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_switch_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_switch_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);
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*) output;
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 size_t bh = block_height;
52 for (; bh >= 8; bh -= 8) {
53 const uint16x8_t v3_0 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_offset);
54 const uint16x8_t v3_1 = vld1q_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset);
55 const uint16x8_t v3_2 = vld1q_u16(i2); i2 = (uint16_t*) ((uintptr_t) i2 + input_offset);
56 const uint16x8_t v3_3 = vld1q_u16(i3); i3 = (uint16_t*) ((uintptr_t) i3 + input_offset);
57 const uint16x8_t v3_4 = vld1q_u16(i4); i4 = (uint16_t*) ((uintptr_t) i4 + input_offset);
58 const uint16x8_t v3_5 = vld1q_u16(i5); i5 = (uint16_t*) ((uintptr_t) i5 + input_offset);
59 const uint16x8_t v3_6 = vld1q_u16(i6); i6 = (uint16_t*) ((uintptr_t) i6 + input_offset);
60 const uint16x8_t v3_7 = vld1q_u16(i7); i7 = (uint16_t*) ((uintptr_t) i7 + input_offset);
61
62 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
63 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
64 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
65 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
66
67 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
68 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
69 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
70 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
71 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
72 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
73 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
74 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
75
76 uint16_t *oN = (uint16_t*) ((uintptr_t) o + oN_stride);
77 switch (rem) {
78 case 7:
79 vst1q_u16(oN, v0_3.val[1]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
80 case 6:
81 vst1q_u16(oN, v0_3.val[0]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
82 case 5:
83 vst1q_u16(oN, v0_2.val[1]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
84 case 4:
85 vst1q_u16(oN, v0_2.val[0]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
86 case 3:
87 vst1q_u16(oN, v0_1.val[1]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
88 case 2:
89 vst1q_u16(oN, v0_1.val[0]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
90 case 1:
91 vst1q_u16(oN, v0_0.val[1]);
92 case 0:
93 vst1q_u16(o, v0_0.val[0]); o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
94 break;
95 default:
96 XNN_UNREACHABLE;
97 }
98 }
99
100 if (bh != 0) {
101 const uint16x8_t v3_0 = vld1q_u16(i0);
102 if XNN_UNPREDICTABLE(bh < 2) {
103 i1 = i0;
104 }
105 const uint16x8_t v3_1 = vld1q_u16(i1);
106 if XNN_UNPREDICTABLE(bh <= 2) {
107 i2 = i0;
108 }
109 const uint16x8_t v3_2 = vld1q_u16(i2);
110 if XNN_UNPREDICTABLE(bh < 4) {
111 i3 = i0;
112 }
113 const uint16x8_t v3_3 = vld1q_u16(i3);
114 if XNN_UNPREDICTABLE(bh <= 4) {
115 i4 = i0;
116 }
117 const uint16x8_t v3_4 = vld1q_u16(i4);
118 if XNN_UNPREDICTABLE(bh < 6) {
119 i5 = i0;
120 }
121 const uint16x8_t v3_5 = vld1q_u16(i5);
122 if XNN_UNPREDICTABLE(bh <= 6) {
123 i6 = i0;
124 }
125 const uint16x8_t v3_6 = vld1q_u16(i6);
126 const uint16x8_t v3_7 = vmovq_n_u16(0);
127
128 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
129 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
130 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
131 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
132
133 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
134 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
135 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
136 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
137 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
138 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
139 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
140 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
141
142 uint16x4_t v0_low = vget_low_u16(v0_0.val[0]);
143 uint16x4_t v1_low = vget_low_u16(v0_0.val[1]);
144 uint16x4_t v2_low = vget_low_u16(v0_1.val[0]);
145 uint16x4_t v3_low = vget_low_u16(v0_1.val[1]);
146 uint16x4_t v4_low = vget_low_u16(v0_2.val[0]);
147 uint16x4_t v5_low = vget_low_u16(v0_2.val[1]);
148 uint16x4_t v6_low = vget_low_u16(v0_3.val[0]);
149 uint16x4_t v7_low = vget_low_u16(v0_3.val[1]);
150
151 if (bh & 4) {
152 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
153 switch (rem) {
154 case 7:
155 vst1_u16(oN, v7_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
156 case 6:
157 vst1_u16(oN, v6_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
158 case 5:
159 vst1_u16(oN, v5_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
160 case 4:
161 vst1_u16(oN, v4_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
162 case 3:
163 vst1_u16(oN, v3_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
164 case 2:
165 vst1_u16(oN, v2_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
166 case 1:
167 vst1_u16(oN, v1_low);
168 case 0:
169 vst1_u16(o, v0_low); o += 4;
170 break;
171 default:
172 XNN_UNREACHABLE;
173 }
174 v0_low = vget_high_u16(v0_0.val[0]);
175 v1_low = vget_high_u16(v0_0.val[1]);
176 v2_low = vget_high_u16(v0_1.val[0]);
177 v3_low = vget_high_u16(v0_1.val[1]);
178 v4_low = vget_high_u16(v0_2.val[0]);
179 v5_low = vget_high_u16(v0_2.val[1]);
180 v6_low = vget_high_u16(v0_3.val[0]);
181 v7_low = vget_high_u16(v0_3.val[1]);
182 }
183
184 if (bh & 2) {
185 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
186 switch (rem) {
187 case 7:
188 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v7_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
189 case 6:
190 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v6_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
191 case 5:
192 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v5_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
193 case 4:
194 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v4_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
195 case 3:
196 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v3_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
197 case 2:
198 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v2_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
199 case 1:
200 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v1_low), 0);
201 case 0:
202 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
203 break;
204 default:
205 XNN_UNREACHABLE;
206 }
207 v0_low = vext_u16(v0_low, v0_low, 2);
208 v1_low = vext_u16(v1_low, v1_low, 2);
209 v2_low = vext_u16(v2_low, v2_low, 2);
210 v3_low = vext_u16(v3_low, v3_low, 2);
211 v4_low = vext_u16(v4_low, v4_low, 2);
212 v5_low = vext_u16(v5_low, v5_low, 2);
213 v6_low = vext_u16(v6_low, v6_low, 2);
214 v7_low = vext_u16(v7_low, v7_low, 2);
215 }
216 if (bh & 1) {
217 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
218 switch (rem) {
219 case 7:
220 vst1_lane_u16(oN, v7_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
221 case 6:
222 vst1_lane_u16(oN, v6_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
223 case 5:
224 vst1_lane_u16(oN, v5_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
225 case 4:
226 vst1_lane_u16(oN, v4_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
227 case 3:
228 vst1_lane_u16(oN, v3_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
229 case 2:
230 vst1_lane_u16(oN, v2_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
231 case 1:
232 vst1_lane_u16(oN, v1_low, 0);
233 case 0:
234 vst1_lane_u16(o, v0_low, 0);
235 break;
236 default:
237 XNN_UNREACHABLE;
238 }
239 }
240 }
241
242 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
243 i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
244 i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
245 i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
246 i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
247 i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
248 i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
249 i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
250 o = (uint16_t*) ((uintptr_t) o + output_reset);
251 block_width = doz(block_width, tile_width);
252 } while (block_width != 0);
253 }
254