1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/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/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__neon_4x16(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride)18 void xnn_f32_prelu_ukernel__neon_4x16(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(channels != 0);
29 assert(channels % sizeof(float) == 0);
30
31 const float* i0 = input;
32 float* o0 = output;
33 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
34 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
35 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
36 float* o2 = (float*) ((uintptr_t) o1 + output_stride);
37 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
38 float* o3 = (float*) ((uintptr_t) o2 + output_stride);
39
40 const size_t input_increment = input_stride * 4 - channels;
41 const size_t output_increment = output_stride * 4 - channels;
42
43 do {
44 if XNN_UNPREDICTABLE(rows < 2) {
45 i1 = i0;
46 o1 = o0;
47 }
48 if XNN_UNPREDICTABLE(rows <= 2) {
49 i2 = i1;
50 o2 = o1;
51 }
52 if XNN_UNPREDICTABLE(rows < 4) {
53 i3 = i2;
54 o3 = o2;
55 }
56
57 const float* w = weights;
58 size_t c = channels;
59 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
60 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
61 const float32x4_t vw4567 = vld1q_f32(w); w += 4;
62 const float32x4_t vw89AB = vld1q_f32(w); w += 4;
63 const float32x4_t vwCDEF = vld1q_f32(w); w += 4;
64
65 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
66 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
67 const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
68 const float32x4_t vi0xCDEF = vld1q_f32(i0); i0 += 4;
69 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
70 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
71 const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
72 const float32x4_t vi1xCDEF = vld1q_f32(i1); i1 += 4;
73 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
74 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
75 const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
76 const float32x4_t vi2xCDEF = vld1q_f32(i2); i2 += 4;
77 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
78 const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
79 const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
80 const float32x4_t vi3xCDEF = vld1q_f32(i3); i3 += 4;
81
82 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
83 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
84 float32x4_t vacc0x4567 = vmulq_f32(vi0x4567, vw4567);
85 const uint32x4_t vm0x4567 = vcltq_s32(vreinterpretq_s32_f32(vi0x4567), vmovq_n_s32(0));
86 float32x4_t vacc0x89AB = vmulq_f32(vi0x89AB, vw89AB);
87 const uint32x4_t vm0x89AB = vcltq_s32(vreinterpretq_s32_f32(vi0x89AB), vmovq_n_s32(0));
88 float32x4_t vacc0xCDEF = vmulq_f32(vi0xCDEF, vwCDEF);
89 const uint32x4_t vm0xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi0xCDEF), vmovq_n_s32(0));
90 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
91 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
92 float32x4_t vacc1x4567 = vmulq_f32(vi1x4567, vw4567);
93 const uint32x4_t vm1x4567 = vcltq_s32(vreinterpretq_s32_f32(vi1x4567), vmovq_n_s32(0));
94 float32x4_t vacc1x89AB = vmulq_f32(vi1x89AB, vw89AB);
95 const uint32x4_t vm1x89AB = vcltq_s32(vreinterpretq_s32_f32(vi1x89AB), vmovq_n_s32(0));
96 float32x4_t vacc1xCDEF = vmulq_f32(vi1xCDEF, vwCDEF);
