xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x16c2-minmax-rndnu-neon-mlal-ld1r.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_ld1r(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_ld1r(
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 int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
62       const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
63       const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
64       const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
65       const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
66       const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
67       const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
68       const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
69       const int16x4_t va10x0 = vld1_dup_s16((const void*)a1);
70       const int16x4_t va11x0 = vld1_dup_s16((const void*)(a1 + 2));
71       const int16x4_t va12x0 = vld1_dup_s16((const void*)(a1 + 4));
72       const int16x4_t va13x0 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
73       const int16x4_t va10x1 = vld1_dup_s16((const void*)a1);
74       const int16x4_t va11x1 = vld1_dup_s16((const void*)(a1 + 2));
75       const int16x4_t va12x1 = vld1_dup_s16((const void*)(a1 + 4));
76       const int16x4_t va13x1 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
77       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 
94       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
95       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
96       const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0);
97       const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1);
98 
99       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
100       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
101       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
103       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
104       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
105       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
106       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
107       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
108       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
110       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
111       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
112       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
113       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
114       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
115       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
117       vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
118       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
119       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
120       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
121       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
122       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
124       vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
125       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
126       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
127       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
128       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
129       const int8x8_t va1c1x0 = vreinterpret_s8_s16(va11x0);
130       const int8x8_t va1c1x1 = vreinterpret_s8_s16(va11x1);
131 
132       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
133       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
134       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
136       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
137       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
138       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
139       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
140       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
141       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
143       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
144       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
145       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
146       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
147       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
148       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
150       vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
151       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
152       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
153       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
154       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
155       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
157       vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
158       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
159       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
160       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
161       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
162       const int8x8_t va1c2x0 = vreinterpret_s8_s16(va12x0);
163       const int8x8_t va1c2x1 = vreinterpret_s8_s16(va12x1);
164 
165       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
166       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
167       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
169       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
170       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
171       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
172       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
173       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
174       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
176       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
177       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
178       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
179       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
180       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
181       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
183       vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
184       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
185       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
186       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
187       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
188       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
190       vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
191       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
192       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
193       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
194       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
195       const int8x8_t va1c3x0 = vreinterpret_s8_s16(va13x0);
196       const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1);
197 
198       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
199       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
200       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
201       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
202       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
203       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
204       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
205       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
206       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
207       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
209       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
210       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
211       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
212       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
213       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
214       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
216       vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
217       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
218       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
219       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
220       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
221       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
223       vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
224       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
225       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
226 
227       k -= 16 * sizeof(int8_t);
228     }
229 
230     if (k >= 8 * sizeof(int8_t)) {
231       const int16x4_t va00 = vld1_dup_s16((const void*)a0);
232       const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
233       const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
234       const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
235       const int16x4_t va10 = vld1_dup_s16((const void*)a1);
236       const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
237       const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
238       const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
239 
240       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256 
257       const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
258       const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
259 
260       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
261       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
262       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
263       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
264       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
265       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
266       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
267       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
268       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
269       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
270       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
271       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
272       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
273       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
274       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
275       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
276       const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
277       const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
278 
279       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
280       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
281       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
282       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
283       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
284       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
285       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
286       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
287       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
288       const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
289       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
290       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
291       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
292       const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
293       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
294       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
295       const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
296       const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
297 
298       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
299       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
300       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
301       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
302       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
303       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
304       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
305       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
306       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
307       const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
308       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
309       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
310       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
311       const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
312       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
313       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
314       const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
315       const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
316 
317       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
318       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
319       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
320       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
321       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
322       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
323       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
324       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
325       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
326       const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
327       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
328       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
329       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
330       const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
331       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
332       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
333 
334       k -= 8 * sizeof(int8_t);
335     }
336 
337     if XNN_UNLIKELY(k != 0) {
338       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
339       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
340 
341       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
342       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
343       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
344       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
345 
346       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
347       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
348       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
349       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
350       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
351       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
352       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
353       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
354       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
355       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
356       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
357       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
358       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
359       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
360       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
361       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
362       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
363       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
364 
365       if (k > 2 * sizeof(int8_t)) {
366         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
367         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
368         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
369         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
370 
371         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
372         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
373         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
374         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
375         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
376         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
377         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
378         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
379         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
380         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
381         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
382         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
383         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
384         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
385         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
386         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
387         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
388         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
389 
390         if (k > 4 * sizeof(int8_t)) {
391           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
392           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
393           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
394           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
395 
396           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
397           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
398           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
399           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
400           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
401           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
402           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
403           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
404           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
405           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
406           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
407           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
408           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
409           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
410           const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
411           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
412           const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
413           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
414         }
415       }
416     }
417 
418     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
419     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
420     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
421 
422     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
423     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
424     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
425     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
426     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
427     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
428     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
429     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
430 
431     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
432     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
433     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
434     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
435     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
436     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
437     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
438     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
439 
440     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
441     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
442     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
443     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
444     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
445     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
446     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
447     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
448 
449     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
450 #if XNN_ARCH_ARM64
451     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
452     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
453     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
454     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
455 
456     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
457     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
458     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
459     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
460 
461     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
462     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
463 #else
464     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
465     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
466     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
467     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
468 
469     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
470     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
471     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
472     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
473 
474     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
475     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
476 #endif
477 
478     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
479     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
480     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
481 
482     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
483     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
484     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
485 
486     if (nc >= 16) {
487       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
488       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
489 
490       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
491       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
492 
493       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
494       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
495 
496       nc -= 16;
497     } else {
498       // Final case where not all of the 16 columns fit in the destination.
499       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
500       if (nc & 8) {
501         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
502         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
503         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
504       }
505       if (nc & 4) {
506         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
507         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
508         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
509       }
510       if (nc & 2) {
511         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
512         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
513         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
514       }
515       if (nc & 1) {
516         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
517         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
518       }
519 
520       nc = 0;
521     }
522   } while (nc != 0);
523 }
524