Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x16-minmax-neon-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__neon_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 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123, vget_low_f32(va0123), 0);
93 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123, vget_low_f32(va0123), 1);
94 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123, vget_high_f32(va0123), 0);
95 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123, vget_high_f32(va0123), 1);
96 xnn_prefetch_to_l1((const int8_t*) w + 192);
97 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567, vget_low_f32(va0123), 0);
98 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567, vget_low_f32(va0123), 1);
99 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567, vget_high_f32(va0123), 0);
100 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567, vget_high_f32(va0123), 1);
101 vacc0x89AB = vmlaq_lane_f32(vacc0x89AB, vb89AB, vget_low_f32(va0123), 0);
102 vacc1x89AB = vmlaq_lane_f32(vacc1x89AB, vb89AB, vget_low_f32(va0123), 1);
103 vacc2x89AB = vmlaq_lane_f32(vacc2x89AB, vb89AB, vget_high_f32(va0123), 0);
104 vacc3x89AB = vmlaq_lane_f32(vacc3x89AB, vb89AB, vget_high_f32(va0123), 1);
105 vacc0xCDEF = vmlaq_lane_f32(vacc0xCDEF, vbCDEF, vget_low_f32(va0123), 0);
106 vacc1xCDEF = vmlaq_lane_f32(vacc1xCDEF, vbCDEF, vget_low_f32(va0123), 1);
107 vacc2xCDEF = vmlaq_lane_f32(vacc2xCDEF, vbCDEF, vget_high_f32(va0123), 0);
108 vacc3xCDEF = vmlaq_lane_f32(vacc3xCDEF, vbCDEF, vget_high_f32(va0123), 1);
109
110 k -= sizeof(float);
111 } while (k != 0);
112
113 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
114 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
115 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
116 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
117 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
118 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
119 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
120 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
121 vacc0x89AB = vminq_f32(vacc0x89AB, vmax);
122 vacc1x89AB = vminq_f32(vacc1x89AB, vmax);
123 vacc2x89AB = vminq_f32(vacc2x89AB, vmax);
124 vacc3x89AB = vminq_f32(vacc3x89AB, vmax);
125 vacc0xCDEF = vminq_f32(vacc0xCDEF, vmax);
126 vacc1xCDEF = vminq_f32(vacc1xCDEF, vmax);
127 vacc2xCDEF = vminq_f32(vacc2xCDEF, vmax);
128 vacc3xCDEF = vminq_f32(vacc3xCDEF, vmax);
129
130 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
131 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
132 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
133 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
134 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
135 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
136 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
137 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
138 vacc0x89AB = vmaxq_f32(vacc0x89AB, vmin);
139 vacc1x89AB = vmaxq_f32(vacc1x89AB, vmin);
