Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-8x8-minmax-aarch64-neonfma-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-ppmm/neon.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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/ppmm.h"
16 #include "xnnpack/prefetch.h"
17
18
19 void xnn_f32_ppmm_minmax_ukernel_8x8__aarch64_neonfma_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
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 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 8);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35
36 #if XNN_ARCH_ARM64
37 const float32x4x2_t vminmax = vld2q_dup_f32(&params->scalar.min);
38 const float32x4_t vmin = vminmax.val[0];
39 const float32x4_t vmax = vminmax.val[1];
40 #else
41 const float32x2x2_t vminmax = vld2_dup_f32(&params->scalar.min);
42 const float32x4_t vmin = vcombine_f32(vminmax.val[0], vminmax.val[0]);
43 const float32x4_t vmax = vcombine_f32(vminmax.val[1], vminmax.val[1]);
44 #endif
45
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
72 if XNN_UNPREDICTABLE(mr != 8) {
73 c7 = c6;
74 }
75
76 xnn_prefetch_to_l1((const int8_t*) w + 0);
77 xnn_prefetch_to_l1((const int8_t*) w + 64);
78 xnn_prefetch_to_l1((const int8_t*) w + 128);
79 xnn_prefetch_to_l1((const int8_t*) w + 192);
80
81 do {
82 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
83 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
84 float32x4_t vacc1x0123 = vacc0x0123;
85 float32x4_t vacc1x4567 = vacc0x4567;
86 float32x4_t vacc2x0123 = vacc0x0123;
87 float32x4_t vacc2x4567 = vacc0x4567;
88 float32x4_t vacc3x0123 = vacc0x0123;
89 float32x4_t vacc3x4567 = vacc0x4567;
90 float32x4_t vacc4x0123 = vacc0x0123;
91 float32x4_t vacc4x4567 = vacc0x4567;
92 float32x4_t vacc5x0123 = vacc0x0123;
93 float32x4_t vacc5x4567 = vacc0x4567;
94 float32x4_t vacc6x0123 = vacc0x0123;
95 float32x4_t vacc6x4567 = vacc0x4567;
96 float32x4_t vacc7x0123 = vacc0x0123;
97 float32x4_t vacc7x4567 = vacc0x4567;
98
99 size_t k = kc;
100 do {
101 const float32x4_t va0123 = vld1q_f32(a); a += 4;
102 const float32x4_t va4567 = vld1q_f32(a); a += 4;
103
104 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
105 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
106
107 #if XNN_ARCH_ARM64
108 vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0);
109 vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1);
110 vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2);
111 vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3);
112 vacc4x0123 = vfmaq_laneq_f32(vacc4x0123, vb0123, va4567, 0);
113 vacc5x0123 = vfmaq_laneq_f32(vacc5x0123, vb0123, va4567, 1);
114 vacc6x0123 = vfmaq_laneq_f32(vacc6x0123, vb0123, va4567, 2);
115 vacc7x0123 = vfmaq_laneq_f32(vacc7x0123, vb0123, va4567, 3);
116 xnn_prefetch_to_l1((const int8_t*) w + 192);
117 vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0);
118 vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1);
119 vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2);
120 vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3);
121 vacc4x4567 = vfmaq_laneq_f32(vacc4x4567, vb4567, va4567, 0);
122 vacc5x4567 = vfmaq_laneq_f32(vacc5x4567, vb4567, va4567, 1);
123 vacc6x4567 = vfmaq_laneq_f32(vacc6x4567, vb4567, va4567, 2);
124 vacc7x4567 = vfmaq_laneq_f32(vacc7x4567, vb4567, va4567, 3);
125 #else
126 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0);
127 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1);
128 const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0);
129 const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1);
130 const float32x4_t va4444 = vdupq_lane_f32(vget_low_f32(va4567), 0);
131 const float32x4_t va5555 = vdupq_lane_f32(vget_low_f32(va4567), 1);
132 const float32x4_t va6666 = vdupq_lane_f32(vget_high_f32(va4567), 0);
133 const float32x4_t va7777 = vdupq_lane_f32(vget_high_f32(va4567), 1);
134
135 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123);
136 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123);
137 vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123);
138 vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123);
139 vacc4x0123 = vfmaq_f32(vacc4x0123, va4444, vb0123);
140 vacc5x0123 = vfmaq_f32(vacc5x0123, va5555, vb0123);
141 vacc6x0123 = vfmaq_f32(vacc6x0123, va6666, vb0123);
142 vacc7x0123 = vfmaq_f32(vacc7x0123, va7777, vb0123);
143 xnn_prefetch_to_l1((const int8_t*) w + 192);
144 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567);
145 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567);
146 vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567);
147 vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567);
148 vacc4x4567 = vfmaq_f32(vacc4x4567, va4444, vb4567);
149 vacc5x4567 = vfmaq_f32(vacc5x4567, va5555, vb4567);
