xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/6x16c4-minmax-fp32-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 6);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (6 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 4 * sizeof(int8_t));
45   int8_t* c0 = c;
46   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr < 2) {
48     c1 = c0;
49   }
50   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     c2 = c1;
53   }
54   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr < 4) {
56     c3 = c2;
57   }
58   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59   if XNN_UNPREDICTABLE(mr <= 4) {
60     c4 = c3;
61   }
62   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63   if XNN_UNPREDICTABLE(mr != 6) {
64     c5 = c4;
65   }
66 
67   do {
68     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
69     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
70     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
71     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
72     int32x4_t vacc1x0123 = vacc0x0123;
73     int32x4_t vacc1x4567 = vacc0x4567;
74     int32x4_t vacc1x89AB = vacc0x89AB;
75     int32x4_t vacc1xCDEF = vacc0xCDEF;
76     int32x4_t vacc2x0123 = vacc0x0123;
77     int32x4_t vacc2x4567 = vacc0x4567;
78     int32x4_t vacc2x89AB = vacc0x89AB;
79     int32x4_t vacc2xCDEF = vacc0xCDEF;
80     int32x4_t vacc3x0123 = vacc0x0123;
81     int32x4_t vacc3x4567 = vacc0x4567;
82     int32x4_t vacc3x89AB = vacc0x89AB;
83     int32x4_t vacc3xCDEF = vacc0xCDEF;
84     int32x4_t vacc4x0123 = vacc0x0123;
85     int32x4_t vacc4x4567 = vacc0x4567;
86     int32x4_t vacc4x89AB = vacc0x89AB;
87     int32x4_t vacc4xCDEF = vacc0xCDEF;
88     int32x4_t vacc5x0123 = vacc0x0123;
89     int32x4_t vacc5x4567 = vacc0x4567;
90     int32x4_t vacc5x89AB = vacc0x89AB;
91     int32x4_t vacc5xCDEF = vacc0xCDEF;
92 
93     size_t p = ks;
94     do {
95       const int8_t* restrict a0 = a[0];
96       if XNN_UNPREDICTABLE(a0 != zero) {
97         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
98       }
99       const int8_t* restrict a1 = a[1];
100       if XNN_UNPREDICTABLE(a1 != zero) {
101         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
102       }
103       const int8_t* restrict a2 = a[2];
104       if XNN_UNPREDICTABLE(a2 != zero) {
105         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
106       }
107       const int8_t* restrict a3 = a[3];
108       if XNN_UNPREDICTABLE(a3 != zero) {
109         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
110       }
111       const int8_t* restrict a4 = a[4];
112       if XNN_UNPREDICTABLE(a4 != zero) {
113         a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
114       }
115       const int8_t* restrict a5 = a[5];
116       if XNN_UNPREDICTABLE(a5 != zero) {
117         a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
118       }
119       a += 6;
120 
121       // Inner accumulation loop along the 16 columns.
122       size_t k = kc;
123       // 2x partial unrolled loop to load 8 bytes at a time.
124       while (k >= 8 * sizeof(int8_t)) {
125         // Load a 6x8 block of activations.
126         const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
127         const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
128         const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
129         const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
130         const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
131         const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
132 
133         // Load a 8x16 block of weights.
134         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
135         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
136         const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
137         const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
138         const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
139         const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
140         const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
141         const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
142 
143         // Multiply-accumulate: 6x8 * 8x16 --> 6x16.
144         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
145         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
146         vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
147         vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
148         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
149         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
150         vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
151         vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
152         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
153         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
154         vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
155         vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
156         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
157         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
158         vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
159         vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
160         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
161         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
162         vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
163         vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
164         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
165         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
166         vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
167         vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
168         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
169         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
170         vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
171         vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
172         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
173         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
174         vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
175         vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
176         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
177         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
178         vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
179         vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
180         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
181         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
182         vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
183         vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
184         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
185         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
186         vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
187         vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
188         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
189         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
190         vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
191         vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
192 
193         k -= 8 * sizeof(int8_t);
194       }
195       // Handle up to 4 final positions of `k`
196       if XNN_UNLIKELY(k != 0) {
197         // Load a 6x4 block of activations.
