Debian Codegraph @

src/x32-packw/gen/x32-packw-x16s4-gemm-goi-sse2-u4-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/x32-packw/s4-sse2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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
11 #include <assert.h>
12 #include <stddef.h>
13 #include <stdint.h>
14
15 #include <immintrin.h>
16
17 #include "xnnpack/packw.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_x32_packw_gemm_goi_ukernel_x16s4__sse2_u4_prfm(
22 size_t g,
23 size_t nc,
24 size_t kc,
25 size_t nr,
26 size_t kr,
27 size_t sr,
28 const uint32_t* weights,
29 const uint32_t* bias,
30 const void* scale,
31 uint32_t* packed_weights,
32 size_t extra_bytes,
33 const void* params)
34 {
35 assert(g != 0);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(nr == 16);
39 assert(kr == 1);
40 assert(sr == 4);
41 assert(weights != NULL);
42 assert(packed_weights != NULL);
43
44 const float* b = (const float*) bias;
45 float* packed_w = (float*) packed_weights;
46 do {
47 // NC main loop multiple of 16
48 const float* w0 = (const float*) weights;
49 size_t n = nc;
50
51 for (; n >= 16; n -= 16) {
52 if XNN_LIKELY(b != NULL) {
53 const __m128 vb0 = _mm_loadu_ps(b);
54 const __m128 vb4 = _mm_loadu_ps(b + 4);
55 const __m128 vb8 = _mm_loadu_ps(b + 8);
56 const __m128 vb12 = _mm_loadu_ps(b + 12);
57 _mm_store_ps(packed_w, vb0);
58 _mm_store_ps(packed_w + 4, vb4);
59 _mm_store_ps(packed_w + 8, vb8);
60 _mm_store_ps(packed_w + 12, vb12);
61 b += 16;
62 } else {
63 const __m128 vzero = _mm_setzero_ps();
64 _mm_store_ps(packed_w, vzero);
65 _mm_store_ps(packed_w + 4, vzero);
66 _mm_store_ps(packed_w + 8, vzero);
67 _mm_store_ps(packed_w + 12, vzero);
68 }
69 packed_w += 16;
70
71 const float* w1 = w0 + kc;
72 const float* w2 = w1 + kc;
73 const float* w3 = w2 + kc;
74 const float* w4 = w3 + kc;
75 const float* w5 = w4 + kc;
76 const float* w6 = w5 + kc;
77 const float* w7 = w6 + kc;
78 const float* w8 = w7 + kc;
79 const float* w9 = w8 + kc;
80 const float* w10 = w9 + kc;
81 const float* w11 = w10 + kc;
82 const float* w12 = w11 + kc;
83 const float* w13 = w12 + kc;
84 const float* w14 = w13 + kc;
85 const float* w15 = w14 + kc;
86 xnn_prefetch_to_l1((const int8_t*) w0);
87 xnn_prefetch_to_l1((const int8_t*) w0 + 64);
88 xnn_prefetch_to_l1((const int8_t*) w1);
89 xnn_prefetch_to_l1((const int8_t*) w1 + 64);
90 xnn_prefetch_to_l1((const int8_t*) w2);
91 xnn_prefetch_to_l1((const int8_t*) w2 + 64);
92 xnn_prefetch_to_l1((const int8_t*) w3);
93 xnn_prefetch_to_l1((const int8_t*) w3 + 64);
94 xnn_prefetch_to_l1((const int8_t*) w4);
95 xnn_prefetch_to_l1((const int8_t*) w4 + 64);
96 xnn_prefetch_to_l1((const int8_t*) w5);
97 xnn_prefetch_to_l1((const int8_t*) w5 + 64);
98 xnn_prefetch_to_l1((const int8_t*) w6);
99 xnn_prefetch_to_l1((const int8_t*) w6 + 64);
100 xnn_prefetch_to_l1((const int8_t*) w7);
101 xnn_prefetch_to_l1((const int8_t*) w7 + 64);
102 xnn_prefetch_to_l1((const int8_t*) w8);
103 xnn_prefetch_to_l1((const int8_t*) w8 + 64);
104 xnn_prefetch_to_l1((const int8_t*) w9);
105 xnn_prefetch_to_l1((const int8_t*) w9 + 64);
106 xnn_prefetch_to_l1((const int8_t*) w10);
107 xnn_prefetch_to_l1((const int8_t*) w10 + 64);
108 xnn_prefetch_to_l1((const int8_t*) w11);
109 xnn_prefetch_to_l1((const int8_t*) w11 + 64);
110 xnn_prefetch_to_l1((const int8_t*) w12);
111 xnn_prefetch_to_l1((const int8_t*) w12 + 64);
112 xnn_prefetch_to_l1((const int8_t*) w13);
113 xnn_prefetch_to_l1((const int8_t*) w13 + 64);
114 xnn_prefetch_to_l1((const int8_t*) w14);
115 xnn_prefetch_to_l1((const int8_t*) w14 + 64);
116 xnn_prefetch_to_l1((const int8_t*) w15);
117 xnn_prefetch_to_l1((const int8_t*) w15 + 64);
118
119 size_t k = kc;
120
121 // KC multiple of 4
122 for (; k >= 4; k -= 4) {
123 // Read blocks of 4x4
124 // a b c d
125 // e f g h
126 // i j k l
127 // m n o p
128 __m128 v0x0123 = _mm_loadu_ps(w0);
129 w0 += 4;
130 __m128 v1x0123 = _mm_loadu_ps(w1);
131 w1 += 4;
132 __m128 v2x0123 = _mm_loadu_ps(w2);
133 w2 += 4;
134 __m128 v3x0123 = _mm_loadu_ps(w3);
135 w3 += 4;
136 __m128 v4x0123 = _mm_loadu_ps(w4);
137 w4 += 4;
138 __m128 v5x0123 = _mm_loadu_ps(w5);
139 w5 += 4;
140 __m128 v6x0123 = _mm_loadu_ps(w6);
141 w6 += 4;
142 __m128 v7x0123 = _mm_loadu_ps(w7);
143 w7 += 4;
144 __m128 v8x0123 = _mm_loadu_ps(w8);
145 w8 += 4;
146 __m128 v9x0123 = _mm_loadu_ps(w9);
147 w9 += 4;
148 __m128 v10x0123 = _mm_loadu_ps(w10);
149 w10 += 4;
150 __m128 v11x0123 = _mm_loadu_ps(w11);
151 w11 += 4;
152 __m128 v12x0123 = _mm_loadu_ps(w12);
153 w12 += 4;
154 __m128 v13x0123 = _mm_loadu_ps(w13);
155 w13 += 4;
156 __m128 v14x0123 = _mm_loadu_ps(w14);
157 w14 += 4;
158 __m128 v15x0123 = _mm_loadu_ps(w15);
