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_2x16(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_2x16(
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
36 const size_t input_increment = input_stride * 2 - channels;
37 const size_t output_increment = output_stride * 2 - channels;
38
39 do {
40 if XNN_UNPREDICTABLE(rows < 2) {
41 i1 = i0;
42 o1 = o0;
43 }
44
45 const float* w = weights;
46 size_t c = channels;
47 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
48 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
49 const float32x4_t vw4567 = vld1q_f32(w); w += 4;
50 const float32x4_t vw89AB = vld1q_f32(w); w += 4;
51 const float32x4_t vwCDEF = vld1q_f32(w); w += 4;
52
53 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
54 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
55 const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
56 const float32x4_t vi0xCDEF = vld1q_f32(i0); i0 += 4;
57 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
58 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
59 const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
60 const float32x4_t vi1xCDEF = vld1q_f32(i1); i1 += 4;
61
62 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
63 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
64 float32x4_t vacc0x4567 = vmulq_f32(vi0x4567, vw4567);
65 const uint32x4_t vm0x4567 = vcltq_s32(vreinterpretq_s32_f32(vi0x4567), vmovq_n_s32(0));
66 float32x4_t vacc0x89AB = vmulq_f32(vi0x89AB, vw89AB);
67 const uint32x4_t vm0x89AB = vcltq_s32(vreinterpretq_s32_f32(vi0x89AB), vmovq_n_s32(0));
68 float32x4_t vacc0xCDEF = vmulq_f32(vi0xCDEF, vwCDEF);
69 const uint32x4_t vm0xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi0xCDEF), vmovq_n_s32(0));
70 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
71 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
72 float32x4_t vacc1x4567 = vmulq_f32(vi1x4567, vw4567);
73 const uint32x4_t vm1x4567 = vcltq_s32(vreinterpretq_s32_f32(vi1x4567), vmovq_n_s32(0));
74 float32x4_t vacc1x89AB = vmulq_f32(vi1x89AB, vw89AB);
75 const uint32x4_t vm1x89AB = vcltq_s32(vreinterpretq_s32_f32(vi1x89AB), vmovq_n_s32(0));
76 float32x4_t vacc1xCDEF = vmulq_f32(vi1xCDEF, vwCDEF);
77 const uint32x4_t vm1xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi1xCDEF), vmovq_n_s32(0));
78
79 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
80 vacc0x4567 = vbslq_f32(vm0x4567, vacc0x4567, vi0x4567);
81 vacc0x89AB = vbslq_f32(vm0x89AB, vacc0x89AB, vi0x89AB);
82 vacc0xCDEF = vbslq_f32(vm0xCDEF, vacc0xCDEF, vi0xCDEF);
83 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
84 vacc1x4567 = vbslq_f32(vm1x4567, vacc1x4567, vi1x4567);
85 vacc1x89AB = vbslq_f32(vm1x89AB, vacc1x89AB, vi1x89AB);
86 vacc1xCDEF = vbslq_f32(vm1xCDEF, vacc1xCDEF, vi1xCDEF);
87
88 vst1q_f32(o0, vacc0x0123); o0 += 4;
89 vst1q_f32(o0, vacc0x4567); o0 += 4;
90 vst1q_f32(o0, vacc0x89AB); o0 += 4;
91 vst1q_f32(o0, vacc0xCDEF); o0 += 4;
92 vst1q_f32(o1, vacc1x0123); o1 += 4;
93 vst1q_f32(o1, vacc1x4567); o1 += 4;
94 vst1q_f32(o1, vacc1x89AB); o1 += 4;
95 vst1q_f32(o1, vacc1xCDEF); o1 += 4;
96 }
97 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
98 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
99
100 const float32x4_t vi0x0123 = vld1q_f32(i0);
101 i0 += 4;
102 const float32x4_t vi1x0123 = vld1q_f32(i1);
103 i1 += 4;
104
105 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
106 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
107 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
108 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
109
110 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
111 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
112
113 vst1q_f32(o0, vacc0x0123); o0 += 4;
114 vst1q_f32(o1, vacc1x0123); o1 += 4;
115 }
116 if XNN_UNLIKELY(c != 0) {
117 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
118
119 const float32x4_t vi0x0123 = vld1q_f32(i0);
120 i0 = (const float*) ((uintptr_t) i0 + c);
121 const float32x4_t vi1x0123 = vld1q_f32(i1);
122 i1 = (const float*) ((uintptr_t) i1 + c);
123
124 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
125 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
126 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
127 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
128
129 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
130 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
131
132 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
133 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
134 if (c & (2 * sizeof(float))) {
135 vst1_f32(o0, vacc0x01); o0 += 2;
136 vst1_f32(o1, vacc1x01); o1 += 2;
137
138 vacc0x01 = vget_high_f32(vacc0x0123);
139 vacc1x01 = vget_high_f32(vacc1x0123);
140 }
141 if (c & (1 * sizeof(float))) {
142 vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
143 vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
144 }
145 }
146 i0 = (const float*) ((uintptr_t) i0 + input_increment);
147 o0 = (float*) ((uintptr_t) o0 + output_increment);
148 i1 = (const float*) ((uintptr_t) i1 + input_increment);
149 o1 = (float*) ((uintptr_t) o1 + output_increment);
150 rows = doz(rows, 2);
151 } while (rows != 0);
152 }
153