140 vacc2x89AB = vmaxq_f32(vacc2x89AB, vmin);
141 vacc3x89AB = vmaxq_f32(vacc3x89AB, vmin);
142 vacc0xCDEF = vmaxq_f32(vacc0xCDEF, vmin);
143 vacc1xCDEF = vmaxq_f32(vacc1xCDEF, vmin);
144 vacc2xCDEF = vmaxq_f32(vacc2xCDEF, vmin);
145 vacc3xCDEF = vmaxq_f32(vacc3xCDEF, vmin);
146
147 if XNN_LIKELY(nc >= 16) {
148 vst1q_f32(c3, vacc3x0123);
149 vst1q_f32(c3 + 4, vacc3x4567);
150 vst1q_f32(c3 + 8, vacc3x89AB);
151 vst1q_f32(c3 + 12, vacc3xCDEF);
152 c3 = (float*) ((uintptr_t) c3 + cn_stride);
153 vst1q_f32(c2, vacc2x0123);
154 vst1q_f32(c2 + 4, vacc2x4567);
155 vst1q_f32(c2 + 8, vacc2x89AB);
156 vst1q_f32(c2 + 12, vacc2xCDEF);
157 c2 = (float*) ((uintptr_t) c2 + cn_stride);
158 vst1q_f32(c1, vacc1x0123);
159 vst1q_f32(c1 + 4, vacc1x4567);
160 vst1q_f32(c1 + 8, vacc1x89AB);
161 vst1q_f32(c1 + 12, vacc1xCDEF);
162 c1 = (float*) ((uintptr_t) c1 + cn_stride);
163 vst1q_f32(c0, vacc0x0123);
164 vst1q_f32(c0 + 4, vacc0x4567);
165 vst1q_f32(c0 + 8, vacc0x89AB);
166 vst1q_f32(c0 + 12, vacc0xCDEF);
167 c0 = (float*) ((uintptr_t) c0 + cn_stride);
168
169 a = (const float*) ((uintptr_t) a - kc * 4);
170
171 nc -= 16;
172 } else {
173 if (nc & 8) {
174 vst1q_f32(c3, vacc3x0123); c3 += 4;
175 vst1q_f32(c3, vacc3x4567); c3 += 4;
176 vst1q_f32(c2, vacc2x0123); c2 += 4;
177 vst1q_f32(c2, vacc2x4567); c2 += 4;
178 vst1q_f32(c1, vacc1x0123); c1 += 4;
179 vst1q_f32(c1, vacc1x4567); c1 += 4;
180 vst1q_f32(c0, vacc0x0123); c0 += 4;
181 vst1q_f32(c0, vacc0x4567); c0 += 4;
182
183 vacc3x0123 = vacc3x89AB;
184 vacc3x4567 = vacc3xCDEF;
185 vacc2x0123 = vacc2x89AB;
186 vacc2x4567 = vacc2xCDEF;
187 vacc1x0123 = vacc1x89AB;
188 vacc1x4567 = vacc1xCDEF;
189 vacc0x0123 = vacc0x89AB;
190 vacc0x4567 = vacc0xCDEF;
191 }
192 if (nc & 4) {
193 vst1q_f32(c3, vacc3x0123); c3 += 4;
194 vst1q_f32(c2, vacc2x0123); c2 += 4;
195 vst1q_f32(c1, vacc1x0123); c1 += 4;
196 vst1q_f32(c0, vacc0x0123); c0 += 4;
197
198 vacc3x0123 = vacc3x4567;
199 vacc3x4567 = vacc3x89AB;
200 vacc3x89AB = vacc3xCDEF;
201 vacc2x0123 = vacc2x4567;
202 vacc2x4567 = vacc2x89AB;
203 vacc2x89AB = vacc2xCDEF;
204 vacc1x0123 = vacc1x4567;
205 vacc1x4567 = vacc1x89AB;
206 vacc1x89AB = vacc1xCDEF;
207 vacc0x0123 = vacc0x4567;
208 vacc0x4567 = vacc0x89AB;
209 vacc0x89AB = vacc0xCDEF;
210 }
211 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
212 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
213 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
214 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
215 if (nc & 2) {
216 vst1_f32(c3, vacc3x01); c3 += 2;
217 vst1_f32(c2, vacc2x01); c2 += 2;
218 vst1_f32(c1, vacc1x01); c1 += 2;
219 vst1_f32(c0, vacc0x01); c0 += 2;
220
221 vacc3x01 = vget_high_f32(vacc3x0123);
222 vacc2x01 = vget_high_f32(vacc2x0123);
223 vacc1x01 = vget_high_f32(vacc1x0123);
224 vacc0x01 = vget_high_f32(vacc0x0123);
225 }
226 if (nc & 1) {
227 vst1_lane_f32(c3, vacc3x01, 0);
228 vst1_lane_f32(c2, vacc2x01, 0);
229 vst1_lane_f32(c1, vacc1x01, 0);
230 vst1_lane_f32(c0, vacc0x01, 0);
231 }
232
233 nc = 0;
234 }
235 } while (nc != 0);
236 }