159 w15 += 4;
160
161 // Apply SR4 shuffle
162 v1x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v1x0123), _MM_SHUFFLE(0, 3, 2, 1)));
163 v2x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v2x0123), _MM_SHUFFLE(1, 0, 3, 2)));
164 v3x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v3x0123), _MM_SHUFFLE(2, 1, 0, 3)));
165 v5x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v5x0123), _MM_SHUFFLE(0, 3, 2, 1)));
166 v6x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v6x0123), _MM_SHUFFLE(1, 0, 3, 2)));
167 v7x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v7x0123), _MM_SHUFFLE(2, 1, 0, 3)));
168 v9x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v9x0123), _MM_SHUFFLE(0, 3, 2, 1)));
169 v10x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v10x0123), _MM_SHUFFLE(1, 0, 3, 2)));
170 v11x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v11x0123), _MM_SHUFFLE(2, 1, 0, 3)));
171 v13x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v13x0123), _MM_SHUFFLE(0, 3, 2, 1)));
172 v14x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v14x0123), _MM_SHUFFLE(1, 0, 3, 2)));
173 v15x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v15x0123), _MM_SHUFFLE(2, 1, 0, 3)));
174
175 // Transpose 2x2
176 const __m128 vtmp0x0123 = _mm_unpacklo_ps(v0x0123, v1x0123); // a e b f from row 0, 1
177 const __m128 vtmp1x0123 = _mm_unpacklo_ps(v2x0123, v3x0123); // i m j n from row 2, 3
178 const __m128 vtmp2x0123 = _mm_unpackhi_ps(v0x0123, v1x0123); // c g d h from row 0, 1
179 const __m128 vtmp3x0123 = _mm_unpackhi_ps(v2x0123, v3x0123); // k o l p from row 2, 3
180 const __m128 vtmp4x0123 = _mm_unpacklo_ps(v4x0123, v5x0123); // a e b f from row 0, 1
181 const __m128 vtmp5x0123 = _mm_unpacklo_ps(v6x0123, v7x0123); // i m j n from row 2, 3
182 const __m128 vtmp6x0123 = _mm_unpackhi_ps(v4x0123, v5x0123); // c g d h from row 0, 1
183 const __m128 vtmp7x0123 = _mm_unpackhi_ps(v6x0123, v7x0123); // k o l p from row 2, 3
184 const __m128 vtmp8x0123 = _mm_unpacklo_ps(v8x0123, v9x0123); // a e b f from row 0, 1
185 const __m128 vtmp9x0123 = _mm_unpacklo_ps(v10x0123, v11x0123); // i m j n from row 2, 3
186 const __m128 vtmp10x0123 = _mm_unpackhi_ps(v8x0123, v9x0123); // c g d h from row 0, 1
187 const __m128 vtmp11x0123 = _mm_unpackhi_ps(v10x0123, v11x0123); // k o l p from row 2, 3
188 const __m128 vtmp12x0123 = _mm_unpacklo_ps(v12x0123, v13x0123); // a e b f from row 0, 1
189 const __m128 vtmp13x0123 = _mm_unpacklo_ps(v14x0123, v15x0123); // i m j n from row 2, 3
190 const __m128 vtmp14x0123 = _mm_unpackhi_ps(v12x0123, v13x0123); // c g d h from row 0, 1
191 const __m128 vtmp15x0123 = _mm_unpackhi_ps(v14x0123, v15x0123); // k o l p from row 2, 3
192 xnn_prefetch_to_l1((const int8_t*) w0 + 128);
193 xnn_prefetch_to_l1((const int8_t*) w1 + 128);
194 xnn_prefetch_to_l1((const int8_t*) w2 + 128);
195 xnn_prefetch_to_l1((const int8_t*) w3 + 128);
196 xnn_prefetch_to_l1((const int8_t*) w4 + 128);
197 xnn_prefetch_to_l1((const int8_t*) w5 + 128);
198 xnn_prefetch_to_l1((const int8_t*) w6 + 128);
199 xnn_prefetch_to_l1((const int8_t*) w7 + 128);
200 xnn_prefetch_to_l1((const int8_t*) w8 + 128);
201 xnn_prefetch_to_l1((const int8_t*) w9 + 128);
202 xnn_prefetch_to_l1((const int8_t*) w10 + 128);
203 xnn_prefetch_to_l1((const int8_t*) w11 + 128);
204 xnn_prefetch_to_l1((const int8_t*) w12 + 128);
205 xnn_prefetch_to_l1((const int8_t*) w13 + 128);
206 xnn_prefetch_to_l1((const int8_t*) w14 + 128);
207 xnn_prefetch_to_l1((const int8_t*) w15 + 128);
208 // Transpose 4x4
209 v0x0123 = _mm_movelh_ps(vtmp0x0123, vtmp1x0123); // a e i m from row 0, 1
210 v1x0123 = _mm_movehl_ps(vtmp1x0123, vtmp0x0123); // b f j n from row 0, 1
211 v2x0123 = _mm_movelh_ps(vtmp2x0123, vtmp3x0123); // c g k o from row 2, 3
212 v3x0123 = _mm_movehl_ps(vtmp3x0123, vtmp2x0123); // d h l p from row 2, 3
213 v4x0123 = _mm_movelh_ps(vtmp4x0123, vtmp5x0123); // a e i m from row 0, 1
214 v5x0123 = _mm_movehl_ps(vtmp5x0123, vtmp4x0123); // b f j n from row 0, 1
215 v6x0123 = _mm_movelh_ps(vtmp6x0123, vtmp7x0123); // c g k o from row 2, 3
216 v7x0123 = _mm_movehl_ps(vtmp7x0123, vtmp6x0123); // d h l p from row 2, 3
217 v8x0123 = _mm_movelh_ps(vtmp8x0123, vtmp9x0123); // a e i m from row 0, 1
218 v9x0123 = _mm_movehl_ps(vtmp9x0123, vtmp8x0123); // b f j n from row 0, 1
219 v10x0123 = _mm_movelh_ps(vtmp10x0123, vtmp11x0123); // c g k o from row 2, 3
220 v11x0123 = _mm_movehl_ps(vtmp11x0123, vtmp10x0123); // d h l p from row 2, 3
221 v12x0123 = _mm_movelh_ps(vtmp12x0123, vtmp13x0123); // a e i m from row 0, 1
222 v13x0123 = _mm_movehl_ps(vtmp13x0123, vtmp12x0123); // b f j n from row 0, 1
