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_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_reuse_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 output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint16_t);
35
36 const uint16_t* i0 = input;
37 uint16_t* o = (uint16_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 uint16x8_t v3_0 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
46 const uint16x8_t v3_1 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
47 const uint16x8_t v3_2 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
48 const uint16x8_t v3_3 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
49 const uint16x8_t v3_4 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
50 const uint16x8_t v3_5 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
51 const uint16x8_t v3_6 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
52 const uint16x8_t v3_7 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride);
53
54 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
55 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
56 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
57 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
58
59 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
60 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
61 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
62 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
63 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
64 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
65 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
66 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
67
68 uint16_t *oN = (uint16_t*) ((uintptr_t) o + oN_stride);
69 switch (rem) {
70 case 7:
71 vst1q_u16(oN, v0_3.val[1]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
72 case 6:
73 vst1q_u16(oN, v0_3.val[0]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
74 case 5:
75 vst1q_u16(oN, v0_2.val[1]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
76 case 4:
77 vst1q_u16(oN, v0_2.val[0]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
78 case 3:
79 vst1q_u16(oN, v0_1.val[1]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
80 case 2:
81 vst1q_u16(oN, v0_1.val[0]); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
82 case 1:
83 vst1q_u16(oN, v0_0.val[1]);
84 case 0:
85 vst1q_u16(o, v0_0.val[0]); o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
86 break;
87 default:
88 XNN_UNREACHABLE;
89 }
90 }
91
92 if (bh != 0) {
93 const uint16x8_t v3_0 = vld1q_u16(i0);
94 const uint16_t *i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
95 if XNN_UNPREDICTABLE(bh < 2) {
96 i1 = i0;
97 }
98 const uint16x8_t v3_1 = vld1q_u16(i1);
99 const uint16_t *i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
100 if XNN_UNPREDICTABLE(bh <= 2) {
101 i2 = i1;
102 }
103 const uint16x8_t v3_2 = vld1q_u16(i2);
104 const uint16_t *i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
105 if XNN_UNPREDICTABLE(bh < 4) {
106 i3 = i2;
107 }
108 const uint16x8_t v3_3 = vld1q_u16(i3);
109 const uint16_t *i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
110 if XNN_UNPREDICTABLE(bh <= 4) {
111 i4 = i3;
112 }
113 const uint16x8_t v3_4 = vld1q_u16(i4);
114 const uint16_t *i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
115 if XNN_UNPREDICTABLE(bh < 6) {
116 i5 = i4;
117 }
118 const uint16x8_t v3_5 = vld1q_u16(i5);
119 const uint16_t *i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
120 if XNN_UNPREDICTABLE(bh <= 6) {
121 i6 = i5;
122 }
123 const uint16x8_t v3_6 = vld1q_u16(i6);
124 const uint16x8_t v3_7 = vmovq_n_u16(0);
125
126 const uint16x8x2_t v2_0 = vzipq_u16(v3_0, v3_4);
127 const uint16x8x2_t v2_1 = vzipq_u16(v3_1, v3_5);
128 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6);
129 const uint16x8x2_t v2_3 = vzipq_u16(v3_3, v3_7);
130
131 const uint16x8x2_t v1_0 = vzipq_u16(v2_0.val[0], v2_2.val[0]);
132 const uint16x8x2_t v1_1 = vzipq_u16(v2_0.val[1], v2_2.val[1]);
133 const uint16x8x2_t v1_2 = vzipq_u16(v2_1.val[0], v2_3.val[0]);
134 const uint16x8x2_t v1_3 = vzipq_u16(v2_1.val[1], v2_3.val[1]);
