Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-8x16c4-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_8x16c4__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 <= 8);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (8 * 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 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
72 if XNN_UNPREDICTABLE(mr != 8) {
73 c7 = c6;
74 }
75
76 do {
77 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
78 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
79 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
80 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
81 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
82 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
83 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
84 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
85 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
86 int32x4_t vacc1x0123 = vacc0x0123;
87 int32x4_t vacc1x4567 = vacc0x4567;
88 int32x4_t vacc1x89AB = vacc0x89AB;
89 int32x4_t vacc1xCDEF = vacc0xCDEF;
90 int32x4_t vacc2x0123 = vacc0x0123;
91 int32x4_t vacc2x4567 = vacc0x4567;
92 int32x4_t vacc2x89AB = vacc0x89AB;
93 int32x4_t vacc2xCDEF = vacc0xCDEF;
94 int32x4_t vacc3x0123 = vacc0x0123;
95 int32x4_t vacc3x4567 = vacc0x4567;
96 int32x4_t vacc3x89AB = vacc0x89AB;
97 int32x4_t vacc3xCDEF = vacc0xCDEF;
98 int32x4_t vacc4x0123 = vacc0x0123;
99 int32x4_t vacc4x4567 = vacc0x4567;
100 int32x4_t vacc4x89AB = vacc0x89AB;
101 int32x4_t vacc4xCDEF = vacc0xCDEF;
102 int32x4_t vacc5x0123 = vacc0x0123;
103 int32x4_t vacc5x4567 = vacc0x4567;
104 int32x4_t vacc5x89AB = vacc0x89AB;
105 int32x4_t vacc5xCDEF = vacc0xCDEF;
106 int32x4_t vacc6x0123 = vacc0x0123;
107 int32x4_t vacc6x4567 = vacc0x4567;
108 int32x4_t vacc6x89AB = vacc0x89AB;
109 int32x4_t vacc6xCDEF = vacc0xCDEF;
110 int32x4_t vacc7x0123 = vacc0x0123;
111 int32x4_t vacc7x4567 = vacc0x4567;
112 int32x4_t vacc7x89AB = vacc0x89AB;
113 int32x4_t vacc7xCDEF = vacc0xCDEF;
114
115 size_t p = ks;
116 do {
117 const int8_t* restrict a0 = a[0];
118 if XNN_UNPREDICTABLE(a0 != zero) {
119 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
120 } else {
121 a0 = zero_data;
122 }
123 const int8_t* restrict a1 = a[1];
124 if XNN_UNPREDICTABLE(a1 != zero) {
125 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
126 } else {
127 a1 = zero_data;
128 }
129 const int8_t* restrict a2 = a[2];
130 if XNN_UNPREDICTABLE(a2 != zero) {
131 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
132 } else {
133 a2 = zero_data;
134 }
135 const int8_t* restrict a3 = a[3];
136 if XNN_UNPREDICTABLE(a3 != zero) {
137 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
138 } else {
139 a3 = zero_data;
140 }
141 const int8_t* restrict a4 = a[4];
142 if XNN_UNPREDICTABLE(a4 != zero) {
143 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
144 } else {
145 a4 = zero_data;
146 }
147 const int8_t* restrict a5 = a[5];
148 if XNN_UNPREDICTABLE(a5 != zero) {
149 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
150 } else {
151 a5 = zero_data;
152 }
153 const int8_t* restrict a6 = a[6];
154 if XNN_UNPREDICTABLE(a6 != zero) {
155 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
156 } else {
157 a6 = zero_data;
158 }
159 const int8_t* restrict a7 = a[7];
160 if XNN_UNPREDICTABLE(a7 != zero) {
161 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
162 } else {
163 a7 = zero_data;
164 }
165 a += 8;
166
167 // Inner accumulation loop along the 16 columns.
168 size_t k = kc;
169 // 2x partial unrolled loop to load 8 bytes at a time.
170 while (k >= 8 * sizeof(int8_t)) {
171 // Load a 8x8 block of activations.
172 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
173 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
174 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
175 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
176 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
177 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
178 const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8;
179 const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8;
180
181 // Load a 8x16 block of weights.
182 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
183 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
184 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
185 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
186 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
187 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
188 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
189 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
190
191 // Multiply-accumulate: 8x8 * 8x16 --> 8x16.
