1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemminc_ukernel_6x8__psimd_splat(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const float * restrict acc,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_gemminc_ukernel_6x8__psimd_splat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const float*restrict acc,
28 const union xnn_f32_output_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
67 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr != 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72
73 do {
74 psimd_f32 vacc0x0123 = psimd_load_f32(acc + 0);
75 psimd_f32 vacc0x4567 = psimd_load_f32(acc + 4);
76 psimd_f32 vacc1x0123 = psimd_load_f32(acc + 8);
77 psimd_f32 vacc1x4567 = psimd_load_f32(acc + 12);
78 psimd_f32 vacc2x0123 = psimd_load_f32(acc + 16);
79 psimd_f32 vacc2x4567 = psimd_load_f32(acc + 20);
80 psimd_f32 vacc3x0123 = psimd_load_f32(acc + 24);
81 psimd_f32 vacc3x4567 = psimd_load_f32(acc + 28);
82 psimd_f32 vacc4x0123 = psimd_load_f32(acc + 32);
83 psimd_f32 vacc4x4567 = psimd_load_f32(acc + 36);
84 psimd_f32 vacc5x0123 = psimd_load_f32(acc + 40);
85 psimd_f32 vacc5x4567 = psimd_load_f32(acc + 44);
86 acc += 48;
87
88 size_t k = kc;
89 while (k >= 4 * sizeof(float)) {
90 const psimd_f32 va0 = psimd_load_f32(a0);
91 a0 += 4;
92 const psimd_f32 va1 = psimd_load_f32(a1);
93 a1 += 4;
94 const psimd_f32 va2 = psimd_load_f32(a2);
95 a2 += 4;
96 const psimd_f32 va3 = psimd_load_f32(a3);
97 a3 += 4;
98 const psimd_f32 va4 = psimd_load_f32(a4);
99 a4 += 4;
100 const psimd_f32 va5 = psimd_load_f32(a5);
101 a5 += 4;
102
103 const psimd_f32 va0c0 = psimd_splat0_f32(va0);
104 const psimd_f32 va1c0 = psimd_splat0_f32(va1);
105 const psimd_f32 va2c0 = psimd_splat0_f32(va2);
106 const psimd_f32 va3c0 = psimd_splat0_f32(va3);
107 const psimd_f32 va4c0 = psimd_splat0_f32(va4);
108 const psimd_f32 va5c0 = psimd_splat0_f32(va5);
109
110 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
111 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
112
113 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c0, vb0123c0);
114 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c0, vb0123c0);
115 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0);
116 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c0, vb0123c0);
117 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c0, vb0123c0);
118 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c0, vb0123c0);
119 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c0, vb4567c0);
120 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c0, vb4567c0);
121 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0);
122 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c0, vb4567c0);
123 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c0, vb4567c0);
124 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c0, vb4567c0);
125 const psimd_f32 va0c1 = psimd_splat1_f32(va0);
126 const psimd_f32 va1c1 = psimd_splat1_f32(va1);
127 const psimd_f32 va2c1 = psimd_splat1_f32(va2);
128 const psimd_f32 va3c1 = psimd_splat1_f32(va3);
129 const psimd_f32 va4c1 = psimd_splat1_f32(va4);
130 const psimd_f32 va5c1 = psimd_splat1_f32(va5);
131
132 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
133 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
134
135 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c1, vb0123c1);
136 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1);
137 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1);
138 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c1, vb0123c1);
139 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c1, vb0123c1);
140 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c1, vb0123c1);
141 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c1, vb4567c1);
142 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1);
143 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1);
144 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c1, vb4567c1);
145 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c1, vb4567c1);
