Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-6x16c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_6x16c8__neoni8mm(
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 <= 6);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (6 * 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, 8 * 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
68 do {
69 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
70 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
71 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
72 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
73 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
74 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
75 const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
76 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
77 const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
78 const int32x4_t vksumzp1x0123 = vksumzp0x0123;
79 const int32x4_t vksumzp1x4567 = vksumzp0x4567;
80 const int32x4_t vksumzp1x89AB = vksumzp0x89AB;
81 const int32x4_t vksumzp1xCDEF = vksumzp0xCDEF;
82 const int32x4_t vksumzp2x0123 = vksumzp0x0123;
83 const int32x4_t vksumzp2x4567 = vksumzp0x4567;
84 const int32x4_t vksumzp2x89AB = vksumzp0x89AB;
85 const int32x4_t vksumzp2xCDEF = vksumzp0xCDEF;
86 const int32x4_t vksumzp3x0123 = vksumzp0x0123;
87 const int32x4_t vksumzp3x4567 = vksumzp0x4567;
88 const int32x4_t vksumzp3x89AB = vksumzp0x89AB;
89 const int32x4_t vksumzp3xCDEF = vksumzp0xCDEF;
90 const int32x4_t vksumzp4x0123 = vksumzp0x0123;
91 const int32x4_t vksumzp4x4567 = vksumzp0x4567;
92 const int32x4_t vksumzp4x89AB = vksumzp0x89AB;
93 const int32x4_t vksumzp4xCDEF = vksumzp0xCDEF;
94 const int32x4_t vksumzp5x0123 = vksumzp0x0123;
95 const int32x4_t vksumzp5x4567 = vksumzp0x4567;
96 const int32x4_t vksumzp5x89AB = vksumzp0x89AB;
97 const int32x4_t vksumzp5xCDEF = vksumzp0xCDEF;
98
99 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
100 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
101 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
102 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
103 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
104 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
105 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
106 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
107 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
108 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
109 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
110 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
111 int32x4_t vacc23x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp3x89AB)));
112 int32x4_t vacc23xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp3x89AB)));
113 int32x4_t vacc23xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp3xCDEF)));
114 int32x4_t vacc23xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp3xCDEF)));
115 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
116 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
117 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
118 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
119 int32x4_t vacc45x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x89AB), vreinterpretq_u64_s32(vksumzp5x89AB)));
120 int32x4_t vacc45xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x89AB), vreinterpretq_u64_s32(vksumzp5x89AB)));
121 int32x4_t vacc45xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4xCDEF), vreinterpretq_u64_s32(vksumzp5xCDEF)));
122 int32x4_t vacc45xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4xCDEF), vreinterpretq_u64_s32(vksumzp5xCDEF)));
123
124 size_t p = ks;
125 do {
126 const int8_t* restrict a0 = a[0];
127 if XNN_UNPREDICTABLE(a0 != zero) {
128 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
129 } else {
130 a0 = zero_data;
131 }
132 const int8_t* restrict a1 = a[1];
133 if XNN_UNPREDICTABLE(a1 != zero) {
134 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
135 } else {
136 a1 = zero_data;
137 }
138 const int8_t* restrict a2 = a[2];
139 if XNN_UNPREDICTABLE(a2 != zero) {
140 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
141 } else {
142 a2 = zero_data;
143 }
144 const int8_t* restrict a3 = a[3];
145 if XNN_UNPREDICTABLE(a3 != zero) {
146 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
147 } else {
148 a3 = zero_data;
149 }
150 const int8_t* restrict a4 = a[4];
151 if XNN_UNPREDICTABLE(a4 != zero) {
152 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
153 } else {
154 a4 = zero_data;
155 }
156 const int8_t* restrict a5 = a[5];
157 if XNN_UNPREDICTABLE(a5 != zero) {
158 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
159 } else {
160 a5 = zero_data;
161 }
162 a += 6;
163
164 uint64x2x2_t va01x0123456789ABCDEF;
165 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
166 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
167 uint64x2x2_t va23x0123456789ABCDEF;
168 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
169 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
170 uint64x2x2_t va45x0123456789ABCDEF;
171 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
172 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
173
174 // Inner accumulation loop along the 16 columns.
175 size_t k = kc;
176 // 2x partial unrolled loop to load 8 bytes at a time.
