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_reuse_switch_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_reuse_switch_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint8_t);
35
36 const uint8_t* i0 = input;
37 uint8_t* o = (uint8_t*) output;
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 size_t bh = block_height;
44 for (; bh >= 8; bh -= 8) {
45 const uint8x8_t v3_0 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
46 const uint8x8_t v3_1 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
47 const uint8x8_t v3_2 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
48 const uint8x8_t v3_3 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
49 const uint8x8_t v3_4 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
50 const uint8x8_t v3_5 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
51 const uint8x8_t v3_6 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
52 const uint8x8_t v3_7 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride);
53
54 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
55 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
56 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
57 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
58
59 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
60 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
61 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
62 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
63 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
64 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
65 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
66 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
67
68 uint8_t *oN = (uint8_t*) ((uintptr_t) o + oN_stride);
69 switch (rem) {
70 case 7:
71 vst1_u8(oN, v0_3.val[1]); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
72 case 6:
73 vst1_u8(oN, v0_3.val[0]); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
74 case 5:
75 vst1_u8(oN, v0_2.val[1]); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
76 case 4:
77 vst1_u8(oN, v0_2.val[0]); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
78 case 3:
79 vst1_u8(oN, v0_1.val[1]); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
80 case 2:
81 vst1_u8(oN, v0_1.val[0]); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
82 case 1:
83 vst1_u8(oN, v0_0.val[1]);
84 case 0:
85 vst1_u8(o, v0_0.val[0]); o = (uint8_t*) ((uintptr_t) o + tile_hbytes);
86 break;
87 default:
88 XNN_UNREACHABLE;
89 }
90 }
91
92 if (bh != 0) {
93 const uint8x8_t v3_0 = vld1_u8(i0);
94 const uint8_t *i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
95 if XNN_UNPREDICTABLE(bh < 2) {
96 i1 = i0;
97 }
98 const uint8x8_t v3_1 = vld1_u8(i1);
99 const uint8_t *i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
100 if XNN_UNPREDICTABLE(bh <= 2) {
101 i2 = i1;
102 }
103 const uint8x8_t v3_2 = vld1_u8(i2);
104 const uint8_t *i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
105 if XNN_UNPREDICTABLE(bh < 4) {
106 i3 = i2;
107 }
108 const uint8x8_t v3_3 = vld1_u8(i3);
109 const uint8_t *i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
110 if XNN_UNPREDICTABLE(bh <= 4) {
111 i4 = i3;
112 }
113 const uint8x8_t v3_4 = vld1_u8(i4);
114 const uint8_t *i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
115 if XNN_UNPREDICTABLE(bh < 6) {
116 i5 = i4;
117 }
118 const uint8x8_t v3_5 = vld1_u8(i5);
119 const uint8_t *i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
120 if XNN_UNPREDICTABLE(bh <= 6) {
121 i6 = i5;
122 }
123 const uint8x8_t v3_6 = vld1_u8(i6);
124 const uint8x8_t v3_7 = vmov_n_u8(0);
125
126 const uint8x8x2_t v2_0 = vzip_u8(v3_0, v3_4);
127 const uint8x8x2_t v2_1 = vzip_u8(v3_1, v3_5);
128 const uint8x8x2_t v2_2 = vzip_u8(v3_2, v3_6);
129 const uint8x8x2_t v2_3 = vzip_u8(v3_3, v3_7);
130
131 const uint8x8x2_t v1_0 = vzip_u8(v2_0.val[0], v2_2.val[0]);
132 const uint8x8x2_t v1_1 = vzip_u8(v2_0.val[1], v2_2.val[1]);
133 const uint8x8x2_t v1_2 = vzip_u8(v2_1.val[0], v2_3.val[0]);
134 const uint8x8x2_t v1_3 = vzip_u8(v2_1.val[1], v2_3.val[1]);
135 const uint8x8x2_t v0_0 = vzip_u8(v1_0.val[0], v1_2.val[0]);
136 const uint8x8x2_t v0_1 = vzip_u8(v1_0.val[1], v1_2.val[1]);
