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_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
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
60 do {
61 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
62 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
63 float32x4_t vacc1x0123 = vacc0x0123;
64 float32x4_t vacc1x4567 = vacc0x4567;
65 float32x4_t vacc2x0123 = vacc0x0123;
66 float32x4_t vacc2x4567 = vacc0x4567;
67 float32x4_t vacc3x0123 = vacc0x0123;
68 float32x4_t vacc3x4567 = vacc0x4567;
69
70 size_t k = kc;
71 while (k >= 4 * sizeof(float)) {
72 float32x4_t va0 = vld1q_f32(a0); a0 += 4;
73 float32x4_t va1 = vld1q_f32(a1); a1 += 4;
74 float32x4_t va2 = vld1q_f32(a2); a2 += 4;
75 float32x4_t va3 = vld1q_f32(a3); a3 += 4;
76
77
78 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
79 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
80
81 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
82 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);
83 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);
84 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);
85 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
86 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);
87 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);
88 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);
89
90 va0 = vextq_f32(va0, va0, 1);
91 va1 = vextq_f32(va1, va1, 1);
92 va2 = vextq_f32(va2, va2, 1);
93 va3 = vextq_f32(va3, va3, 1);
94
95 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
96 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
97
98 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
99 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);
100 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);
101 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);
102 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
103 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);
104 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);
105 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);
106
107 va0 = vextq_f32(va0, va0, 1);
108 va1 = vextq_f32(va1, va1, 1);
109 va2 = vextq_f32(va2, va2, 1);
110 va3 = vextq_f32(va3, va3, 1);
111
112 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
113 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
114
115 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
116 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);
117 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);
118 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);
119 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
120 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);
121 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);
122 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);
123
124 va0 = vextq_f32(va0, va0, 1);
125 va1 = vextq_f32(va1, va1, 1);
126 va2 = vextq_f32(va2, va2, 1);
127 va3 = vextq_f32(va3, va3, 1);
128
129 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
130 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
131
132 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
133 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);
134 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);
135 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);
136 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
137 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);
138 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);
139 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);
140
141
142 k -= 4 * sizeof(float);
143 }
144 if XNN_UNLIKELY(k != 0) {
145 float32x4_t va0 = vld1q_f32(a0); a0 = (const float*) ((uintptr_t) a0 + k);
146 float32x4_t va1 = vld1q_f32(a1); a1 = (const float*) ((uintptr_t) a1 + k);
147 float32x4_t va2 = vld1q_f32(a2); a2 = (const float*) ((uintptr_t) a2 + k);
148 float32x4_t va3 = vld1q_f32(a3); a3 = (const float*) ((uintptr_t) a3 + k);
149
150
151 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
152 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
153
154 const float32x4_t vmska0x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
155 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c0, vb0123c0);
156 const float32x4_t vmska1x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
157 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c0, vb0123c0);
158 const float32x4_t vmska2x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
159 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c0, vb0123c0);
160 const float32x4_t vmska3x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
161 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c0, vb0123c0);
162 const float32x4_t vmska0x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
163 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c0, vb4567c0);
164 const float32x4_t vmska1x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
165 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c0, vb4567c0);
166 const float32x4_t vmska2x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
167 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c0, vb4567c0);
168 const float32x4_t vmska3x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
169 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c0, vb4567c0);
170
171 va0 = vextq_f32(va0, va0, 1);
172 va1 = vextq_f32(va1, va1, 1);
173 va2 = vextq_f32(va2, va2, 1);
174 va3 = vextq_f32(va3, va3, 1);
175
176 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
177 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
178
179 const float32x4_t vmska0x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
180 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c1, vb0123c1);
181 const float32x4_t vmska1x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
182 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c1, vb0123c1);
183 const float32x4_t vmska2x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
184 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c1, vb0123c1);
185 const float32x4_t vmska3x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
186 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c1, vb0123c1);
187 const float32x4_t vmska0x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
