1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transposec/wasmsimd.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 <wasm_simd128.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_wasmsimd(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_wasmsimd(
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 v128_t v3_0 = wasm_v128_load(i0);
54 i0 = (uint16_t*) ((uintptr_t) i0 + input_offset);
55 const v128_t v3_1 = wasm_v128_load(i1);
56 i1 = (uint16_t*) ((uintptr_t) i1 + input_offset);
57 const v128_t v3_2 = wasm_v128_load(i2);
58 i2 = (uint16_t*) ((uintptr_t) i2 + input_offset);
59 const v128_t v3_3 = wasm_v128_load(i3);
60 i3 = (uint16_t*) ((uintptr_t) i3 + input_offset);
61 const v128_t v3_4 = wasm_v128_load(i4);
62 i4 = (uint16_t*) ((uintptr_t) i4 + input_offset);
63 const v128_t v3_5 = wasm_v128_load(i5);
64 i5 = (uint16_t*) ((uintptr_t) i5 + input_offset);
65 const v128_t v3_6 = wasm_v128_load(i6);
66 i6 = (uint16_t*) ((uintptr_t) i6 + input_offset);
67 const v128_t v3_7 = wasm_v128_load(i7);
68 i7 = (uint16_t*) ((uintptr_t) i7 + input_offset);
69
70 const v128_t v2_0 = wasm_v16x8_shuffle(v3_0, v3_4, 0, 8, 1, 9, 2, 10, 3, 11);
71 const v128_t v2_1 = wasm_v16x8_shuffle(v3_0, v3_4, 4, 12, 5, 13, 6, 14, 7, 15);
72 const v128_t v2_2 = wasm_v16x8_shuffle(v3_1, v3_5, 0, 8, 1, 9, 2, 10, 3, 11);
73 const v128_t v2_3 = wasm_v16x8_shuffle(v3_1, v3_5, 4, 12, 5, 13, 6, 14, 7, 15);
74 const v128_t v2_4 = wasm_v16x8_shuffle(v3_2, v3_6, 0, 8, 1, 9, 2, 10, 3, 11);
75 const v128_t v2_5 = wasm_v16x8_shuffle(v3_2, v3_6, 4, 12, 5, 13, 6, 14, 7, 15);
76 const v128_t v2_6 = wasm_v16x8_shuffle(v3_3, v3_7, 0, 8, 1, 9, 2, 10, 3, 11);
77 const v128_t v2_7 = wasm_v16x8_shuffle(v3_3, v3_7, 4, 12, 5, 13, 6, 14, 7, 15);
78 const v128_t v1_0 = wasm_v16x8_shuffle(v2_0, v2_4, 0, 8, 1, 9, 2, 10, 3, 11);
79 const v128_t v1_1 = wasm_v16x8_shuffle(v2_0, v2_4, 4, 12, 5, 13, 6, 14, 7, 15);
80 const v128_t v1_2 = wasm_v16x8_shuffle(v2_1, v2_5, 0, 8, 1, 9, 2, 10, 3, 11);
81 const v128_t v1_3 = wasm_v16x8_shuffle(v2_1, v2_5, 4, 12, 5, 13, 6, 14, 7, 15);
82 const v128_t v1_4 = wasm_v16x8_shuffle(v2_2, v2_6, 0, 8, 1, 9, 2, 10, 3, 11);
83 const v128_t v1_5 = wasm_v16x8_shuffle(v2_2, v2_6, 4, 12, 5, 13, 6, 14, 7, 15);
84 const v128_t v1_6 = wasm_v16x8_shuffle(v2_3, v2_7, 0, 8, 1, 9, 2, 10, 3, 11);
85 const v128_t v1_7 = wasm_v16x8_shuffle(v2_3, v2_7, 4, 12, 5, 13, 6, 14, 7, 15);
86 const v128_t v0_0 = wasm_v16x8_shuffle(v1_0, v1_4, 0, 8, 1, 9, 2, 10, 3, 11);
87 const v128_t v0_1 = wasm_v16x8_shuffle(v1_0, v1_4, 4, 12, 5, 13, 6, 14, 7, 15);
88 const v128_t v0_2 = wasm_v16x8_shuffle(v1_1, v1_5, 0, 8, 1, 9, 2, 10, 3, 11);
89 const v128_t v0_3 = wasm_v16x8_shuffle(v1_1, v1_5, 4, 12, 5, 13, 6, 14, 7, 15);
