Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x16-minmax-neon.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
17
18 void xnn_f32_ppmm_minmax_ukernel_4x16__neon(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const float* restrict a,
23 const float* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34
35 #if XNN_ARCH_ARM64
36 const float32x4x2_t vminmax = vld2q_dup_f32(&params->scalar.min);
37 const float32x4_t vmin = vminmax.val[0];
38 const float32x4_t vmax = vminmax.val[1];
39 #else
40 const float32x2x2_t vminmax = vld2_dup_f32(&params->scalar.min);
41 const float32x4_t vmin = vcombine_f32(vminmax.val[0], vminmax.val[0]);
42 const float32x4_t vmax = vcombine_f32(vminmax.val[1], vminmax.val[1]);
43 #endif
44
45 float* c0 = c;
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59
60 do {
61 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
62 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
63 float32x4_t vacc0x89AB = vld1q_f32(w); w += 4;
64 float32x4_t vacc0xCDEF = vld1q_f32(w); w += 4;
65 float32x4_t vacc1x0123 = vacc0x0123;
66 float32x4_t vacc1x4567 = vacc0x4567;
67 float32x4_t vacc1x89AB = vacc0x89AB;
68 float32x4_t vacc1xCDEF = vacc0xCDEF;
69 float32x4_t vacc2x0123 = vacc0x0123;
70 float32x4_t vacc2x4567 = vacc0x4567;
71 float32x4_t vacc2x89AB = vacc0x89AB;
72 float32x4_t vacc2xCDEF = vacc0xCDEF;
73 float32x4_t vacc3x0123 = vacc0x0123;
74 float32x4_t vacc3x4567 = vacc0x4567;
75 float32x4_t vacc3x89AB = vacc0x89AB;
76 float32x4_t vacc3xCDEF = vacc0xCDEF;
77
78 size_t k = kc;
79 do {
80 const float32x4_t va0123 = vld1q_f32(a); a += 4;
81
82 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
83 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
84 const float32x4_t vb89AB = vld1q_f32(w); w += 4;
85 const float32x4_t vbCDEF = vld1q_f32(w); w += 4;
86
87 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123, vget_low_f32(va0123), 0);
88 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123, vget_low_f32(va0123), 1);
89 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123, vget_high_f32(va0123), 0);
90 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123, vget_high_f32(va0123), 1);
91 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567, vget_low_f32(va0123), 0);
92 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567, vget_low_f32(va0123), 1);
93 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567, vget_high_f32(va0123), 0);
94 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567, vget_high_f32(va0123), 1);
95 vacc0x89AB = vmlaq_lane_f32(vacc0x89AB, vb89AB, vget_low_f32(va0123), 0);
96 vacc1x89AB = vmlaq_lane_f32(vacc1x89AB, vb89AB, vget_low_f32(va0123), 1);
97 vacc2x89AB = vmlaq_lane_f32(vacc2x89AB, vb89AB, vget_high_f32(va0123), 0);
98 vacc3x89AB = vmlaq_lane_f32(vacc3x89AB, vb89AB, vget_high_f32(va0123), 1);
99 vacc0xCDEF = vmlaq_lane_f32(vacc0xCDEF, vbCDEF, vget_low_f32(va0123), 0);
100 vacc1xCDEF = vmlaq_lane_f32(vacc1xCDEF, vbCDEF, vget_low_f32(va0123), 1);
101 vacc2xCDEF = vmlaq_lane_f32(vacc2xCDEF, vbCDEF, vget_high_f32(va0123), 0);
102 vacc3xCDEF = vmlaq_lane_f32(vacc3xCDEF, vbCDEF, vget_high_f32(va0123), 1);
103
104 k -= sizeof(float);
105 } while (k != 0);
106
107 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
108 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
109 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
110 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
111 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
112 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
113 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
114 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
115 vacc0x89AB = vminq_f32(vacc0x89AB, vmax);
116 vacc1x89AB = vminq_f32(vacc1x89AB, vmax);
117 vacc2x89AB = vminq_f32(vacc2x89AB, vmax);