188 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c1, vb4567c1);
189 const float32x4_t vmska1x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
190 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c1, vb4567c1);
191 const float32x4_t vmska2x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
192 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c1, vb4567c1);
193 const float32x4_t vmska3x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
194 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c1, vb4567c1);
195
196 va0 = vextq_f32(va0, va0, 1);
197 va1 = vextq_f32(va1, va1, 1);
198 va2 = vextq_f32(va2, va2, 1);
199 va3 = vextq_f32(va3, va3, 1);
200
201 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
202 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
203
204 const float32x4_t vmska0x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
205 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c2, vb0123c2);
206 const float32x4_t vmska1x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
207 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c2, vb0123c2);
208 const float32x4_t vmska2x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
209 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c2, vb0123c2);
210 const float32x4_t vmska3x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
211 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c2, vb0123c2);
212 const float32x4_t vmska0x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
213 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c2, vb4567c2);
214 const float32x4_t vmska1x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
215 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c2, vb4567c2);
216 const float32x4_t vmska2x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
217 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c2, vb4567c2);
218 const float32x4_t vmska3x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
219 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c2, vb4567c2);
220
221 va0 = vextq_f32(va0, va0, 1);
222 va1 = vextq_f32(va1, va1, 1);
223 va2 = vextq_f32(va2, va2, 1);
224 va3 = vextq_f32(va3, va3, 1);
225
226 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
227 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
228
229 const float32x4_t vmska0x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
230 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c3, vb0123c3);
231 const float32x4_t vmska1x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
232 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c3, vb0123c3);
233 const float32x4_t vmska2x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
234 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c3, vb0123c3);
235 const float32x4_t vmska3x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
236 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c3, vb0123c3);
237 const float32x4_t vmska0x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
238 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c3, vb4567c3);
239 const float32x4_t vmska1x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
240 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c3, vb4567c3);
241 const float32x4_t vmska2x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
242 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c3, vb4567c3);
243 const float32x4_t vmska3x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
244 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c3, vb4567c3);
245
246 }
247 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
248 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
249 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
250 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
251 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
252 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
253 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
254 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
255 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
256
257 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
258 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
259 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
260 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
261 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
262 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
263 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
264 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
265 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
266
267 if XNN_LIKELY(nc >= 8) {
268 vst1q_f32(c3, vacc3x0123);
269 vst1q_f32(c3 + 4, vacc3x4567);
270 c3 = (float*) ((uintptr_t) c3 + cn_stride);
271 vst1q_f32(c2, vacc2x0123);
272 vst1q_f32(c2 + 4, vacc2x4567);
273 c2 = (float*) ((uintptr_t) c2 + cn_stride);
274 vst1q_f32(c1, vacc1x0123);
275 vst1q_f32(c1 + 4, vacc1x4567);
276 c1 = (float*) ((uintptr_t) c1 + cn_stride);
277 vst1q_f32(c0, vacc0x0123);
278 vst1q_f32(c0 + 4, vacc0x4567);
279 c0 = (float*) ((uintptr_t) c0 + cn_stride);
280
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
288 } else {
289 if (nc & 4) {
290 vst1q_f32(c3, vacc3x0123); c3 += 4;
291 vst1q_f32(c2, vacc2x0123); c2 += 4;
292 vst1q_f32(c1, vacc1x0123); c1 += 4;
293 vst1q_f32(c0, vacc0x0123); c0 += 4;
294
295 vacc3x0123 = vacc3x4567;
296 vacc2x0123 = vacc2x4567;
297 vacc1x0123 = vacc1x4567;
298 vacc0x0123 = vacc0x4567;
299 }
300 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
301 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
302 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
303 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
304 if (nc & 2) {
305 vst1_f32(c3, vacc3x01); c3 += 2;
306 vst1_f32(c2, vacc2x01); c2 += 2;
307 vst1_f32(c1, vacc1x01); c1 += 2;
308 vst1_f32(c0, vacc0x01); c0 += 2;
309
310 vacc3x01 = vget_high_f32(vacc3x0123);
311 vacc2x01 = vget_high_f32(vacc2x0123);
312 vacc1x01 = vget_high_f32(vacc1x0123);
313 vacc0x01 = vget_high_f32(vacc0x0123);
314 }
315 if (nc & 1) {
316 vst1_lane_f32(c3, vacc3x01, 0);
317 vst1_lane_f32(c2, vacc2x01, 0);
318 vst1_lane_f32(c1, vacc1x01, 0);
319 vst1_lane_f32(c0, vacc0x01, 0);
320 }
321
322 nc = 0;
323 }
324 } while (nc != 0);
325 }
326