xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x9-minmax-fp32-neonv8-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul16.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/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_mul16(
19     size_t channels,
20     size_t output_width,
21     const int8_t** input,
22     const void* weights,
23     int8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const int8_t* zero,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
34   const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
36   do {
37     const int8_t* i0 = input[0];
38     assert(i0 != NULL);
39     if XNN_UNPREDICTABLE(i0 != zero) {
40       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41     }
42     const int8_t* i1 = input[1];
43     assert(i1 != NULL);
44     if XNN_UNPREDICTABLE(i1 != zero) {
45       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46     }
47     const int8_t* i2 = input[2];
48     assert(i2 != NULL);
49     if XNN_UNPREDICTABLE(i2 != zero) {
50       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51     }
52     const int8_t* i3 = input[3];
53     assert(i3 != NULL);
54     if XNN_UNPREDICTABLE(i3 != zero) {
55       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56     }
57     const int8_t* i4 = input[4];
58     assert(i4 != NULL);
59     if XNN_UNPREDICTABLE(i4 != zero) {
60       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61     }
62     const int8_t* i5 = input[5];
63     assert(i5 != NULL);
64     if XNN_UNPREDICTABLE(i5 != zero) {
65       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66     }
67     const int8_t* i6 = input[6];
68     assert(i6 != NULL);
69     if XNN_UNPREDICTABLE(i6 != zero) {
70       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71     }
72     const int8_t* i7 = input[7];
73     assert(i7 != NULL);
74     if XNN_UNPREDICTABLE(i7 != zero) {
75       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76     }
77     const int8_t* i8 = input[8];
78     assert(i8 != NULL);
79     if XNN_UNPREDICTABLE(i8 != zero) {
80       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81     }
82     input = (const int8_t**) ((uintptr_t) input + input_stride);
83 
84     size_t c = channels;
85     const void* w = weights;
86     for (; c >= 16; c -= 16) {
87       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
88       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
89       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91 
92 
93       const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
94       const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
95       const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
96       const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
97 
98       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
99       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
100       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
101       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
102 
103       const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
104       const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
105       const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
106       const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
107 
108       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
109       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
110       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
111       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
112 
113       const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
114       const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
115       const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
116       const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
117 
118       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
119       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
120       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
121       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
122 
123       const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
124       const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
125       const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
126       const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
127 
128       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
129       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
130       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
131       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
132 
133       const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
134       const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
135       const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
136       const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
137 
138       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
139       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
140       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
141       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
142 
143       const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
144       const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
145       const int16x8_t vi5x89ABCDEF = vmovl_s8(vld1_s8(i5)); i5 += 8;
146       const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
147 
148       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
149       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
150       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
151       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
152 
153       const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
154       const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
155       const int16x8_t vi6x89ABCDEF = vmovl_s8(vld1_s8(i6)); i6 += 8;
156       const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
157 
158       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
159       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
160       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
161       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
162 
163       const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
164       const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
165       const int16x8_t vi7x89ABCDEF = vmovl_s8(vld1_s8(i7)); i7 += 8;
166       const int16x8_t vk7x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
167 
168       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
169       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
170       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
171       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
172 
173       const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
174       const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
175       const int16x8_t vi8x89ABCDEF = vmovl_s8(vld1_s8(i8)); i8 += 8;
176       const int16x8_t vk8x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
177 
178       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
179       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
180       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
181       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
182 
183       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
184       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
185       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
186       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
187 
188       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
189       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
190       const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
191       const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
192 
193       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
194       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
195       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
196       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
197 
198       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
199       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
200       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
201       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
202 
203 #if XNN_ARCH_ARM64
204       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
205       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
206 
207       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
208       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
209 
210       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
211 #else  // !XNN_ARCH_ARM64
212       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
213       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
214 
215       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
216       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
217 
218       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
219 #endif  // !XNN_ARCH_ARM64
220 
221       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
222 
223       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
224 
225       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
226     }
227     if XNN_UNLIKELY(c != 0) {
228       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
229       do {
230         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
231         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
232 
233         const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
234         const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(k)); k += 8;
235 
236         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
237         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
238         const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
239         const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) (k + 8)));
240 
241         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
242         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
243         const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
244         const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) (k + 24)));
245 
246         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
247         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
248         const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
249         const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) (k + 40)));
250 
251         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
252         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
253         const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
254         const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) (k + 56)));
255 
256         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
257         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
258         const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
259         const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) (k + 72)));
260 
261         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
262         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
263         const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
264         const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) (k + 88)));
265 
266         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
267         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
268         const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
269         const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) (k + 104)));
270 
271         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
272         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
273         const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
274         const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) (k + 120)));
275 
276         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
277         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
278 
279         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
280         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
281 
282         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
283         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
284         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
285         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
286 
287         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
288         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
289 
290 #if XNN_ARCH_ARM64
291         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
292 #else
293         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
294 #endif
295         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
296 
297         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
298         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
299         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
300 
301         if XNN_LIKELY(c >= 8) {
302           vst1_s8(output, vout01234567); output += 8;
303           c -= 8;
304         } else {
305           if (c & 4) {
306             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
307             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
308           }
309           if (c & 2) {
310             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
311             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
312           }
313           if (c & 1) {
314             vst1_lane_s8(output, vout01234567, 0); output += 1;
315           }
316           c = 0;
317         }
318       } while (c != 0);
319     }
320 
321     output = (int8_t*) ((uintptr_t) output + output_increment);
322   } while (--output_width != 0);
323 }
324