146 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c1, vb4567c1);
147 const psimd_f32 va0c2 = psimd_splat2_f32(va0);
148 const psimd_f32 va1c2 = psimd_splat2_f32(va1);
149 const psimd_f32 va2c2 = psimd_splat2_f32(va2);
150 const psimd_f32 va3c2 = psimd_splat2_f32(va3);
151 const psimd_f32 va4c2 = psimd_splat2_f32(va4);
152 const psimd_f32 va5c2 = psimd_splat2_f32(va5);
153
154 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
155 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
156
157 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c2, vb0123c2);
158 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c2, vb0123c2);
159 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c2, vb0123c2);
160 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c2, vb0123c2);
161 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c2, vb0123c2);
162 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c2, vb0123c2);
163 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c2, vb4567c2);
164 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c2, vb4567c2);
165 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c2, vb4567c2);
166 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c2, vb4567c2);
167 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c2, vb4567c2);
168 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c2, vb4567c2);
169 const psimd_f32 va0c3 = psimd_splat3_f32(va0);
170 const psimd_f32 va1c3 = psimd_splat3_f32(va1);
171 const psimd_f32 va2c3 = psimd_splat3_f32(va2);
172 const psimd_f32 va3c3 = psimd_splat3_f32(va3);
173 const psimd_f32 va4c3 = psimd_splat3_f32(va4);
174 const psimd_f32 va5c3 = psimd_splat3_f32(va5);
175
176 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
177 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
178
179 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c3, vb0123c3);
180 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c3, vb0123c3);
181 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c3, vb0123c3);
182 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c3, vb0123c3);
183 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4c3, vb0123c3);
184 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5c3, vb0123c3);
185 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c3, vb4567c3);
186 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c3, vb4567c3);
187 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c3, vb4567c3);
188 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c3, vb4567c3);
189 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4c3, vb4567c3);
190 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5c3, vb4567c3);
191
192 w += 32;
193 k -= 4 * sizeof(float);
194 }
195 if XNN_UNLIKELY(k != 0) {
196 do {
197 const psimd_f32 va0 = psimd_load_splat_f32(a0);
198 a0 += 1;
199 const psimd_f32 va1 = psimd_load_splat_f32(a1);
200 a1 += 1;
201 const psimd_f32 va2 = psimd_load_splat_f32(a2);
202 a2 += 1;
203 const psimd_f32 va3 = psimd_load_splat_f32(a3);
204 a3 += 1;
205 const psimd_f32 va4 = psimd_load_splat_f32(a4);
206 a4 += 1;
207 const psimd_f32 va5 = psimd_load_splat_f32(a5);
208 a5 += 1;
209
210 const psimd_f32 vb0123 = psimd_load_f32(w);
211 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
212 w += 8;
213
214 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
215 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123);
216 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);
217 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123);
218 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4, vb0123);
219 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123);
220 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
221 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567);
222 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);
223 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567);
224 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4, vb4567);
225 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567);
226
227 k -= sizeof(float);
