld1q_u16(i0);
112       if XNN_UNPREDICTABLE(bh < 2) {
113         i1 = i0;
114       }
115       const uint16x8_t v3_1 = vld1q_u16(i1);
116       if XNN_UNPREDICTABLE(bh <= 2) {
117         i2 = i0;
118       }
119       const uint16x8_t v3_2 = vld1q_u16(i2);
120       if XNN_UNPREDICTABLE(bh < 4) {
121         i3 = i0;
122       }
123       const uint16x8_t v3_3 = vld1q_u16(i3);
124       if XNN_UNPREDICTABLE(bh <= 4) {
125         i4 = i0;
126       }
127       const uint16x8_t v3_4 = vld1q_u16(i4);
128       if XNN_UNPREDICTABLE(bh < 6) {
129         i5 = i0;
130       }
131       const uint16x8_t v3_5 = vld1q_u16(i5);
132       if XNN_UNPREDICTABLE(bh <= 6) {
133         i6 = i0;
134       }
135       const uint16x8_t v3_6 = vld1q_u16(i6);
136       const uint16x8_t v3_7 = vmovq_n_u16(0);
137 
138       const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
139       const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
140       const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
141       const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
142 
143       const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
144       const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
145       const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
146       const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
147       const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
148       const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
149       const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
150       const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
151 
152       uint16x4_t v0_low = vget_low_u16(v0_0.val[0]);
153       uint16x4_t v1_low = vget_low_u16(v0_0.val[1]);
154       uint16x4_t v2_low = vget_low_u16(v0_1.val[0]);
155       uint16x4_t v3_low = vget_low_u16(v0_1.val[1]);
156       uint16x4_t v4_low = vget_low_u16(v0_2.val[0]);
157       uint16x4_t v5_low = vget_low_u16(v0_2.val[1]);
158       uint16x4_t v6_low = vget_low_u16(v0_3.val[0]);
159       uint16x4_t v7_low = vget_low_u16(v0_3.val[1]);
160 
161       if (bh & 4) {
162         o = (uint16_t*) ((uintptr_t) o + oN_stride);
163         vst1_u16(o, v7_low);
164         if XNN_UNPREDICTABLE(block_width > 7) {
165           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
166         }
167         vst1_u16(o, v6_low);
168         if XNN_UNPREDICTABLE(block_width >= 7) {
169           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
170         }
171         vst1_u16(o, v5_low);
172         if XNN_UNPREDICTABLE(block_width > 5) {
173           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
174         }
175         vst1_u16(o, v4_low);
176         if XNN_UNPREDICTABLE(block_width >= 5) {
177           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
178         }
179         vst1_u16(o, v3_low);
180         if XNN_UNPREDICTABLE(block_width > 3) {
181           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
182         }
183         vst1_u16(o, v2_low);
184         if XNN_UNPREDICTABLE(block_width >= 3) {
185           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
186         }
187         vst1_u16(o, v1_low);
188         if XNN_UNPREDICTABLE(block_width > 1) {
189           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
190         }
191         vst1_u16(o, v0_low); o += 4;
192         v0_low = vget_high_u16(v0_0.val[0]);
193         v1_low = vget_high_u16(v0_0.val[1]);
194         v2_low = vget_high_u16(v0_1.val[0]);
195         v3_low = vget_high_u16(v0_1.val[1]);
196         v4_low = vget_high_u16(v0_2.val[0]);
197         v5_low = vget_high_u16(v0_2.val[1]);
198         v6_low = vget_high_u16(v0_3.val[0]);
199         v7_low = vget_high_u16(v0_3.val[1]);
200       }
201 
202       if (bh & 2) {
203         o = (uint16_t*) ((uintptr_t) o + oN_stride);
204         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v7_low), 0);
205         if XNN_UNPREDICTABLE(block_width > 7) {
206           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
207         }
208         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v6_low), 0);
209         if XNN_UNPREDICTABLE(block_width >= 7) {
210           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
211         }
212         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v5_low), 0);
213         if XNN_UNPREDICTABLE(block_width > 5) {
214           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
215         }
216         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v4_low), 0);
217         if XNN_UNPREDICTABLE(block_width >= 5) {
218           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
219         }
220         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v3_low), 0);
221         if XNN_UNPREDICTABLE(block_width > 3) {
222           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
223         }
224         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v2_low), 0);
225         if XNN_UNPREDICTABLE(block_width >= 3) {
226           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
227         }
228         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v1_low), 0);
229         if XNN_UNPREDICTABLE(block_width > 1) {
230           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
231         }
232         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
233         v0_low = vext_u16(v0_low, v0_low, 2);
234         v1_low = vext_u16(v1_low, v1_low, 2);
235         v2_low = vext_u16(v2_low, v2_low, 2);
236         v3_low = vext_u16(v3_low, v3_low, 2);
237         v4_low = vext_u16(v4_low, v4_low, 2);
238         v5_low = vext_u16(v5_low, v5_low, 2);
239         v6_low = vext_u16(v6_low, v6_low, 2);
240         v7_low = vext_u16(v7_low, v7_low, 2);
241       }
242       if (bh & 1) {
243         o = (uint16_t*) ((uintptr_t) o + oN_stride);
244         vst1_lane_u16(o, v7_low, 0);
245         if XNN_UNPREDICTABLE(block_width > 7) {
246           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
247         }
248         vst1_lane_u16(o, v6_low, 0);
249         if XNN_UNPREDICTABLE(block_width >= 7) {
250           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
251         }
252         vst1_lane_u16(o, v5_low, 0);
253         if XNN_UNPREDICTABLE(block_width > 5) {
254           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
255         }
256         vst1_lane_u16(o, v4_low, 0);
257         if XNN_UNPREDICTABLE(block_width >= 5) {
258           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
259         }
260         vst1_lane_u16(o, v3_low, 0);
261         if XNN_UNPREDICTABLE(block_width > 3) {
262           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
263         }
264         vst1_lane_u16(o, v2_low, 0);
265         if XNN_UNPREDICTABLE(block_width >= 3) {
266           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
267         }
268         vst1_lane_u16(o, v1_low, 0);
269         if XNN_UNPREDICTABLE(block_width > 1) {
270           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
271         }
272         vst1_lane_u16(o, v0_low, 0);
273       }
274     }
275 
276     i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
277     i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
278     i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
279     i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
280     i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
281     i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
282     i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
283     i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
284     o = (uint16_t*) ((uintptr_t) o + output_reset);
285     block_width = doz(block_width, tile_width);
286   } while (block_width != 0);
287 }
288