1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/neon-ld128.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_6x8__neonfma_lane_ld128(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_6x8__neonfma_lane_ld128(
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 <= 6);
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
74 do {
75 float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
76 float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
77 float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
78 float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
79 float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
80 float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
81 float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
82 float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
83 float32x4_t vacc4x0123 = vld1q_f32(acc); acc += 4;
84 float32x4_t vacc4x4567 = vld1q_f32(acc); acc += 4;
85 float32x4_t vacc5x0123 = vld1q_f32(acc); acc += 4;
86 float32x4_t vacc5x4567 = vld1q_f32(acc); acc += 4;
87
88 size_t k = kc;
89 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
90 const float32x4_t va0 = vld1q_f32(a0); a0 += 4;
91 const float32x4_t va1 = vld1q_f32(a1); a1 += 4;
92 const float32x4_t va2 = vld1q_f32(a2); a2 += 4;
93 const float32x4_t va3 = vld1q_f32(a3); a3 += 4;
94 const float32x4_t va4 = vld1q_f32(a4); a4 += 4;
95 const float32x4_t va5 = vld1q_f32(a5); a5 += 4;
96
97
98 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
99 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
100
101 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);
102 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);
103 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);
104 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);
105 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, vget_low_f32(va4), 0);
106 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c0, vget_low_f32(va5), 0);
107 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);
108 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);
109 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);
110 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);
111 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, vget_low_f32(va4), 0);
112 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c0, vget_low_f32(va5), 0);
113
114 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
115 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
116
117 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);
118 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);
119 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);
120 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);
121 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, vget_low_f32(va4), 1);
122 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c1, vget_low_f32(va5), 1);
123 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);
124 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);
125 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);
126 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);
127 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c1, vget_low_f32(va4), 1);
128 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c1, vget_low_f32(va5), 1);
129
130 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
131 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
132
133 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c2, vget_high_f32(va0), 0);
134 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c2, vget_high_f32(va1), 0);
135 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c2, vget_high_f32(va2), 0);
136 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c2, vget_high_f32(va3), 0);
137 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c2, vget_high_f32(va4), 0);
138 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c2, vget_high_f32(va5), 0);
139 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c2, vget_high_f32(va0), 0);
140 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c2, vget_high_f32(va1), 0);
141 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c2, vget_high_f32(va2), 0);
142 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c2, vget_high_f32(va3), 0);
143 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c2, vget_high_f32(va4), 0);
144 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c2, vget_high_f32(va5), 0);
145
146 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
147 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
148
149 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c3, vget_high_f32(va0), 1);
150 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c3, vget_high_f32(va1), 1);
151 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c3, vget_high_f32(va2), 1);
152 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c3, vget_high_f32(va3), 1);
153 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c3, vget_high_f32(va4), 1);
154 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c3, vget_high_f32(va5), 1);
155 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c3, vget_high_f32(va0), 1);
156 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c3, vget_high_f32(va1), 1);
157 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c3, vget_high_f32(va2), 1);
158 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c3, vget_high_f32(va3), 1);
159 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c3, vget_high_f32(va4), 1);