228 } while (k != 0);
229 }
230
231 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
232 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
233 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
234 vacc2x0123 = psimd_min_f32(vacc2x0123, vmax);
235 vacc3x0123 = psimd_min_f32(vacc3x0123, vmax);
236 vacc4x0123 = psimd_min_f32(vacc4x0123, vmax);
237 vacc5x0123 = psimd_min_f32(vacc5x0123, vmax);
238 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
239 vacc1x4567 = psimd_min_f32(vacc1x4567, vmax);
240 vacc2x4567 = psimd_min_f32(vacc2x4567, vmax);
241 vacc3x4567 = psimd_min_f32(vacc3x4567, vmax);
242 vacc4x4567 = psimd_min_f32(vacc4x4567, vmax);
243 vacc5x4567 = psimd_min_f32(vacc5x4567, vmax);
244
245 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
246 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
247 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
248 vacc2x0123 = psimd_max_f32(vacc2x0123, vmin);
249 vacc3x0123 = psimd_max_f32(vacc3x0123, vmin);
250 vacc4x0123 = psimd_max_f32(vacc4x0123, vmin);
251 vacc5x0123 = psimd_max_f32(vacc5x0123, vmin);
252 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
253 vacc1x4567 = psimd_max_f32(vacc1x4567, vmin);
254 vacc2x4567 = psimd_max_f32(vacc2x4567, vmin);
255 vacc3x4567 = psimd_max_f32(vacc3x4567, vmin);
256 vacc4x4567 = psimd_max_f32(vacc4x4567, vmin);
257 vacc5x4567 = psimd_max_f32(vacc5x4567, vmin);
258
259 if XNN_LIKELY(nc >= 8) {
260 psimd_store_f32(c5, vacc5x0123);
261 psimd_store_f32(c5 + 4, vacc5x4567);
262 c5 = (float*) ((uintptr_t) c5 + cn_stride);
263 psimd_store_f32(c4, vacc4x0123);
264 psimd_store_f32(c4 + 4, vacc4x4567);
265 c4 = (float*) ((uintptr_t) c4 + cn_stride);
266 psimd_store_f32(c3, vacc3x0123);
267 psimd_store_f32(c3 + 4, vacc3x4567);
268 c3 = (float*) ((uintptr_t) c3 + cn_stride);
269 psimd_store_f32(c2, vacc2x0123);
270 psimd_store_f32(c2 + 4, vacc2x4567);
271 c2 = (float*) ((uintptr_t) c2 + cn_stride);
272 psimd_store_f32(c1, vacc1x0123);
273 psimd_store_f32(c1 + 4, vacc1x4567);
274 c1 = (float*) ((uintptr_t) c1 + cn_stride);
275 psimd_store_f32(c0, vacc0x0123);
276 psimd_store_f32(c0 + 4, vacc0x4567);
277 c0 = (float*) ((uintptr_t) c0 + cn_stride);
278
279 a5 = (const float*) ((uintptr_t) a5 - kc);
280 a4 = (const float*) ((uintptr_t) a4 - kc);
281 a3 = (const float*) ((uintptr_t) a3 - kc);
282 a2 = (const float*) ((uintptr_t) a2 - kc);
283 a1 = (const float*) ((uintptr_t) a1 - kc);
284 a0 = (const float*) ((uintptr_t) a0 - kc);
285
286 nc -= 8;
287 } else {
288 if (nc & 4) {
289 psimd_store_f32(c5, vacc5x0123);
290 psimd_store_f32(c4, vacc4x0123);
291 psimd_store_f32(c3, vacc3x0123);
292 psimd_store_f32(c2, vacc2x0123);
293 psimd_store_f32(c1, vacc1x0123);
294 psimd_store_f32(c0, vacc0x0123);
295
296 vacc5x0123 = vacc5x4567;
297 vacc4x0123 = vacc4x4567;
298 vacc3x0123 = vacc3x4567;
299 vacc2x0123 = vacc2x4567;
300 vacc1x0123 = vacc1x4567;
301 vacc0x0123 = vacc0x4567;
302
303 c5 += 4;
304 c4 += 4;
305 c3 += 4;
306 c2 += 4;
307 c1 += 4;
308 c0 += 4;
309 }
310 if (nc & 2) {
311 psimd_store2_f32(c5, vacc5x0123);
312 psimd_store2_f32(c4, vacc4x0123);
313 psimd_store2_f32(c3, vacc3x0123);
314 psimd_store2_f32(c2, vacc2x0123);
315 psimd_store2_f32(c1, vacc1x0123);
316 psimd_store2_f32(c0, vacc0x0123);
317
318 vacc5x0123 = psimd_concat_hi_f32(vacc5x0123, vacc5x0123);
319 vacc4x0123 = psimd_concat_hi_f32(vacc4x0123, vacc4x0123);
320 vacc3x0123 = psimd_concat_hi_f32(vacc3x0123, vacc3x0123);
321 vacc2x0123 = psimd_concat_hi_f32(vacc2x0123, vacc2x0123);
322 vacc1x0123 = psimd_concat_hi_f32(vacc1x0123, vacc1x0123);
323 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
324
325 c5 += 2;
326 c4 += 2;
327 c3 += 2;
328 c2 += 2;
329 c1 += 2;
330 c0 += 2;
331 }
332 if (nc & 1) {
333 psimd_store1_f32(c5, vacc5x0123);
334 psimd_store1_f32(c4, vacc4x0123);
335 psimd_store1_f32(c3, vacc3x0123);
336 psimd_store1_f32(c2, vacc2x0123);
337 psimd_store1_f32(c1, vacc1x0123);
338 psimd_store1_f32(c0, vacc0x0123);
339 }
340
341 nc = 0;
342 }
343 } while (nc != 0);
344 }
345