137 const uint8x8x2_t v0_2 = vzip_u8(v1_1.val[0], v1_3.val[0]);
138 const uint8x8x2_t v0_3 = vzip_u8(v1_1.val[1], v1_3.val[1]);
139
140 uint8x8_t v0_low = v0_0.val[0];
141 uint8x8_t v1_low = v0_0.val[1];
142 uint8x8_t v2_low = v0_1.val[0];
143 uint8x8_t v3_low = v0_1.val[1];
144 uint8x8_t v4_low = v0_2.val[0];
145 uint8x8_t v5_low = v0_2.val[1];
146 uint8x8_t v6_low = v0_3.val[0];
147 uint8x8_t v7_low = v0_3.val[1];
148
149 if (bh & 4) {
150 uint8_t* oN = (uint8_t*) ((uintptr_t) o + oN_stride);
151 switch (rem) {
152 case 7:
153 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v7_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
154 case 6:
155 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v6_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
156 case 5:
157 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v5_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
158 case 4:
159 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v4_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
160 case 3:
161 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v3_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
162 case 2:
163 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v2_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
164 case 1:
165 vst1_lane_u32((void*) oN, vreinterpret_u32_u8(v1_low), 0);
166 case 0:
167 vst1_lane_u32((void*) o, vreinterpret_u32_u8(v0_low), 0); o += 4;
168 break;
169 default:
170 XNN_UNREACHABLE;
171 }
172 v0_low = vext_u8(v0_low, v0_low, 4);
173 v1_low = vext_u8(v1_low, v1_low, 4);
174 v2_low = vext_u8(v2_low, v2_low, 4);
175 v3_low = vext_u8(v3_low, v3_low, 4);
176 v4_low = vext_u8(v4_low, v4_low, 4);
177 v5_low = vext_u8(v5_low, v5_low, 4);
178 v6_low = vext_u8(v6_low, v6_low, 4);
179 v7_low = vext_u8(v7_low, v7_low, 4);
180 }
181 if (bh & 2) {
182 uint8_t* oN = (uint8_t*) ((uintptr_t) o + oN_stride);
183 switch (rem) {
184 case 7:
185 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v7_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
186 case 6:
187 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v6_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
188 case 5:
189 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v5_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
190 case 4:
191 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v4_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
192 case 3:
193 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v3_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
194 case 2:
195 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v2_low), 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
196 case 1:
197 vst1_lane_u16((void*) oN, vreinterpret_u16_u8(v1_low), 0);
198 case 0:
199 vst1_lane_u16((void*) o, vreinterpret_u16_u8(v0_low), 0); o += 2;
200 break;
201 default:
202 XNN_UNREACHABLE;
203 }
204 v0_low = vext_u8(v0_low, v0_low, 2);
205 v1_low = vext_u8(v1_low, v1_low, 2);
206 v2_low = vext_u8(v2_low, v2_low, 2);
207 v3_low = vext_u8(v3_low, v3_low, 2);
208 v4_low = vext_u8(v4_low, v4_low, 2);
209 v5_low = vext_u8(v5_low, v5_low, 2);
210 v6_low = vext_u8(v6_low, v6_low, 2);
211 v7_low = vext_u8(v7_low, v7_low, 2);
212 }
213 if (bh & 1) {
214 uint8_t* oN = (uint8_t*) ((uintptr_t) o + oN_stride);
215 switch (rem) {
216 case 7:
217 vst1_lane_u8(oN, v7_low, 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
218 case 6:
219 vst1_lane_u8(oN, v6_low, 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
220 case 5:
221 vst1_lane_u8(oN, v5_low, 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
222 case 4:
223 vst1_lane_u8(oN, v4_low, 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
224 case 3:
225 vst1_lane_u8(oN, v3_low, 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
226 case 2:
227 vst1_lane_u8(oN, v2_low, 0); oN = (uint8_t*) ((uintptr_t) oN + minus_output_stride);
228 case 1:
229 vst1_lane_u8(oN, v1_low, 0);
230 case 0:
231 vst1_lane_u8(o, v0_low, 0);
232 break;
233 default:
234 XNN_UNREACHABLE;
235 }
236 }
237 }
238
239 i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
240 o = (uint8_t*) ((uintptr_t) o + output_reset);
241 block_width = doz(block_width, tile_width);
242 } while (block_width != 0);
243 }
244