• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-prelu/neonfp16arith.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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_f16_prelu_ukernel__neonfp16arith_2x8(size_t rows,size_t channels,const void * restrict input,size_t input_stride,const void * restrict weights,void * restrict output,size_t output_stride)18 void xnn_f16_prelu_ukernel__neonfp16arith_2x8(
19     size_t rows,
20     size_t channels,
21     const void* restrict input,
22     size_t input_stride,
23     const void* restrict weights,
24     void* restrict output,
25     size_t output_stride) XNN_DISABLE_TSAN
26 {
27   assert(rows != 0);
28   assert(channels != 0);
29   assert(channels % sizeof(__fp16) == 0);
30 
31   const __fp16* i0 = (const __fp16*) input;
32   __fp16* o0 = (__fp16*) output;
33 
34   const __fp16* i1 = (const __fp16*) ((uintptr_t) i0 + input_stride);
35   __fp16* o1 = (__fp16*) ((uintptr_t) o0 + output_stride);
36   if XNN_UNPREDICTABLE(rows < 2) {
37     i1 = i0;
38     o1 = o0;
39   }
40 
41   const size_t input_increment = input_stride * 2 - channels;
42   const size_t output_increment = output_stride * 2 - channels;
43 
44   do {
45     const __fp16* w = (const __fp16*) weights;
46     size_t c = channels;
47     for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
48       const float16x8_t vw01234567 = vld1q_f16(w); w += 8;
49 
50       const float16x8_t vi0x001234567 = vld1q_f16(i0); i0 += 8;
51       const float16x8_t vi1x001234567 = vld1q_f16(i1); i1 += 8;
52 
53       float16x8_t vacc0x001234567 = vmulq_f16(vi0x001234567, vw01234567);
54       const uint16x8_t vm0x001234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x001234567), vmovq_n_s16(0));
55       float16x8_t vacc1x001234567 = vmulq_f16(vi1x001234567, vw01234567);
56       const uint16x8_t vm1x001234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x001234567), vmovq_n_s16(0));
57 
58       vacc0x001234567 = vbslq_f16(vm0x001234567, vacc0x001234567, vi0x001234567);
59       vacc1x001234567 = vbslq_f16(vm1x001234567, vacc1x001234567, vi1x001234567);
60 
61       vst1q_f16(o0, vacc0x001234567); o0 += 8;
62       vst1q_f16(o1, vacc1x001234567); o1 += 8;
63     }
64     if XNN_UNLIKELY(c != 0) {
65       const float16x8_t vw01234567 = vld1q_f16(w); w += 8;
66 
67       const float16x8_t vi0x01234567 = vld1q_f16(i0);
68       i0 = (const __fp16*) ((uintptr_t) i0 + c);
69       const float16x8_t vi1x01234567 = vld1q_f16(i1);
70       i1 = (const __fp16*) ((uintptr_t) i1 + c);
71 
72       float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567);
73       const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0));
74       float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567);
75       const uint16x8_t vm1x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x01234567), vmovq_n_s16(0));
76 
77       vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567);
78       vacc1x01234567 = vbslq_f16(vm1x01234567, vacc1x01234567, vi1x01234567);
79 
80       float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
81       float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
82       if (c & (4 * sizeof(__fp16))) {
83         vst1_f16(o0, vacc0x0123); o0 += 4;
84         vst1_f16(o1, vacc1x0123); o1 += 4;
85 
86         vacc0x0123 = vget_high_f16(vacc0x01234567);
87         vacc1x0123 = vget_high_f16(vacc1x01234567);
88       }
89       if (c & (2 * sizeof(__fp16))) {
90         vst1_lane_u32(__builtin_assume_aligned(o0, 1), vreinterpret_u32_f16(vacc0x0123), 0); o0 += 2;
91         vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
92         vst1_lane_u32(__builtin_assume_aligned(o1, 1), vreinterpret_u32_f16(vacc1x0123), 0); o1 += 2;
93         vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
94       }
95       if (c & (1 * sizeof(__fp16))) {
96         vst1_lane_f16(o0, vacc0x0123, 0); o0 += 1;
97         vst1_lane_f16(o1, vacc1x0123, 0); o1 += 1;
98       }
99     }
100     i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
101     o0 = (__fp16*) ((uintptr_t) o0 + output_increment);
102     i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
103     o1 = (__fp16*) ((uintptr_t) o1 + output_increment);
104     if XNN_UNPREDICTABLE(rows < 4) {
105       i1 = i0;
106       o1 = o0;
107     }
108     rows = doz(rows, 2);
109   } while (rows != 0);
110 }
111