97 const uint32x4_t vm1xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi1xCDEF), vmovq_n_s32(0));
98 float32x4_t vacc2x0123 = vmulq_f32(vi2x0123, vw0123);
99 const uint32x4_t vm2x0123 = vcltq_s32(vreinterpretq_s32_f32(vi2x0123), vmovq_n_s32(0));
100 float32x4_t vacc2x4567 = vmulq_f32(vi2x4567, vw4567);
101 const uint32x4_t vm2x4567 = vcltq_s32(vreinterpretq_s32_f32(vi2x4567), vmovq_n_s32(0));
102 float32x4_t vacc2x89AB = vmulq_f32(vi2x89AB, vw89AB);
103 const uint32x4_t vm2x89AB = vcltq_s32(vreinterpretq_s32_f32(vi2x89AB), vmovq_n_s32(0));
104 float32x4_t vacc2xCDEF = vmulq_f32(vi2xCDEF, vwCDEF);
105 const uint32x4_t vm2xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi2xCDEF), vmovq_n_s32(0));
106 float32x4_t vacc3x0123 = vmulq_f32(vi3x0123, vw0123);
107 const uint32x4_t vm3x0123 = vcltq_s32(vreinterpretq_s32_f32(vi3x0123), vmovq_n_s32(0));
108 float32x4_t vacc3x4567 = vmulq_f32(vi3x4567, vw4567);
109 const uint32x4_t vm3x4567 = vcltq_s32(vreinterpretq_s32_f32(vi3x4567), vmovq_n_s32(0));
110 float32x4_t vacc3x89AB = vmulq_f32(vi3x89AB, vw89AB);
111 const uint32x4_t vm3x89AB = vcltq_s32(vreinterpretq_s32_f32(vi3x89AB), vmovq_n_s32(0));
112 float32x4_t vacc3xCDEF = vmulq_f32(vi3xCDEF, vwCDEF);
113 const uint32x4_t vm3xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi3xCDEF), vmovq_n_s32(0));
114
115 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
116 vacc0x4567 = vbslq_f32(vm0x4567, vacc0x4567, vi0x4567);
117 vacc0x89AB = vbslq_f32(vm0x89AB, vacc0x89AB, vi0x89AB);
118 vacc0xCDEF = vbslq_f32(vm0xCDEF, vacc0xCDEF, vi0xCDEF);
119 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
120 vacc1x4567 = vbslq_f32(vm1x4567, vacc1x4567, vi1x4567);
121 vacc1x89AB = vbslq_f32(vm1x89AB, vacc1x89AB, vi1x89AB);
122 vacc1xCDEF = vbslq_f32(vm1xCDEF, vacc1xCDEF, vi1xCDEF);
123 vacc2x0123 = vbslq_f32(vm2x0123, vacc2x0123, vi2x0123);
124 vacc2x4567 = vbslq_f32(vm2x4567, vacc2x4567, vi2x4567);
125 vacc2x89AB = vbslq_f32(vm2x89AB, vacc2x89AB, vi2x89AB);
126 vacc2xCDEF = vbslq_f32(vm2xCDEF, vacc2xCDEF, vi2xCDEF);
127 vacc3x0123 = vbslq_f32(vm3x0123, vacc3x0123, vi3x0123);
128 vacc3x4567 = vbslq_f32(vm3x4567, vacc3x4567, vi3x4567);
129 vacc3x89AB = vbslq_f32(vm3x89AB, vacc3x89AB, vi3x89AB);
130 vacc3xCDEF = vbslq_f32(vm3xCDEF, vacc3xCDEF, vi3xCDEF);
131
132 vst1q_f32(o0, vacc0x0123); o0 += 4;
133 vst1q_f32(o0, vacc0x4567); o0 += 4;
134 vst1q_f32(o0, vacc0x89AB); o0 += 4;
135 vst1q_f32(o0, vacc0xCDEF); o0 += 4;
136 vst1q_f32(o1, vacc1x0123); o1 += 4;
137 vst1q_f32(o1, vacc1x4567); o1 += 4;
138 vst1q_f32(o1, vacc1x89AB); o1 += 4;
139 vst1q_f32(o1, vacc1xCDEF); o1 += 4;
140 vst1q_f32(o2, vacc2x0123); o2 += 4;
141 vst1q_f32(o2, vacc2x4567); o2 += 4;
142 vst1q_f32(o2, vacc2x89AB); o2 += 4;
143 vst1q_f32(o2, vacc2xCDEF); o2 += 4;
144 vst1q_f32(o3, vacc3x0123); o3 += 4;
145 vst1q_f32(o3, vacc3x4567); o3 += 4;
146 vst1q_f32(o3, vacc3x89AB); o3 += 4;
147 vst1q_f32(o3, vacc3xCDEF); o3 += 4;
148 }
149 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
150 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
151
152 const float32x4_t vi0x0123 = vld1q_f32(i0);
153 i0 += 4;
154 const float32x4_t vi1x0123 = vld1q_f32(i1);
155 i1 += 4;
156 const float32x4_t vi2x0123 = vld1q_f32(i2);
157 i2 += 4;
158 const float32x4_t vi3x0123 = vld1q_f32(i3);
159 i3 += 4;
160
161 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
162 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