150 vacc6x4567 = vfmaq_f32(vacc6x4567, va6666, vb4567);
151 vacc7x4567 = vfmaq_f32(vacc7x4567, va7777, vb4567);
152 #endif
153
154 k -= sizeof(float);
155 } while (k != 0);
156
157 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
158 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
159 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
160 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
161 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
162 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
163 vacc6x0123 = vminq_f32(vacc6x0123, vmax);
164 vacc7x0123 = vminq_f32(vacc7x0123, vmax);
165 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
166 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
167 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
168 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
169 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
170 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
171 vacc6x4567 = vminq_f32(vacc6x4567, vmax);
172 vacc7x4567 = vminq_f32(vacc7x4567, vmax);
173
174 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
175 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
176 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
177 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
178 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
179 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
180 vacc6x0123 = vmaxq_f32(vacc6x0123, vmin);
181 vacc7x0123 = vmaxq_f32(vacc7x0123, vmin);
182 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
183 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
184 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
185 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
186 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
187 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
188 vacc6x4567 = vmaxq_f32(vacc6x4567, vmin);
189 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin);
190
191 if XNN_LIKELY(nc >= 8) {
192 vst1q_f32(c7, vacc7x0123);
193 vst1q_f32(c7 + 4, vacc7x4567);
194 c7 = (float*) ((uintptr_t) c7 + cn_stride);
195 vst1q_f32(c6, vacc6x0123);
196 vst1q_f32(c6 + 4, vacc6x4567);
197 c6 = (float*) ((uintptr_t) c6 + cn_stride);
198 vst1q_f32(c5, vacc5x0123);
199 vst1q_f32(c5 + 4, vacc5x4567);
200 c5 = (float*) ((uintptr_t) c5 + cn_stride);
201 vst1q_f32(c4, vacc4x0123);
202 vst1q_f32(c4 + 4, vacc4x4567);
203 c4 = (float*) ((uintptr_t) c4 + cn_stride);
204 vst1q_f32(c3, vacc3x0123);
205 vst1q_f32(c3 + 4, vacc3x4567);
206 c3 = (float*) ((uintptr_t) c3 + cn_stride);
207 vst1q_f32(c2, vacc2x0123);
208 vst1q_f32(c2 + 4, vacc2x4567);
209 c2 = (float*) ((uintptr_t) c2 + cn_stride);
210 vst1q_f32(c1, vacc1x0123);
211 vst1q_f32(c1 + 4, vacc1x4567);
212 c1 = (float*) ((uintptr_t) c1 + cn_stride);
213 vst1q_f32(c0, vacc0x0123);
214 vst1q_f32(c0 + 4, vacc0x4567);
215 c0 = (float*) ((uintptr_t) c0 + cn_stride);
216
217 a = (const float*) ((uintptr_t) a - kc * 8);
218
219 nc -= 8;
220 } else {
221 if (nc & 4) {
222 vst1q_f32(c7, vacc7x0123); c7 += 4;
223 vst1q_f32(c6, vacc6x0123); c6 += 4;
224 vst1q_f32(c5, vacc5x0123); c5 += 4;
225 vst1q_f32(c4, vacc4x0123); c4 += 4;
226 vst1q_f32(c3, vacc3x0123); c3 += 4;
227 vst1q_f32(c2, vacc2x0123); c2 += 4;
228 vst1q_f32(c1, vacc1x0123); c1 += 4;
229 vst1q_f32(c0, vacc0x0123); c0 += 4;
230
231 vacc7x0123 = vacc7x4567;
232 vacc6x0123 = vacc6x4567;
233 vacc5x0123 = vacc5x4567;
234 vacc4x0123 = vacc4x4567;
235 vacc3x0123 = vacc3x4567;
236 vacc2x0123 = vacc2x4567;
237 vacc1x0123 = vacc1x4567;
238 vacc0x0123 = vacc0x4567;
239 }
240 float32x2_t vacc7x01 = vget_low_f32(vacc7x0123);
241 float32x2_t vacc6x01 = vget_low_f32(vacc6x0123);
242 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
243 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
244 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
245 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
246 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
247 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
248 if (nc & 2) {
249 vst1_f32(c7, vacc7x01); c7 += 2;
250 vst1_f32(c6, vacc6x01); c6 += 2;
251 vst1_f32(c5, vacc5x01); c5 += 2;
252 vst1_f32(c4, vacc4x01); c4 += 2;
253 vst1_f32(c3, vacc3x01); c3 += 2;
254 vst1_f32(c2, vacc2x01); c2 += 2;
255 vst1_f32(c1, vacc1x01); c1 += 2;
256 vst1_f32(c0, vacc0x01); c0 += 2;
257
258 vacc7x01 = vget_high_f32(vacc7x0123);
259 vacc6x01 = vget_high_f32(vacc6x0123);
260 vacc5x01 = vget_high_f32(vacc5x0123);
261 vacc4x01 = vget_high_f32(vacc4x0123);
262 vacc3x01 = vget_high_f32(vacc3x0123);
263 vacc2x01 = vget_high_f32(vacc2x0123);
264 vacc1x01 = vget_high_f32(vacc1x0123);
265 vacc0x01 = vget_high_f32(vacc0x0123);
266 }
267 if (nc & 1) {
268 vst1_lane_f32(c7, vacc7x01, 0);
269 vst1_lane_f32(c6, vacc6x01, 0);
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 }