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_1x16(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_1x16(
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
34 const size_t input_increment = input_stride * 1 - channels;
35 const size_t output_increment = output_stride * 1 - channels;
36
37 do {
38
39 const float* w = weights;
40 size_t c = channels;
41 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
42 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
43 const float32x4_t vw4567 = vld1q_f32(w); w += 4;
44 const float32x4_t vw89AB = vld1q_f32(w); w += 4;
45 const float32x4_t vwCDEF = vld1q_f32(w); w += 4;
46
47 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
48 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
49 const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
50 const float32x4_t vi0xCDEF = vld1q_f32(i0); i0 += 4;
51
52 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
53 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
54 float32x4_t vacc0x4567 = vmulq_f32(vi0x4567, vw4567);
55 const uint32x4_t vm0x4567 = vcltq_s32(vreinterpretq_s32_f32(vi0x4567), vmovq_n_s32(0));
56 float32x4_t vacc0x89AB = vmulq_f32(vi0x89AB, vw89AB);
57 const uint32x4_t vm0x89AB = vcltq_s32(vreinterpretq_s32_f32(vi0x89AB), vmovq_n_s32(0));
58 float32x4_t vacc0xCDEF = vmulq_f32(vi0xCDEF, vwCDEF);
59 const uint32x4_t vm0xCDEF = vcltq_s32(vreinterpretq_s32_f32(vi0xCDEF), vmovq_n_s32(0));
60
61 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
62 vacc0x4567 = vbslq_f32(vm0x4567, vacc0x4567, vi0x4567);
63 vacc0x89AB = vbslq_f32(vm0x89AB, vacc0x89AB, vi0x89AB);
64 vacc0xCDEF = vbslq_f32(vm0xCDEF, vacc0xCDEF, vi0xCDEF);
65
66 vst1q_f32(o0, vacc0x0123); o0 += 4;
67 vst1q_f32(o0, vacc0x4567); o0 += 4;
68 vst1q_f32(o0, vacc0x89AB); o0 += 4;
69 vst1q_f32(o0, vacc0xCDEF); o0 += 4;
70 }
71 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
72 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
73
74 const float32x4_t vi0x0123 = vld1q_f32(i0);
75 i0 += 4;
76
77 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
78 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
79
80 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
81
82 vst1q_f32(o0, vacc0x0123); o0 += 4;
83 }
84 if XNN_UNLIKELY(c != 0) {
85 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
86
87 const float32x4_t vi0x0123 = vld1q_f32(i0);
88 i0 = (const float*) ((uintptr_t) i0 + c);
89
90 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
91 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
92
93 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
94
95 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
96 if (c & (2 * sizeof(float))) {
97 vst1_f32(o0, vacc0x01); o0 += 2;
98
99 vacc0x01 = vget_high_f32(vacc0x0123);
100 }
101 if (c & (1 * sizeof(float))) {
102 vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
103 }
104 }
105 i0 = (const float*) ((uintptr_t) i0 + input_increment);
106 o0 = (float*) ((uintptr_t) o0 + output_increment);
107 rows = doz(rows, 1);
108 } while (rows != 0);
109 }
110