90 const v128_t v0_4 = wasm_v16x8_shuffle(v1_2, v1_6, 0, 8, 1, 9, 2, 10, 3, 11);
91 const v128_t v0_5 = wasm_v16x8_shuffle(v1_2, v1_6, 4, 12, 5, 13, 6, 14, 7, 15);
92 const v128_t v0_6 = wasm_v16x8_shuffle(v1_3, v1_7, 0, 8, 1, 9, 2, 10, 3, 11);
93 const v128_t v0_7 = wasm_v16x8_shuffle(v1_3, v1_7, 4, 12, 5, 13, 6, 14, 7, 15);
94
95 uint16_t *oN = (uint16_t*) ((uintptr_t) o + oN_stride);
96 switch (rem) {
97 case 7:
98 wasm_v128_store(oN, v0_7);
99 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
100 case 6:
101 wasm_v128_store(oN, v0_6);
102 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
103 case 5:
104 wasm_v128_store(oN, v0_5);
105 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
106 case 4:
107 wasm_v128_store(oN, v0_4);
108 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
109 case 3:
110 wasm_v128_store(oN, v0_3);
111 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
112 case 2:
113 wasm_v128_store(oN, v0_2);
114 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
115 case 1:
116 wasm_v128_store(oN, v0_1);
117 case 0:
118 wasm_v128_store(o, v0_0);
119 o = (uint16_t*) ((uintptr_t) o + tile_hbytes);
120 break;
121 default:
122 XNN_UNREACHABLE;
123 }
124 }
125
126 if (bh != 0) {
127 const v128_t v3_0 = wasm_v128_load(i0);
128 if XNN_UNPREDICTABLE(bh < 2) {
129 i1 = i0;
130 }
131 const v128_t v3_1 = wasm_v128_load(i1);
132 if XNN_UNPREDICTABLE(bh <= 2) {
133 i2 = i0;
134 }
135 const v128_t v3_2 = wasm_v128_load(i2);
136 if XNN_UNPREDICTABLE(bh < 4) {
137 i3 = i0;
138 }
139 const v128_t v3_3 = wasm_v128_load(i3);
140 if XNN_UNPREDICTABLE(bh <= 4) {
141 i4 = i0;
142 }
143 const v128_t v3_4 = wasm_v128_load(i4);
144 if XNN_UNPREDICTABLE(bh < 6) {
145 i5 = i0;
146 }
147 const v128_t v3_5 = wasm_v128_load(i5);
148 if XNN_UNPREDICTABLE(bh <= 6) {
149 i6 = i0;
150 }
151 const v128_t v3_6 = wasm_v128_load(i6);
152 const v128_t v3_7 = wasm_v128_xor(v3_0, v3_0);
153
154 const v128_t v2_0 = wasm_v16x8_shuffle(v3_0, v3_4, 0, 8, 1, 9, 2, 10, 3, 11);
155 const v128_t v2_1 = wasm_v16x8_shuffle(v3_0, v3_4, 4, 12, 5, 13, 6, 14, 7, 15);
156 const v128_t v2_2 = wasm_v16x8_shuffle(v3_1, v3_5, 0, 8, 1, 9, 2, 10, 3, 11);
157 const v128_t v2_3 = wasm_v16x8_shuffle(v3_1, v3_5, 4, 12, 5, 13, 6, 14, 7, 15);
158 const v128_t v2_4 = wasm_v16x8_shuffle(v3_2, v3_6, 0, 8, 1, 9, 2, 10, 3, 11);
159 const v128_t v2_5 = wasm_v16x8_shuffle(v3_2, v3_6, 4, 12, 5, 13, 6, 14, 7, 15);
160 const v128_t v2_6 = wasm_v16x8_shuffle(v3_3, v3_7, 0, 8, 1, 9, 2, 10, 3, 11);
161 const v128_t v2_7 = wasm_v16x8_shuffle(v3_3, v3_7, 4, 12, 5, 13, 6, 14, 7, 15);
162 const v128_t v1_0 = wasm_v16x8_shuffle(v2_0, v2_4, 0, 8, 1, 9, 2, 10, 3, 11);
163 const v128_t v1_1 = wasm_v16x8_shuffle(v2_0, v2_4, 4, 12, 5, 13, 6, 14, 7, 15);
