Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x16-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_4x16__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 <= 4);
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
60 xnn_prefetch_to_l1((const int8_t*) w + 0);
61 xnn_prefetch_to_l1((const int8_t*) w + 64);
62 xnn_prefetch_to_l1((const int8_t*) w + 128);
63 xnn_prefetch_to_l1((const int8_t*) w + 192);
64
65 do {
66 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
67 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
68 float32x4_t vacc0x89AB = vld1q_f32(w); w += 4;
69 float32x4_t vacc0xCDEF = vld1q_f32(w); w += 4;
70 float32x4_t vacc1x0123 = vacc0x0123;
71 float32x4_t vacc1x4567 = vacc0x4567;
72 float32x4_t vacc1x89AB = vacc0x89AB;
73 float32x4_t vacc1xCDEF = vacc0xCDEF;
74 float32x4_t vacc2x0123 = vacc0x0123;
75 float32x4_t vacc2x4567 = vacc0x4567;
76 float32x4_t vacc2x89AB = vacc0x89AB;
77 float32x4_t vacc2xCDEF = vacc0xCDEF;
78 float32x4_t vacc3x0123 = vacc0x0123;
79 float32x4_t vacc3x4567 = vacc0x4567;
80 float32x4_t vacc3x89AB = vacc0x89AB;
81 float32x4_t vacc3xCDEF = vacc0xCDEF;
82
83 size_t k = kc;
84 do {
85 const float32x4_t va0123 = vld1q_f32(a); a += 4;
86
87 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
88 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
89 const float32x4_t vb89AB = vld1q_f32(w); w += 4;
90 const float32x4_t vbCDEF = vld1q_f32(w); w += 4;
91
92 #if XNN_ARCH_ARM64
93 vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0);
94 vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1);
95 vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2);
96 vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3);
97 xnn_prefetch_to_l1((const int8_t*) w + 192);
98 vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0);
99 vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1);
100 vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2);
101 vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3);
102 vacc0x89AB = vfmaq_laneq_f32(vacc0x89AB, vb89AB, va0123, 0);
103 vacc1x89AB = vfmaq_laneq_f32(vacc1x89AB, vb89AB, va0123, 1);
104 vacc2x89AB = vfmaq_laneq_f32(vacc2x89AB, vb89AB, va0123, 2);
105 vacc3x89AB = vfmaq_laneq_f32(vacc3x89AB, vb89AB, va0123, 3);
106 vacc0xCDEF = vfmaq_laneq_f32(vacc0xCDEF, vbCDEF, va0123, 0);
107 vacc1xCDEF = vfmaq_laneq_f32(vacc1xCDEF, vbCDEF, va0123, 1);
108 vacc2xCDEF = vfmaq_laneq_f32(vacc2xCDEF, vbCDEF, va0123, 2);
109 vacc3xCDEF = vfmaq_laneq_f32(vacc3xCDEF, vbCDEF, va0123, 3);
110 #else
111 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0);
112 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1);
113 const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0);
114 const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1);
115
116 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123);
117 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123);
118 vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123);
119 vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123);
120 xnn_prefetch_to_l1((const int8_t*) w + 192);
121 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567);
122 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567);
123 vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567);
124 vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567);
125 vacc0x89AB = vfmaq_f32(vacc0x89AB, va0000, vb89AB);
126 vacc1x89AB = vfmaq_f32(vacc1x89AB, va1111, vb89AB);
127 vacc2x89AB = vfmaq_f32(vacc2x89AB, va2222, vb89AB);
128 vacc3x89AB = vfmaq_f32(vacc3x89AB, va3333, vb89AB);
129 vacc0xCDEF = vfmaq_f32(vacc0xCDEF, va0000, vbCDEF);
130 vacc1xCDEF = vfmaq_f32(vacc1xCDEF, va1111, vbCDEF);
131 vacc2xCDEF = vfmaq_f32(vacc2xCDEF, va2222, vbCDEF);
132 vacc3xCDEF = vfmaq_f32(vacc3xCDEF, va3333, vbCDEF);
133 #endif
134
135 k -= sizeof(float);
136 } while (k != 0);
137
138 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
139 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
140 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
141 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
142 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
