Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x16c4-minmax-neondotfp16arith.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neondot.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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_4x16c4__neondotfp16arith(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 uint16_t* c0 = (uint16_t*) c;
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 do {
61 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
62 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
64 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
65 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
66 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
67 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
68 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
69 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
70 int32x4_t vacc1x0123 = vacc0x0123;
71 int32x4_t vacc1x4567 = vacc0x4567;
72 int32x4_t vacc1x89AB = vacc0x89AB;
73 int32x4_t vacc1xCDEF = vacc0xCDEF;
74 int32x4_t vacc2x0123 = vacc0x0123;
75 int32x4_t vacc2x4567 = vacc0x4567;
76 int32x4_t vacc2x89AB = vacc0x89AB;
77 int32x4_t vacc2xCDEF = vacc0xCDEF;
78 int32x4_t vacc3x0123 = vacc0x0123;
79 int32x4_t vacc3x4567 = vacc0x4567;
80 int32x4_t vacc3x89AB = vacc0x89AB;
81 int32x4_t vacc3xCDEF = vacc0xCDEF;
82
83 size_t p = ks;
84 do {
85 const int8_t* restrict a0 = a[0];
86 if XNN_UNPREDICTABLE(a0 != zero) {
87 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
88 } else {
89 a0 = zero_data;
90 }
91 const int8_t* restrict a1 = a[1];
92 if XNN_UNPREDICTABLE(a1 != zero) {
93 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
94 } else {
95 a1 = zero_data;
96 }
97 const int8_t* restrict a2 = a[2];
98 if XNN_UNPREDICTABLE(a2 != zero) {
99 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
100 } else {
101 a2 = zero_data;
102 }
103 const int8_t* restrict a3 = a[3];
104 if XNN_UNPREDICTABLE(a3 != zero) {
105 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
106 } else {
107 a3 = zero_data;
108 }
109 a += 4;
110
111 // Inner accumulation loop along the 16 columns.
112 size_t k = kc;
113 // 2x partial unrolled loop to load 8 bytes at a time.
114 while (k >= 8 * sizeof(int8_t)) {
115 // Load a 4x8 block of activations.
116 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
117 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
118 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
119 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
120
121 // Load a 8x16 block of weights.
122 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
123 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
127 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
128 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
130
131 // Multiply-accumulate: 4x8 * 8x16 --> 4x16.
132 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
133 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
134 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
135 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
136 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
137 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
138 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
139 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
140 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
141 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
142 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
143 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
144 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
145 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
146 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
147 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
148 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
149 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
150 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
151 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
152 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
153 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
154 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
155 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
156 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
157 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
158 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
159 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
160 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
161 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
162 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
163 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
164
165 k -= 8 * sizeof(int8_t);
166 }
167 // Handle up to 4 final positions of `k`
168 if XNN_UNLIKELY(k != 0) {
169 // Load a 4x4 block of activations.
170 const int8x8_t va0x01234567 = vld1_s8(a0);
171 const int8x8_t va1x01234567 = vld1_s8(a1);
172 const int8x8_t va2x01234567 = vld1_s8(a2);
173 const int8x8_t va3x01234567 = vld1_s8(a3);
174
175 // Load a 4x16 block of weights.
176 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
177 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
178 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
179 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
180
181 // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
182 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
183 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
184 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
185 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
186 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
187 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
188 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
189 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
190 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
191 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
192 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
193 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
194 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
195 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
196 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
197 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
198 }
199 p -= 4 * sizeof(void*);
200 } while (p != 0);
201
202 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
203 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
204 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
205 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
206 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
207 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
208 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
209 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
210 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
211 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
212 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
213 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
214 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
215 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
216 float32x4_t vout3x89AB = vcvtq_f32_s32(vacc3x89AB);
217 float32x4_t vout3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
218 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
219 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
220 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
221 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
222 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
223 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
224 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
225 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
226 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
227 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
228 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
229 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale);
230 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale);
231 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
232 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
233 vout3x89AB = vmulq_f32(vout3x89AB, vinput_scale);
234 vout3xCDEF = vmulq_f32(vout3xCDEF, vinput_scale);
235
236 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
237 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
238 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
239 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
240
241 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
242 #if XNN_ARCH_ARM64
243 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
244 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
245 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
246 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
247 #else
248 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
249 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
250 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
251 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
252 #endif
253 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
254 #if XNN_ARCH_ARM64
255 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
256 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
257 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
258 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
259 #else
260 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
261 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
262 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
263 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
264 #endif
265 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
266 #if XNN_ARCH_ARM64
267 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
268 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
269 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
270 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
271 #else
272 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
273 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
274 vout2x89AB = vmlaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
275 vout3x89AB = vmlaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
276 #endif
277 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
278 #if XNN_ARCH_ARM64
279 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
280 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
281 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
282 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
283 #else
284 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
285 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
286 vout2xCDEF = vmlaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
287 vout3xCDEF = vmlaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
288 #endif
289
290 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
291 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
292 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
293 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
294 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
295 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
296 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
297 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
298
299 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
300 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
301 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
302 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
303 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
304 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
305 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
306 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
307 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
308 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
309 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
310 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
311 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
312 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
313 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
314 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
315 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
316 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
317 if XNN_LIKELY(nc >= 16) {
318 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
319 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
320 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
321 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
322 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
323 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
324 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
325 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
326
327 a = (const int8_t**restrict) ((uintptr_t) a - ks);
328
329 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
330 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
331 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
332 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
333
334 nc -= 16;
335 } else {
336 if (nc & 8) {
337 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
338 vfp16out3x01234567 = vfp16out3x89ABCDEF;
339 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
340 vfp16out2x01234567 = vfp16out2x89ABCDEF;
341 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
342 vfp16out1x01234567 = vfp16out1x89ABCDEF;
343 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
344 vfp16out0x01234567 = vfp16out0x89ABCDEF;
345 }
346 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
347 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
348 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
349 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
350 if (nc & 4) {
351 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
352 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
353 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
354 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
355 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
356 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
357 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
358 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
359 }
360 if (nc & 2) {
361 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
362 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
363 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
364 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
365 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
366 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
367 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
368 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
369 }
370 if (nc & 1) {
371 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
372 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
373 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
374 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
375 }
376 nc = 0;
377 }
378 } while (nc != 0);
379 }