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_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_output_params params[restrict static1])18 void xnn_f32_gemminc_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_output_params params[restrict static 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 + 0);
81 const float32x4_t vb4567c0 = vld1q_f32(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 + 8);
98 const float32x4_t vb4567c1 = vld1q_f32(w + 12);
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 + 16);
115 const float32x4_t vb4567c2 = vld1q_f32(w + 20);
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 + 24);
132 const float32x4_t vb4567c3 = vld1q_f32(w + 28);
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 w += 32;
145 k -= 4 * sizeof(float);
146 }
147 if XNN_UNLIKELY(k != 0) {
148 do {
149 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
150 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
151 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
152 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
153
154 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
155 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
156
157 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
158 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
159 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
160 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
161 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
162 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
163 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
164 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
165
166 k -= sizeof(float);
167 } while (k != 0);
168 }
169 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
170 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
171 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
172 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
173 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
174 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
175 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
176 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
177 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
178
179 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
180 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
181 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
182 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
183 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
184 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
185 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
186 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
187 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
188
189 if XNN_LIKELY(nc >= 8) {
190 vst1q_f32(c3, vacc3x0123);
191 vst1q_f32(c3 + 4, vacc3x4567);
192 c3 = (float*) ((uintptr_t) c3 + cn_stride);
193 vst1q_f32(c2, vacc2x0123);
194 vst1q_f32(c2 + 4, vacc2x4567);
195 c2 = (float*) ((uintptr_t) c2 + cn_stride);
196 vst1q_f32(c1, vacc1x0123);
197 vst1q_f32(c1 + 4, vacc1x4567);
198 c1 = (float*) ((uintptr_t) c1 + cn_stride);
199 vst1q_f32(c0, vacc0x0123);
200 vst1q_f32(c0 + 4, vacc0x4567);
201 c0 = (float*) ((uintptr_t) c0 + cn_stride);
202
203 a3 = (const float*) ((uintptr_t) a3 - kc);
204 a2 = (const float*) ((uintptr_t) a2 - kc);
205 a1 = (const float*) ((uintptr_t) a1 - kc);
206 a0 = (const float*) ((uintptr_t) a0 - kc);
207
208 nc -= 8;
209
210 } else {
211 if (nc & 4) {
212 vst1q_f32(c3, vacc3x0123); c3 += 4;
213 vst1q_f32(c2, vacc2x0123); c2 += 4;
214 vst1q_f32(c1, vacc1x0123); c1 += 4;
215 vst1q_f32(c0, vacc0x0123); c0 += 4;
216
217 vacc3x0123 = vacc3x4567;
218 vacc2x0123 = vacc2x4567;
219 vacc1x0123 = vacc1x4567;
220 vacc0x0123 = vacc0x4567;
221 }
222 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
223 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
224 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
225 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
226 if (nc & 2) {
227 vst1_f32(c3, vacc3x01); c3 += 2;
228 vst1_f32(c2, vacc2x01); c2 += 2;
229 vst1_f32(c1, vacc1x01); c1 += 2;
230 vst1_f32(c0, vacc0x01); c0 += 2;
231
232 vacc3x01 = vget_high_f32(vacc3x0123);
233 vacc2x01 = vget_high_f32(vacc2x0123);
234 vacc1x01 = vget_high_f32(vacc1x0123);
235 vacc0x01 = vget_high_f32(vacc0x0123);
236 }
237 if (nc & 1) {
238 vst1_lane_f32(c3, vacc3x01, 0);
239 vst1_lane_f32(c2, vacc2x01, 0);
240 vst1_lane_f32(c1, vacc1x01, 0);
241 vst1_lane_f32(c0, vacc0x01, 0);
242 }
243
244 nc = 0;
245 }
246 } while (nc != 0);
247 }
248