1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/psimd-splat.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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 <assert.h>
11
12 #include <psimd.h>
13
14 #include <xnnpack/igemm.h>
15
16
xnn_f32_igemm_ukernel_6x8__psimd_splat(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_igemm_ukernel_6x8__psimd_splat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_output_params params[restrict static 1])
30 {
31 assert(mr != 0);
32 assert(mr <= 6);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (6 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr < 4) {
54 c3 = c2;
55 }
56 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 6) {
62 c5 = c4;
63 }
64
65 do {
66 psimd_f32 vacc0x0123 = psimd_load_f32(w);
67 psimd_f32 vacc0x4567 = psimd_load_f32(w + 4);
68 psimd_f32 vacc1x0123 = vacc0x0123;
69 psimd_f32 vacc1x4567 = vacc0x4567;
70 psimd_f32 vacc2x0123 = vacc0x0123;
71 psimd_f32 vacc2x4567 = vacc0x4567;
72 psimd_f32 vacc3x0123 = vacc0x0123;
73 psimd_f32 vacc3x4567 = vacc0x4567;
74 psimd_f32 vacc4x0123 = vacc0x0123;
75 psimd_f32 vacc4x4567 = vacc0x4567;
76 psimd_f32 vacc5x0123 = vacc0x0123;
77 psimd_f32 vacc5x4567 = vacc0x4567;
78 w += 8;
79
80 size_t p = ks;
81 do {
82 const float* restrict a0 = a[0];
83 assert(a0 != NULL);
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const float*) ((uintptr_t) a0 + a_offset);
86 }
87 const float* restrict a1 = a[1];
88 assert(a1 != NULL);
89 if XNN_UNPREDICTABLE(a1 != zero) {
90 a1 = (const float*) ((uintptr_t) a1 + a_offset);
91 }
92 const float* restrict a2 = a[2];
93 assert(a2 != NULL);
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const float*) ((uintptr_t) a2 + a_offset);
96 }
97 const float* restrict a3 = a[3];
98 assert(a3 != NULL);
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const float*) ((uintptr_t) a3 + a_offset);
101 }
102 const float* restrict a4 = a[4];
103 assert(a4 != NULL);
104 if XNN_UNPREDICTABLE(a4 != zero) {
105 a4 = (const float*) ((uintptr_t) a4 + a_offset);
106 }
107 const float* restrict a5 = a[5];
108 assert(a5 != NULL);
109 if XNN_UNPREDICTABLE(a5 != zero) {
110 a5 = (const float*) ((uintptr_t) a5 + a_offset);
111 }
112 a += 6;
113
114 size_t k = kc;
115 while (k >= 4 * sizeof(float)) {
116 const psimd_f32 va0 = psimd_load_f32(a0);
117 a0 += 4;
118 const psimd_f32 va1 = psimd_load_f32(a1);
119 a1 += 4;
120 const psimd_f32 va2 = psimd_load_f32(a2);
121 a2 += 4;
122 const psimd_f32 va3 = psimd_load_f32(a3);
123 a3 += 4;
124 const psimd_f32 va4 = psimd_load_f32(a4);
125 a4 += 4;
126 const psimd_f32 va5 = psimd_load_f32(a5);
127 a5 += 4;
128
129 const psimd_f32 va0c0 = psimd_splat0_f32(va0);
130 const psimd_f32 va1c0 = psimd_splat0_f32(va1);
131 const psimd_f32 va2c0 = psimd_splat0_f32(va2);
132 const psimd_f32 va3c0 = psimd_splat0_f32(va3);
133 const psimd_f32 va4c0 = psimd_splat0_f32(va4);
134 const psimd_f32 va5c0 = psimd_splat0_f32(va5);
135
136 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
137 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
138
139 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c0, vb0123c0);
140 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c0, vb0123c0);
141 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0);
142 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c0, vb0123c0);
143 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c0, vb0123c0);
144 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c0, vb0123c0);
145 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c0, vb4567c0);
146 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c0, vb4567c0);
147 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0);
148 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c0, vb4567c0);
149 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c0, vb4567c0);
150 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c0, vb4567c0);
151 const psimd_f32 va0c1 = psimd_splat1_f32(va0);
152 const psimd_f32 va1c1 = psimd_splat1_f32(va1);
153 const psimd_f32 va2c1 = psimd_splat1_f32(va2);
154 const psimd_f32 va3c1 = psimd_splat1_f32(va3);
155 const psimd_f32 va4c1 = psimd_splat1_f32(va4);
156 const psimd_f32 va5c1 = psimd_splat1_f32(va5);
157
158 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
159 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