223 v14x0123 = _mm_movelh_ps(vtmp14x0123, vtmp15x0123); // c g k o from row 2, 3
224 v15x0123 = _mm_movehl_ps(vtmp15x0123, vtmp14x0123); // d h l p from row 2, 3
225
226 _mm_store_ps(packed_w, v0x0123);
227 _mm_store_ps(packed_w + 4, v4x0123);
228 _mm_store_ps(packed_w + 8, v8x0123);
229 _mm_store_ps(packed_w + 12, v12x0123);
230 _mm_store_ps(packed_w + 16, v1x0123);
231 _mm_store_ps(packed_w + 20, v5x0123);
232 _mm_store_ps(packed_w + 24, v9x0123);
233 _mm_store_ps(packed_w + 28, v13x0123);
234 _mm_store_ps(packed_w + 32, v2x0123);
235 _mm_store_ps(packed_w + 36, v6x0123);
236 _mm_store_ps(packed_w + 40, v10x0123);
237 _mm_store_ps(packed_w + 44, v14x0123);
238 _mm_store_ps(packed_w + 48, v3x0123);
239 _mm_store_ps(packed_w + 52, v7x0123);
240 _mm_store_ps(packed_w + 56, v11x0123);
241 _mm_store_ps(packed_w + 60, v15x0123);
242 packed_w += 64;
243 }
244
245 // KC remainder (1..3)
246 if XNN_UNLIKELY(k != 0) {
247 assert(k >= 1);
248 assert(k <= 3);
249 __m128 v0 = _mm_undefined_ps();
250 __m128 v1 = _mm_undefined_ps();
251 __m128 v2 = _mm_undefined_ps();
252 __m128 v3 = _mm_undefined_ps();
253 __m128 v4 = _mm_undefined_ps();
254 __m128 v5 = _mm_undefined_ps();
255 __m128 v6 = _mm_undefined_ps();
256 __m128 v7 = _mm_undefined_ps();
257 __m128 v8 = _mm_undefined_ps();
258 __m128 v9 = _mm_undefined_ps();
259 __m128 v10 = _mm_undefined_ps();
260 __m128 v11 = _mm_undefined_ps();
261 __m128 v12 = _mm_undefined_ps();
262 __m128 v13 = _mm_undefined_ps();
263 __m128 v14 = _mm_undefined_ps();
264 __m128 v15 = _mm_undefined_ps();
265
266 switch (k) {
267 case 1:
268 // Read blocks of 4x1
269 // a
270 // e
271 // i
272 // m
273 v0 = _mm_load_ss(w0);
274 w0 += 1;
275 v1 = _mm_load_ss(w1);
276 w1 += 1;
277 v2 = _mm_load_ss(w2);
278 w2 += 1;
279 v3 = _mm_load_ss(w3);
280 w3 += 1;
281 v4 = _mm_load_ss(w4);
282 w4 += 1;
283 v5 = _mm_load_ss(w5);
284 w5 += 1;
285 v6 = _mm_load_ss(w6);
286 w6 += 1;
287 v7 = _mm_load_ss(w7);
288 w7 += 1;
289 v8 = _mm_load_ss(w8);
290 w8 += 1;
291 v9 = _mm_load_ss(w9);
292 w9 += 1;
293 v10 = _mm_load_ss(w10);
294 w10 += 1;
295 v11 = _mm_load_ss(w11);
296 w11 += 1;
297 v12 = _mm_load_ss(w12);
298 w12 += 1;
299 v13 = _mm_load_ss(w13);
300 w13 += 1;
301 v14 = _mm_load_ss(w14);
302 w14 += 1;
303 v15 = _mm_load_ss(w15);
304 w15 += 1;
305 break;
306 case 2:
307 // Read blocks of 4x2
308 // a b
309 // e f
310 // i j
311 // m n
312 v0 = _mm_castpd_ps(_mm_load_sd((const double*) w0));
313 w0 += 2;
314 v1 = _mm_castpd_ps(_mm_load_sd((const double*) w1));
315 w1 += 2;
316 v2 = _mm_castpd_ps(_mm_load_sd((const double*) w2));
317 w2 += 2;
318 v3 = _mm_castpd_ps(_mm_load_sd((const double*) w3));
319 w3 += 2;
320 v4 = _mm_castpd_ps(_mm_load_sd((const double*) w4));
321 w4 += 2;
322 v5 = _mm_castpd_ps(_mm_load_sd((const double*) w5));
323 w5 += 2;
324 v6 = _mm_castpd_ps(_mm_load_sd((const double*) w6));
325 w6 += 2;
326 v7 = _mm_castpd_ps(_mm_load_sd((const double*) w7));
327 w7 += 2;
328 v8 = _mm_castpd_ps(_mm_load_sd((const double*) w8));
329 w8 += 2;
330 v9 = _mm_castpd_ps(_mm_load_sd((const double*) w9));
331 w9 += 2;
332 v10 = _mm_castpd_ps(_mm_load_sd((const double*) w10));
333 w10 += 2;
334 v11 = _mm_castpd_ps(_mm_load_sd((const double*) w11));
335 w11 += 2;
336 v12 = _mm_castpd_ps(_mm_load_sd((const double*) w12));
337 w12 += 2;
338 v13 = _mm_castpd_ps(_mm_load_sd((const double*) w13));
339 w13 += 2;
340 v14 = _mm_castpd_ps(_mm_load_sd((const double*) w14));
341 w14 += 2;
342 v15 = _mm_castpd_ps(_mm_load_sd((const double*) w15));
343 w15 += 2;
344 break;
345 case 3:
346 {
347 // Read blocks of 4x3
348 // a b c
349 // e f g
350 // i j k
351 // m n o
352 const __m128 v0lo = _mm_castpd_ps(_mm_load_sd((const double*) w0));
353 const __m128 v0hi = _mm_load_ss(w0 + 2);
354 v0 = _mm_movelh_ps(v0lo, v0hi);
355 w0 += 3;
356 const __m128 v1lo = _mm_castpd_ps(_mm_load_sd((const double*) w1));
357 const __m128 v1hi = _mm_load_ss(w1 + 2);
358 v1 = _mm_movelh_ps(v1lo, v1hi);
359 w1 += 3;
360 const __m128 v2lo = _mm_castpd_ps(_mm_load_sd((const double*) w2));
361 const __m128 v2hi = _mm_load_ss(w2 + 2);
362 v2 = _mm_movelh_ps(v2lo, v2hi);
363 w2 += 3;
364 const __m128 v3lo = _mm_castpd_ps(_mm_load_sd((const double*) w3));
365 const __m128 v3hi = _mm_load_ss(w3 + 2);
366 v3 = _mm_movelh_ps(v3lo, v3hi);
367 w3 += 3;
368 const __m128 v4lo = _mm_castpd_ps(_mm_load_sd((const double*) w4));
369 const __m128 v4hi = _mm_load_ss(w4 + 2);
370 v4 = _mm_movelh_ps(v4lo, v4hi);
371 w4 += 3;
372 const __m128 v5lo = _mm_castpd_ps(_mm_load_sd((const double*) w5));
373 const __m128 v5hi = _mm_load_ss(w5 + 2);
374 v5 = _mm_movelh_ps(v5lo, v5hi);
375 w5 += 3;
376 const __m128 v6lo = _mm_castpd_ps(_mm_load_sd((const double*) w6));