192 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
193 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
194 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
195 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
196 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
197 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
198 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
199 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
200 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
201 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
202 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
203 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
204 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
205 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
206 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
207 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
208 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
209 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
210 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
211 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
212 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
213 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
214 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
215 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
216 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
217 vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
218 vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb0123x89AB, va6x01234567, 0);
219 vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb0123xCDEF, va6x01234567, 0);
220 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
221 vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
222 vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb0123x89AB, va7x01234567, 0);
223 vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb0123xCDEF, va7x01234567, 0);
224 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
225 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
226 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
227 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
228 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
229 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
230 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
231 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
232 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
233 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
234 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
235 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
236 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
237 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
238 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
239 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
240 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
241 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
242 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
243 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
244 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
245 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
246 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
247 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
248 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1);
249 vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1);
250 vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb4567x89AB, va6x01234567, 1);
251 vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb4567xCDEF, va6x01234567, 1);
252 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1);
253 vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1);
254 vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb4567x89AB, va7x01234567, 1);
255 vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb4567xCDEF, va7x01234567, 1);
256
257 k -= 8 * sizeof(int8_t);
258 }
259 // Handle up to 4 final positions of `k`
260 if XNN_UNLIKELY(k != 0) {
261 // Load a 8x4 block of activations.
262 const int8x8_t va0x01234567 = vld1_s8(a0);
263 const int8x8_t va1x01234567 = vld1_s8(a1);
264 const int8x8_t va2x01234567 = vld1_s8(a2);
265 const int8x8_t va3x01234567 = vld1_s8(a3);
266 const int8x8_t va4x01234567 = vld1_s8(a4);
267 const int8x8_t va5x01234567 = vld1_s8(a5);
268 const int8x8_t va6x01234567 = vld1_s8(a6);
269 const int8x8_t va7x01234567 = vld1_s8(a7);
270
271 // Load a 4x16 block of weights.
272 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
273 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
274 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
275 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
276
277 // Multiply-accumulate: 8x4 * 4x16 --> 8x16.
278 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
279 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
280 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
281 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
282 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
283 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
284 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
285 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
286 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
287 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
288 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
289 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
290 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
291 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
292 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
293 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
294 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
295 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
296 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
297 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
298 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
299 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
300 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
301 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
302 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
303 vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
304 vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb0123x89AB, va6x01234567, 0);
305 vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb0123xCDEF, va6x01234567, 0);
306 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
307 vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
308 vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb0123x89AB, va7x01234567, 0);
309 vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb0123xCDEF, va7x01234567, 0);
310 }
311 p -= 8 * sizeof(void*);
312 } while (p != 0);
313
314 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
315 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
316 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
317 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
318 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
319 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
320 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
321 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
322 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
323 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
324 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
325 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
326 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
327 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
328 float32x4_t vout3x89AB = vcvtq_f32_s32(vacc3x89AB);
329 float32x4_t vout3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
330 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
331 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
332 float32x4_t vout4x89AB = vcvtq_f32_s32(vacc4x89AB);
333 float32x4_t vout4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