164 const v128_t v1_2 = wasm_v16x8_shuffle(v2_1, v2_5, 0, 8, 1, 9, 2, 10, 3, 11);
165 const v128_t v1_3 = wasm_v16x8_shuffle(v2_1, v2_5, 4, 12, 5, 13, 6, 14, 7, 15);
166 const v128_t v1_4 = wasm_v16x8_shuffle(v2_2, v2_6, 0, 8, 1, 9, 2, 10, 3, 11);
167 const v128_t v1_5 = wasm_v16x8_shuffle(v2_2, v2_6, 4, 12, 5, 13, 6, 14, 7, 15);
168 const v128_t v1_6 = wasm_v16x8_shuffle(v2_3, v2_7, 0, 8, 1, 9, 2, 10, 3, 11);
169 const v128_t v1_7 = wasm_v16x8_shuffle(v2_3, v2_7, 4, 12, 5, 13, 6, 14, 7, 15);
170
171 v128_t v0_0 = wasm_v16x8_shuffle(v1_0, v1_4, 0, 8, 1, 9, 2, 10, 3, 11);
172 v128_t v0_1 = wasm_v16x8_shuffle(v1_0, v1_4, 4, 12, 5, 13, 6, 14, 7, 15);
173 v128_t v0_2 = wasm_v16x8_shuffle(v1_1, v1_5, 0, 8, 1, 9, 2, 10, 3, 11);
174 v128_t v0_3 = wasm_v16x8_shuffle(v1_1, v1_5, 4, 12, 5, 13, 6, 14, 7, 15);
175 v128_t v0_4 = wasm_v16x8_shuffle(v1_2, v1_6, 0, 8, 1, 9, 2, 10, 3, 11);
176 v128_t v0_5 = wasm_v16x8_shuffle(v1_2, v1_6, 4, 12, 5, 13, 6, 14, 7, 15);
177 v128_t v0_6 = wasm_v16x8_shuffle(v1_3, v1_7, 0, 8, 1, 9, 2, 10, 3, 11);
178 v128_t v0_7 = wasm_v16x8_shuffle(v1_3, v1_7, 4, 12, 5, 13, 6, 14, 7, 15);
179
180 if (bh & 4) {
181 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
182 switch (rem) {
183 case 7:
184 *((double*) oN) = wasm_f64x2_extract_lane(v0_7, 0);
185 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
186 case 6:
187 *((double*) oN) = wasm_f64x2_extract_lane(v0_6, 0);
188 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
189 case 5:
190 *((double*) oN) = wasm_f64x2_extract_lane(v0_5, 0);
191 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
192 case 4:
193 *((double*) oN) = wasm_f64x2_extract_lane(v0_4, 0);
194 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
195 case 3:
196 *((double*) oN) = wasm_f64x2_extract_lane(v0_3, 0);
197 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
198 case 2:
199 *((double*) oN) = wasm_f64x2_extract_lane(v0_2, 0);
200 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
201 case 1:
202 *((double*) oN) = wasm_f64x2_extract_lane(v0_1, 0);
203 case 0:
204 *((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
205 o += 4;
206 break;
207 default:
208 XNN_UNREACHABLE;
209 }
210 v0_0 = wasm_v64x2_shuffle(v0_0, v0_0, 1, 1);
211 v0_1 = wasm_v64x2_shuffle(v0_1, v0_1, 1, 1);
212 v0_2 = wasm_v64x2_shuffle(v0_2, v0_2, 1, 1);
213 v0_3 = wasm_v64x2_shuffle(v0_3, v0_3, 1, 1);
214 v0_4 = wasm_v64x2_shuffle(v0_4, v0_4, 1, 1);
215 v0_5 = wasm_v64x2_shuffle(v0_5, v0_5, 1, 1);
216 v0_6 = wasm_v64x2_shuffle(v0_6, v0_6, 1, 1);
217 v0_7 = wasm_v64x2_shuffle(v0_7, v0_7, 1, 1);
218 }
219
220 if (bh & 2) {
221 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
222 switch (rem) {
223 case 7:
224 *((float*) oN) = wasm_f32x4_extract_lane(v0_7, 0);