160
161 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c1, vb0123c1);
162 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1);
163 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1);
164 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c1, vb0123c1);
165 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c1, vb0123c1);
166 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c1, vb0123c1);
167 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c1, vb4567c1);
168 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1);
169 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1);
170 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c1, vb4567c1);
171 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c1, vb4567c1);
172 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c1, vb4567c1);
173 const psimd_f32 va0c2 = psimd_splat2_f32(va0);
174 const psimd_f32 va1c2 = psimd_splat2_f32(va1);
175 const psimd_f32 va2c2 = psimd_splat2_f32(va2);
176 const psimd_f32 va3c2 = psimd_splat2_f32(va3);
177 const psimd_f32 va4c2 = psimd_splat2_f32(va4);
178 const psimd_f32 va5c2 = psimd_splat2_f32(va5);
179
180 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
181 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
182
183 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c2, vb0123c2);
184 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c2, vb0123c2);
185 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c2, vb0123c2);
186 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c2, vb0123c2);
187 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c2, vb0123c2);
188 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c2, vb0123c2);
189 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c2, vb4567c2);
190 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c2, vb4567c2);
191 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c2, vb4567c2);
192 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c2, vb4567c2);
193 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c2, vb4567c2);
194 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c2, vb4567c2);
195 const psimd_f32 va0c3 = psimd_splat3_f32(va0);
196 const psimd_f32 va1c3 = psimd_splat3_f32(va1);
197 const psimd_f32 va2c3 = psimd_splat3_f32(va2);
198 const psimd_f32 va3c3 = psimd_splat3_f32(va3);
199 const psimd_f32 va4c3 = psimd_splat3_f32(va4);
200 const psimd_f32 va5c3 = psimd_splat3_f32(va5);
201
202 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
203 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
204
205 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c3, vb0123c3);
206 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c3, vb0123c3);
207 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c3, vb0123c3);
208 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c3, vb0123c3);
209 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c3, vb0123c3);
210 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c3, vb0123c3);
211 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c3, vb4567c3);
212 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c3, vb4567c3);
213 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c3, vb4567c3);
214 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c3, vb4567c3);
215 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c3, vb4567c3);
216 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c3, vb4567c3);
217
218 w += 32;
219 k -= 4 * sizeof(float);
220 }
221 if XNN_UNLIKELY(k != 0) {
222 do {
223 const psimd_f32 vb0123 = psimd_load_f32(w);
224 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
225 w += 8;
226
227 const psimd_f32 va0 = psimd_load_splat_f32(a0);
228 a0 += 1;
229 const psimd_f32 va1 = psimd_load_splat_f32(a1);
230 a1 += 1;
231 const psimd_f32 va2 = psimd_load_splat_f32(a2);
232 a2 += 1;
233 const psimd_f32 va3 = psimd_load_splat_f32(a3);
234 a3 += 1;
235 const psimd_f32 va4 = psimd_load_splat_f32(a4);
236 a4 += 1;
237 const psimd_f32 va5 = psimd_load_splat_f32(a5);
238 a5 += 1;
239
240 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
241 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
242 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123);
243 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567);
244 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);