334 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
335 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
336 float32x4_t vout5x89AB = vcvtq_f32_s32(vacc5x89AB);
337 float32x4_t vout5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
338 float32x4_t vout6x0123 = vcvtq_f32_s32(vacc6x0123);
339 float32x4_t vout6x4567 = vcvtq_f32_s32(vacc6x4567);
340 float32x4_t vout6x89AB = vcvtq_f32_s32(vacc6x89AB);
341 float32x4_t vout6xCDEF = vcvtq_f32_s32(vacc6xCDEF);
342 float32x4_t vout7x0123 = vcvtq_f32_s32(vacc7x0123);
343 float32x4_t vout7x4567 = vcvtq_f32_s32(vacc7x4567);
344 float32x4_t vout7x89AB = vcvtq_f32_s32(vacc7x89AB);
345 float32x4_t vout7xCDEF = vcvtq_f32_s32(vacc7xCDEF);
346 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
347 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
348 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
349 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
350 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
351 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
352 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
353 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
354 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
355 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
356 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
357 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale);
358 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale);
359 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
360 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
361 vout3x89AB = vmulq_f32(vout3x89AB, vinput_scale);
362 vout3xCDEF = vmulq_f32(vout3xCDEF, vinput_scale);
363 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale);
364 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale);
365 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale);
366 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale);
367 vout5x0123 = vmulq_f32(vout5x0123, vinput_scale);
368 vout5x4567 = vmulq_f32(vout5x4567, vinput_scale);
369 vout5x89AB = vmulq_f32(vout5x89AB, vinput_scale);
370 vout5xCDEF = vmulq_f32(vout5xCDEF, vinput_scale);
371 vout6x0123 = vmulq_f32(vout6x0123, vinput_scale);
372 vout6x4567 = vmulq_f32(vout6x4567, vinput_scale);
373 vout6x89AB = vmulq_f32(vout6x89AB, vinput_scale);
374 vout6xCDEF = vmulq_f32(vout6xCDEF, vinput_scale);
375 vout7x0123 = vmulq_f32(vout7x0123, vinput_scale);
376 vout7x4567 = vmulq_f32(vout7x4567, vinput_scale);
377 vout7x89AB = vmulq_f32(vout7x89AB, vinput_scale);
378 vout7xCDEF = vmulq_f32(vout7xCDEF, vinput_scale);
379
380 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
381 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
382 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
383 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
384
385 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
386 #if XNN_ARCH_ARM64
387 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
388 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
389 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
390 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
391 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
392 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
393 vout6x0123 = vfmaq_f32(vbias0123, vout6x0123, vfilter_output_scale0123);
394 vout7x0123 = vfmaq_f32(vbias0123, vout7x0123, vfilter_output_scale0123);
395 #else
396 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
397 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
398 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
399 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
400 vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
401 vout5x0123 = vmlaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
402 vout6x0123 = vmlaq_f32(vbias0123, vout6x0123, vfilter_output_scale0123);
403 vout7x0123 = vmlaq_f32(vbias0123, vout7x0123, vfilter_output_scale0123);
404 #endif
405 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
406 #if XNN_ARCH_ARM64
407 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
408 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
409 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
410 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
411 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
412 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
413 vout6x4567 = vfmaq_f32(vbias4567, vout6x4567, vfilter_output_scale4567);
414 vout7x4567 = vfmaq_f32(vbias4567, vout7x4567, vfilter_output_scale4567);
415 #else
416 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
417 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
418 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
419 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
420 vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
421 vout5x4567 = vmlaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
422 vout6x4567 = vmlaq_f32(vbias4567, vout6x4567, vfilter_output_scale4567);
423 vout7x4567 = vmlaq_f32(vbias4567, vout7x4567, vfilter_output_scale4567);
424 #endif
425 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
426 #if XNN_ARCH_ARM64
427 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
428 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
429 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
430 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
431 vout4x89AB = vfmaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
432 vout5x89AB = vfmaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
433 vout6x89AB = vfmaq_f32(vbias89AB, vout6x89AB, vfilter_output_scale89AB);
434 vout7x89AB = vfmaq_f32(vbias89AB, vout7x89AB, vfilter_output_scale89AB);
435 #else
436 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
437 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
438 vout2x89AB = vmlaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
439 vout3x89AB = vmlaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
440 vout4x89AB = vmlaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
441 vout5x89AB = vmlaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
442 vout6x89AB = vmlaq_f32(vbias89AB, vout6x89AB, vfilter_output_scale89AB);
443 vout7x89AB = vmlaq_f32(vbias89AB, vout7x89AB, vfilter_output_scale89AB);
444 #endif
445 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
446 #if XNN_ARCH_ARM64
447 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
448 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
449 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
450 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
451 vout4xCDEF = vfmaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
452 vout5xCDEF = vfmaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
453 vout6xCDEF = vfmaq_f32(vbiasCDEF, vout6xCDEF, vfilter_output_scaleCDEF);
454 vout7xCDEF = vfmaq_f32(vbiasCDEF, vout7xCDEF, vfilter_output_scaleCDEF);
455 #else
456 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
457 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
458 vout2xCDEF = vmlaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
459 vout3xCDEF = vmlaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
460 vout4xCDEF = vmlaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
461 vout5xCDEF = vmlaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
462 vout6xCDEF = vmlaq_f32(vbiasCDEF, vout6xCDEF, vfilter_output_scaleCDEF);
463 vout7xCDEF = vmlaq_f32(vbiasCDEF, vout7xCDEF, vfilter_output_scaleCDEF);
464 #endif
465
466 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
467 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
468 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
469 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
470 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
471 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
472 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
473 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
474 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
475 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