118 vacc3x89AB = vminq_f32(vacc3x89AB, vmax);
119 vacc0xCDEF = vminq_f32(vacc0xCDEF, vmax);
120 vacc1xCDEF = vminq_f32(vacc1xCDEF, vmax);
121 vacc2xCDEF = vminq_f32(vacc2xCDEF, vmax);
122 vacc3xCDEF = vminq_f32(vacc3xCDEF, vmax);
123
124 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
125 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
126 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
127 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
128 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
129 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
130 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
131 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
132 vacc0x89AB = vmaxq_f32(vacc0x89AB, vmin);
133 vacc1x89AB = vmaxq_f32(vacc1x89AB, vmin);
134 vacc2x89AB = vmaxq_f32(vacc2x89AB, vmin);
135 vacc3x89AB = vmaxq_f32(vacc3x89AB, vmin);
136 vacc0xCDEF = vmaxq_f32(vacc0xCDEF, vmin);
137 vacc1xCDEF = vmaxq_f32(vacc1xCDEF, vmin);
138 vacc2xCDEF = vmaxq_f32(vacc2xCDEF, vmin);
139 vacc3xCDEF = vmaxq_f32(vacc3xCDEF, vmin);
140
141 if XNN_LIKELY(nc >= 16) {
142 vst1q_f32(c3, vacc3x0123);
143 vst1q_f32(c3 + 4, vacc3x4567);
144 vst1q_f32(c3 + 8, vacc3x89AB);
145 vst1q_f32(c3 + 12, vacc3xCDEF);
146 c3 = (float*) ((uintptr_t) c3 + cn_stride);
147 vst1q_f32(c2, vacc2x0123);
148 vst1q_f32(c2 + 4, vacc2x4567);
149 vst1q_f32(c2 + 8, vacc2x89AB);
150 vst1q_f32(c2 + 12, vacc2xCDEF);
151 c2 = (float*) ((uintptr_t) c2 + cn_stride);
152 vst1q_f32(c1, vacc1x0123);
153 vst1q_f32(c1 + 4, vacc1x4567);
154 vst1q_f32(c1 + 8, vacc1x89AB);
155 vst1q_f32(c1 + 12, vacc1xCDEF);
156 c1 = (float*) ((uintptr_t) c1 + cn_stride);
157 vst1q_f32(c0, vacc0x0123);
158 vst1q_f32(c0 + 4, vacc0x4567);
159 vst1q_f32(c0 + 8, vacc0x89AB);
160 vst1q_f32(c0 + 12, vacc0xCDEF);
161 c0 = (float*) ((uintptr_t) c0 + cn_stride);
162
163 a = (const float*) ((uintptr_t) a - kc * 4);
164
165 nc -= 16;
166 } else {
167 if (nc & 8) {
168 vst1q_f32(c3, vacc3x0123); c3 += 4;
169 vst1q_f32(c3, vacc3x4567); c3 += 4;
170 vst1q_f32(c2, vacc2x0123); c2 += 4;
171 vst1q_f32(c2, vacc2x4567); c2 += 4;
172 vst1q_f32(c1, vacc1x0123); c1 += 4;
173 vst1q_f32(c1, vacc1x4567); c1 += 4;
174 vst1q_f32(c0, vacc0x0123); c0 += 4;
175 vst1q_f32(c0, vacc0x4567); c0 += 4;
176
177 vacc3x0123 = vacc3x89AB;
178 vacc3x4567 = vacc3xCDEF;
179 vacc2x0123 = vacc2x89AB;
180 vacc2x4567 = vacc2xCDEF;
181 vacc1x0123 = vacc1x89AB;
182 vacc1x4567 = vacc1xCDEF;
183 vacc0x0123 = vacc0x89AB;
184 vacc0x4567 = vacc0xCDEF;
185 }
186 if (nc & 4) {
187 vst1q_f32(c3, vacc3x0123); c3 += 4;
188 vst1q_f32(c2, vacc2x0123); c2 += 4;
189 vst1q_f32(c1, vacc1x0123); c1 += 4;
190 vst1q_f32(c0, vacc0x0123); c0 += 4;
191
192 vacc3x0123 = vacc3x4567;
193 vacc3x4567 = vacc3x89AB;
194 vacc3x89AB = vacc3xCDEF;
195 vacc2x0123 = vacc2x4567;
196 vacc2x4567 = vacc2x89AB;
197 vacc2x89AB = vacc2xCDEF;
198 vacc1x0123 = vacc1x4567;
199 vacc1x4567 = vacc1x89AB;
200 vacc1x89AB = vacc1xCDEF;
201 vacc0x0123 = vacc0x4567;
202 vacc0x4567 = vacc0x89AB;
203 vacc0x89AB = vacc0xCDEF;
204 }
205 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
206 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
207 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
208 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
209 if (nc & 2) {
210 vst1_f32(c3, vacc3x01); c3 += 2;
211 vst1_f32(c2, vacc2x01); c2 += 2;
212 vst1_f32(c1, vacc1x01); c1 += 2;
213 vst1_f32(c0, vacc0x01); c0 += 2;
214
215 vacc3x01 = vget_high_f32(vacc3x0123);
216 vacc2x01 = vget_high_f32(vacc2x0123);
217 vacc1x01 = vget_high_f32(vacc1x0123);
218 vacc0x01 = vget_high_f32(vacc0x0123);
219 }
220 if (nc & 1) {
221 vst1_lane_f32(c3, vacc3x01, 0);
222 vst1_lane_f32(c2, vacc2x01, 0);
223 vst1_lane_f32(c1, vacc1x01, 0);
224 vst1_lane_f32(c0, vacc0x01, 0);
225 }
226
227 nc = 0;
228 }
229 } while (nc != 0);
230 }