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_4x8s4__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_4x8s4__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 <= 4);
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
62 do {
63 float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
64 float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
65 float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
66 float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
67 float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
68 float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
69 float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
70 float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
71
72 size_t k = kc;
73 while (k >= 4 * sizeof(float)) {
74 float32x4_t va0 = vld1q_f32(a0); a0 += 4;
75 float32x4_t va1 = vld1q_f32(a1); a1 += 4;
76 float32x4_t va2 = vld1q_f32(a2); a2 += 4;
77 float32x4_t va3 = vld1q_f32(a3); a3 += 4;
78
79
80 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
81 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
82
83 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
84 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);
85 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);
86 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);
87 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
88 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);
89 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);
90 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);
91
92 va0 = vextq_f32(va0, va0, 1);
93 va1 = vextq_f32(va1, va1, 1);
94 va2 = vextq_f32(va2, va2, 1);
95 va3 = vextq_f32(va3, va3, 1);
96
97 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
98 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
99
100 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
101 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);
102 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);
103 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);
104 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
105 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);
106 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);
107 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);
108
109 va0 = vextq_f32(va0, va0, 1);
110 va1 = vextq_f32(va1, va1, 1);
111 va2 = vextq_f32(va2, va2, 1);
112 va3 = vextq_f32(va3, va3, 1);
113
114 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
115 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
116
117 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
118 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);
119 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);
120 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);
121 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
122 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);
123 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);
124 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);
125
126 va0 = vextq_f32(va0, va0, 1);
127 va1 = vextq_f32(va1, va1, 1);
128 va2 = vextq_f32(va2, va2, 1);
129 va3 = vextq_f32(va3, va3, 1);
130
131 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
132 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
133
134 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
135 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);
136 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);
137 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);
138 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
139 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);
140 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);
141 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);
142
143
144 k -= 4 * sizeof(float);
145 }
146 if XNN_UNLIKELY(k != 0) {
147 float32x4_t va0 = vld1q_f32(a0); a0 = (const float*) ((uintptr_t) a0 + k);
148 float32x4_t va1 = vld1q_f32(a1); a1 = (const float*) ((uintptr_t) a1 + k);
149 float32x4_t va2 = vld1q_f32(a2); a2 = (const float*) ((uintptr_t) a2 + k);
150 float32x4_t va3 = vld1q_f32(a3); a3 = (const float*) ((uintptr_t) a3 + k);
151
152
153 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
154 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
155
156 const float32x4_t vmska0x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
157 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c0, vb0123c0);
158 const float32x4_t vmska1x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
159 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c0, vb0123c0);
160 const float32x4_t vmska2x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
161 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c0, vb0123c0);
162 const float32x4_t vmska3x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
163 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c0, vb0123c0);
164 const float32x4_t vmska0x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
165 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c0, vb4567c0);
166 const float32x4_t vmska1x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
167 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c0, vb4567c0);
168 const float32x4_t vmska2x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
169 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c0, vb4567c0);
170 const float32x4_t vmska3x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
171 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c0, vb4567c0);
172
173 va0 = vextq_f32(va0, va0, 1);
174 va1 = vextq_f32(va1, va1, 1);
175 va2 = vextq_f32(va2, va2, 1);
176 va3 = vextq_f32(va3, va3, 1);
177
178 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
179 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
180
181 const float32x4_t vmska0x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
182 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c1, vb0123c1);
183 const float32x4_t vmska1x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
184 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c1, vb0123c1);
185 const float32x4_t vmska2x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
186 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c1, vb0123c1);
187 const float32x4_t vmska3x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
188 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c1, vb0123c1);
