Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x16-minmax-neonfp16arith-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mlal-lane.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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/prefetch.h"
18
19
20 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_1x16__neonfp16arith_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 const int8_t* a0 = a;
43 uint16_t* c0 = (uint16_t*) c;
44
45 const int8x8_t vmask = vmov_n_s8(INT8_C(0xF0));
46 kc = round_up_po2(kc, 2);
47 do {
48 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
49 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
50 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
52 const int32x4_t vzp0 = vld1q_dup_s32(&quantization_params[0].zero_point);
53 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vzp0);
54 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vzp0);
55 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vzp0);
56 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vzp0);
57
58 size_t k = kc;
59 while (k >= 8 * sizeof(int8_t)) {
60 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
61 const int16x8_t vxa0 = vmovl_s8(va0);
62
63 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const uint8_t*) w + 8;
64 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const uint8_t*) w + 8;
65 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const uint8_t*) w + 8;
66 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const uint8_t*) w + 8;
67 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const uint8_t*) w + 8;
68 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const uint8_t*) w + 8;
69 const int8x8_t vb01234567c67 = vld1_s8(w); w = (const uint8_t*) w + 8;
70 const int8x8_t vb89ABCDEFc67 = vld1_s8(w); w = (const uint8_t*) w + 8;
71 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
72 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
73 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
74 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
75 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
76 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
77 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
78 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
79 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
80 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
81 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
82 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
83 const int8x8_t vb01234567c6 = vshl_n_s8(vb01234567c67, 4);
84 const int8x8_t vb01234567c7 = vand_s8(vb01234567c67, vmask);
85 const int8x8_t vb89ABCDEFc6 = vshl_n_s8(vb89ABCDEFc67, 4);
86 const int8x8_t vb89ABCDEFc7 = vand_s8(vb89ABCDEFc67, vmask);
87
88 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
89 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
90 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
91 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
92 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
93 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
94 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
95 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
96 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
97 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
98 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
99 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
100 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
101 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
102 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
103 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
104
105 xnn_prefetch_to_l1((const int8_t*) w + 448);
106 xnn_prefetch_to_l1((const int8_t*) w + 512);
107
108 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
109 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
110 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
111 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
112 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
113 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
114 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
115 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
116 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
117 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
118 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
119 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
120 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
121 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
122 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
123 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
124
125 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
126 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
127 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
128 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
129 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
130 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
131 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
132 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
133 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
134 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
135 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
136 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
137 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
138 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
139 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
140 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
141
142 k -= 8 * sizeof(int8_t);
143 }
144 if XNN_UNLIKELY(k != 0) {
145 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
146 const int16x8_t vxa0 = vmovl_s8(va0);
147
148 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const int8_t*) w + 8;
149 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const int8_t*) w + 8;
150 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
151 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
152 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
153 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
154
155 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
156 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
157
158 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
159 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
160 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
161 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
162
163 if (k >= 2 * sizeof(int8_t)) {
164 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
165 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
166
167 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
168 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
169 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
170 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
171
172 if (k > 2 * sizeof(int8_t)) {
173 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const int8_t*) w + 8;
174 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const int8_t*) w + 8;
175 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
176 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
177 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
178 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
179 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
180 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
181
182 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
183 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
184 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
185 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
186
187 if (k >= 4 * sizeof(int8_t)) {
188 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
189 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
190
191 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
192 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
193 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
194 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
195
196 if (k > 4 * sizeof(int8_t)) {
197 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const int8_t*) w + 8;
198 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const int8_t*) w + 8;
199 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
200 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
201 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
202 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
203 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
204 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
205
206 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
207 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
208 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
209 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
210
211 if (k >= 6 * sizeof(int8_t)) {
212 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
213 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
214
215 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
216 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
217 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
218 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
219
220 }
221 }
222 }
223 }
224 }
225 }
226
227 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
228 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
229 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
230 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
231
232 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
233 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
234 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
235 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
236 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
237
238 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
239 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
240 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
241 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
242
243 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
244 #if XNN_ARCH_ARM64
245 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
246 #else
247 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
248 #endif
249 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
250 #if XNN_ARCH_ARM64
251 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
252 #else
253 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
254 #endif
255 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
256 #if XNN_ARCH_ARM64
257 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
258 #else
259 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
260 #endif
261 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
262 #if XNN_ARCH_ARM64
263 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
264 #else
265 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
266 #endif
267
268 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
269 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
270 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
271 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
272 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
273 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
274 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
275 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
276
277 if XNN_LIKELY(nc >= 16) {
278 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
279 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
280
281 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
282
283 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
284
285 nc -= 16;
286 } else {
287 if (nc & 8) {
288 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
289 vfp16out0x01234567 = vfp16out0x89ABCDEF;
290 }
291 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
292 if (nc & 4) {
293 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
294 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
295 }
296 if (nc & 2) {
297 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
298 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
299 }
300 if (nc & 1) {
301 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
302 }
303 nc = 0;
304 }
305 } while (nc != 0);
306 }