377 const __m128 v6hi = _mm_load_ss(w6 + 2);
378 v6 = _mm_movelh_ps(v6lo, v6hi);
379 w6 += 3;
380 const __m128 v7lo = _mm_castpd_ps(_mm_load_sd((const double*) w7));
381 const __m128 v7hi = _mm_load_ss(w7 + 2);
382 v7 = _mm_movelh_ps(v7lo, v7hi);
383 w7 += 3;
384 const __m128 v8lo = _mm_castpd_ps(_mm_load_sd((const double*) w8));
385 const __m128 v8hi = _mm_load_ss(w8 + 2);
386 v8 = _mm_movelh_ps(v8lo, v8hi);
387 w8 += 3;
388 const __m128 v9lo = _mm_castpd_ps(_mm_load_sd((const double*) w9));
389 const __m128 v9hi = _mm_load_ss(w9 + 2);
390 v9 = _mm_movelh_ps(v9lo, v9hi);
391 w9 += 3;
392 const __m128 v10lo = _mm_castpd_ps(_mm_load_sd((const double*) w10));
393 const __m128 v10hi = _mm_load_ss(w10 + 2);
394 v10 = _mm_movelh_ps(v10lo, v10hi);
395 w10 += 3;
396 const __m128 v11lo = _mm_castpd_ps(_mm_load_sd((const double*) w11));
397 const __m128 v11hi = _mm_load_ss(w11 + 2);
398 v11 = _mm_movelh_ps(v11lo, v11hi);
399 w11 += 3;
400 const __m128 v12lo = _mm_castpd_ps(_mm_load_sd((const double*) w12));
401 const __m128 v12hi = _mm_load_ss(w12 + 2);
402 v12 = _mm_movelh_ps(v12lo, v12hi);
403 w12 += 3;
404 const __m128 v13lo = _mm_castpd_ps(_mm_load_sd((const double*) w13));
405 const __m128 v13hi = _mm_load_ss(w13 + 2);
406 v13 = _mm_movelh_ps(v13lo, v13hi);
407 w13 += 3;
408 const __m128 v14lo = _mm_castpd_ps(_mm_load_sd((const double*) w14));
409 const __m128 v14hi = _mm_load_ss(w14 + 2);
410 v14 = _mm_movelh_ps(v14lo, v14hi);
411 w14 += 3;
412 const __m128 v15lo = _mm_castpd_ps(_mm_load_sd((const double*) w15));
413 const __m128 v15hi = _mm_load_ss(w15 + 2);
414 v15 = _mm_movelh_ps(v15lo, v15hi);
415 w15 += 3;
416 break;
417 }
418 default:
419 XNN_UNREACHABLE;
420 }
421
422 // Apply SR4 shuffle
423 v1 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v1), _MM_SHUFFLE(0, 3, 2, 1)));
424 v2 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v2), _MM_SHUFFLE(1, 0, 3, 2)));
425 v3 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v3), _MM_SHUFFLE(2, 1, 0, 3)));
426 v5 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v5), _MM_SHUFFLE(0, 3, 2, 1)));
427 v6 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v6), _MM_SHUFFLE(1, 0, 3, 2)));
428 v7 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v7), _MM_SHUFFLE(2, 1, 0, 3)));
429 v9 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v9), _MM_SHUFFLE(0, 3, 2, 1)));
430 v10 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v10), _MM_SHUFFLE(1, 0, 3, 2)));
431 v11 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v11), _MM_SHUFFLE(2, 1, 0, 3)));
432 v13 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v13), _MM_SHUFFLE(0, 3, 2, 1)));
433 v14 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v14), _MM_SHUFFLE(1, 0, 3, 2)));
434 v15 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v15), _MM_SHUFFLE(2, 1, 0, 3)));
435 // Transpose 2x2
436 const __m128 vtmp0 = _mm_unpacklo_ps(v0, v1); // a e b f from row 0, 1
437 const __m128 vtmp1 = _mm_unpacklo_ps(v2, v3); // i m j n from row 2, 3
438 const __m128 vtmp2 = _mm_unpackhi_ps(v0, v1); // c g d h from row 0, 1
439 const __m128 vtmp3 = _mm_unpackhi_ps(v2, v3); // k o l p from row 2, 3
440 const __m128 vtmp4 = _mm_unpacklo_ps(v4, v5); // a e b f from row 0, 1
441 const __m128 vtmp5 = _mm_unpacklo_ps(v6, v7); // i m j n from row 2, 3
442 const __m128 vtmp6 = _mm_unpackhi_ps(v4, v5); // c g d h from row 0, 1
443 const __m128 vtmp7 = _mm_unpackhi_ps(v6, v7); // k o l p from row 2, 3
444 const __m128 vtmp8 = _mm_unpacklo_ps(v8, v9); // a e b f from row 0, 1
445 const __m128 vtmp9 = _mm_unpacklo_ps(v10, v11); // i m j n from row 2, 3
446 const __m128 vtmp10 = _mm_unpackhi_ps(v8, v9); // c g d h from row 0, 1
447 const __m128 vtmp11 = _mm_unpackhi_ps(v10, v11); // k o l p from row 2, 3
448 const __m128 vtmp12 = _mm_unpacklo_ps(v12, v13); // a e b f from row 0, 1
449 const __m128 vtmp13 = _mm_unpacklo_ps(v14, v15); // i m j n from row 2, 3
450 const __m128 vtmp14 = _mm_unpackhi_ps(v12, v13); // c g d h from row 0, 1
451 const __m128 vtmp15 = _mm_unpackhi_ps(v14, v15); // k o l p from row 2, 3
452 // Transpose 4x4
453 v0 = _mm_movelh_ps(vtmp0, vtmp1); // a e i m from row 0, 1
454 v1 = _mm_movehl_ps(vtmp1, vtmp0); // b f j n from row 0, 1
455 v2 = _mm_movelh_ps(vtmp2, vtmp3); // c g k o from row 2, 3
456 v3 = _mm_movehl_ps(vtmp3, vtmp2); // d h l p from row 2, 3
457 v4 = _mm_movelh_ps(vtmp4, vtmp5); // a e i m from row 0, 1
458 v5 = _mm_movehl_ps(vtmp5, vtmp4); // b f j n from row 0, 1
459 v6 = _mm_movelh_ps(vtmp6, vtmp7); // c g k o from row 2, 3
460 v7 = _mm_movehl_ps(vtmp7, vtmp6); // d h l p from row 2, 3
