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_2x8(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_2x8(
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 >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
48 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
49 const float32x4_t vw4567 = vld1q_f32(w); w += 4;
50
51 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
52 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
53 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
54 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
55
56 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
57 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
58 float32x4_t vacc0x4567 = vmulq_f32(vi0x4567, vw4567);
59 const uint32x4_t vm0x4567 = vcltq_s32(vreinterpretq_s32_f32(vi0x4567), vmovq_n_s32(0));
60 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
61 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
62 float32x4_t vacc1x4567 = vmulq_f32(vi1x4567, vw4567);
63 const uint32x4_t vm1x4567 = vcltq_s32(vreinterpretq_s32_f32(vi1x4567), vmovq_n_s32(0));
64
65 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
66 vacc0x4567 = vbslq_f32(vm0x4567, vacc0x4567, vi0x4567);
67 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
68 vacc1x4567 = vbslq_f32(vm1x4567, vacc1x4567, vi1x4567);
69
70 vst1q_f32(o0, vacc0x0123); o0 += 4;
71 vst1q_f32(o0, vacc0x4567); o0 += 4;
72 vst1q_f32(o1, vacc1x0123); o1 += 4;
73 vst1q_f32(o1, vacc1x4567); o1 += 4;
74 }
75 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
76 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
77
78 const float32x4_t vi0x0123 = vld1q_f32(i0);
79 i0 += 4;
80 const float32x4_t vi1x0123 = vld1q_f32(i1);
81 i1 += 4;
82
83 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
84 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
85 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
86 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
87
88 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
89 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
90
91 vst1q_f32(o0, vacc0x0123); o0 += 4;
92 vst1q_f32(o1, vacc1x0123); o1 += 4;
93 }
94 if XNN_UNLIKELY(c != 0) {
95 const float32x4_t vw0123 = vld1q_f32(w); w += 4;
96
97 const float32x4_t vi0x0123 = vld1q_f32(i0);
98 i0 = (const float*) ((uintptr_t) i0 + c);
99 const float32x4_t vi1x0123 = vld1q_f32(i1);
100 i1 = (const float*) ((uintptr_t) i1 + c);
101
102 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
103 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
104 float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
105 const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
106
107 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
108 vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
109
110 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
111 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
112 if (c & (2 * sizeof(float))) {
113 vst1_f32(o0, vacc0x01); o0 += 2;
114 vst1_f32(o1, vacc1x01); o1 += 2;
115
116 vacc0x01 = vget_high_f32(vacc0x0123);
117 vacc1x01 = vget_high_f32(vacc1x0123);
118 }
119 if (c & (1 * sizeof(float))) {
120 vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
121 vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
122 }
123 }
124 i0 = (const float*) ((uintptr_t) i0 + input_increment);
125 o0 = (float*) ((uintptr_t) o0 + output_increment);
126 i1 = (const float*) ((uintptr_t) i1 + input_increment);
127 o1 = (float*) ((uintptr_t) o1 + output_increment);
128 rows = doz(rows, 2);
129 } while (rows != 0);
130 }
131