225 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
226 case 6:
227 *((float*) oN) = wasm_f32x4_extract_lane(v0_6, 0);
228 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
229 case 5:
230 *((float*) oN) = wasm_f32x4_extract_lane(v0_5, 0);
231 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
232 case 4:
233 *((float*) oN) = wasm_f32x4_extract_lane(v0_4, 0);
234 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
235 case 3:
236 *((float*) oN) = wasm_f32x4_extract_lane(v0_3, 0);
237 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
238 case 2:
239 *((float*) oN) = wasm_f32x4_extract_lane(v0_2, 0);
240 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
241 case 1:
242 *((float*) oN) = wasm_f32x4_extract_lane(v0_1, 0);
243 case 0:
244 *((float*) o) = wasm_f32x4_extract_lane(v0_0, 0);
245 o += 2;
246 break;
247 default:
248 XNN_UNREACHABLE;
249 }
250 v0_0 = wasm_u64x2_shr(v0_0, 32);
251 v0_1 = wasm_u64x2_shr(v0_1, 32);
252 v0_2 = wasm_u64x2_shr(v0_2, 32);
253 v0_3 = wasm_u64x2_shr(v0_3, 32);
254 v0_4 = wasm_u64x2_shr(v0_4, 32);
255 v0_5 = wasm_u64x2_shr(v0_5, 32);
256 v0_6 = wasm_u64x2_shr(v0_6, 32);
257 v0_7 = wasm_u64x2_shr(v0_7, 32);
258 }
259 if (bh & 1) {
260 uint16_t* oN = (uint16_t*) ((uintptr_t) o + oN_stride);
261 switch (rem) {
262 case 7:
263 *oN = wasm_i16x8_extract_lane(v0_7, 0);
264 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
265 case 6:
266 *oN = wasm_i16x8_extract_lane(v0_6, 0);
267 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
268 case 5:
269 *oN = wasm_i16x8_extract_lane(v0_5, 0);
270 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
271 case 4:
272 *oN = wasm_i16x8_extract_lane(v0_4, 0);
273 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
274 case 3:
275 *oN = wasm_i16x8_extract_lane(v0_3, 0);
276 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
277 case 2:
278 *oN = wasm_i16x8_extract_lane(v0_2, 0);
279 oN = (uint16_t*) ((uintptr_t) oN + minus_output_stride);
280 case 1:
281 *oN = wasm_i16x8_extract_lane(v0_1, 0);
282 case 0:
283 *o = wasm_i16x8_extract_lane(v0_0, 0);
284 break;
285 default:
286 XNN_UNREACHABLE;
287 }
288 }
289 }
290
291 i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset);
292 i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
293 i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
294 i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
295 i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
296 i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
297 i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
298 i7 = (const uint16_t*) ((uintptr_t) i6 + input_stride);
299 o = (uint16_t*) ((uintptr_t) o + output_reset);
300 block_width = doz(block_width, tile_width);
301 } while (block_width != 0);
302 }
303