xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x16c2-minmax-rndnu-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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/math.h>
16 
xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 2);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr != 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47 
48   do {
49     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53     int32x4_t vacc1x0123 = vacc0x0123;
54     int32x4_t vacc1x4567 = vacc0x4567;
55     int32x4_t vacc1x89AB = vacc0x89AB;
56     int32x4_t vacc1xCDEF = vacc0xCDEF;
57 
58     size_t k = kc;
59 
60     while (k >= 16 * sizeof(int8_t)) {
61       const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
62       const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
63       const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
64       const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
65       const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
66       const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
67       const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
68       const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
69       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 
86       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
87       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
88       const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
89       const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
90 
91       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
92       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
93       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
95       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
96       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
97       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
98       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
99       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
100       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
102       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
103       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
104       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
105       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
106       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
107       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
109       vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
110       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
111       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
112       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
113       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
114       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
116       vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
117       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
118       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
119       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
120       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
121       const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
122       const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
123 
124       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
125       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
126       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
128       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
129       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
130       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
131       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
132       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
133       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
135       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
136       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
137       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
138       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
139       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
140       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
142       vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
143       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
144       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
145       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
146       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
147       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
149       vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
150       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
151       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
152       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
153       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
154       const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
155       const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
156 
157       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
158       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
159       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
161       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
162       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
163       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
164       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
165       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
166       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
168       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
169       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
170       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
171       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
172       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
173       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
175       vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
176       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
177       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
178       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
179       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
180       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
182       vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
183       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
184       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
185       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
186       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
187       const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
188       const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
189 
190       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
191       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
192       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
194       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
195       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
196       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
197       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
198       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
199       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
201       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
202       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
203       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
204       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
205       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
206       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
208       vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
209       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
210       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
211       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
212       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
213       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
215       vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
216       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
217       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
218 
219       k -= 16 * sizeof(int8_t);
220     }
221 
222     if (k >= 8 * sizeof(int8_t)) {
223       const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
224       const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
225       const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
226       const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
227 
228       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244 
245       const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
246       const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
247 
248       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
249       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
250       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
251       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
252       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
253       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
254       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
255       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
256       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
257       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
258       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
259       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
260       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
261       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
262       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
263       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
264       const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
265       const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
266 
267       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
268       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
269       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
270       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
271       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
272       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
273       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
274       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
275       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
276       const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
277       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
278       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
279       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
280       const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
281       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
282       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
283       const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
284       const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
285 
286       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
287       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
288       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
289       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
290       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
291       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
292       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
293       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
294       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
295       const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
296       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
297       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
298       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
299       const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
300       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
301       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
302       const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
303       const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
304 
305       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
306       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
307       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
308       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
309       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
310       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
311       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
312       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
313       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
314       const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
315       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
316       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
317       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
318       const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
319       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
320       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
321 
322       k -= 8 * sizeof(int8_t);
323     }
324 
325     if XNN_UNLIKELY(k != 0) {
326       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
327       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
328 
329       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
330       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
331       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
332       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
333 
334       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
335       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
336       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
337       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
338       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
339       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
340       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
341       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
342       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
343       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
344       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
345       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
346       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
347       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
348       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
349       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
350       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
351       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
352 
353       if (k > 2 * sizeof(int8_t)) {
354         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
355         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
356         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
357         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
358 
359         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
360         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
361         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
362         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
363         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
364         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
365         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
366         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
367         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
368         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
369         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
370         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
371         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
372         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
373         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
374         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
375         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
376         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
377 
378         if (k > 4 * sizeof(int8_t)) {
379           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
380           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
382           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
383 
384           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
385           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
386           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
387           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
388           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
389           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
390           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
391           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
392           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
393           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
394           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
395           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
396           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
397           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
398           const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
399           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
400           const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
401           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
402         }
403       }
404     }
405 
406     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
407     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
408     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
409 
410     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
411     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
412     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
413     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
414     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
415     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
416     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
417     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
418 
419     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
420     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
421     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
422     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
423     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
424     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
425     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
426     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
427 
428     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
429     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
430     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
431     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
432     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
433     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
434     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
435     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
436 
437     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
438 #if XNN_ARCH_ARM64
439     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
440     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
441     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
442     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
443 
444     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
445     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
446     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
447     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
448 
449     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
450     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
451 #else
452     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
453     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
454     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
455     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
456 
457     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
458     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
459     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
460     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
461 
462     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
463     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
464 #endif
465 
466     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
467     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
468     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
469 
470     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
471     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
472     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
473 
474     if (nc >= 16) {
475       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
476       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
477 
478       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
479       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
480 
481       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
482       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
483 
484       nc -= 16;
485     } else {
486       // Final case where not all of the 16 columns fit in the destination.
487       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
488       if (nc & 8) {
489         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
490         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
491         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
492       }
493       if (nc & 4) {
494         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
495         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
496         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
497       }
498       if (nc & 2) {
499         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
500         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
501         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
502       }
503       if (nc & 1) {
504         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
505         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
506       }
507 
508       nc = 0;
509     }
510   } while (nc != 0);
511 }
512