461 v8 = _mm_movelh_ps(vtmp8, vtmp9); // a e i m from row 0, 1
462 v9 = _mm_movehl_ps(vtmp9, vtmp8); // b f j n from row 0, 1
463 v10 = _mm_movelh_ps(vtmp10, vtmp11); // c g k o from row 2, 3
464 v11 = _mm_movehl_ps(vtmp11, vtmp10); // d h l p from row 2, 3
465 v12 = _mm_movelh_ps(vtmp12, vtmp13); // a e i m from row 0, 1
466 v13 = _mm_movehl_ps(vtmp13, vtmp12); // b f j n from row 0, 1
467 v14 = _mm_movelh_ps(vtmp14, vtmp15); // c g k o from row 2, 3
468 v15 = _mm_movehl_ps(vtmp15, vtmp14); // d h l p from row 2, 3
469 _mm_store_ps(packed_w, v0);
470 _mm_store_ps(packed_w + 4, v4);
471 _mm_store_ps(packed_w + 8, v8);
472 _mm_store_ps(packed_w + 12, v12);
473 _mm_store_ps(packed_w + 16, v1);
474 _mm_store_ps(packed_w + 20, v5);
475 _mm_store_ps(packed_w + 24, v9);
476 _mm_store_ps(packed_w + 28, v13);
477 _mm_store_ps(packed_w + 32, v2);
478 _mm_store_ps(packed_w + 36, v6);
479 _mm_store_ps(packed_w + 40, v10);
480 _mm_store_ps(packed_w + 44, v14);
481 _mm_store_ps(packed_w + 48, v3);
482 _mm_store_ps(packed_w + 52, v7);
483 _mm_store_ps(packed_w + 56, v11);
484 _mm_store_ps(packed_w + 60, v15);
485 packed_w += 64;
486 }
487 packed_w = (float*) ((uintptr_t) packed_w + extra_bytes);
488 w0 = w15;
489 }
490
491 // NC remainder (1..15)
492 if XNN_UNLIKELY(n != 0) {
493 assert(n >= 1);
494 assert(n <= 15);
495 if XNN_LIKELY(b != NULL) {
496 size_t nb = n;
497 do {
498 *packed_w++ = *b++;
499 } while (--nb != 0);
500 packed_w += (16 - n);
501 } else {
502 const __m128 vzero = _mm_setzero_ps();
503 _mm_store_ps(packed_w, vzero);
504 _mm_store_ps(packed_w + 4, vzero);
505 _mm_store_ps(packed_w + 8, vzero);
506 _mm_store_ps(packed_w + 12, vzero);
507 packed_w += 16;
508 }
509
510 // NR remainder has less than 16 rows so last row is not loaded
511 const float* w1 = w0 + kc;
512 if XNN_UNPREDICTABLE(n < 2) {
513 w1 = w0;
514 }
515 const float* w2 = w1 + kc;
516 if XNN_UNPREDICTABLE(n <= 2) {
517 w2 = w1;
518 }
519 const float* w3 = w2 + kc;
520 if XNN_UNPREDICTABLE(n < 4) {
521 w3 = w2;
522 }
523 const float* w4 = w3 + kc;
524 if XNN_UNPREDICTABLE(n <= 4) {
525 w4 = w3;
526 }
527 const float* w5 = w4 + kc;
528 if XNN_UNPREDICTABLE(n < 6) {
529 w5 = w4;
530 }
531 const float* w6 = w5 + kc;
532 if XNN_UNPREDICTABLE(n <= 6) {
533 w6 = w5;
534 }
535 const float* w7 = w6 + kc;
536 if XNN_UNPREDICTABLE(n < 8) {
537 w7 = w6;
538 }
539 const float* w8 = w7 + kc;
540 if XNN_UNPREDICTABLE(n <= 8) {
541 w8 = w7;
542 }
543 const float* w9 = w8 + kc;
544 if XNN_UNPREDICTABLE(n < 10) {
545 w9 = w8;
546 }
547 const float* w10 = w9 + kc;
548 if XNN_UNPREDICTABLE(n <= 10) {
549 w10 = w9;
550 }
551 const float* w11 = w10 + kc;
552 if XNN_UNPREDICTABLE(n < 12) {
553 w11 = w10;
554 }
555 const float* w12 = w11 + kc;
556 if XNN_UNPREDICTABLE(n <= 12) {
557 w12 = w11;
558 }
559 const float* w13 = w12 + kc;
560 if XNN_UNPREDICTABLE(n < 14) {
561 w13 = w12;
562 }
563 const float* w14 = w13 + kc;
564 if XNN_UNPREDICTABLE(n <= 14) {
565 w14 = w13;
566 }
567
568 size_t k = kc;
569
570 // KC multiple of 4
571 for (; k >= 4; k -= 4) {
572 // Read blocks of 4x4
573 // a b c d
574 // e f g h
575 // i j k l
576 // m n o p
577 __m128 v0x0123 = _mm_loadu_ps(w0);
578 w0 += 4;
579 __m128 v1x0123 = _mm_loadu_ps(w1);
580 w1 += 4;
581 __m128 v2x0123 = _mm_loadu_ps(w2);
582 w2 += 4;
583 __m128 v3x0123 = _mm_loadu_ps(w3);
584 w3 += 4;
585 __m128 v4x0123 = _mm_loadu_ps(w4);
586 w4 += 4;
587 __m128 v5x0123 = _mm_loadu_ps(w5);
588 w5 += 4;
589 __m128 v6x0123 = _mm_loadu_ps(w6);
590 w6 += 4;
591 __m128 v7x0123 = _mm_loadu_ps(w7);
592 w7 += 4;
593 __m128 v8x0123 = _mm_loadu_ps(w8);
594 w8 += 4;
595 __m128 v9x0123 = _mm_loadu_ps(w9);
596 w9 += 4;
597 __m128 v10x0123 = _mm_loadu_ps(w10);
598 w10 += 4;
599 __m128 v11x0123 = _mm_loadu_ps(w11);
600 w11 += 4;
601 __m128 v12x0123 = _mm_loadu_ps(w12);
602 w12 += 4;
603 __m128 v13x0123 = _mm_loadu_ps(w13);
604 w13 += 4;
605 __m128 v14x0123 = _mm_loadu_ps(w14);
606 w14 += 4;
607 __m128 v15x0123 = _mm_undefined_ps();
608
609 // Apply SR4 shuffle
610 v1x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v1x0123), _MM_SHUFFLE(0, 3, 2, 1)));
611 v2x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v2x0123), _MM_SHUFFLE(1, 0, 3, 2)));
612 v3x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v3x0123), _MM_SHUFFLE(2, 1, 0, 3)));
613 v5x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v5x0123), _MM_SHUFFLE(0, 3, 2, 1)));
614 v6x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v6x0123), _MM_SHUFFLE(1, 0, 3, 2)));
615 v7x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v7x0123), _MM_SHUFFLE(2, 1, 0, 3)));
616 v9x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v9x0123), _MM_SHUFFLE(0, 3, 2, 1)));
