• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_DISABLE_TSAN
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   if XNN_UNPREDICTABLE(rows < 2) {
36     i1 = i0;
37     o1 = o0;
38   }
39 
40   const size_t input_increment = input_stride * 2 - channels;
41   const size_t output_increment = output_stride * 2 - channels;
42 
43   do {
44     const float* w = weights;
45     size_t c = channels;
46     for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
47       const float32x4_t vw0123 = vld1q_f32(w); w += 4;
48       const float32x4_t vw4567 = vld1q_f32(w); w += 4;
49 
50       const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
51       const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
52       const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
53       const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
54 
55       float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
56       const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
57       float32x4_t vacc0x4567 = vmulq_f32(vi0x4567, vw4567);
58       const uint32x4_t vm0x4567 = vcltq_s32(vreinterpretq_s32_f32(vi0x4567), vmovq_n_s32(0));
59       float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
60       const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
61       float32x4_t vacc1x4567 = vmulq_f32(vi1x4567, vw4567);
62       const uint32x4_t vm1x4567 = vcltq_s32(vreinterpretq_s32_f32(vi1x4567), vmovq_n_s32(0));
63 
64       vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
65       vacc0x4567 = vbslq_f32(vm0x4567, vacc0x4567, vi0x4567);
66       vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
67       vacc1x4567 = vbslq_f32(vm1x4567, vacc1x4567, vi1x4567);
68 
69       vst1q_f32(o0, vacc0x0123); o0 += 4;
70       vst1q_f32(o0, vacc0x4567); o0 += 4;
71       vst1q_f32(o1, vacc1x0123); o1 += 4;
72       vst1q_f32(o1, vacc1x4567); o1 += 4;
73     }
74     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
75       const float32x4_t vw0123 = vld1q_f32(w); w += 4;
76 
77       const float32x4_t vi0x0123 = vld1q_f32(i0);
78       i0 += 4;
79       const float32x4_t vi1x0123 = vld1q_f32(i1);
80       i1 += 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 vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
85       const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
86 
87       vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
88       vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
89 
90       vst1q_f32(o0, vacc0x0123); o0 += 4;
91       vst1q_f32(o1, vacc1x0123); o1 += 4;
92     }
93     if XNN_UNLIKELY(c != 0) {
94       const float32x4_t vw0123 = vld1q_f32(w); w += 4;
95 
96       const float32x4_t vi0x0123 = vld1q_f32(i0);
97       i0 = (const float*) ((uintptr_t) i0 + c);
98       const float32x4_t vi1x0123 = vld1q_f32(i1);
99       i1 = (const float*) ((uintptr_t) i1 + c);
100 
101       float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
102       const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
103       float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);
104       const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));
105 
106       vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
107       vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);
108 
109       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
110       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
111       if (c & (2 * sizeof(float))) {
112         vst1_f32(o0, vacc0x01); o0 += 2;
113         vst1_f32(o1, vacc1x01); o1 += 2;
114 
115         vacc0x01 = vget_high_f32(vacc0x0123);
116         vacc1x01 = vget_high_f32(vacc1x0123);
117       }
118       if (c & (1 * sizeof(float))) {
119         vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
120         vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
121       }
122     }
123     i0 = (const float*) ((uintptr_t) i0 + input_increment);
124     o0 = (float*) ((uintptr_t) o0 + output_increment);
125     i1 = (const float*) ((uintptr_t) i1 + input_increment);
126     o1 = (float*) ((uintptr_t) o1 + output_increment);
127     if XNN_UNPREDICTABLE(rows < 4) {
128       i1 = i0;
129       o1 = o0;
130     }
131     rows = doz(rows, 2);
132   } while (rows != 0);
133 }
134