476 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
477 float16x8_t vfp16out5x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout5x89AB), vcvt_f16_f32(vout5xCDEF));
478 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
479 float16x8_t vfp16out6x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout6x89AB), vcvt_f16_f32(vout6xCDEF));
480 float16x8_t vfp16out7x01234567 = vcombine_f16(vcvt_f16_f32(vout7x0123), vcvt_f16_f32(vout7x4567));
481 float16x8_t vfp16out7x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout7x89AB), vcvt_f16_f32(vout7xCDEF));
482
483 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
484 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
485 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
486 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
487 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
488 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
489 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
490 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
491 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
492 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
493 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
494 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
495 vfp16out5x89ABCDEF = vmaxq_f16(vfp16out5x89ABCDEF, voutput_min);
496 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
497 vfp16out6x89ABCDEF = vmaxq_f16(vfp16out6x89ABCDEF, voutput_min);
498 vfp16out7x01234567 = vmaxq_f16(vfp16out7x01234567, voutput_min);
499 vfp16out7x89ABCDEF = vmaxq_f16(vfp16out7x89ABCDEF, voutput_min);
500 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
501 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
502 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
503 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
504 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
505 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
506 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
507 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
508 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
509 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
510 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
511 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
512 vfp16out5x89ABCDEF = vminq_f16(vfp16out5x89ABCDEF, voutput_max);
513 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
514 vfp16out6x89ABCDEF = vminq_f16(vfp16out6x89ABCDEF, voutput_max);
515 vfp16out7x01234567 = vminq_f16(vfp16out7x01234567, voutput_max);
516 vfp16out7x89ABCDEF = vminq_f16(vfp16out7x89ABCDEF, voutput_max);
517 if XNN_LIKELY(nc >= 16) {
518 vst1q_u16(c7, vreinterpretq_u16_f16(vfp16out7x01234567));
519 vst1q_u16(c7 + 8, vreinterpretq_u16_f16(vfp16out7x89ABCDEF));
520 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
521 vst1q_u16(c6 + 8, vreinterpretq_u16_f16(vfp16out6x89ABCDEF));
522 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
523 vst1q_u16(c5 + 8, vreinterpretq_u16_f16(vfp16out5x89ABCDEF));
524 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
525 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
526 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
527 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
528 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
529 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
530 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
531 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
532 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
533 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
534
535 a = (const int8_t**restrict) ((uintptr_t) a - ks);
536
537 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
538 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
539 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
540 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
541 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
542 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
543 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
544 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
545
546 nc -= 16;
547 } else {
548 if (nc & 8) {
549 vst1q_u16(c7, vreinterpretq_u16_f16(vfp16out7x01234567)); c7 += 8;
550 vfp16out7x01234567 = vfp16out7x89ABCDEF;
551 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567)); c6 += 8;
552 vfp16out6x01234567 = vfp16out6x89ABCDEF;
553 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567)); c5 += 8;
554 vfp16out5x01234567 = vfp16out5x89ABCDEF;
555 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
556 vfp16out4x01234567 = vfp16out4x89ABCDEF;
557 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
558 vfp16out3x01234567 = vfp16out3x89ABCDEF;
559 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
560 vfp16out2x01234567 = vfp16out2x89ABCDEF;
561 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
562 vfp16out1x01234567 = vfp16out1x89ABCDEF;
563 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
564 vfp16out0x01234567 = vfp16out0x89ABCDEF;
565 }
566 float16x4_t vfp16out7x0123 = vget_low_f16(vfp16out7x01234567);
567 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
568 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
569 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
570 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
571 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
572 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
573 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
574 if (nc & 4) {
575 vst1_u16(c7, vreinterpret_u16_f16(vfp16out7x0123)); c7 += 4;
576 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
577 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
578 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
579 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
580 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
581 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
582 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
583 vfp16out7x0123 = vget_high_f16(vfp16out7x01234567);
584 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
585 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
586 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
587 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
588 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
589 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
590 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
591 }
592 if (nc & 2) {
593 vst1_lane_u32((void*) c7, vreinterpret_u32_f16(vfp16out7x0123), 0); c7 += 2;
594 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
595 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
596 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
597 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
598 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
599 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
600 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
601 vfp16out7x0123 = vext_f16(vfp16out7x0123, vfp16out7x0123, 2);
602 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
603 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
604 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
605 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
606 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
607 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
608 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
609 }
610 if (nc & 1) {
611 vst1_lane_u16(c7, vreinterpret_u16_f16(vfp16out7x0123), 0);
612 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
613 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
614 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
615 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
616 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
617 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
618 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
619 }
620 nc = 0;
621 }
622 } while (nc != 0);
623 }