143 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
144 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
145 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
146 vacc0x89AB = vminq_f32(vacc0x89AB, vmax);
147 vacc1x89AB = vminq_f32(vacc1x89AB, vmax);
148 vacc2x89AB = vminq_f32(vacc2x89AB, vmax);
149 vacc3x89AB = vminq_f32(vacc3x89AB, vmax);
150 vacc0xCDEF = vminq_f32(vacc0xCDEF, vmax);
151 vacc1xCDEF = vminq_f32(vacc1xCDEF, vmax);
152 vacc2xCDEF = vminq_f32(vacc2xCDEF, vmax);
153 vacc3xCDEF = vminq_f32(vacc3xCDEF, vmax);
154
155 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
156 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
157 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
158 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
159 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
160 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
161 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
162 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
163 vacc0x89AB = vmaxq_f32(vacc0x89AB, vmin);
164 vacc1x89AB = vmaxq_f32(vacc1x89AB, vmin);
165 vacc2x89AB = vmaxq_f32(vacc2x89AB, vmin);
166 vacc3x89AB = vmaxq_f32(vacc3x89AB, vmin);
167 vacc0xCDEF = vmaxq_f32(vacc0xCDEF, vmin);
168 vacc1xCDEF = vmaxq_f32(vacc1xCDEF, vmin);
169 vacc2xCDEF = vmaxq_f32(vacc2xCDEF, vmin);
170 vacc3xCDEF = vmaxq_f32(vacc3xCDEF, vmin);
171
172 if XNN_LIKELY(nc >= 16) {
173 vst1q_f32(c3, vacc3x0123);
174 vst1q_f32(c3 + 4, vacc3x4567);
175 vst1q_f32(c3 + 8, vacc3x89AB);
176 vst1q_f32(c3 + 12, vacc3xCDEF);
177 c3 = (float*) ((uintptr_t) c3 + cn_stride);
178 vst1q_f32(c2, vacc2x0123);
179 vst1q_f32(c2 + 4, vacc2x4567);
180 vst1q_f32(c2 + 8, vacc2x89AB);
181 vst1q_f32(c2 + 12, vacc2xCDEF);
182 c2 = (float*) ((uintptr_t) c2 + cn_stride);
183 vst1q_f32(c1, vacc1x0123);
184 vst1q_f32(c1 + 4, vacc1x4567);
185 vst1q_f32(c1 + 8, vacc1x89AB);
186 vst1q_f32(c1 + 12, vacc1xCDEF);
187 c1 = (float*) ((uintptr_t) c1 + cn_stride);
188 vst1q_f32(c0, vacc0x0123);
189 vst1q_f32(c0 + 4, vacc0x4567);
190 vst1q_f32(c0 + 8, vacc0x89AB);
191 vst1q_f32(c0 + 12, vacc0xCDEF);
192 c0 = (float*) ((uintptr_t) c0 + cn_stride);
193
194 a = (const float*) ((uintptr_t) a - kc * 4);
195
196 nc -= 16;
197 } else {
198 if (nc & 8) {
199 vst1q_f32(c3, vacc3x0123); c3 += 4;
200 vst1q_f32(c3, vacc3x4567); c3 += 4;
201 vst1q_f32(c2, vacc2x0123); c2 += 4;
202 vst1q_f32(c2, vacc2x4567); c2 += 4;
203 vst1q_f32(c1, vacc1x0123); c1 += 4;
204 vst1q_f32(c1, vacc1x4567); c1 += 4;
205 vst1q_f32(c0, vacc0x0123); c0 += 4;
206 vst1q_f32(c0, vacc0x4567); c0 += 4;
207
208 vacc3x0123 = vacc3x89AB;
209 vacc3x4567 = vacc3xCDEF;
210 vacc2x0123 = vacc2x89AB;
211 vacc2x4567 = vacc2xCDEF;
212 vacc1x0123 = vacc1x89AB;
213 vacc1x4567 = vacc1xCDEF;
214 vacc0x0123 = vacc0x89AB;
215 vacc0x4567 = vacc0xCDEF;
216 }
217 if (nc & 4) {
218 vst1q_f32(c3, vacc3x0123); c3 += 4;
219 vst1q_f32(c2, vacc2x0123); c2 += 4;
220 vst1q_f32(c1, vacc1x0123); c1 += 4;
221 vst1q_f32(c0, vacc0x0123); c0 += 4;
222
223 vacc3x0123 = vacc3x4567;
224 vacc3x4567 = vacc3x89AB;
225 vacc3x89AB = vacc3xCDEF;
226 vacc2x0123 = vacc2x4567;
227 vacc2x4567 = vacc2x89AB;
228 vacc2x89AB = vacc2xCDEF;
229 vacc1x0123 = vacc1x4567;
230 vacc1x4567 = vacc1x89AB;
231 vacc1x89AB = vacc1xCDEF;
232 vacc0x0123 = vacc0x4567;
233 vacc0x4567 = vacc0x89AB;
234 vacc0x89AB = vacc0xCDEF;
235 }
236 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
237 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
238 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
239 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
240 if (nc & 2) {
241 vst1_f32(c3, vacc3x01); c3 += 2;
242 vst1_f32(c2, vacc2x01); c2 += 2;
243 vst1_f32(c1, vacc1x01); c1 += 2;
244 vst1_f32(c0, vacc0x01); c0 += 2;
245
246 vacc3x01 = vget_high_f32(vacc3x0123);
247 vacc2x01 = vget_high_f32(vacc2x0123);
248 vacc1x01 = vget_high_f32(vacc1x0123);
249 vacc0x01 = vget_high_f32(vacc0x0123);
250 }
251 if (nc & 1) {
252 vst1_lane_f32(c3, vacc3x01, 0);
253 vst1_lane_f32(c2, vacc2x01, 0);
254 vst1_lane_f32(c1, vacc1x01, 0);
255 vst1_lane_f32(c0, vacc0x01, 0);
256 }
257
258 nc = 0;
259 }
260 } while (nc != 0);
261 }