xref: /aosp_15_r20/external/XNNPACK/src/x16-transposec/gen/8x8-reuse-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_reuse_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_reuse_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 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       if XNN_UNPREDICTABLE(block_width > 7) {
72         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
73       }
74       vst1q_u16(o, v0_3.val[0]);
75       if XNN_UNPREDICTABLE(block_width >= 7) {
76         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
77       }
78       vst1q_u16(o, v0_2.val[1]);
79       if XNN_UNPREDICTABLE(block_width > 5) {
80         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
81       }
82       vst1q_u16(o, v0_2.val[0]);
83       if XNN_UNPREDICTABLE(block_width >= 5) {
84         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
85       }
86       vst1q_u16(o, v0_1.val[1]);
87       if XNN_UNPREDICTABLE(block_width > 3) {
88         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
89       }
90       vst1q_u16(o, v0_1.val[0]);
91       if XNN_UNPREDICTABLE(block_width >= 3) {
92         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
93       }
94       vst1q_u16(o, v0_0.val[1]);
95       if XNN_UNPREDICTABLE(block_width > 1) {
96         o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
97       }
98       vst1q_u16(o, v0_0.val[0]);
99     }
100     o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
101 
102     if (bh != 0) {
103       const uint16x8_t v3_0 = vld1q_u16(i0);
104       const uint16_t *i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
105       if XNN_UNPREDICTABLE(bh < 2) {
106         i1 = i0;
107       }
108       const uint16x8_t v3_1 = vld1q_u16(i1);
109       const uint16_t *i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
110       if XNN_UNPREDICTABLE(bh <= 2) {
111         i2 = i1;
112       }
113       const uint16x8_t v3_2 = vld1q_u16(i2);
114       const uint16_t *i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
115       if XNN_UNPREDICTABLE(bh < 4) {
116         i3 = i2;
117       }
118       const uint16x8_t v3_3 = vld1q_u16(i3);
119       const uint16_t *i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
120       if XNN_UNPREDICTABLE(bh <= 4) {
121         i4 = i3;
122       }
123       const uint16x8_t v3_4 = vld1q_u16(i4);
124       const uint16_t *i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
125       if XNN_UNPREDICTABLE(bh < 6) {
126         i5 = i4;
127       }
128       const uint16x8_t v3_5 = vld1q_u16(i5);
129       const uint16_t *i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
130       if XNN_UNPREDICTABLE(bh <= 6) {
131         i6 = i5;
132       }
133       const uint16x8_t v3_6 = vld1q_u16(i6);
134       const uint16x8_t v3_7 = vmovq_n_u16(0);
135 
136       const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
137       const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
138       const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
139       const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
140 
141       const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
142       const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
143       const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
144       const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
145       const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
146       const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
147       const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
148       const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
149 
150       uint16x4_t v0_low = vget_low_u16(v0_0.val[0]);
151       uint16x4_t v1_low = vget_low_u16(v0_0.val[1]);
152       uint16x4_t v2_low = vget_low_u16(v0_1.val[0]);
153       uint16x4_t v3_low = vget_low_u16(v0_1.val[1]);
154       uint16x4_t v4_low = vget_low_u16(v0_2.val[0]);
155       uint16x4_t v5_low = vget_low_u16(v0_2.val[1]);
156       uint16x4_t v6_low = vget_low_u16(v0_3.val[0]);
157       uint16x4_t v7_low = vget_low_u16(v0_3.val[1]);
158 
159       if (bh & 4) {
160         o = (uint16_t*) ((uintptr_t) o + oN_stride);
161         vst1_u16(o, v7_low);
162         if XNN_UNPREDICTABLE(block_width > 7) {
163           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
164         }
165         vst1_u16(o, v6_low);
166         if XNN_UNPREDICTABLE(block_width >= 7) {
167           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
168         }
169         vst1_u16(o, v5_low);
170         if XNN_UNPREDICTABLE(block_width > 5) {
171           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
172         }
