xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/8x8c4-minmax-fp32-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot(
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     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 8);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 4 * sizeof(int8_t));
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     a1 = a0;
47     c1 = c0;
48   }
49   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     a2 = a1;
53     c2 = c1;
54   }
55   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
57   if XNN_UNPREDICTABLE(mr < 4) {
58     a3 = a2;
59     c3 = c2;
60   }
61   const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
63   if XNN_UNPREDICTABLE(mr <= 4) {
64     a4 = a3;
65     c4 = c3;
66   }
67   const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
69   if XNN_UNPREDICTABLE(mr < 6) {
70     a5 = a4;
71     c5 = c4;
72   }
73   const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
74   int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
75   if XNN_UNPREDICTABLE(mr <= 6) {
76     a6 = a5;
77     c6 = c5;
78   }
79   const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
80   int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
81   if XNN_UNPREDICTABLE(mr != 8) {
82     a7 = a6;
83     c7 = c6;
84   }
85 
86   // Loop over groups of 8 columns.
87   do {
88     // Initialize accumulators with bias. 8 bias values are loaded from the
89     // weight matrix, at the start of the group of 8 columns.
90     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
92     int32x4_t vacc1x0123 = vacc0x0123;
93     int32x4_t vacc1x4567 = vacc0x4567;
94     int32x4_t vacc2x0123 = vacc0x0123;
95     int32x4_t vacc2x4567 = vacc0x4567;
96     int32x4_t vacc3x0123 = vacc0x0123;
97     int32x4_t vacc3x4567 = vacc0x4567;
98     int32x4_t vacc4x0123 = vacc0x0123;
99     int32x4_t vacc4x4567 = vacc0x4567;
100     int32x4_t vacc5x0123 = vacc0x0123;
101     int32x4_t vacc5x4567 = vacc0x4567;
102     int32x4_t vacc6x0123 = vacc0x0123;
103     int32x4_t vacc6x4567 = vacc0x4567;
104     int32x4_t vacc7x0123 = vacc0x0123;
105     int32x4_t vacc7x4567 = vacc0x4567;
106 
107     // Inner accumulation loop along the 8 columns.
108     size_t k = kc;
109     // 2x partial unrolled loop to load 8 bytes at a time.
110     while (k >= 8 * sizeof(int8_t)) {
111       // Load a 8x8 block of activations.
112       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
113       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
114       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
115       const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
116       const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
117       const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
118       const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8;
119       const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8;
120 
121       // Load a 8x8 block of weights.
122       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
123       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
124       const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
125       const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
126 
127       // Multiply-accumulate: 8x8 * 8x8 --> 8x8.
128       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
129       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
130       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
131       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
132       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
133       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
134       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
135       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
136       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
137       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
138       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
139       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
140       vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
141       vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
142       vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
143       vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
144       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
145       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
146       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
147       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
148       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
149       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
150       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
151       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
152       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
153       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
154       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
155       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
156       vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1);
157       vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1);
158       vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1);
159       vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1);
160 
161       k -= 8 * sizeof(int8_t);
162     }
163     // Handle up to 4 final positions of `k`
164     if XNN_UNLIKELY(k != 0) {
165       // Load a 8x4 block of activations.
166       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
167       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
168       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
169       const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
170       const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
171       const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
172       const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 4;
173       const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 4;
174 
175       // Load a 4x8 block of weights.
176       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
177       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
178 
179       // Multiply-accumulate: 8x4 * 4x8 --> 8x8.
180       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
181       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
182       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
183       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
184       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
185       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
186       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
187       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
188       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
189       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
190       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
191       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
192       vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
193       vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
194       vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
195       vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
196     }
197 
198     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
199     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
200     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
201     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
202     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
203     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
204     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
205     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
206     float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
207     float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
208     float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
209     float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
210     float32x4_t vfpacc6x0123 = vcvtq_f32_s32(vacc6x0123);
211     float32x4_t vfpacc6x4567 = vcvtq_f32_s32(vacc6x4567);
212     float32x4_t vfpacc7x0123 = vcvtq_f32_s32(vacc7x0123);
213     float32x4_t vfpacc7x4567 = vcvtq_f32_s32(vacc7x4567);
214 
215     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
216     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
217     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
218     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
219     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
220     vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
221     vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
222     vfpacc6x0123 = vmulq_f32(vfpacc6x0123, vscale0123);
223     vfpacc7x0123 = vmulq_f32(vfpacc7x0123, vscale0123);
224     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
225     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
226     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
227     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
228     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
229     vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
230     vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
231     vfpacc6x4567 = vmulq_f32(vfpacc6x4567, vscale4567);
232     vfpacc7x4567 = vmulq_f32(vfpacc7x4567, vscale4567);
233 
234     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
235     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
236     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
237     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
238     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
239     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
240     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
241     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
242     vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
243     vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
244     vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
245     vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
246     vacc6x0123 = vcvtnq_s32_f32(vfpacc6x0123);
247     vacc6x4567 = vcvtnq_s32_f32(vfpacc6x4567);
248     vacc7x0123 = vcvtnq_s32_f32(vfpacc7x0123);
249     vacc7x4567 = vcvtnq_s32_f32(vfpacc7x4567);
250 
251     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
252 #if XNN_ARCH_ARM64
253     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
254     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
255     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
256     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
257     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
258     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
259     const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point);
260     const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point);
261 
262     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
263     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
264     int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
265     int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567);
266 #else
267     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
268     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
269     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
270     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
271     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
272     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
273     const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);
274     const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point);
275 
276     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
277     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
278     int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
279     int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x01234567));
280 #endif
281     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
282     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
283 
284     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
285     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
286     vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
287     vout6x01234567_7x01234567 = vmaxq_s8(vout6x01234567_7x01234567, voutput_min);
288 
289     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
290     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
291     vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
292     vout6x01234567_7x01234567 = vminq_s8(vout6x01234567_7x01234567, voutput_max);
293 
294     if (nc >= 8) {
295       // Main case where there the 8 columns fit in the destination.
296       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
297       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
298       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
299       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
300       vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
301       vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
302       vst1_s8(c6 + 0, vget_low_s8(vout6x01234567_7x01234567));
303       vst1_s8(c7 + 0, vget_high_s8(vout6x01234567_7x01234567));
304 
305       // Advance to the next 8 columns.
306       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
307       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
308       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
309       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
310       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
311       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
312       c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
313       c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
314 
315       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
316       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
317       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
318       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
319       a4 = (const int8_t*) ((uintptr_t) a4 - kc);
320       a5 = (const int8_t*) ((uintptr_t) a5 - kc);
321       a6 = (const int8_t*) ((uintptr_t) a6 - kc);
322       a7 = (const int8_t*) ((uintptr_t) a7 - kc);
323 
324       nc -= 8;
325     } else {
326       // Final case where not all of the 8 columns fit in the destination.
327       if (nc & 4) {
328         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
329         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
330         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
331         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
332         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
333         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
334         vst1q_lane_u32((void*) c6, vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4;
335         vst1q_lane_u32((void*) c7, vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4;
336         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
337         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
338         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
339         vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4);
340       }
341       if (nc & 2) {
342         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
343         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
344         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
345         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
346         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
347         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
348         vst1q_lane_u16((void*) c6, vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2;
349         vst1q_lane_u16((void*) c7, vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2;
350         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
351         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
352         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
353         vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2);
354       }
355       if (nc & 1) {
356         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
357         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
358         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
359         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
360         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
361         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
362         vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0);
363         vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8);
364       }
365 
366       nc = 0;
367     }
368   } while (nc != 0);
369 }
370