617 v10x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v10x0123), _MM_SHUFFLE(1, 0, 3, 2)));
618 v11x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v11x0123), _MM_SHUFFLE(2, 1, 0, 3)));
619 v13x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v13x0123), _MM_SHUFFLE(0, 3, 2, 1)));
620 v14x0123 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v14x0123), _MM_SHUFFLE(1, 0, 3, 2)));
621
622 // Transpose 2x2
623 const __m128 vtmp0x0123 = _mm_unpacklo_ps(v0x0123, v1x0123); // a e b f from row 0, 1
624 const __m128 vtmp1x0123 = _mm_unpacklo_ps(v2x0123, v3x0123); // i m j n from row 2, 3
625 const __m128 vtmp2x0123 = _mm_unpackhi_ps(v0x0123, v1x0123); // c g d h from row 0, 1
626 const __m128 vtmp3x0123 = _mm_unpackhi_ps(v2x0123, v3x0123); // k o l p from row 2, 3
627 const __m128 vtmp4x0123 = _mm_unpacklo_ps(v4x0123, v5x0123); // a e b f from row 0, 1
628 const __m128 vtmp5x0123 = _mm_unpacklo_ps(v6x0123, v7x0123); // i m j n from row 2, 3
629 const __m128 vtmp6x0123 = _mm_unpackhi_ps(v4x0123, v5x0123); // c g d h from row 0, 1
630 const __m128 vtmp7x0123 = _mm_unpackhi_ps(v6x0123, v7x0123); // k o l p from row 2, 3
631 const __m128 vtmp8x0123 = _mm_unpacklo_ps(v8x0123, v9x0123); // a e b f from row 0, 1
632 const __m128 vtmp9x0123 = _mm_unpacklo_ps(v10x0123, v11x0123); // i m j n from row 2, 3
633 const __m128 vtmp10x0123 = _mm_unpackhi_ps(v8x0123, v9x0123); // c g d h from row 0, 1
634 const __m128 vtmp11x0123 = _mm_unpackhi_ps(v10x0123, v11x0123); // k o l p from row 2, 3
635 const __m128 vtmp12x0123 = _mm_unpacklo_ps(v12x0123, v13x0123); // a e b f from row 0, 1
636 const __m128 vtmp13x0123 = _mm_unpacklo_ps(v14x0123, v14x0123); // i m j n from row 2, 3
637 const __m128 vtmp14x0123 = _mm_unpackhi_ps(v12x0123, v13x0123); // c g d h from row 0, 1
638 const __m128 vtmp15x0123 = _mm_unpackhi_ps(v14x0123, v14x0123); // k o l p from row 2, 3
639 xnn_prefetch_to_l1((const int8_t*) w0 + 128);
640 xnn_prefetch_to_l1((const int8_t*) w1 + 128);
641 xnn_prefetch_to_l1((const int8_t*) w2 + 128);
642 xnn_prefetch_to_l1((const int8_t*) w3 + 128);
643 xnn_prefetch_to_l1((const int8_t*) w4 + 128);
644 xnn_prefetch_to_l1((const int8_t*) w5 + 128);
645 xnn_prefetch_to_l1((const int8_t*) w6 + 128);
646 xnn_prefetch_to_l1((const int8_t*) w7 + 128);
647 xnn_prefetch_to_l1((const int8_t*) w8 + 128);
648 xnn_prefetch_to_l1((const int8_t*) w9 + 128);
649 xnn_prefetch_to_l1((const int8_t*) w10 + 128);
650 xnn_prefetch_to_l1((const int8_t*) w11 + 128);
651 xnn_prefetch_to_l1((const int8_t*) w12 + 128);
652 xnn_prefetch_to_l1((const int8_t*) w13 + 128);
653 xnn_prefetch_to_l1((const int8_t*) w14 + 128);
654 // Transpose 4x4
655 v0x0123 = _mm_movelh_ps(vtmp0x0123, vtmp1x0123); // a e i m from row 0, 1
656 v1x0123 = _mm_movehl_ps(vtmp1x0123, vtmp0x0123); // b f j n from row 0, 1
657 v2x0123 = _mm_movelh_ps(vtmp2x0123, vtmp3x0123); // c g k o from row 2, 3
658 v3x0123 = _mm_movehl_ps(vtmp3x0123, vtmp2x0123); // d h l p from row 2, 3
659 v4x0123 = _mm_movelh_ps(vtmp4x0123, vtmp5x0123); // a e i m from row 0, 1
660 v5x0123 = _mm_movehl_ps(vtmp5x0123, vtmp4x0123); // b f j n from row 0, 1
661 v6x0123 = _mm_movelh_ps(vtmp6x0123, vtmp7x0123); // c g k o from row 2, 3
662 v7x0123 = _mm_movehl_ps(vtmp7x0123, vtmp6x0123); // d h l p from row 2, 3
663 v8x0123 = _mm_movelh_ps(vtmp8x0123, vtmp9x0123); // a e i m from row 0, 1
664 v9x0123 = _mm_movehl_ps(vtmp9x0123, vtmp8x0123); // b f j n from row 0, 1
665 v10x0123 = _mm_movelh_ps(vtmp10x0123, vtmp11x0123); // c g k o from row 2, 3
666 v11x0123 = _mm_movehl_ps(vtmp11x0123, vtmp10x0123); // d h l p from row 2, 3
667 v12x0123 = _mm_movelh_ps(vtmp12x0123, vtmp13x0123); // a e i m from row 0, 1
668 v13x0123 = _mm_movehl_ps(vtmp13x0123, vtmp12x0123); // b f j n from row 0, 1
669 v14x0123 = _mm_movelh_ps(vtmp14x0123, vtmp15x0123); // c g k o from row 2, 3
670 v15x0123 = _mm_movehl_ps(vtmp15x0123, vtmp14x0123); // d h l p from row 2, 3
671
672 _mm_store_ps(packed_w, v0x0123);
673 _mm_store_ps(packed_w + 4, v4x0123);
674 _mm_store_ps(packed_w + 8, v8x0123);
675 _mm_store_ps(packed_w + 12, v12x0123);
676 _mm_store_ps(packed_w + 16, v1x0123);
677 _mm_store_ps(packed_w + 20, v5x0123);
678 _mm_store_ps(packed_w + 24, v9x0123);
679 _mm_store_ps(packed_w + 28, v13x0123);
680 _mm_store_ps(packed_w + 32, v2x0123);
681 _mm_store_ps(packed_w + 36, v6x0123);
682 _mm_store_ps(packed_w + 40, v10x0123);
683 _mm_store_ps(packed_w + 44, v14x0123);
684 _mm_store_ps(packed_w + 48, v3x0123);
685 _mm_store_ps(packed_w + 52, v7x0123);
686 _mm_store_ps(packed_w + 56, v11x0123);
687 _mm_store_ps(packed_w + 60, v15x0123);
688 packed_w += 64;
689 }
690
