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_OOB_READS
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 const __fp16* i1 = (const __fp16*) ((uintptr_t) i0 + input_stride);
34 __fp16* o1 = (__fp16*) ((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 __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 vi0x01234567 = vld1q_f16(i0);
51 i0 += 8;
52 const float16x8_t vi1x01234567 = vld1q_f16(i1);
53 i1 += 8;
54
55 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567);
56 const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0));
57 float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567);
58 const uint16x8_t vm1x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x01234567), vmovq_n_s16(0));
59
60 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567);
61 vacc1x01234567 = vbslq_f16(vm1x01234567, vacc1x01234567, vi1x01234567);
62
63 vst1q_f16(o0, vacc0x01234567); o0 += 8;
64 vst1q_f16(o1, vacc1x01234567); o1 += 8;
65 }
66 if XNN_UNLIKELY(c != 0) {
67 const float16x8_t vw01234567 = vld1q_f16(w);
68
69 const float16x8_t vi0x01234567 = vld1q_f16(i0);
70 i0 = (const __fp16*) ((uintptr_t) i0 + c);
71 const float16x8_t vi1x01234567 = vld1q_f16(i1);
72 i1 = (const __fp16*) ((uintptr_t) i1 + c);
73
74 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567);
75 const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0));
76 float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567);
77 const uint16x8_t vm1x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x01234567), vmovq_n_s16(0));
78
79 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567);
80 vacc1x01234567 = vbslq_f16(vm1x01234567, vacc1x01234567, vi1x01234567);
81
82 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
83 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
84 if (c & (4 * sizeof(__fp16))) {
85 vst1_f16(o0, vacc0x0123); o0 += 4;
86 vst1_f16(o1, vacc1x0123); o1 += 4;
87
88 vacc0x0123 = vget_high_f16(vacc0x01234567);
89 vacc1x0123 = vget_high_f16(vacc1x01234567);
90 }
91 if (c & (2 * sizeof(__fp16))) {
92 vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vacc0x0123), 0); o0 += 2;
93 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
94 vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vacc1x0123), 0); o1 += 2;
95 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
96 }
97 if (c & (1 * sizeof(__fp16))) {
98 vst1_lane_f16(o0, vacc0x0123, 0); o0 += 1;
99 vst1_lane_f16(o1, vacc1x0123, 0); o1 += 1;
100 }
101 }
102 i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
103 o0 = (__fp16*) ((uintptr_t) o0 + output_increment);
104 i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
105 o1 = (__fp16*) ((uintptr_t) o1 + output_increment);
106 rows = doz(rows, 2);
107 } while (rows != 0);
108 }
109