1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/neon-shuffle.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include <xnnpack/gemm.h>
16
17
xnn_f32_gemminc_minmax_ukernel_8x8s4__neon(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemminc_minmax_ukernel_8x8s4__neon(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const float* restrict a,
23 size_t a_stride,
24 const float* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const float*restrict acc,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 8);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39 assert(acc != NULL);
40
41 const float* a0 = a;
42 float* c0 = c;
43 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
62 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
68 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr < 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73 const float* a6 = (const float*) ((uintptr_t) a5 + a_stride);
74 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 6) {
76 a6 = a5;
77 c6 = c5;
78 }
79 const float* a7 = (const float*) ((uintptr_t) a6 + a_stride);
80 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
81 if XNN_UNPREDICTABLE(mr != 8) {
82 a7 = a6;
83 c7 = c6;
84 }
85
86 do {
87 float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
88 float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
89 float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
90 float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
91 float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
92 float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
93 float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
94 float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
95 float32x4_t vacc4x0123 = vld1q_f32(acc); acc += 4;
96 float32x4_t vacc4x4567 = vld1q_f32(acc); acc += 4;
97 float32x4_t vacc5x0123 = vld1q_f32(acc); acc += 4;
98 float32x4_t vacc5x4567 = vld1q_f32(acc); acc += 4;
99 float32x4_t vacc6x0123 = vld1q_f32(acc); acc += 4;
100 float32x4_t vacc6x4567 = vld1q_f32(acc); acc += 4;
101 float32x4_t vacc7x0123 = vld1q_f32(acc); acc += 4;
102 float32x4_t vacc7x4567 = vld1q_f32(acc); acc += 4;
103
104 size_t k = kc;
105 while (k >= 4 * sizeof(float)) {
106 float32x4_t va0 = vld1q_f32(a0); a0 += 4;
107 float32x4_t va1 = vld1q_f32(a1); a1 += 4;
108 float32x4_t va2 = vld1q_f32(a2); a2 += 4;
109 float32x4_t va3 = vld1q_f32(a3); a3 += 4;
110 float32x4_t va4 = vld1q_f32(a4); a4 += 4;
111 float32x4_t va5 = vld1q_f32(a5); a5 += 4;
112 float32x4_t va6 = vld1q_f32(a6); a6 += 4;
113 float32x4_t va7 = vld1q_f32(a7); a7 += 4;
114
115
116 const float32x4_t vb0123c0 = vld1q_f32(w + 0);
117 const float32x4_t vb4567c0 = vld1q_f32(w + 4);
118
119 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
120 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);
121 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);
122 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);
123 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c0);
124 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c0);
125 vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c0);
126 vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c0);
127 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
128 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);
129 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);
130 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);
131 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c0);
132 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c0);
133 vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c0);
134 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c0);
135
136 va0 = vextq_f32(va0, va0, 1);
137 va1 = vextq_f32(va1, va1, 1);
138 va2 = vextq_f32(va2, va2, 1);
139 va3 = vextq_f32(va3, va3, 1);
140 va4 = vextq_f32(va4, va4, 1);
141 va5 = vextq_f32(va5, va5, 1);
142 va6 = vextq_f32(va6, va6, 1);
143 va7 = vextq_f32(va7, va7, 1);
144
145 const float32x4_t vb0123c1 = vld1q_f32(w + 8);
146 const float32x4_t vb4567c1 = vld1q_f32(w + 12);
147
148 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
149 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);
150 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);
151 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);
152 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c1);
153 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c1);
154 vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c1);
155 vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c1);
156 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
157 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);
158 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);
159 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);
160 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c1);
161 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c1);
