Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x16-minmax-neonfp16arith-mlal-lane.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
18
19 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_1x16__neonfp16arith_mlal_lane(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 xnn_float16* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43
44 const int8x8_t vmask = vmov_n_s8(INT8_C(0xF0));
45 kc = round_up_po2(kc, 2);
46 do {
47 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
48 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
49 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
50 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
51 const int32x4_t vzp0 = vld1q_dup_s32(&quantization_params[0].zero_point);
52 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vzp0);
53 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vzp0);
54 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vzp0);
55 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vzp0);
56
57 size_t k = kc;
58 while (k >= 8 * sizeof(int8_t)) {
59 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
60 const int16x8_t vxa0 = vmovl_s8(va0);
61
62 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const uint8_t*) w + 8;
63 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const uint8_t*) w + 8;
64 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const uint8_t*) w + 8;
65 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const uint8_t*) w + 8;
66 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const uint8_t*) w + 8;
67 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const uint8_t*) w + 8;
68 const int8x8_t vb01234567c67 = vld1_s8(w); w = (const uint8_t*) w + 8;
69 const int8x8_t vb89ABCDEFc67 = vld1_s8(w); w = (const uint8_t*) w + 8;
70 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
71 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
72 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
73 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
74 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
75 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
76 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
77 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
78 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
79 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
80 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
81 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
82 const int8x8_t vb01234567c6 = vshl_n_s8(vb01234567c67, 4);
83 const int8x8_t vb01234567c7 = vand_s8(vb01234567c67, vmask);
84 const int8x8_t vb89ABCDEFc6 = vshl_n_s8(vb89ABCDEFc67, 4);
85 const int8x8_t vb89ABCDEFc7 = vand_s8(vb89ABCDEFc67, vmask);
86
87 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
88 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
89 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
90 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
91 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
92 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
93 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
94 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
95 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
96 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
97 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
98 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
99 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
100 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
101 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
102 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
103
104
105 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
106 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
107 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
108 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
110 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
111 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
112 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
113 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
114 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
115 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
116 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
117 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
118 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
119 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
120 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
121
122 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
123 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
124 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
125 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
128 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
129 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
130 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
131 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
132 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
133 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
134 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
135 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
136 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
137 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
138
139 k -= 8 * sizeof(int8_t);
140 }
141 if XNN_UNLIKELY(k != 0) {
142 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
143 const int16x8_t vxa0 = vmovl_s8(va0);
144
145 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const int8_t*) w + 8;
146 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
148 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
149 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
150 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
151
152 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
153 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
154
155 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
156 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
157 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
158 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
159
160 if (k >= 2 * sizeof(int8_t)) {
161 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
162 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
163
164 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
165 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
166 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
167 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
168
169 if (k > 2 * sizeof(int8_t)) {
170 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const int8_t*) w + 8;
171 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
173 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
174 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
175 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
176 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
177 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
178
179 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
180 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
181 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
182 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
183
184 if (k >= 4 * sizeof(int8_t)) {
185 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
186 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
187
188 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
189 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
190 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
191 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
192
193 if (k > 4 * sizeof(int8_t)) {
194 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const int8_t*) w + 8;
195 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const int8_t*) w + 8;
196 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
197 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
198 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
199 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
200 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
201 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
202
203 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
204 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
205 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
206 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
207
208 if (k >= 6 * sizeof(int8_t)) {
209 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
210 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
211
212 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
213 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
214 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
215 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
216
217 }
218 }
219 }
220 }
221 }
222 }
223
224 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
225 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
226 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
227 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
228
229 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
230 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
231 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
232 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
233 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
234
235 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
236 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
237 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
238 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
239
240 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
241 #if XNN_ARCH_ARM64
242 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
243 #else
244 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
245 #endif
246 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
247 #if XNN_ARCH_ARM64
248 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
249 #else
250 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
251 #endif
252 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
253 #if XNN_ARCH_ARM64
254 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
255 #else
256 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
257 #endif
258 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
259 #if XNN_ARCH_ARM64
260 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
261 #else
262 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
263 #endif
264
265 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
266 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
267 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
268 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
269 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
270 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
271 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
272 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
273
274 if XNN_LIKELY(nc >= 16) {
275 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
276 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
277
278 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
279
280 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
281
282 nc -= 16;
283 } else {
284 if (nc & 8) {
285 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
286 vfp16out0x01234567 = vfp16out0x89ABCDEF;
287 }
288 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
289 if (nc & 4) {
290 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
291 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
292 }
293 if (nc & 2) {
294 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
295 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
296 }
297 if (nc & 1) {
298 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
299 }
300 nc = 0;
301 }
302 } while (nc != 0);
303 }