135 const uint16x8x2_t v0_0 = vzipq_u16(v1_0.val[0], v1_2.val[0]);
136 const uint16x8x2_t v0_1 = vzipq_u16(v1_0.val[1], v1_2.val[1]);
137 const uint16x8x2_t v0_2 = vzipq_u16(v1_1.val[0], v1_3.val[0]);
138 const uint16x8x2_t v0_3 = vzipq_u16(v1_1.val[1], v1_3.val[1]);
139
140 uint16x4_t v0_low = vget_low_u16(v0_0.val[0]);
141 uint16x4_t v1_low = vget_low_u16(v0_0.val[1]);
142 uint16x4_t v2_low = vget_low_u16(v0_1.val[0]);
143 uint16x4_t v3_low = vget_low_u16(v0_1.val[1]);
144 uint16x4_t v4_low = vget_low_u16(v0_2.val[0]);
145 uint16x4_t v5_low = vget_low_u16(v0_2.val[1]);
146 uint16x4_t v6_low = vget_low_u16(v0_3.val[0]);
147 uint16x4_t v7_low = vget_low_u16(v0_3.val[1]);
148
149 if (bh & 4) {
150 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
151 switch (rem) {
152 case 7:
153 vst1_u16(oN, v7_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
154 case 6:
155 vst1_u16(oN, v6_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
156 case 5:
157 vst1_u16(oN, v5_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
158 case 4:
159 vst1_u16(oN, v4_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
160 case 3:
161 vst1_u16(oN, v3_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
162 case 2:
163 vst1_u16(oN, v2_low); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
164 case 1:
165 vst1_u16(oN, v1_low);
166 case 0:
167 vst1_u16(o, v0_low); o += 4;
168 break;
169 default:
170 XNN_UNREACHABLE;
171 }
172 v0_low = vget_high_u16(v0_0.val[0]);
173 v1_low = vget_high_u16(v0_0.val[1]);
174 v2_low = vget_high_u16(v0_1.val[0]);
175 v3_low = vget_high_u16(v0_1.val[1]);
176 v4_low = vget_high_u16(v0_2.val[0]);
177 v5_low = vget_high_u16(v0_2.val[1]);
178 v6_low = vget_high_u16(v0_3.val[0]);
179 v7_low = vget_high_u16(v0_3.val[1]);
180 }
181
182 if (bh & 2) {
183 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
184 switch (rem) {
185 case 7:
186 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v7_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
187 case 6:
188 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v6_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
189 case 5:
190 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v5_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
191 case 4:
192 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v4_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
193 case 3:
194 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v3_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
195 case 2:
196 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v2_low), 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
197 case 1:
198 vst1_lane_u32((void*) oN, vreinterpret_u32_u16(v1_low), 0);
199 case 0:
200 vst1_lane_u32((void*) o, vreinterpret_u32_u16(v0_low), 0); o += 2;
201 break;
202 default:
203 XNN_UNREACHABLE;
204 }
205 v0_low = vext_u16(v0_low, v0_low, 2);
206 v1_low = vext_u16(v1_low, v1_low, 2);
207 v2_low = vext_u16(v2_low, v2_low, 2);
208 v3_low = vext_u16(v3_low, v3_low, 2);
209 v4_low = vext_u16(v4_low, v4_low, 2);
210 v5_low = vext_u16(v5_low, v5_low, 2);
211 v6_low = vext_u16(v6_low, v6_low, 2);
212 v7_low = vext_u16(v7_low, v7_low, 2);
213 }
214 if (bh & 1) {
215 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
216 switch (rem) {
217 case 7:
218 vst1_lane_u16(oN, v7_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
219 case 6:
220 vst1_lane_u16(oN, v6_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
221 case 5:
222 vst1_lane_u16(oN, v5_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
223 case 4:
224 vst1_lane_u16(oN, v4_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
225 case 3:
226 vst1_lane_u16(oN, v3_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
227 case 2:
228 vst1_lane_u16(oN, v2_low, 0); oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
229 case 1:
230 vst1_lane_u16(oN, v1_low, 0);
231 case 0:
232 vst1_lane_u16(o, v0_low, 0);
233 break;
234 default:
235 XNN_UNREACHABLE;
236 }
237 }
238 }
239
240 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
241 o = (uint16_t*) ((uintptr_t) o + output_reset);
242 block_width = doz(block_width, tile_width);
243 } while (block_width != 0);
244 }
245