162 vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c1);
163 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c1);
164
165 va0 = vextq_f32(va0, va0, 1);
166 va1 = vextq_f32(va1, va1, 1);
167 va2 = vextq_f32(va2, va2, 1);
168 va3 = vextq_f32(va3, va3, 1);
169 va4 = vextq_f32(va4, va4, 1);
170 va5 = vextq_f32(va5, va5, 1);
171 va6 = vextq_f32(va6, va6, 1);
172 va7 = vextq_f32(va7, va7, 1);
173
174 const float32x4_t vb0123c2 = vld1q_f32(w + 16);
175 const float32x4_t vb4567c2 = vld1q_f32(w + 20);
176
177 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
178 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);
179 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);
180 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);
181 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c2);
182 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c2);
183 vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c2);
184 vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c2);
185 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
186 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);
187 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);
188 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);
189 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c2);
190 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c2);
191 vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c2);
192 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c2);
193
194 va0 = vextq_f32(va0, va0, 1);
195 va1 = vextq_f32(va1, va1, 1);
196 va2 = vextq_f32(va2, va2, 1);
197 va3 = vextq_f32(va3, va3, 1);
198 va4 = vextq_f32(va4, va4, 1);
199 va5 = vextq_f32(va5, va5, 1);
200 va6 = vextq_f32(va6, va6, 1);
201 va7 = vextq_f32(va7, va7, 1);
202
203 const float32x4_t vb0123c3 = vld1q_f32(w + 24);
204 const float32x4_t vb4567c3 = vld1q_f32(w + 28);
205
206 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
207 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);
208 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);
209 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);
210 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c3);
211 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c3);
212 vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c3);
213 vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c3);
214 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
215 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);
216 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);
217 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);
218 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c3);
219 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c3);
220 vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c3);
221 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c3);
222
223
224 w += 32;
225 k -= 4 * sizeof(float);
226 }
227 if XNN_UNLIKELY(k != 0) {
228 do {
229 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
230 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
231 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
232 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
233 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
234 const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
235 const float32x4_t va6 = vld1q_dup_f32(a6); a6 += 1;
236 const float32x4_t va7 = vld1q_dup_f32(a7); a7 += 1;
237
238 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
239 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
240
241 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
242 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
243 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
244 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
245 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123);
246 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123);
247 vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123);
248 vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123);
249 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
250 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
251 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
252 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
253 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567);
254 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567);
255 vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567);
256 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567);
257
258 k -= sizeof(float);
259 } while (k != 0);
260 }
261 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
262 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
263 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
264 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
265 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
266 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
267 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
