• 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_1x4(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_1x4(
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 
34   const size_t input_increment = input_stride * 1 - channels;
35   const size_t output_increment = output_stride * 1 - channels;
36 
37   do {
38     const float* w = weights;
39     size_t c = channels;
40     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
41       const float32x4_t vw0123 = vld1q_f32(w); w += 4;
42 
43       const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
44 
45       float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
46       const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
47 
48       vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
49 
50       vst1q_f32(o0, vacc0x0123); o0 += 4;
51     }
52     if XNN_UNLIKELY(c != 0) {
53       const float32x4_t vw0123 = vld1q_f32(w); w += 4;
54 
55       const float32x4_t vi0x0123 = vld1q_f32(i0);
56       i0 = (const float*) ((uintptr_t) i0 + c);
57 
58       float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123);
59       const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0));
60 
61       vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123);
62 
63       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
64       if (c & (2 * sizeof(float))) {
65         vst1_f32(o0, vacc0x01); o0 += 2;
66 
67         vacc0x01 = vget_high_f32(vacc0x0123);
68       }
69       if (c & (1 * sizeof(float))) {
70         vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
71       }
72     }
73     i0 = (const float*) ((uintptr_t) i0 + input_increment);
74     o0 = (float*) ((uintptr_t) o0 + output_increment);
75     rows = doz(rows, 1);
76   } while (rows != 0);
77 }
78