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_gemm_minmax_ukernel_6x8s4__neonfma(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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma(
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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(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
39 const float* a0 = a;
40 float* c0 = c;
41 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
42 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
54 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 a4 = a3;
63 c4 = c3;
64 }
65 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
66 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
67 if XNN_UNPREDICTABLE(mr != 6) {
68 a5 = a4;
69 c5 = c4;
70 }
71
72 do {
73 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
74 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
75 float32x4_t vacc1x0123 = vacc0x0123;
76 float32x4_t vacc1x4567 = vacc0x4567;
77 float32x4_t vacc2x0123 = vacc0x0123;
78 float32x4_t vacc2x4567 = vacc0x4567;
79 float32x4_t vacc3x0123 = vacc0x0123;
80 float32x4_t vacc3x4567 = vacc0x4567;
81 float32x4_t vacc4x0123 = vacc0x0123;
82 float32x4_t vacc4x4567 = vacc0x4567;
83 float32x4_t vacc5x0123 = vacc0x0123;
84 float32x4_t vacc5x4567 = vacc0x4567;
85
86 size_t k = kc;
87 while (k >= 4 * sizeof(float)) {
88 float32x4_t va0 = vld1q_f32(a0); a0 += 4;
89 float32x4_t va1 = vld1q_f32(a1); a1 += 4;
90 float32x4_t va2 = vld1q_f32(a2); a2 += 4;
91 float32x4_t va3 = vld1q_f32(a3); a3 += 4;
92 float32x4_t va4 = vld1q_f32(a4); a4 += 4;
93 float32x4_t va5 = vld1q_f32(a5); a5 += 4;
94
95
96 const float32x4_t vb0123c0 = vld1q_f32(w + 0);
97 const float32x4_t vb4567c0 = vld1q_f32(w + 4);
98
99 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c0);
100 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c0);
101 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c0);
102 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c0);
103 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c0);
104 vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c0);
105 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c0);
106 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c0);
107 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c0);
108 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c0);
109 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c0);
110 vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c0);
111
112 va0 = vextq_f32(va0, va0, 1);
113 va1 = vextq_f32(va1, va1, 1);
114 va2 = vextq_f32(va2, va2, 1);
115 va3 = vextq_f32(va3, va3, 1);
116 va4 = vextq_f32(va4, va4, 1);
117 va5 = vextq_f32(va5, va5, 1);
118
119 const float32x4_t vb0123c1 = vld1q_f32(w + 8);
120 const float32x4_t vb4567c1 = vld1q_f32(w + 12);
121
122 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c1);
123 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c1);
124 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c1);
125 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c1);
126 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c1);
127 vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c1);
128 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c1);
129 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c1);
130 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c1);
131 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c1);
132 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c1);
133 vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c1);
134
135 va0 = vextq_f32(va0, va0, 1);
136 va1 = vextq_f32(va1, va1, 1);
137 va2 = vextq_f32(va2, va2, 1);
138 va3 = vextq_f32(va3, va3, 1);
139 va4 = vextq_f32(va4, va4, 1);
140 va5 = vextq_f32(va5, va5, 1);
141
142 const float32x4_t vb0123c2 = vld1q_f32(w + 16);
143 const float32x4_t vb4567c2 = vld1q_f32(w + 20);
144
145 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c2);
146 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c2);
147 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c2);
148 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c2);
149 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c2);
150 vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c2);
151 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c2);
152 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c2);
153 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c2);
154 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c2);
155 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c2);
156 vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c2);
157
158 va0 = vextq_f32(va0, va0, 1);
159 va1 = vextq_f32(va1, va1, 1);
160 va2 = vextq_f32(va2, va2, 1);
161 va3 = vextq_f32(va3, va3, 1);
162 va4 = vextq_f32(va4, va4, 1);
163 va5 = vextq_f32(va5, va5, 1);
164
165 const float32x4_t vb0123c3 = vld1q_f32(w + 24);
166 const float32x4_t vb4567c3 = vld1q_f32(w + 28);
167
168 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c3);
169 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c3);
170 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c3);
171 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c3);
172 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c3);
173 vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c3);
174 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c3);
175 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c3);
176 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c3);
177 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c3);