691 // KC remainder (1..3)
692 if XNN_UNLIKELY(k != 0) {
693 assert(k >= 1);
694 assert(k <= 3);
695 __m128 v0 = _mm_undefined_ps();
696 __m128 v1 = _mm_undefined_ps();
697 __m128 v2 = _mm_undefined_ps();
698 __m128 v3 = _mm_undefined_ps();
699 __m128 v4 = _mm_undefined_ps();
700 __m128 v5 = _mm_undefined_ps();
701 __m128 v6 = _mm_undefined_ps();
702 __m128 v7 = _mm_undefined_ps();
703 __m128 v8 = _mm_undefined_ps();
704 __m128 v9 = _mm_undefined_ps();
705 __m128 v10 = _mm_undefined_ps();
706 __m128 v11 = _mm_undefined_ps();
707 __m128 v12 = _mm_undefined_ps();
708 __m128 v13 = _mm_undefined_ps();
709 __m128 v14 = _mm_undefined_ps();
710 __m128 v15 = _mm_undefined_ps();
711
712 switch (k) {
713 case 1:
714 // Read blocks of 4x1
715 // a
716 // e
717 // i
718 // m
719 v0 = _mm_load_ss(w0);
720 w0 += 1;
721 v1 = _mm_load_ss(w1);
722 w1 += 1;
723 v2 = _mm_load_ss(w2);
724 w2 += 1;
725 v3 = _mm_load_ss(w3);
726 w3 += 1;
727 v4 = _mm_load_ss(w4);
728 w4 += 1;
729 v5 = _mm_load_ss(w5);
730 w5 += 1;
731 v6 = _mm_load_ss(w6);
732 w6 += 1;
733 v7 = _mm_load_ss(w7);
734 w7 += 1;
735 v8 = _mm_load_ss(w8);
736 w8 += 1;
737 v9 = _mm_load_ss(w9);
738 w9 += 1;
739 v10 = _mm_load_ss(w10);
740 w10 += 1;
741 v11 = _mm_load_ss(w11);
742 w11 += 1;
743 v12 = _mm_load_ss(w12);
744 w12 += 1;
745 v13 = _mm_load_ss(w13);
746 w13 += 1;
747 v14 = _mm_load_ss(w14);
748 w14 += 1;
749 break;
750 case 2:
751 // Read blocks of 4x2
752 // a b
753 // e f
754 // i j
755 // m n
756 v0 = _mm_castpd_ps(_mm_load_sd((const double*) w0));
757 w0 += 2;
758 v1 = _mm_castpd_ps(_mm_load_sd((const double*) w1));
759 w1 += 2;
760 v2 = _mm_castpd_ps(_mm_load_sd((const double*) w2));
761 w2 += 2;
762 v3 = _mm_castpd_ps(_mm_load_sd((const double*) w3));
763 w3 += 2;
764 v4 = _mm_castpd_ps(_mm_load_sd((const double*) w4));
765 w4 += 2;
766 v5 = _mm_castpd_ps(_mm_load_sd((const double*) w5));
767 w5 += 2;
768 v6 = _mm_castpd_ps(_mm_load_sd((const double*) w6));
769 w6 += 2;
770 v7 = _mm_castpd_ps(_mm_load_sd((const double*) w7));
771 w7 += 2;
772 v8 = _mm_castpd_ps(_mm_load_sd((const double*) w8));
773 w8 += 2;
774 v9 = _mm_castpd_ps(_mm_load_sd((const double*) w9));
775 w9 += 2;
776 v10 = _mm_castpd_ps(_mm_load_sd((const double*) w10));
777 w10 += 2;
778 v11 = _mm_castpd_ps(_mm_load_sd((const double*) w11));
779 w11 += 2;
780 v12 = _mm_castpd_ps(_mm_load_sd((const double*) w12));
781 w12 += 2;
782 v13 = _mm_castpd_ps(_mm_load_sd((const double*) w13));
783 w13 += 2;
784 v14 = _mm_castpd_ps(_mm_load_sd((const double*) w14));
785 w14 += 2;
786 break;
787 case 3:
788 {
789 // Read blocks of 4x3
790 // a b c
791 // e f g
792 // i j k
793 // m n o
794 const __m128 v0lo = _mm_castpd_ps(_mm_load_sd((const double*) w0));
795 const __m128 v0hi = _mm_load_ss(w0 + 2);
796 v0 = _mm_movelh_ps(v0lo, v0hi);
797 w0 += 3;
798 const __m128 v1lo = _mm_castpd_ps(_mm_load_sd((const double*) w1));
799 const __m128 v1hi = _mm_load_ss(w1 + 2);
800 v1 = _mm_movelh_ps(v1lo, v1hi);
801 w1 += 3;
802 const __m128 v2lo = _mm_castpd_ps(_mm_load_sd((const double*) w2));
803 const __m128 v2hi = _mm_load_ss(w2 + 2);
804 v2 = _mm_movelh_ps(v2lo, v2hi);
805 w2 += 3;
806 const __m128 v3lo = _mm_castpd_ps(_mm_load_sd((const double*) w3));
807 const __m128 v3hi = _mm_load_ss(w3 + 2);
808 v3 = _mm_movelh_ps(v3lo, v3hi);
809 w3 += 3;
810 const __m128 v4lo = _mm_castpd_ps(_mm_load_sd((const double*) w4));
811 const __m128 v4hi = _mm_load_ss(w4 + 2);
812 v4 = _mm_movelh_ps(v4lo, v4hi);
813 w4 += 3;
814 const __m128 v5lo = _mm_castpd_ps(_mm_load_sd((const double*) w5));
815 const __m128 v5hi = _mm_load_ss(w5 + 2);
816 v5 = _mm_movelh_ps(v5lo, v5hi);
817 w5 += 3;
818 const __m128 v6lo = _mm_castpd_ps(_mm_load_sd((const double*) w6));
819 const __m128 v6hi = _mm_load_ss(w6 + 2);
820 v6 = _mm_movelh_ps(v6lo, v6hi);
821 w6 += 3;
822 const __m128 v7lo = _mm_castpd_ps(_mm_load_sd((const double*) w7));
823 const __m128 v7hi = _mm_load_ss(w7 + 2);
824 v7 = _mm_movelh_ps(v7lo, v7hi);
825 w7 += 3;
826 const __m128 v8lo = _mm_castpd_ps(_mm_load_sd((const double*) w8));
827 const __m128 v8hi = _mm_load_ss(w8 + 2);
828 v8 = _mm_movelh_ps(v8lo, v8hi);
829 w8 += 3;
830 const __m128 v9lo = _mm_castpd_ps(_mm_load_sd((const double*) w9));
831 const __m128 v9hi = _mm_load_ss(w9 + 2);
832 v9 = _mm_movelh_ps(v9lo, v9hi);
833 w9 += 3;
834 const __m128 v10lo = _mm_castpd_ps(_mm_load_sd((const double*) w10));
835 const __m128 v10hi = _mm_load_ss(w10 + 2);
836 v10 = _mm_movelh_ps(v10lo, v10hi);
837 w10 += 3;
838 const __m128 v11lo = _mm_castpd_ps(_mm_load_sd((const double*) w11));
839 const __m128 v11hi = _mm_load_ss(w11 + 2);
840 v11 = _mm_movelh_ps(v11lo, v11hi);
841 w11 += 3;
