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_2x16(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_2x16(
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 >= 16 * sizeof(__fp16); c -= 16 * sizeof(__fp16)) {
48 const float16x8_t vw01234567 = vld1q_f16(w); w += 8;
49 const float16x8_t vw89ABCDEF = vld1q_f16(w); w += 8;
50
51 const float16x8_t vi0x001234567 = vld1q_f16(i0); i0 += 8;
52 const float16x8_t vi0x089ABCDEF = vld1q_f16(i0); i0 += 8;
53 const float16x8_t vi1x001234567 = vld1q_f16(i1); i1 += 8;
54 const float16x8_t vi1x089ABCDEF = vld1q_f16(i1); i1 += 8;
55
56 float16x8_t vacc0x001234567 = vmulq_f16(vi0x001234567, vw01234567);
57 const uint16x8_t vm0x001234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x001234567), vmovq_n_s16(0));
58 float16x8_t vacc0x089ABCDEF = vmulq_f16(vi0x089ABCDEF, vw89ABCDEF);
59 const uint16x8_t vm0x089ABCDEF = vcltq_s16(vreinterpretq_s16_f16(vi0x089ABCDEF), vmovq_n_s16(0));
60 float16x8_t vacc1x001234567 = vmulq_f16(vi1x001234567, vw01234567);
61 const uint16x8_t vm1x001234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x001234567), vmovq_n_s16(0));
62 float16x8_t vacc1x089ABCDEF = vmulq_f16(vi1x089ABCDEF, vw89ABCDEF);
63 const uint16x8_t vm1x089ABCDEF = vcltq_s16(vreinterpretq_s16_f16(vi1x089ABCDEF), vmovq_n_s16(0));
64
65 vacc0x001234567 = vbslq_f16(vm0x001234567, vacc0x001234567, vi0x001234567);
66 vacc0x089ABCDEF = vbslq_f16(vm0x089ABCDEF, vacc0x089ABCDEF, vi0x089ABCDEF);
67 vacc1x001234567 = vbslq_f16(vm1x001234567, vacc1x001234567, vi1x001234567);
68 vacc1x089ABCDEF = vbslq_f16(vm1x089ABCDEF, vacc1x089ABCDEF, vi1x089ABCDEF);
69
70 vst1q_f16(o0, vacc0x001234567); o0 += 8;
71 vst1q_f16(o0, vacc0x089ABCDEF); o0 += 8;
72 vst1q_f16(o1, vacc1x001234567); o1 += 8;
73 vst1q_f16(o1, vacc1x089ABCDEF); o1 += 8;
74 }
75 for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
76 const float16x8_t vw01234567 = vld1q_f16(w); w += 8;
77
78 const float16x8_t vi0x01234567 = vld1q_f16(i0);
79 i0 += 8;
80 const float16x8_t vi1x01234567 = vld1q_f16(i1);
81 i1 += 8;
82
83 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567);
84 const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0));
85 float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567);
86 const uint16x8_t vm1x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x01234567), vmovq_n_s16(0));
87
88 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567);
89 vacc1x01234567 = vbslq_f16(vm1x01234567, vacc1x01234567, vi1x01234567);
90
91 vst1q_f16(o0, vacc0x01234567); o0 += 8;
92 vst1q_f16(o1, vacc1x01234567); o1 += 8;
93 }
94 if XNN_UNLIKELY(c != 0) {
95 const float16x8_t vw01234567 = vld1q_f16(w);
96
97 const float16x8_t vi0x01234567 = vld1q_f16(i0);
98 i0 = (const __fp16*) ((uintptr_t) i0 + c);
99 const float16x8_t vi1x01234567 = vld1q_f16(i1);
100 i1 = (const __fp16*) ((uintptr_t) i1 + c);
101
102 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567);
103 const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0));
104 float16x8_t vacc1x01234567 = vmulq_f16(vi1x01234567, vw01234567);
105 const uint16x8_t vm1x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi1x01234567), vmovq_n_s16(0));
106
107 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567);
108 vacc1x01234567 = vbslq_f16(vm1x01234567, vacc1x01234567, vi1x01234567);
109
110 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
111 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
112 if (c & (4 * sizeof(__fp16))) {
113 vst1_f16(o0, vacc0x0123); o0 += 4;
114 vst1_f16(o1, vacc1x0123); o1 += 4;
115
116 vacc0x0123 = vget_high_f16(vacc0x01234567);
117 vacc1x0123 = vget_high_f16(vacc1x01234567);
118 }
119 if (c & (2 * sizeof(__fp16))) {
120 vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vacc0x0123), 0); o0 += 2;
121 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
122 vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vacc1x0123), 0); o1 += 2;
123 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
124 }
125 if (c & (1 * sizeof(__fp16))) {
126 vst1_lane_f16(o0, vacc0x0123, 0); o0 += 1;
127 vst1_lane_f16(o1, vacc1x0123, 0); o1 += 1;
128 }
129 }
130 i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
131 o0 = (__fp16*) ((uintptr_t) o0 + output_increment);
132 i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
133 o1 = (__fp16*) ((uintptr_t) o1 + output_increment);
134 rows = doz(rows, 2);
135 } while (rows != 0);
136 }
137