268 vacc6x0123 = vminq_f32(vacc6x0123, vmax);
269 vacc7x0123 = vminq_f32(vacc7x0123, vmax);
270 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
271 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
272 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
273 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
274 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
275 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
276 vacc6x4567 = vminq_f32(vacc6x4567, vmax);
277 vacc7x4567 = vminq_f32(vacc7x4567, vmax);
278
279 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
280 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
281 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
282 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
283 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
284 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
285 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
286 vacc6x0123 = vmaxq_f32(vacc6x0123, vmin);
287 vacc7x0123 = vmaxq_f32(vacc7x0123, vmin);
288 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
289 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
290 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
291 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
292 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
293 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
294 vacc6x4567 = vmaxq_f32(vacc6x4567, vmin);
295 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin);
296
297 if XNN_LIKELY(nc >= 8) {
298 vst1q_f32(c7, vacc7x0123);
299 vst1q_f32(c7 + 4, vacc7x4567);
300 c7 = (float*) ((uintptr_t) c7 + cn_stride);
301 vst1q_f32(c6, vacc6x0123);
302 vst1q_f32(c6 + 4, vacc6x4567);
303 c6 = (float*) ((uintptr_t) c6 + cn_stride);
304 vst1q_f32(c5, vacc5x0123);
305 vst1q_f32(c5 + 4, vacc5x4567);
306 c5 = (float*) ((uintptr_t) c5 + cn_stride);
307 vst1q_f32(c4, vacc4x0123);
308 vst1q_f32(c4 + 4, vacc4x4567);
309 c4 = (float*) ((uintptr_t) c4 + cn_stride);
310 vst1q_f32(c3, vacc3x0123);
311 vst1q_f32(c3 + 4, vacc3x4567);
312 c3 = (float*) ((uintptr_t) c3 + cn_stride);
313 vst1q_f32(c2, vacc2x0123);
314 vst1q_f32(c2 + 4, vacc2x4567);
315 c2 = (float*) ((uintptr_t) c2 + cn_stride);
316 vst1q_f32(c1, vacc1x0123);
317 vst1q_f32(c1 + 4, vacc1x4567);
318 c1 = (float*) ((uintptr_t) c1 + cn_stride);
319 vst1q_f32(c0, vacc0x0123);
320 vst1q_f32(c0 + 4, vacc0x4567);
321 c0 = (float*) ((uintptr_t) c0 + cn_stride);
322
323 a7 = (const float*) ((uintptr_t) a7 - kc);
324 a6 = (const float*) ((uintptr_t) a6 - kc);
325 a5 = (const float*) ((uintptr_t) a5 - kc);
326 a4 = (const float*) ((uintptr_t) a4 - kc);
327 a3 = (const float*) ((uintptr_t) a3 - kc);
328 a2 = (const float*) ((uintptr_t) a2 - kc);
329 a1 = (const float*) ((uintptr_t) a1 - kc);
330 a0 = (const float*) ((uintptr_t) a0 - kc);
331
332 nc -= 8;
333
334 } else {
335 if (nc & 4) {
336 vst1q_f32(c7, vacc7x0123); c7 += 4;
337 vst1q_f32(c6, vacc6x0123); c6 += 4;
338 vst1q_f32(c5, vacc5x0123); c5 += 4;
339 vst1q_f32(c4, vacc4x0123); c4 += 4;
340 vst1q_f32(c3, vacc3x0123); c3 += 4;
341 vst1q_f32(c2, vacc2x0123); c2 += 4;
342 vst1q_f32(c1, vacc1x0123); c1 += 4;
343 vst1q_f32(c0, vacc0x0123); c0 += 4;
344
345 vacc7x0123 = vacc7x4567;
346 vacc6x0123 = vacc6x4567;
347 vacc5x0123 = vacc5x4567;
348 vacc4x0123 = vacc4x4567;
349 vacc3x0123 = vacc3x4567;
350 vacc2x0123 = vacc2x4567;
351 vacc1x0123 = vacc1x4567;
352 vacc0x0123 = vacc0x4567;
353 }
354 float32x2_t vacc7x01 = vget_low_f32(vacc7x0123);
355 float32x2_t vacc6x01 = vget_low_f32(vacc6x0123);
356 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
357 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
358 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
359 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
360 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
361 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
362 if (nc & 2) {
363 vst1_f32(c7, vacc7x01); c7 += 2;
364 vst1_f32(c6, vacc6x01); c6 += 2;
365 vst1_f32(c5, vacc5x01); c5 += 2;
366 vst1_f32(c4, vacc4x01); c4 += 2;
367 vst1_f32(c3, vacc3x01); c3 += 2;
368 vst1_f32(c2, vacc2x01); c2 += 2;
369 vst1_f32(c1, vacc1x01); c1 += 2;
370 vst1_f32(c0, vacc0x01); c0 += 2;
371
372 vacc7x01 = vget_high_f32(vacc7x0123);
373 vacc6x01 = vget_high_f32(vacc6x0123);
374 vacc5x01 = vget_high_f32(vacc5x0123);
375 vacc4x01 = vget_high_f32(vacc4x0123);
376 vacc3x01 = vget_high_f32(vacc3x0123);
377 vacc2x01 = vget_high_f32(vacc2x0123);
378 vacc1x01 = vget_high_f32(vacc1x0123);
379 vacc0x01 = vget_high_f32(vacc0x0123);
380 }
381 if (nc & 1) {
382 vst1_lane_f32(c7, vacc7x01, 0);
383 vst1_lane_f32(c6, vacc6x01, 0);
384 vst1_lane_f32(c5, vacc5x01, 0);
385 vst1_lane_f32(c4, vacc4x01, 0);
386 vst1_lane_f32(c3, vacc3x01, 0);
387 vst1_lane_f32(c2, vacc2x01, 0);
388 vst1_lane_f32(c1, vacc1x01, 0);
389 vst1_lane_f32(c0, vacc0x01, 0);
390 }
391
392 nc = 0;
393 }
394 } while (nc != 0);
395 }
396