1 // Auto-generated file. Do not edit!
2 // Template: src/f16-dwconv/up-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/dwconv.h>
15
16
xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith(size_t channels,size_t output_width,const void ** input,const void * weights,void * output_ptr,size_t input_stride,size_t output_increment,size_t input_offset,const void * zero,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith(
18 size_t channels,
19 size_t output_width,
20 const void** input,
21 const void* weights,
22 void* output_ptr,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const void* zero,
27 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 __fp16* output = (__fp16*) output_ptr;
33 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
34 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
35 do {
36 const __fp16* i0 = (const __fp16*) input[0];
37 assert(i0 != NULL);
38 if XNN_UNPREDICTABLE(i0 != (const __fp16*) zero) {
39 i0 = (const __fp16*) ((uintptr_t) i0 + input_offset);
40 }
41 const __fp16* i1 = (const __fp16*) input[1];
42 assert(i1 != NULL);
43 if XNN_UNPREDICTABLE(i1 != (const __fp16*) zero) {
44 i1 = (const __fp16*) ((uintptr_t) i1 + input_offset);
45 }
46 const __fp16* i2 = (const __fp16*) input[2];
47 assert(i2 != NULL);
48 if XNN_UNPREDICTABLE(i2 != (const __fp16*) zero) {
49 i2 = (const __fp16*) ((uintptr_t) i2 + input_offset);
50 }
51 const __fp16* i3 = (const __fp16*) input[3];
52 assert(i3 != NULL);
53 if XNN_UNPREDICTABLE(i3 != (const __fp16*) zero) {
54 i3 = (const __fp16*) ((uintptr_t) i3 + input_offset);
55 }
56
57 input = (const void**) ((uintptr_t) input + input_stride);
58
59 size_t c = channels;
60 const __fp16* w = (const __fp16*) weights;
61 for (; c >= 16; c -= 16) {
62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
63 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8;
64
65
66 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
67 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
68 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
69 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8;
70 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
71 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);
72
73 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
74 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
75 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
76 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8;
77 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
78 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF);
79
80 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
81 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
82 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
83 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8;
84 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
85 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);
86
87 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
88 const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
89 const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
90 const float16x8_t vk3x89ABCDEF = vld1q_f16(w); w += 8;
91 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
92 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi3x89ABCDEF, vk3x89ABCDEF);
93
94
95 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
96 float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
97 vacc01234567 = vminq_f16(vacc01234567, vmax);
98 vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
99
100 vst1q_f16(output, vacc01234567); output += 8;
101 vst1q_f16(output, vacc89ABCDEF); output += 8;
102 }
103 for (; c >= 8; c -= 8) {
104 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
105
106
107 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
108 const float16x8_t vk0x01234567 = vld1q_f16(w + 8);
109 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
110
111 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
112 const float16x8_t vk1x01234567 = vld1q_f16(w + 24);
113 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
114
115 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
116 const float16x8_t vk2x01234567 = vld1q_f16(w + 40);
117 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
118
119 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
120 const float16x8_t vk3x01234567 = vld1q_f16(w + 56);
121 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
122
123
124 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
125 vacc01234567 = vminq_f16(vacc01234567, vmax);
126
127 vst1q_f16(output, vacc01234567); output += 8;
128 }
129 if XNN_UNLIKELY(c != 0) {
130 float16x8_t vacc01234567p0 = vld1q_f16(w);
131
132
133 const float16x8_t vi0x01234567 = vld1q_f16(i0);
134 const float16x8_t vk0x01234567 = vld1q_f16(w + 16);
135 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
136
137 const float16x8_t vi1x01234567 = vld1q_f16(i1);
138 const float16x8_t vk1x01234567 = vld1q_f16(w + 32);
139 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
140
141 const float16x8_t vi2x01234567 = vld1q_f16(i2);
142 const float16x8_t vk2x01234567 = vld1q_f16(w + 48);
143 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
144
145 const float16x8_t vi3x01234567 = vld1q_f16(i3);
146 const float16x8_t vk3x01234567 = vld1q_f16(w + 64);
147 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
148
149
150 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
151 vacc01234567 = vminq_f16(vacc01234567, vmax);
152
153 float16x4_t vacc0123 = vget_low_f16(vacc01234567);
154 if (c & 4) {
155 vst1_f16(output, vacc0123); output += 4;
156 vacc0123 = vget_high_f16(vacc01234567);
157 }
158 if (c & 2) {
159 vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
160 vacc0123 = vext_f16(vacc0123, vacc0123, 2);
161 }
162 if (c & 1) {
163 vst1_lane_f16(output, vacc0123, 0); output += 1;
164 }
165 }
166
167 output = (__fp16*) ((uintptr_t) output + output_increment);
168 } while (--output_width != 0);
169 }
170