198         const int8x8_t va0x01234567 = vld1_s8(a0);
199         const int8x8_t va1x01234567 = vld1_s8(a1);
200         const int8x8_t va2x01234567 = vld1_s8(a2);
201         const int8x8_t va3x01234567 = vld1_s8(a3);
202         const int8x8_t va4x01234567 = vld1_s8(a4);
203         const int8x8_t va5x01234567 = vld1_s8(a5);
204 
205         // Load a 4x16 block of weights.
206         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
207         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
208         const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
209         const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
210 
211         // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
212         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
213         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
214         vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
215         vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
216         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
217         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
218         vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
219         vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
220         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
221         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
222         vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
223         vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
224         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
225         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
226         vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
227         vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
228         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
229         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
230         vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
231         vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
232         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
233         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
234         vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
235         vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
236       }
237       p -= 6 * sizeof(void*);
238     } while (p != 0);
239 
240     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
241     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
242     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
243     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
244     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
245     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
246     float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
247     float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
248     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
249     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
250     float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
251     float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
252     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
253     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
254     float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
255     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
256     float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
257     float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
258     float32x4_t vfpacc4x89AB = vcvtq_f32_s32(vacc4x89AB);
259     float32x4_t vfpacc4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
260     float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
261     float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
262     float32x4_t vfpacc5x89AB = vcvtq_f32_s32(vacc5x89AB);
263     float32x4_t vfpacc5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
264 
265     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
266     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
267     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
268     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
269     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
270     vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
271     vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
272     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
273     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
274     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
275     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
276     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
277     vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
278     vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
279     const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
280     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
281     vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
282     vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
283     vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
284     vfpacc4x89AB = vmulq_f32(vfpacc4x89AB, vscale89AB);
285     vfpacc5x89AB = vmulq_f32(vfpacc5x89AB, vscale89AB);
286     const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
287     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
288     vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
289     vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
290     vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
291     vfpacc4xCDEF = vmulq_f32(vfpacc4xCDEF, vscaleCDEF);
292     vfpacc5xCDEF = vmulq_f32(vfpacc5xCDEF, vscaleCDEF);
293 
294     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
295     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
296     vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
297     vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
298     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
299     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
300     vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
301     vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
302     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
303     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
304     vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
305     vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
306     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
307     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
308     vacc3x89AB = vcvtnq_s32_f32(vfpacc3x89AB);
309     vacc3xCDEF = vcvtnq_s32_f32(vfpacc3xCDEF);
310     vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
311     vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
312     vacc4x89AB = vcvtnq_s32_f32(vfpacc4x89AB);
313     vacc4xCDEF = vcvtnq_s32_f32(vfpacc4xCDEF);
314     vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
315     vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
316     vacc5x89AB = vcvtnq_s32_f32(vfpacc5x89AB);
317     vacc5xCDEF = vcvtnq_s32_f32(vfpacc5xCDEF);
318 
319     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
320 #if XNN_ARCH_ARM64
321     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
322     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
323     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
324     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
325     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
326     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
327     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
328     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
329     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
330     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
331     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
332     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF), voutput_zero_point);
333 
334     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
335     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
336     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
337     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
338     int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
339     int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
340 #else
341     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
342     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
343     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
344     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
345     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
346     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
347     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
348     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
349     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
350     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
351     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
352     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF)), voutput_zero_point);
353 
354     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
355     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
356     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
357     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
358     int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
359     int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
360 #endif
361     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
362     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
363 
364     vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
365     vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
366     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
367     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
368     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
369     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
370 
371     vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
372     vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
373     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
374     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
375     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
376     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
377 
378     if (nc >= 16) {
379       vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
380       vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
381       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
382       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
383       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
384       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
385 
386       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
387       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
388       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
389       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
390       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
391       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
392 
393       a = (const int8_t**restrict) ((uintptr_t) a - ks);
394 
395       nc -= 16;
396     } else {
397       int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
398       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
399       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
400       if (nc & 8) {
401         vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
402         vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
403         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
404         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
405         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
406         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
407         vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
408         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
409         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
410       }
411       if (nc & 4) {
412         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
413         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
414         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
415         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
416         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
417         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
418         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
419         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
420         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
421       }
422       if (nc & 2) {
423         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
424         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
425         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
426         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
427         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
428         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
429         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
430         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
431         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
432       }
433       if (nc & 1) {
434         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
435         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
436         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
437         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
438         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
439         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
440       }
441 
442       nc = 0;
443     }
444   } while (nc != 0);
445 }
446