160 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c3, vget_high_f32(va5), 1);
161 }
162 if XNN_UNLIKELY(k != 0) {
163 do {
164 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
165 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
166 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
167 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
168 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
169 const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
170
171 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
172 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
173
174 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
175 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
176 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
177 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
178 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123);
179 vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123);
180 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
181 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
182 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
183 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
184 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567);
185 vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567);
186
187 k -= sizeof(float);
188 } while (k != 0);
189 }
190 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
191 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
192 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
193 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
194 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
195 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
196 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
197 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
198 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
199 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
200 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
201 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
202 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
203
204 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
205 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
206 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
207 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
208 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
209 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
210 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
211 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
212 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
213 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
214 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
215 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
216 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
217
218 if XNN_LIKELY(nc >= 8) {
219 vst1q_f32(c5, vacc5x0123);
220 vst1q_f32(c5 + 4, vacc5x4567);
221 c5 = (float*) ((uintptr_t) c5 + cn_stride);
222 vst1q_f32(c4, vacc4x0123);
223 vst1q_f32(c4 + 4, vacc4x4567);
224 c4 = (float*) ((uintptr_t) c4 + cn_stride);
225 vst1q_f32(c3, vacc3x0123);
226 vst1q_f32(c3 + 4, vacc3x4567);
227 c3 = (float*) ((uintptr_t) c3 + cn_stride);
228 vst1q_f32(c2, vacc2x0123);
229 vst1q_f32(c2 + 4, vacc2x4567);
230 c2 = (float*) ((uintptr_t) c2 + cn_stride);
231 vst1q_f32(c1, vacc1x0123);
232 vst1q_f32(c1 + 4, vacc1x4567);
233 c1 = (float*) ((uintptr_t) c1 + cn_stride);
234 vst1q_f32(c0, vacc0x0123);
235 vst1q_f32(c0 + 4, vacc0x4567);
236 c0 = (float*) ((uintptr_t) c0 + cn_stride);
237
238 a5 = (const float*) ((uintptr_t) a5 - kc);
239 a4 = (const float*) ((uintptr_t) a4 - kc);
240 a3 = (const float*) ((uintptr_t) a3 - kc);
241 a2 = (const float*) ((uintptr_t) a2 - kc);
242 a1 = (const float*) ((uintptr_t) a1 - kc);
243 a0 = (const float*) ((uintptr_t) a0 - kc);
244
245 nc -= 8;
246
247 } else {
248 if (nc & 4) {
249 vst1q_f32(c5, vacc5x0123); c5 += 4;
250 vst1q_f32(c4, vacc4x0123); c4 += 4;
251 vst1q_f32(c3, vacc3x0123); c3 += 4;
252 vst1q_f32(c2, vacc2x0123); c2 += 4;
253 vst1q_f32(c1, vacc1x0123); c1 += 4;
254 vst1q_f32(c0, vacc0x0123); c0 += 4;
255
256 vacc5x0123 = vacc5x4567;
257 vacc4x0123 = vacc4x4567;
258 vacc3x0123 = vacc3x4567;
259 vacc2x0123 = vacc2x4567;
260 vacc1x0123 = vacc1x4567;
261 vacc0x0123 = vacc0x4567;
262 }
263 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
264 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
265 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
266 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
267 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
268 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
269 if (nc & 2) {
270 vst1_f32(c5, vacc5x01); c5 += 2;
271 vst1_f32(c4, vacc4x01); c4 += 2;
272 vst1_f32(c3, vacc3x01); c3 += 2;
273 vst1_f32(c2, vacc2x01); c2 += 2;
274 vst1_f32(c1, vacc1x01); c1 += 2;
275 vst1_f32(c0, vacc0x01); c0 += 2;
276
277 vacc5x01 = vget_high_f32(vacc5x0123);
278 vacc4x01 = vget_high_f32(vacc4x0123);
279 vacc3x01 = vget_high_f32(vacc3x0123);
280 vacc2x01 = vget_high_f32(vacc2x0123);
281 vacc1x01 = vget_high_f32(vacc1x0123);
282 vacc0x01 = vget_high_f32(vacc0x0123);
283 }
284 if (nc & 1) {
285 vst1_lane_f32(c5, vacc5x01, 0);
286 vst1_lane_f32(c4, vacc4x01, 0);
287 vst1_lane_f32(c3, vacc3x01, 0);
288 vst1_lane_f32(c2, vacc2x01, 0);
289 vst1_lane_f32(c1, vacc1x01, 0);
290 vst1_lane_f32(c0, vacc0x01, 0);
291 }
292
293 nc = 0;
294 }
295 } while (nc != 0);
296 }
297