177 while (k >= 16 * sizeof(int8_t)) {
178 // Load a 6x16 block of activations.
179 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
180 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
181 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
182 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
183 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
184 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
185
186 // Load a 16x16 block of weights.
187 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
188 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
189 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
190 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
191 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
192 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
193 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
194 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
195 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
196 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
197 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
198 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
199 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
200 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
201 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
202 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
203
204 // Multiply-accumulate: 6x16 * 16x16 --> 6x16.
205 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
206 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
207 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
208 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
209 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
210 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
211 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
212 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
213 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
214 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
215 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
216 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
217 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb89x01234567);
218 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbABx01234567);
219 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbCDx01234567);
220 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbEFx01234567);
221 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb01x01234567);
222 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb23x01234567);
223 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb45x01234567);
224 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb67x01234567);
225 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb89x01234567);
226 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vbABx01234567);
227 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vbCDx01234567);
228 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vbEFx01234567);
229 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
230 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
231 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
232 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
233 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
234 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
235 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
236 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
237 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
238 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
239 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
240 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
241 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
242 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
243 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
244 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
245 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
246 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
247 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
248 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
249 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
250 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
251 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
252 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
253
254 k -= 16 * sizeof(int8_t);
255 }
256 // Handle up to 8 final positions of `k`
257 if XNN_UNLIKELY(k != 0) {
258 // Load a 6x8 block of activations.
259 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
260 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
261 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
262 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
263 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
264 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
265
266 // Load a 16x16 block of weights.
267 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
268 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
269 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
270 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
271 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
272 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
273 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
274 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
275
276 // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
277 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
278 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
279 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
280 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
281 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
282 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
283 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
284 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
285 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
286 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
287 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
288 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
289 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
290 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
291 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
292 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
293 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
294 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
295 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
296 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
297 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x01234567), vb89x01234567);
298 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x01234567), vbABx01234567);
299 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x01234567), vbCDx01234567);
300 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x01234567), vbEFx01234567);
301 }
302
303 p -= 6 * sizeof(void*);
304 } while (p != 0);
305
306 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
307 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
308 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
309 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
310 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
311 int32x4_t vacc1x89AB = vcombine_s32(vget_high_s32(vacc01x89), vget_high_s32(vacc01xAB));
312 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
313 int32x4_t vacc1xCDEF = vcombine_s32(vget_high_s32(vacc01xCD), vget_high_s32(vacc01xEF));
314 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
315 int32x4_t vacc3x0123 = vcombine_s32(vget_high_s32(vacc23x01), vget_high_s32(vacc23x23));
316 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
317 int32x4_t vacc3x4567 = vcombine_s32(vget_high_s32(vacc23x45), vget_high_s32(vacc23x67));
318 int32x4_t vacc2x89AB = vcombine_s32(vget_low_s32(vacc23x89), vget_low_s32(vacc23xAB));
319 int32x4_t vacc3x89AB = vcombine_s32(vget_high_s32(vacc23x89), vget_high_s32(vacc23xAB));
320 int32x4_t vacc2xCDEF = vcombine_s32(vget_low_s32(vacc23xCD), vget_low_s32(vacc23xEF));
321 int32x4_t vacc3xCDEF = vcombine_s32(vget_high_s32(vacc23xCD), vget_high_s32(vacc23xEF));
322 int32x4_t vacc4x0123 = vcombine_s32(vget_low_s32(vacc45x01), vget_low_s32(vacc45x23));
323 int32x4_t vacc5x0123 = vcombine_s32(vget_high_s32(vacc45x01), vget_high_s32(vacc45x23));
324 int32x4_t vacc4x4567 = vcombine_s32(vget_low_s32(vacc45x45), vget_low_s32(vacc45x67));
325 int32x4_t vacc5x4567 = vcombine_s32(vget_high_s32(vacc45x45), vget_high_s32(vacc45x67));
326 int32x4_t vacc4x89AB = vcombine_s32(vget_low_s32(vacc45x89), vget_low_s32(vacc45xAB));
327 int32x4_t vacc5x89AB = vcombine_s32(vget_high_s32(vacc45x89), vget_high_s32(vacc45xAB));
328 int32x4_t vacc4xCDEF = vcombine_s32(vget_low_s32(vacc45xCD), vget_low_s32(vacc45xEF));
329 int32x4_t vacc5xCDEF = vcombine_s32(vget_high_s32(vacc45xCD), vget_high_s32(vacc45xEF));
330
331 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
332 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
333 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
334 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
335 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
336 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
337 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
338 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
339 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
340 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
341 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
342 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
343 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
344 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
345 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
346 float32x4_t vout3x89AB = vcvtq_f32_s32(vacc3x89AB);