245 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);
246 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123);
247 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567);
248 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4, vb0123);
249 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4, vb4567);
250 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123);
251 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567);
252 k -= sizeof(float);
253 } while (k != 0);
254 }
255 p -= 6 * sizeof(void*);
256 } while (p != 0);
257
258 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
259 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
260 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
261 vacc2x0123 = psimd_min_f32(vacc2x0123, vmax);
262 vacc3x0123 = psimd_min_f32(vacc3x0123, vmax);
263 vacc4x0123 = psimd_min_f32(vacc4x0123, vmax);
264 vacc5x0123 = psimd_min_f32(vacc5x0123, vmax);
265 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
266 vacc1x4567 = psimd_min_f32(vacc1x4567, vmax);
267 vacc2x4567 = psimd_min_f32(vacc2x4567, vmax);
268 vacc3x4567 = psimd_min_f32(vacc3x4567, vmax);
269 vacc4x4567 = psimd_min_f32(vacc4x4567, vmax);
270 vacc5x4567 = psimd_min_f32(vacc5x4567, vmax);
271
272 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
273 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
274 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
275 vacc2x0123 = psimd_max_f32(vacc2x0123, vmin);
276 vacc3x0123 = psimd_max_f32(vacc3x0123, vmin);
277 vacc4x0123 = psimd_max_f32(vacc4x0123, vmin);
278 vacc5x0123 = psimd_max_f32(vacc5x0123, vmin);
279 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
280 vacc1x4567 = psimd_max_f32(vacc1x4567, vmin);
281 vacc2x4567 = psimd_max_f32(vacc2x4567, vmin);
282 vacc3x4567 = psimd_max_f32(vacc3x4567, vmin);
283 vacc4x4567 = psimd_max_f32(vacc4x4567, vmin);
284 vacc5x4567 = psimd_max_f32(vacc5x4567, vmin);
285
286 if XNN_LIKELY(nc >= 8) {
287 psimd_store_f32(c5, vacc5x0123);
288 psimd_store_f32(c5 + 4, vacc5x4567);
289 c5 = (float*) ((uintptr_t) c5 + cn_stride);
290 psimd_store_f32(c4, vacc4x0123);
291 psimd_store_f32(c4 + 4, vacc4x4567);
292 c4 = (float*) ((uintptr_t) c4 + cn_stride);
293 psimd_store_f32(c3, vacc3x0123);
294 psimd_store_f32(c3 + 4, vacc3x4567);
295 c3 = (float*) ((uintptr_t) c3 + cn_stride);
296 psimd_store_f32(c2, vacc2x0123);
297 psimd_store_f32(c2 + 4, vacc2x4567);
298 c2 = (float*) ((uintptr_t) c2 + cn_stride);
299 psimd_store_f32(c1, vacc1x0123);
300 psimd_store_f32(c1 + 4, vacc1x4567);
301 c1 = (float*) ((uintptr_t) c1 + cn_stride);
302 psimd_store_f32(c0, vacc0x0123);
303 psimd_store_f32(c0 + 4, vacc0x4567);
304 c0 = (float*) ((uintptr_t) c0 + cn_stride);
305
306 a = (const float**restrict) ((uintptr_t) a - ks);
307 nc -= 8;
308 } else {
309 if (nc & 4) {
310 psimd_store_f32(c5, vacc5x0123);
311 psimd_store_f32(c4, vacc4x0123);
312 psimd_store_f32(c3, vacc3x0123);
313 psimd_store_f32(c2, vacc2x0123);
314 psimd_store_f32(c1, vacc1x0123);
315 psimd_store_f32(c0, vacc0x0123);
316
317 vacc5x0123 = vacc5x4567;
318 vacc4x0123 = vacc4x4567;
319 vacc3x0123 = vacc3x4567;
320 vacc2x0123 = vacc2x4567;
321 vacc1x0123 = vacc1x4567;
322 vacc0x0123 = vacc0x4567;
323
324 c5 += 4;
325 c4 += 4;
326 c3 += 4;
327 c2 += 4;
328 c1 += 4;
329 c0 += 4;
330 }
331 if (nc & 2) {
332 psimd_store2_f32(c5, vacc5x0123);
333 psimd_store2_f32(c4, vacc4x0123);
334 psimd_store2_f32(c3, vacc3x0123);
335 psimd_store2_f32(c2, vacc2x0123);
336 psimd_store2_f32(c1, vacc1x0123);
337 psimd_store2_f32(c0, vacc0x0123);
338
339 vacc5x0123 = psimd_concat_hi_f32(vacc5x0123, vacc5x0123);
340 vacc4x0123 = psimd_concat_hi_f32(vacc4x0123, vacc4x0123);
341 vacc3x0123 = psimd_concat_hi_f32(vacc3x0123, vacc3x0123);
342 vacc2x0123 = psimd_concat_hi_f32(vacc2x0123, vacc2x0123);
343 vacc1x0123 = psimd_concat_hi_f32(vacc1x0123, vacc1x0123);
344 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
345
346 c5 += 2;
347 c4 += 2;
348 c3 += 2;
349 c2 += 2;
350 c1 += 2;
351 c0 += 2;
352 }
353 if (nc & 1) {
354 psimd_store1_f32(c5, vacc5x0123);
355 psimd_store1_f32(c4, vacc4x0123);
356 psimd_store1_f32(c3, vacc3x0123);
357 psimd_store1_f32(c2, vacc2x0123);
358 psimd_store1_f32(c1, vacc1x0123);
359 psimd_store1_f32(c0, vacc0x0123);
360 }
361
362 nc = 0;
363 }
364 } while (nc != 0);
365 }
366