xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-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 
17 
xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 2);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr != 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48 
49   do {
50     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58     int32x4_t vacc1x01 = vacc0x01;
59     int32x4_t vacc1x23 = vacc0x23;
60     int32x4_t vacc1x45 = vacc0x45;
61     int32x4_t vacc1x67 = vacc0x67;
62     int32x4_t vacc1x89 = vacc0x89;
63     int32x4_t vacc1xAB = vacc0xAB;
64     int32x4_t vacc1xCD = vacc0xCD;
65     int32x4_t vacc1xEF = vacc0xEF;
66 
67     size_t k = kc;
68 
69     while (k >= 16 * sizeof(int8_t)) {
70       const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
71       const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
72       const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
73       const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
74       const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
75       const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
76       const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
77       const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
78 
79       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 
96       const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
97       const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
98       const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
99       const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
100 
101       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
102       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
103       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
105       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
106       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
107       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
108       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
109       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
110       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
112       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
113       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
114       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
115       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
116       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
117       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
119       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
120       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
121       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
122       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
123       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
124       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
126       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
127       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
128       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
129       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
130       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
131       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
133       vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
134       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
135       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
136       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
137       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
138       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
140       vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
141       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
142       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
143       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
144       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
145       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
147       vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
148       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
149       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
150       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
151       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
152       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
154       vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
155       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
156       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
157       const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
158       const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
159       const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
160       const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
161 
162       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
163       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
164       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
166       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
167       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
168       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
169       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
170       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
171       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
173       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
174       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
175       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
176       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
177       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
178       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
180       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
181       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
182       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
183       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
184       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
185       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
186       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
187       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
188       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
189       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
190       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
191       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
192       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
194       vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
195       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
196       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
197       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
198       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
199       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
201       vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
202       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
203       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
204       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
205       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
206       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
208       vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
209       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
210       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
211       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
212       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
213       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
215       vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
216       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
217       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
218 
219       k -= 16 * sizeof(int8_t);
220     }
221 
222     if (k >= 8 * sizeof(int8_t)) {
223       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
224       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
225       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
226       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
227 
228       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244 
245       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
246       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
247 
248       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
249       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
250       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
251       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
252       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
253       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
254       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
255       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
256       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
257       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
258       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
259       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
260       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
261       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
262       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
263       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
264       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
265       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
266       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
267       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
268       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
269       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
270       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
271       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
272       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
273       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
274       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
275       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
276       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
277       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
278       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
279       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
280       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
281       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
282 
283       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
284       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
285       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
286       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
287       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
288       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
289       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
290       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
291       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
292       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
293       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
294       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
295       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
296       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
297       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
298       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
299       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
300       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
301       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
302       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
303       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
304       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
305       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
306       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
307       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
308       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
309       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
310       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
311       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
312       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
313       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
314       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
315 
316       k -= 8 * sizeof(int8_t);
317     }
318 
319     if XNN_UNLIKELY(k != 0) {
320       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
321       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
322 
323       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
325       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
326       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
328       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
330       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
331 
332       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
333       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
334       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
335       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
336       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
337       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
338       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
339       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
340       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
341       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
342       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
343       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
344       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
345       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
346       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
347       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
348       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
349       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
350       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
351       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
352       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
353       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
354       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
355       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
356       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
357       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
358       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
359       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
360       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
361       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
362       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
363       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
364       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
365       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
366     }
367 
368 #if XNN_ARCH_ARM64
369     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
370     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
371     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
372     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
373     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
374     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
375     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
376     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
377 #else
378     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
379     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
380     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
381     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
382     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
383     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
384     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
385     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
386     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
387     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
388     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
389     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
390     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
391     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
392     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
393     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
394     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
395     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
396     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
397     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
398     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
399     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
400     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
401     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
402 #endif
403 
404     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
405     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
406     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
407 
408     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
409     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
410     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
411     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
412     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
413     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
414     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
415     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
416 
417     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
418     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
419     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
420     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
421     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
422     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
423     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
424     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
425 
426     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
427     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
428     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
429     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
430     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
431     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
432     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
433     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
434 
435     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
436 #if XNN_ARCH_ARM64
437     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
438     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
439     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
440     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
441 
442     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
443     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
444     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
445     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
446 
447     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
448     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
449 #else
450     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
451     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
452     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
453     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
454 
455     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
456     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
457     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
458     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
459 
460     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
461     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
462 #endif
463 
464     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
465     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
466     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
467 
468     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
469     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
470     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
471 
472     if (nc >= 16) {
473       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
474       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
475 
476       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
477       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
478 
479       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
480       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
481 
482       nc -= 16;
483     } else {
484       // Final case where not all of the 16 columns fit in the destination.
485       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
486       if (nc & 8) {
487         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
488         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
489         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
490       }
491       if (nc & 4) {
492         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
493         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
494         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
495       }
496       if (nc & 2) {
497         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
498         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
499         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
500       }
501       if (nc & 1) {
502         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
503         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
504       }
505 
506       nc = 0;
507     }
508   } while (nc != 0);
509 }
510