189 const float32x4_t vmska0x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
190 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c1, vb4567c1);
191 const float32x4_t vmska1x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
192 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c1, vb4567c1);
193 const float32x4_t vmska2x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
194 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c1, vb4567c1);
195 const float32x4_t vmska3x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
196 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c1, vb4567c1);
197
198 va0 = vextq_f32(va0, va0, 1);
199 va1 = vextq_f32(va1, va1, 1);
200 va2 = vextq_f32(va2, va2, 1);
201 va3 = vextq_f32(va3, va3, 1);
202
203 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
204 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
205
206 const float32x4_t vmska0x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
207 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c2, vb0123c2);
208 const float32x4_t vmska1x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
209 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c2, vb0123c2);
210 const float32x4_t vmska2x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
211 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c2, vb0123c2);
212 const float32x4_t vmska3x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
213 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c2, vb0123c2);
214 const float32x4_t vmska0x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
215 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c2, vb4567c2);
216 const float32x4_t vmska1x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
217 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c2, vb4567c2);
218 const float32x4_t vmska2x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
219 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c2, vb4567c2);
220 const float32x4_t vmska3x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
221 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c2, vb4567c2);
222
223 va0 = vextq_f32(va0, va0, 1);
224 va1 = vextq_f32(va1, va1, 1);
225 va2 = vextq_f32(va2, va2, 1);
226 va3 = vextq_f32(va3, va3, 1);
227
228 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
229 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
230
231 const float32x4_t vmska0x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
232 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c3, vb0123c3);
233 const float32x4_t vmska1x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
234 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c3, vb0123c3);
235 const float32x4_t vmska2x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
236 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c3, vb0123c3);
237 const float32x4_t vmska3x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
238 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c3, vb0123c3);
239 const float32x4_t vmska0x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
240 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c3, vb4567c3);
241 const float32x4_t vmska1x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
242 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c3, vb4567c3);
243 const float32x4_t vmska2x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
244 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c3, vb4567c3);
245 const float32x4_t vmska3x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
246 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c3, vb4567c3);
247
248 }
249 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
250 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
251 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
252 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
253 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
254 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
255 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
256 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
257 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
258
259 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
260 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
261 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
262 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
263 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
264 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
265 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
266 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
267 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
268
269 if XNN_LIKELY(nc >= 8) {
270 vst1q_f32(c3, vacc3x0123);
271 vst1q_f32(c3 + 4, vacc3x4567);
272 c3 = (float*) ((uintptr_t) c3 + cn_stride);
273 vst1q_f32(c2, vacc2x0123);
274 vst1q_f32(c2 + 4, vacc2x4567);
275 c2 = (float*) ((uintptr_t) c2 + cn_stride);
276 vst1q_f32(c1, vacc1x0123);
277 vst1q_f32(c1 + 4, vacc1x4567);
278 c1 = (float*) ((uintptr_t) c1 + cn_stride);
279 vst1q_f32(c0, vacc0x0123);
280 vst1q_f32(c0 + 4, vacc0x4567);
281 c0 = (float*) ((uintptr_t) c0 + cn_stride);
282
283 a3 = (const float*) ((uintptr_t) a3 - kc);
284 a2 = (const float*) ((uintptr_t) a2 - kc);
285 a1 = (const float*) ((uintptr_t) a1 - kc);
286 a0 = (const float*) ((uintptr_t) a0 - kc);
287
288 nc -= 8;
289
290 } else {
291 if (nc & 4) {
292 vst1q_f32(c3, vacc3x0123); c3 += 4;
293 vst1q_f32(c2, vacc2x0123); c2 += 4;
294 vst1q_f32(c1, vacc1x0123); c1 += 4;
295 vst1q_f32(c0, vacc0x0123); c0 += 4;
296
297 vacc3x0123 = vacc3x4567;
298 vacc2x0123 = vacc2x4567;
299 vacc1x0123 = vacc1x4567;
300 vacc0x0123 = vacc0x4567;
301 }
302 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
303 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
304 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
305 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
306 if (nc & 2) {
307 vst1_f32(c3, vacc3x01); c3 += 2;
308 vst1_f32(c2, vacc2x01); c2 += 2;
309 vst1_f32(c1, vacc1x01); c1 += 2;
310 vst1_f32(c0, vacc0x01); c0 += 2;
311
312 vacc3x01 = vget_high_f32(vacc3x0123);
313 vacc2x01 = vget_high_f32(vacc2x0123);
314 vacc1x01 = vget_high_f32(vacc1x0123);
315 vacc0x01 = vget_high_f32(vacc0x0123);
316 }
317 if (nc & 1) {
318 vst1_lane_f32(c3, vacc3x01, 0);
319 vst1_lane_f32(c2, vacc2x01, 0);
320 vst1_lane_f32(c1, vacc1x01, 0);
321 vst1_lane_f32(c0, vacc0x01, 0);
322 }
323
324 nc = 0;
325 }
326 } while (nc != 0);
327 }
328