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_4x8(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_4x8(
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 >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
60 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
61 const float32x4_t vw4567 = vld1q_f32(w); w += 4;
62
63 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
64 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
65 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
66 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
67 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
68 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
69 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
70 const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
71
72 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
73 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
74 float32x4_t vacc0x4567 = vmulq_f32(vi0x4567, vw4567);
75 const uint32x4_t vm0x4567 = vcltq_s32(vreinterpretq_s32_f32(vi0x4567), vmovq_n_s32(0));
76 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
77 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
78 float32x4_t vacc1x4567 = vmulq_f32(vi1x4567, vw4567);
79 const uint32x4_t vm1x4567 = vcltq_s32(vreinterpretq_s32_f32(vi1x4567), vmovq_n_s32(0));
80 float32x4_t vacc2x0123 = vmulq_f32(vi2x0123, vw0123);
81 const uint32x4_t vm2x0123 = vcltq_s32(vreinterpretq_s32_f32(vi2x0123), vmovq_n_s32(0));
82 float32x4_t vacc2x4567 = vmulq_f32(vi2x4567, vw4567);
83 const uint32x4_t vm2x4567 = vcltq_s32(vreinterpretq_s32_f32(vi2x4567), vmovq_n_s32(0));
84 float32x4_t vacc3x0123 = vmulq_f32(vi3x0123, vw0123);
85 const uint32x4_t vm3x0123 = vcltq_s32(vreinterpretq_s32_f32(vi3x0123), vmovq_n_s32(0));
86 float32x4_t vacc3x4567 = vmulq_f32(vi3x4567, vw4567);
87 const uint32x4_t vm3x4567 = vcltq_s32(vreinterpretq_s32_f32(vi3x4567), vmovq_n_s32(0));
88
89 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
90 vacc0x4567 = vbslq_f32(vm0x4567, vacc0x4567, vi0x4567);
91 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
92 vacc1x4567 = vbslq_f32(vm1x4567, vacc1x4567, vi1x4567);
93 vacc2x0123 = vbslq_f32(vm2x0123, vacc2x0123, vi2x0123);
94 vacc2x4567 = vbslq_f32(vm2x4567, vacc2x4567, vi2x4567);
95 vacc3x0123 = vbslq_f32(vm3x0123, vacc3x0123, vi3x0123);
96 vacc3x4567 = vbslq_f32(vm3x4567, vacc3x4567, vi3x4567);
97
98 vst1q_f32(o0, vacc0x0123); o0 += 4;
99 vst1q_f32(o0, vacc0x4567); o0 += 4;
100 vst1q_f32(o1, vacc1x0123); o1 += 4;
101 vst1q_f32(o1, vacc1x4567); o1 += 4;
102 vst1q_f32(o2, vacc2x0123); o2 += 4;
103 vst1q_f32(o2, vacc2x4567); o2 += 4;
104 vst1q_f32(o3, vacc3x0123); o3 += 4;
105 vst1q_f32(o3, vacc3x4567); o3 += 4;
106 }
107 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
108 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
109
110 const float32x4_t vi0x0123 = vld1q_f32(i0);
111 i0 += 4;
112 const float32x4_t vi1x0123 = vld1q_f32(i1);
113 i1 += 4;
114 const float32x4_t vi2x0123 = vld1q_f32(i2);
115 i2 += 4;
116 const float32x4_t vi3x0123 = vld1q_f32(i3);
117 i3 += 4;
118
119 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
120 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
121 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
122 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
123 float32x4_t vacc2x0123 = vmulq_f32(vi2x0123, vw0123);
124 const uint32x4_t vm2x0123 = vcltq_s32(vreinterpretq_s32_f32(vi2x0123), vmovq_n_s32(0));
125 float32x4_t vacc3x0123 = vmulq_f32(vi3x0123, vw0123);
126 const uint32x4_t vm3x0123 = vcltq_s32(vreinterpretq_s32_f32(vi3x0123), vmovq_n_s32(0));
127
128 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
129 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
130 vacc2x0123 = vbslq_f32(vm2x0123, vacc2x0123, vi2x0123);
131 vacc3x0123 = vbslq_f32(vm3x0123, vacc3x0123, vi3x0123);
132
133 vst1q_f32(o0, vacc0x0123); o0 += 4;
134 vst1q_f32(o1, vacc1x0123); o1 += 4;
135 vst1q_f32(o2, vacc2x0123); o2 += 4;
136 vst1q_f32(o3, vacc3x0123); o3 += 4;
137 }
138 if XNN_UNLIKELY(c != 0) {
139 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
140
141 const float32x4_t vi0x0123 = vld1q_f32(i0);
142 i0 = (const float*) ((uintptr_t) i0 + c);
143 const float32x4_t vi1x0123 = vld1q_f32(i1);
144 i1 = (const float*) ((uintptr_t) i1 + c);
145 const float32x4_t vi2x0123 = vld1q_f32(i2);
146 i2 = (const float*) ((uintptr_t) i2 + c);
147 const float32x4_t vi3x0123 = vld1q_f32(i3);
148 i3 = (const float*) ((uintptr_t) i3 + c);
149
150 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
151 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
152 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
153 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
154 float32x4_t vacc2x0123 = vmulq_f32(vi2x0123, vw0123);
155 const uint32x4_t vm2x0123 = vcltq_s32(vreinterpretq_s32_f32(vi2x0123), vmovq_n_s32(0));
156 float32x4_t vacc3x0123 = vmulq_f32(vi3x0123, vw0123);
157 const uint32x4_t vm3x0123 = vcltq_s32(vreinterpretq_s32_f32(vi3x0123), vmovq_n_s32(0));
158
159 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
160 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
161 vacc2x0123 = vbslq_f32(vm2x0123, vacc2x0123, vi2x0123);
162 vacc3x0123 = vbslq_f32(vm3x0123, vacc3x0123, vi3x0123);
163
164 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
165 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
166 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
167 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
168 if (c & (2 * sizeof(float))) {
169 vst1_f32(o0, vacc0x01); o0 += 2;
170 vst1_f32(o1, vacc1x01); o1 += 2;
171 vst1_f32(o2, vacc2x01); o2 += 2;
172 vst1_f32(o3, vacc3x01); o3 += 2;
173
174 vacc0x01 = vget_high_f32(vacc0x0123);
175 vacc1x01 = vget_high_f32(vacc1x0123);
176 vacc2x01 = vget_high_f32(vacc2x0123);
177 vacc3x01 = vget_high_f32(vacc3x0123);
178 }
179 if (c & (1 * sizeof(float))) {
180 vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
181 vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
182 vst1_lane_f32(o2, vacc2x01, 0); o2 += 1;
183 vst1_lane_f32(o3, vacc3x01, 0); o3 += 1;
184 }
185 }
186 i0 = (const float*) ((uintptr_t) i0 + input_increment);
187 o0 = (float*) ((uintptr_t) o0 + output_increment);
188 i1 = (const float*) ((uintptr_t) i1 + input_increment);
189 o1 = (float*) ((uintptr_t) o1 + output_increment);
190 i2 = (const float*) ((uintptr_t) i2 + input_increment);
191 o2 = (float*) ((uintptr_t) o2 + output_increment);
192 i3 = (const float*) ((uintptr_t) i3 + input_increment);
193 o3 = (float*) ((uintptr_t) o3 + output_increment);
194 rows = doz(rows, 4);
195 } while (rows != 0);
196 }
197