163 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
164 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
165 float32x4_t vacc2x0123 = vmulq_f32(vi2x0123, vw0123);
166 const uint32x4_t vm2x0123 = vcltq_s32(vreinterpretq_s32_f32(vi2x0123), vmovq_n_s32(0));
167 float32x4_t vacc3x0123 = vmulq_f32(vi3x0123, vw0123);
168 const uint32x4_t vm3x0123 = vcltq_s32(vreinterpretq_s32_f32(vi3x0123), vmovq_n_s32(0));
169
170 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
171 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
172 vacc2x0123 = vbslq_f32(vm2x0123, vacc2x0123, vi2x0123);
173 vacc3x0123 = vbslq_f32(vm3x0123, vacc3x0123, vi3x0123);
174
175 vst1q_f32(o0, vacc0x0123); o0 += 4;
176 vst1q_f32(o1, vacc1x0123); o1 += 4;
177 vst1q_f32(o2, vacc2x0123); o2 += 4;
178 vst1q_f32(o3, vacc3x0123); o3 += 4;
179 }
180 if XNN_UNLIKELY(c != 0) {
181 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
182
183 const float32x4_t vi0x0123 = vld1q_f32(i0);
184 i0 = (const float*) ((uintptr_t) i0 + c);
185 const float32x4_t vi1x0123 = vld1q_f32(i1);
186 i1 = (const float*) ((uintptr_t) i1 + c);
187 const float32x4_t vi2x0123 = vld1q_f32(i2);
188 i2 = (const float*) ((uintptr_t) i2 + c);
189 const float32x4_t vi3x0123 = vld1q_f32(i3);
190 i3 = (const float*) ((uintptr_t) i3 + c);
191
192 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
193 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
194 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
195 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
196 float32x4_t vacc2x0123 = vmulq_f32(vi2x0123, vw0123);
197 const uint32x4_t vm2x0123 = vcltq_s32(vreinterpretq_s32_f32(vi2x0123), vmovq_n_s32(0));
198 float32x4_t vacc3x0123 = vmulq_f32(vi3x0123, vw0123);
199 const uint32x4_t vm3x0123 = vcltq_s32(vreinterpretq_s32_f32(vi3x0123), vmovq_n_s32(0));
200
201 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
202 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
203 vacc2x0123 = vbslq_f32(vm2x0123, vacc2x0123, vi2x0123);
204 vacc3x0123 = vbslq_f32(vm3x0123, vacc3x0123, vi3x0123);
205
206 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
207 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
208 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
209 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
210 if (c & (2 * sizeof(float))) {
211 vst1_f32(o0, vacc0x01); o0 += 2;
212 vst1_f32(o1, vacc1x01); o1 += 2;
213 vst1_f32(o2, vacc2x01); o2 += 2;
214 vst1_f32(o3, vacc3x01); o3 += 2;
215
216 vacc0x01 = vget_high_f32(vacc0x0123);
217 vacc1x01 = vget_high_f32(vacc1x0123);
218 vacc2x01 = vget_high_f32(vacc2x0123);
219 vacc3x01 = vget_high_f32(vacc3x0123);
220 }
221 if (c & (1 * sizeof(float))) {
222 vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
223 vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
224 vst1_lane_f32(o2, vacc2x01, 0); o2 += 1;
225 vst1_lane_f32(o3, vacc3x01, 0); o3 += 1;
226 }
227 }
228 i0 = (const float*) ((uintptr_t) i0 + input_increment);
229 o0 = (float*) ((uintptr_t) o0 + output_increment);
230 i1 = (const float*) ((uintptr_t) i1 + input_increment);
231 o1 = (float*) ((uintptr_t) o1 + output_increment);
232 i2 = (const float*) ((uintptr_t) i2 + input_increment);
233 o2 = (float*) ((uintptr_t) o2 + output_increment);
234 i3 = (const float*) ((uintptr_t) i3 + input_increment);
235 o3 = (float*) ((uintptr_t) o3 + output_increment);
236 rows = doz(rows, 4);
237 } while (rows != 0);
238 }
239