178 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c3);
179 vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c3);
180
181
182 w += 32;
183 k -= 4 * sizeof(float);
184 }
185 if XNN_UNLIKELY(k != 0) {
186 do {
187 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
188 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
189 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
190 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
191 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
192 const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
193
194 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
195 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
196
197 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
198 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
199 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
200 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
201 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123);
202 vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123);
203 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
204 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
205 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
206 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
207 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567);
208 vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567);
209
210 k -= sizeof(float);
211 } while (k != 0);
212 }
213 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
214 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
215 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
216 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
217 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
218 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
219 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
220 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
221 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
222 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
223 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
224 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
225 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
226
227 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
228 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
229 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
230 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
231 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
232 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
233 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
234 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
235 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
236 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
237 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
238 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
239 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
240
241 if XNN_LIKELY(nc >= 8) {
242 vst1q_f32(c5, vacc5x0123);
243 vst1q_f32(c5 + 4, vacc5x4567);
244 c5 = (float*) ((uintptr_t) c5 + cn_stride);
245 vst1q_f32(c4, vacc4x0123);
246 vst1q_f32(c4 + 4, vacc4x4567);
247 c4 = (float*) ((uintptr_t) c4 + cn_stride);
248 vst1q_f32(c3, vacc3x0123);
249 vst1q_f32(c3 + 4, vacc3x4567);
250 c3 = (float*) ((uintptr_t) c3 + cn_stride);
251 vst1q_f32(c2, vacc2x0123);
252 vst1q_f32(c2 + 4, vacc2x4567);
253 c2 = (float*) ((uintptr_t) c2 + cn_stride);
254 vst1q_f32(c1, vacc1x0123);
255 vst1q_f32(c1 + 4, vacc1x4567);
256 c1 = (float*) ((uintptr_t) c1 + cn_stride);
257 vst1q_f32(c0, vacc0x0123);
258 vst1q_f32(c0 + 4, vacc0x4567);
259 c0 = (float*) ((uintptr_t) c0 + cn_stride);
260
261 a5 = (const float*) ((uintptr_t) a5 - kc);
262 a4 = (const float*) ((uintptr_t) a4 - kc);
263 a3 = (const float*) ((uintptr_t) a3 - kc);
264 a2 = (const float*) ((uintptr_t) a2 - kc);
265 a1 = (const float*) ((uintptr_t) a1 - kc);
266 a0 = (const float*) ((uintptr_t) a0 - kc);
267
268 nc -= 8;
269
270 } else {
271 if (nc & 4) {
272 vst1q_f32(c5, vacc5x0123); c5 += 4;
273 vst1q_f32(c4, vacc4x0123); c4 += 4;
274 vst1q_f32(c3, vacc3x0123); c3 += 4;
275 vst1q_f32(c2, vacc2x0123); c2 += 4;
276 vst1q_f32(c1, vacc1x0123); c1 += 4;
277 vst1q_f32(c0, vacc0x0123); c0 += 4;
278
279 vacc5x0123 = vacc5x4567;
280 vacc4x0123 = vacc4x4567;
281 vacc3x0123 = vacc3x4567;
282 vacc2x0123 = vacc2x4567;
283 vacc1x0123 = vacc1x4567;
284 vacc0x0123 = vacc0x4567;
285 }
286 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
287 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
288 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
289 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
290 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
291 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
292 if (nc & 2) {
293 vst1_f32(c5, vacc5x01); c5 += 2;
294 vst1_f32(c4, vacc4x01); c4 += 2;
295 vst1_f32(c3, vacc3x01); c3 += 2;
296 vst1_f32(c2, vacc2x01); c2 += 2;
297 vst1_f32(c1, vacc1x01); c1 += 2;
298 vst1_f32(c0, vacc0x01); c0 += 2;
299
300 vacc5x01 = vget_high_f32(vacc5x0123);
301 vacc4x01 = vget_high_f32(vacc4x0123);
302 vacc3x01 = vget_high_f32(vacc3x0123);
303 vacc2x01 = vget_high_f32(vacc2x0123);
304 vacc1x01 = vget_high_f32(vacc1x0123);
305 vacc0x01 = vget_high_f32(vacc0x0123);
306 }
307 if (nc & 1) {
308 vst1_lane_f32(c5, vacc5x01, 0);
309 vst1_lane_f32(c4, vacc4x01, 0);
310 vst1_lane_f32(c3, vacc3x01, 0);
311 vst1_lane_f32(c2, vacc2x01, 0);
312 vst1_lane_f32(c1, vacc1x01, 0);
313 vst1_lane_f32(c0, vacc0x01, 0);
314 }
315
316 nc = 0;
317 }
318 } while (nc != 0);
319 }
320