842 const __m128 v12lo = _mm_castpd_ps(_mm_load_sd((const double*) w12));
843 const __m128 v12hi = _mm_load_ss(w12 + 2);
844 v12 = _mm_movelh_ps(v12lo, v12hi);
845 w12 += 3;
846 const __m128 v13lo = _mm_castpd_ps(_mm_load_sd((const double*) w13));
847 const __m128 v13hi = _mm_load_ss(w13 + 2);
848 v13 = _mm_movelh_ps(v13lo, v13hi);
849 w13 += 3;
850 const __m128 v14lo = _mm_castpd_ps(_mm_load_sd((const double*) w14));
851 const __m128 v14hi = _mm_load_ss(w14 + 2);
852 v14 = _mm_movelh_ps(v14lo, v14hi);
853 w14 += 3;
854 break;
855 }
856 default:
857 XNN_UNREACHABLE;
858 }
859
860 // Apply SR4 shuffle
861 v1 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v1), _MM_SHUFFLE(0, 3, 2, 1)));
862 v2 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v2), _MM_SHUFFLE(1, 0, 3, 2)));
863 v3 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v3), _MM_SHUFFLE(2, 1, 0, 3)));
864 v5 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v5), _MM_SHUFFLE(0, 3, 2, 1)));
865 v6 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v6), _MM_SHUFFLE(1, 0, 3, 2)));
866 v7 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v7), _MM_SHUFFLE(2, 1, 0, 3)));
867 v9 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v9), _MM_SHUFFLE(0, 3, 2, 1)));
868 v10 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v10), _MM_SHUFFLE(1, 0, 3, 2)));
869 v11 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v11), _MM_SHUFFLE(2, 1, 0, 3)));
870 v13 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v13), _MM_SHUFFLE(0, 3, 2, 1)));
871 v14 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v14), _MM_SHUFFLE(1, 0, 3, 2)));
872 // Transpose 2x2
873 const __m128 vtmp0 = _mm_unpacklo_ps(v0, v1); // a e b f from row 0, 1
874 const __m128 vtmp1 = _mm_unpacklo_ps(v2, v3); // i m j n from row 2, 3
875 const __m128 vtmp2 = _mm_unpackhi_ps(v0, v1); // c g d h from row 0, 1
876 const __m128 vtmp3 = _mm_unpackhi_ps(v2, v3); // k o l p from row 2, 3
877 const __m128 vtmp4 = _mm_unpacklo_ps(v4, v5); // a e b f from row 0, 1
878 const __m128 vtmp5 = _mm_unpacklo_ps(v6, v7); // i m j n from row 2, 3
879 const __m128 vtmp6 = _mm_unpackhi_ps(v4, v5); // c g d h from row 0, 1
880 const __m128 vtmp7 = _mm_unpackhi_ps(v6, v7); // k o l p from row 2, 3
881 const __m128 vtmp8 = _mm_unpacklo_ps(v8, v9); // a e b f from row 0, 1
882 const __m128 vtmp9 = _mm_unpacklo_ps(v10, v11); // i m j n from row 2, 3
883 const __m128 vtmp10 = _mm_unpackhi_ps(v8, v9); // c g d h from row 0, 1
884 const __m128 vtmp11 = _mm_unpackhi_ps(v10, v11); // k o l p from row 2, 3
885 const __m128 vtmp12 = _mm_unpacklo_ps(v12, v13); // a e b f from row 0, 1
886 const __m128 vtmp13 = _mm_unpacklo_ps(v14, v14); // i m j n from row 2, 3
887 const __m128 vtmp14 = _mm_unpackhi_ps(v12, v13); // c g d h from row 0, 1
888 const __m128 vtmp15 = _mm_unpackhi_ps(v14, v14); // k o l p from row 2, 3
889 // Transpose 4x4
890 v0 = _mm_movelh_ps(vtmp0, vtmp1); // a e i m from row 0, 1
891 v1 = _mm_movehl_ps(vtmp1, vtmp0); // b f j n from row 0, 1
892 v2 = _mm_movelh_ps(vtmp2, vtmp3); // c g k o from row 2, 3
893 v3 = _mm_movehl_ps(vtmp3, vtmp2); // d h l p from row 2, 3
894 v4 = _mm_movelh_ps(vtmp4, vtmp5); // a e i m from row 0, 1
895 v5 = _mm_movehl_ps(vtmp5, vtmp4); // b f j n from row 0, 1
896 v6 = _mm_movelh_ps(vtmp6, vtmp7); // c g k o from row 2, 3
897 v7 = _mm_movehl_ps(vtmp7, vtmp6); // d h l p from row 2, 3
898 v8 = _mm_movelh_ps(vtmp8, vtmp9); // a e i m from row 0, 1
899 v9 = _mm_movehl_ps(vtmp9, vtmp8); // b f j n from row 0, 1
900 v10 = _mm_movelh_ps(vtmp10, vtmp11); // c g k o from row 2, 3
901 v11 = _mm_movehl_ps(vtmp11, vtmp10); // d h l p from row 2, 3
902 v12 = _mm_movelh_ps(vtmp12, vtmp13); // a e i m from row 0, 1
903 v13 = _mm_movehl_ps(vtmp13, vtmp12); // b f j n from row 0, 1
904 v14 = _mm_movelh_ps(vtmp14, vtmp15); // c g k o from row 2, 3
905 v15 = _mm_movehl_ps(vtmp15, vtmp14); // d h l p from row 2, 3
906 _mm_store_ps(packed_w, v0);
907 _mm_store_ps(packed_w + 4, v4);
908 _mm_store_ps(packed_w + 8, v8);
909 _mm_store_ps(packed_w + 12, v12);
910 _mm_store_ps(packed_w + 16, v1);
911 _mm_store_ps(packed_w + 20, v5);
912 _mm_store_ps(packed_w + 24, v9);
913 _mm_store_ps(packed_w + 28, v13);
914 _mm_store_ps(packed_w + 32, v2);
915 _mm_store_ps(packed_w + 36, v6);
916 _mm_store_ps(packed_w + 40, v10);
917 _mm_store_ps(packed_w + 44, v14);
918 _mm_store_ps(packed_w + 48, v3);
919 _mm_store_ps(packed_w + 52, v7);
920 _mm_store_ps(packed_w + 56, v11);
921 _mm_store_ps(packed_w + 60, v15);
922 packed_w += 64;
923 }
924 packed_w = (float*) ((uintptr_t) packed_w + extra_bytes);
925 }
926 weights += nc * kc;
927 } while (--g != 0);
928 }