173         vst1_u16(o, v4_low);
174         if XNN_UNPREDICTABLE(block_width >= 5) {
175           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
176         }
177         vst1_u16(o, v3_low);
178         if XNN_UNPREDICTABLE(block_width > 3) {
179           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
180         }
181         vst1_u16(o, v2_low);
182         if XNN_UNPREDICTABLE(block_width >= 3) {
183           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
184         }
185         vst1_u16(o, v1_low);
186         if XNN_UNPREDICTABLE(block_width > 1) {
187           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
188         }
189         vst1_u16(o, v0_low); o += 4;
190         v0_low = vget_high_u16(v0_0.val[0]);
191         v1_low = vget_high_u16(v0_0.val[1]);
192         v2_low = vget_high_u16(v0_1.val[0]);
193         v3_low = vget_high_u16(v0_1.val[1]);
194         v4_low = vget_high_u16(v0_2.val[0]);
195         v5_low = vget_high_u16(v0_2.val[1]);
196         v6_low = vget_high_u16(v0_3.val[0]);
197         v7_low = vget_high_u16(v0_3.val[1]);
198       }
199 
200       if (bh & 2) {
201         o = (uint16_t*) ((uintptr_t) o + oN_stride);
202         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v7_low), 0);
203         if XNN_UNPREDICTABLE(block_width > 7) {
204           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
205         }
206         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v6_low), 0);
207         if XNN_UNPREDICTABLE(block_width >= 7) {
208           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
209         }
210         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v5_low), 0);
211         if XNN_UNPREDICTABLE(block_width > 5) {
212           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
213         }
214         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v4_low), 0);
215         if XNN_UNPREDICTABLE(block_width >= 5) {
216           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
217         }
218         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v3_low), 0);
219         if XNN_UNPREDICTABLE(block_width > 3) {
220           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
221         }
222         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v2_low), 0);
223         if XNN_UNPREDICTABLE(block_width >= 3) {
224           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
225         }
226         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v1_low), 0);
227         if XNN_UNPREDICTABLE(block_width > 1) {
228           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
229         }
230         vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
231         v0_low = vext_u16(v0_low, v0_low, 2);
232         v1_low = vext_u16(v1_low, v1_low, 2);
233         v2_low = vext_u16(v2_low, v2_low, 2);
234         v3_low = vext_u16(v3_low, v3_low, 2);
235         v4_low = vext_u16(v4_low, v4_low, 2);
236         v5_low = vext_u16(v5_low, v5_low, 2);
237         v6_low = vext_u16(v6_low, v6_low, 2);
238         v7_low = vext_u16(v7_low, v7_low, 2);
239       }
240       if (bh & 1) {
241         o = (uint16_t*) ((uintptr_t) o + oN_stride);
242         vst1_lane_u16(o, v7_low, 0);
243         if XNN_UNPREDICTABLE(block_width > 7) {
244           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
245         }
246         vst1_lane_u16(o, v6_low, 0);
247         if XNN_UNPREDICTABLE(block_width >= 7) {
248           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
249         }
250         vst1_lane_u16(o, v5_low, 0);
251         if XNN_UNPREDICTABLE(block_width > 5) {
252           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
253         }
254         vst1_lane_u16(o, v4_low, 0);
255         if XNN_UNPREDICTABLE(block_width >= 5) {
256           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
257         }
258         vst1_lane_u16(o, v3_low, 0);
259         if XNN_UNPREDICTABLE(block_width > 3) {
260           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
261         }
262         vst1_lane_u16(o, v2_low, 0);
263         if XNN_UNPREDICTABLE(block_width >= 3) {
264           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
265         }
266         vst1_lane_u16(o, v1_low, 0);
267         if XNN_UNPREDICTABLE(block_width > 1) {
268           o = (uint16_t*) ((uintptr_t) o + minus_output_stride);
269         }
270         vst1_lane_u16(o, v0_low, 0);
271       }
272     }
273 
274     i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
275     o = (uint16_t*) ((uintptr_t) o + output_reset);
276     block_width = doz(block_width, tile_width);
277   } while (block_width != 0);
278 }
279