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