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_up16x9__neonfp16arith_acc2(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_up16x9__neonfp16arith_acc2(
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 const __fp16* i4 = (const __fp16*) input[4];
57 assert(i4 != NULL);
58 if XNN_UNPREDICTABLE(i4 != (const __fp16*) zero) {
59 i4 = (const __fp16*) ((uintptr_t) i4 + input_offset);
60 }
61 const __fp16* i5 = (const __fp16*) input[5];
62 assert(i5 != NULL);
63 if XNN_UNPREDICTABLE(i5 != (const __fp16*) zero) {
64 i5 = (const __fp16*) ((uintptr_t) i5 + input_offset);
65 }
66 const __fp16* i6 = (const __fp16*) input[6];
67 assert(i6 != NULL);
68 if XNN_UNPREDICTABLE(i6 != (const __fp16*) zero) {
69 i6 = (const __fp16*) ((uintptr_t) i6 + input_offset);
70 }
71 const __fp16* i7 = (const __fp16*) input[7];
72 assert(i7 != NULL);
73 if XNN_UNPREDICTABLE(i7 != (const __fp16*) zero) {
74 i7 = (const __fp16*) ((uintptr_t) i7 + input_offset);
75 }
76 const __fp16* i8 = (const __fp16*) input[8];
77 assert(i8 != NULL);
78 if XNN_UNPREDICTABLE(i8 != (const __fp16*) zero) {
79 i8 = (const __fp16*) ((uintptr_t) i8 + input_offset);
80 }
81
82 input = (const void**) ((uintptr_t) input + input_stride);
83
84 size_t c = channels;
85 const __fp16* w = (const __fp16*) weights;
86 for (; c >= 16; c -= 16) {
87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
88 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8;
89
90
91 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
92 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
93 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
94 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8;
95 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
96 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);
97
98 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
99 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
100 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
101 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8;
102 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
103 float16x8_t vacc89ABCDEFp1 = vmulq_f16(vi1x89ABCDEF, vk1x89ABCDEF);
104
105 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
106 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
107 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
108 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8;
109 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
110 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);
111
112 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
113 const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
114 const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
115 const float16x8_t vk3x89ABCDEF = vld1q_f16(w); w += 8;
116 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
117 vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi3x89ABCDEF, vk3x89ABCDEF);
118
119 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
120 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
121 const float16x8_t vk4x01234567 = vld1q_f16(w); w += 8;
122 const float16x8_t vk4x89ABCDEF = vld1q_f16(w); w += 8;
123 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
124 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF);
125
126 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
127 const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
128 const float16x8_t vk5x01234567 = vld1q_f16(w); w += 8;
129 const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8;
130 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
131 vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi5x89ABCDEF, vk5x89ABCDEF);
132
133 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
134 const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
135 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8;
136 const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8;
137 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
138 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF);
139
140 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
141 const float16x8_t vi7x89ABCDEF = vld1q_f16(i7); i7 += 8;
142 const float16x8_t vk7x01234567 = vld1q_f16(w); w += 8;
143 const float16x8_t vk7x89ABCDEF = vld1q_f16(w); w += 8;
144 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
145 vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi7x89ABCDEF, vk7x89ABCDEF);
146
147 const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
148 const float16x8_t vi8x89ABCDEF = vld1q_f16(i8); i8 += 8;
149 const float16x8_t vk8x01234567 = vld1q_f16(w); w += 8;
150 const float16x8_t vk8x89ABCDEF = vld1q_f16(w); w += 8;
151 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
152 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF);
153
154 // Add up all accumulators to vacc0123456789ABCDEFp0
155 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
156 vacc89ABCDEFp0 = vaddq_f16(vacc89ABCDEFp0, vacc89ABCDEFp1);
157
158 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
159 float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
160 vacc01234567 = vminq_f16(vacc01234567, vmax);
161 vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
162
163 vst1q_f16(output, vacc01234567); output += 8;
164 vst1q_f16(output, vacc89ABCDEF); output += 8;
165 }
166 for (; c >= 8; c -= 8) {
167 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
168
169
170 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
171 const float16x8_t vk0x01234567 = vld1q_f16(w + 8);
172 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
173
174 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
175 const float16x8_t vk1x01234567 = vld1q_f16(w + 24);
176 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
177
178 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
179 const float16x8_t vk2x01234567 = vld1q_f16(w + 40);
180 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
181
182 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
183 const float16x8_t vk3x01234567 = vld1q_f16(w + 56);
184 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
185
186 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
187 const float16x8_t vk4x01234567 = vld1q_f16(w + 72);
188 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
189
190 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
191 const float16x8_t vk5x01234567 = vld1q_f16(w + 88);
192 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
193
194 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
195 const float16x8_t vk6x01234567 = vld1q_f16(w + 104);
196 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
197
198 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
199 const float16x8_t vk7x01234567 = vld1q_f16(w + 120);
200 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
201
202 const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
203 const float16x8_t vk8x01234567 = vld1q_f16(w + 136);
204 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
205
206 // Add up all accumulators to vacc01234567p0
207 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
208
209 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
210 vacc01234567 = vminq_f16(vacc01234567, vmax);
211
212 vst1q_f16(output, vacc01234567); output += 8;
213 }
214 if XNN_UNLIKELY(c != 0) {
215 float16x8_t vacc01234567p0 = vld1q_f16(w);
216
217
218 const float16x8_t vi0x01234567 = vld1q_f16(i0);
219 const float16x8_t vk0x01234567 = vld1q_f16(w + 16);
220 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
221
222 const float16x8_t vi1x01234567 = vld1q_f16(i1);
223 const float16x8_t vk1x01234567 = vld1q_f16(w + 32);
224 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
225
226 const float16x8_t vi2x01234567 = vld1q_f16(i2);
227 const float16x8_t vk2x01234567 = vld1q_f16(w + 48);
228 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
229
230 const float16x8_t vi3x01234567 = vld1q_f16(i3);
231 const float16x8_t vk3x01234567 = vld1q_f16(w + 64);
232 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
233
234 const float16x8_t vi4x01234567 = vld1q_f16(i4);
235 const float16x8_t vk4x01234567 = vld1q_f16(w + 80);
236 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
237
238 const float16x8_t vi5x01234567 = vld1q_f16(i5);
239 const float16x8_t vk5x01234567 = vld1q_f16(w + 96);
240 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
241
242 const float16x8_t vi6x01234567 = vld1q_f16(i6);
243 const float16x8_t vk6x01234567 = vld1q_f16(w + 112);
244 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
245
246 const float16x8_t vi7x01234567 = vld1q_f16(i7);
247 const float16x8_t vk7x01234567 = vld1q_f16(w + 128);
248 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
249
250 const float16x8_t vi8x01234567 = vld1q_f16(i8);
251 const float16x8_t vk8x01234567 = vld1q_f16(w + 144);
252 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
253
254 // Add up all accumulators to vacc01234567p0
255 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
256
257 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
258 vacc01234567 = vminq_f16(vacc01234567, vmax);
259
260 float16x4_t vacc0123 = vget_low_f16(vacc01234567);
261 if (c & 4) {
262 vst1_f16(output, vacc0123); output += 4;
263 vacc0123 = vget_high_f16(vacc01234567);
264 }
265 if (c & 2) {
266 vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
267 vacc0123 = vext_f16(vacc0123, vacc0123, 2);
268 }
269 if (c & 1) {
270 vst1_lane_f16(output, vacc0123, 0); output += 1;
271 }
272 }
273
274 output = (__fp16*) ((uintptr_t) output + output_increment);
275 } while (--output_width != 0);
276 }
277