347 float32x4_t vout3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
348 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
349 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
350 float32x4_t vout4x89AB = vcvtq_f32_s32(vacc4x89AB);
351 float32x4_t vout4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
352 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
353 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
354 float32x4_t vout5x89AB = vcvtq_f32_s32(vacc5x89AB);
355 float32x4_t vout5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
356 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
357 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
358 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
359 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
360 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
361 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
362 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
363 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
364 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
365 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
366 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
367 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
368 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale);
369 vout3x89AB = vmulq_f32(vout3x89AB, vinput_scale);
370 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale);
371 vout3xCDEF = vmulq_f32(vout3xCDEF, vinput_scale);
372 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale);
373 vout5x0123 = vmulq_f32(vout5x0123, vinput_scale);
374 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale);
375 vout5x4567 = vmulq_f32(vout5x4567, vinput_scale);
376 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale);
377 vout5x89AB = vmulq_f32(vout5x89AB, vinput_scale);
378 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale);
379 vout5xCDEF = vmulq_f32(vout5xCDEF, vinput_scale);
380
381 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
382 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
383 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
384 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
385
386 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
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 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
394 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
395 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
396 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
397 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
398 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
399 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
400 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
401 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
402 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
403 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
404 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
405 vout4x89AB = vfmaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
406 vout5x89AB = vfmaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
407 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
408 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
409 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
410 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
411 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
412 vout4xCDEF = vfmaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
413 vout5xCDEF = vfmaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
414
415 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
416 float16x8_t vfp16out5x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout5x89AB), vcvt_f16_f32(vout5xCDEF));
417 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
418 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
419 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
420 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
421 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
422 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
423 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
424 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
425 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
426 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
427 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
428 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
429 vfp16out5x89ABCDEF = vmaxq_f16(vfp16out5x89ABCDEF, voutput_min);
430 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
431 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
432 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
433 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
434 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
435 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
436 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
437 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
438 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
439 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
440 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
441 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
442 vfp16out5x89ABCDEF = vminq_f16(vfp16out5x89ABCDEF, voutput_max);
443 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
444 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
445 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
446 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
447 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
448 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
449 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
450 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
451 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
452 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
453
454 if XNN_LIKELY(nc >= 16) {
455 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
456 vst1q_u16(c5 + 8, vreinterpretq_u16_f16(vfp16out5x89ABCDEF));
457 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
458 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
459 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
460 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
461 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
462 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
463 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
464 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
465 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
466 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
467
468 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
469 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
470 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
471 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
472 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
473 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
474
475 a = (const int8_t**restrict) ((uintptr_t) a - ks);
476
477 nc -= 16;
478 } else {
479 if (nc & 8) {
480 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567)); c5 += 8;
481 vfp16out5x01234567 = vfp16out5x89ABCDEF;
482 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
483 vfp16out4x01234567 = vfp16out4x89ABCDEF;
484 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
485 vfp16out3x01234567 = vfp16out3x89ABCDEF;
486 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
487 vfp16out2x01234567 = vfp16out2x89ABCDEF;
488 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
489 vfp16out1x01234567 = vfp16out1x89ABCDEF;
490 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
491 vfp16out0x01234567 = vfp16out0x89ABCDEF;
492 }
493 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
494 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
495 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
496 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
497 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
498 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
499 if (nc & 4) {
500 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
501 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
502 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
503 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
504 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
505 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
506 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
507 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
508 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
509 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
510 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
511 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
512 }
513 if (nc & 2) {
514 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
515 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
516 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
517 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
518 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
519 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
520 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
521 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
522 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
523 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
524 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
525 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
526 }
527 if (nc & 1) {
528 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
529 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
530 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
531 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
532 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
533 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
534 }
535 nc = 0;
536 }
537 } while (nc != 0);
538 }