1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/neon-ld64.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/igemm.h>
16
17
xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const float**restrict a,
24 const float*restrict w,
25 float*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const float* zero,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 6);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (6 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(float) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 float* c0 = c;
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr != 6) {
63 c5 = c4;
64 }
65
66 do {
67 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
68 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
69 float32x4_t vacc1x0123 = vacc0x0123;
70 float32x4_t vacc1x4567 = vacc0x4567;
71 float32x4_t vacc2x0123 = vacc0x0123;
72 float32x4_t vacc2x4567 = vacc0x4567;
73 float32x4_t vacc3x0123 = vacc0x0123;
74 float32x4_t vacc3x4567 = vacc0x4567;
75 float32x4_t vacc4x0123 = vacc0x0123;
76 float32x4_t vacc4x4567 = vacc0x4567;
77 float32x4_t vacc5x0123 = vacc0x0123;
78 float32x4_t vacc5x4567 = vacc0x4567;
79
80 size_t p = ks;
81 do {
82 const float* restrict a0 = a[0];
83 assert(a0 != NULL);
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const float*) ((uintptr_t) a0 + a_offset);
86 }
87 const float* restrict a1 = a[1];
88 assert(a1 != NULL);
89 if XNN_UNPREDICTABLE(a1 != zero) {
90 a1 = (const float*) ((uintptr_t) a1 + a_offset);
91 }
92 const float* restrict a2 = a[2];
93 assert(a2 != NULL);
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const float*) ((uintptr_t) a2 + a_offset);
96 }
97 const float* restrict a3 = a[3];
98 assert(a3 != NULL);
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const float*) ((uintptr_t) a3 + a_offset);
101 }
102 const float* restrict a4 = a[4];
103 assert(a4 != NULL);
104 if XNN_UNPREDICTABLE(a4 != zero) {
105 a4 = (const float*) ((uintptr_t) a4 + a_offset);
106 }
107 const float* restrict a5 = a[5];
108 assert(a5 != NULL);
109 if XNN_UNPREDICTABLE(a5 != zero) {
110 a5 = (const float*) ((uintptr_t) a5 + a_offset);
111 }
112 a += 6;
113
114 size_t k = kc;
115 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
116 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
117 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
118 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
119 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
120 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
121 const float32x2_t va5 = vld1_f32(a5); a5 += 2;
122
123 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
124 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
125
126 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0);
127 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0);
128 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0);
129 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0);
130 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0);
131 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c0, va5, 0);
132 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0);
133 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0);
134 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0);
135 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0);
136 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0);
137 vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c0, va5, 0);
138 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
139 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
140
141 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1);
142 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1);
143 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1);
144 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1);
145 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1);
146 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c1, va5, 1);
147 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c1, va0, 1);
148 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c1, va1, 1);
149 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c1, va2, 1);
150 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c1, va3, 1);
151 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c1, va4, 1);
152 vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c1, va5, 1);
153
154 }
155 if XNN_UNLIKELY(k != 0) {
156 const float32x4_t va0 = vld1q_dup_f32(a0);
157 const float32x4_t va1 = vld1q_dup_f32(a1);
158 const float32x4_t va2 = vld1q_dup_f32(a2);
159 const float32x4_t va3 = vld1q_dup_f32(a3);
160 const float32x4_t va4 = vld1q_dup_f32(a4);
161 const float32x4_t va5 = vld1q_dup_f32(a5);
162
163 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
164 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
165
166 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
167 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
168 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
169 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
170 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123);
171 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123);
172 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
173 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
174 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
175 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
176 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567);
177 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567);
178 }
179 p -= 6 * sizeof(void*);
180 } while (p != 0);
181
182 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
183 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
184 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
185 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
186 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
187 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
188 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
189 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
190 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
191 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
192 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
193 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
194 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
195
196 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
197 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
198 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
199 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
200 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
201 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
202 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
203 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
204 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
205 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
206 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
207 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
208 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
209
210 if XNN_LIKELY(nc >= 8) {
211 vst1q_f32(c5, vacc5x0123);
212 vst1q_f32(c5 + 4, vacc5x4567);
213 c5 = (float*) ((uintptr_t) c5 + cn_stride);
214 vst1q_f32(c4, vacc4x0123);
215 vst1q_f32(c4 + 4, vacc4x4567);
216 c4 = (float*) ((uintptr_t) c4 + cn_stride);
217 vst1q_f32(c3, vacc3x0123);
218 vst1q_f32(c3 + 4, vacc3x4567);
219 c3 = (float*) ((uintptr_t) c3 + cn_stride);
220 vst1q_f32(c2, vacc2x0123);
221 vst1q_f32(c2 + 4, vacc2x4567);
222 c2 = (float*) ((uintptr_t) c2 + cn_stride);
223 vst1q_f32(c1, vacc1x0123);
224 vst1q_f32(c1 + 4, vacc1x4567);
225 c1 = (float*) ((uintptr_t) c1 + cn_stride);
226 vst1q_f32(c0, vacc0x0123);
227 vst1q_f32(c0 + 4, vacc0x4567);
228 c0 = (float*) ((uintptr_t) c0 + cn_stride);
229
230 a = (const float**restrict) ((uintptr_t) a - ks);
231 nc -= 8;
232 } else {
233 if (nc & 4) {
234 vst1q_f32(c5, vacc5x0123); c5 += 4;
235 vst1q_f32(c4, vacc4x0123); c4 += 4;
236 vst1q_f32(c3, vacc3x0123); c3 += 4;
237 vst1q_f32(c2, vacc2x0123); c2 += 4;
238 vst1q_f32(c1, vacc1x0123); c1 += 4;
239 vst1q_f32(c0, vacc0x0123); c0 += 4;
240
241 vacc5x0123 = vacc5x4567;
242 vacc4x0123 = vacc4x4567;
243 vacc3x0123 = vacc3x4567;
244 vacc2x0123 = vacc2x4567;
245 vacc1x0123 = vacc1x4567;
246 vacc0x0123 = vacc0x4567;
247 }
248 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
249 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
250 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
251 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
252 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
253 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
254 if (nc & 2) {
255 vst1_f32(c5, vacc5x01); c5 += 2;
256 vst1_f32(c4, vacc4x01); c4 += 2;
257 vst1_f32(c3, vacc3x01); c3 += 2;
258 vst1_f32(c2, vacc2x01); c2 += 2;
259 vst1_f32(c1, vacc1x01); c1 += 2;
260 vst1_f32(c0, vacc0x01); c0 += 2;
261
262 vacc5x01 = vget_high_f32(vacc5x0123);
263 vacc4x01 = vget_high_f32(vacc4x0123);
264 vacc3x01 = vget_high_f32(vacc3x0123);
265 vacc2x01 = vget_high_f32(vacc2x0123);
266 vacc1x01 = vget_high_f32(vacc1x0123);
267 vacc0x01 = vget_high_f32(vacc0x0123);
268 }
269 if (nc & 1) {
270 vst1_lane_f32(c5, vacc5x01, 0);
271 vst1_lane_f32(c4, vacc4x01, 0);
272 vst1_lane_f32(c3, vacc3x01, 0);
273 vst1_lane_f32(c2, vacc2x01, 0);
274 vst1_lane_f32(c1, vacc1x01, 0);
275 vst1_lane_f32(c0, vacc0x01, 0);
276 }
277
278 nc = 0;
279 }
280 } while (nc != 0);
281 }
282