xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.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_dup(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_dup(
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 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
71       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
72       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
73       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
74 
75       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 
92       const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
93       const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
94       const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
95       const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
96 
97       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
98       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
99       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
101       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
102       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
103       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
104       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
105       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
106       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
108       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
109       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
110       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
111       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
112       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
113       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
115       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
116       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
117       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
118       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
119       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
120       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
122       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
123       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
124       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
125       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
126       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
127       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
129       vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
130       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
131       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
132       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
133       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
134       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
136       vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
137       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
138       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
139       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
140       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
141       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
143       vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
144       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
145       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
146       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
147       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
148       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
150       vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
151       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
152       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
153       const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
154       const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
155       const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
156       const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
157 
158       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
159       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
160       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
162       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
163       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
164       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
165       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
166       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
167       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
169       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
170       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
171       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
172       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
173       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
174       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
176       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
177       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
178       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
179       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
180       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
181       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
183       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
184       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
185       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
186       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
187       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
188       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
190       vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
191       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
192       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
193       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
194       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
195       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
196       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
197       vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
198       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
199       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
200       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
201       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
202       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
203       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
204       vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
205       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
206       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
207       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
208       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
209       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
210       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
211       vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
212       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
213       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
214 
215       k -= 16 * sizeof(int8_t);
216     }
217 
218     if (k >= 8 * sizeof(int8_t)) {
219       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
220       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
221 
222       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238 
239       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
240       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
241 
242       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
243       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
244       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
245       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
246       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
247       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
248       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
249       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
250       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
251       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
252       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
253       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
254       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
255       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
256       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
257       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
258       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
259       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
260       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
261       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
262       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
263       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
264       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
265       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
266       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
267       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
268       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
269       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
270       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
271       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
272       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
273       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
274       const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
275       const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
276 
277       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
278       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
279       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
280       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
281       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
282       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
283       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
284       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
285       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
286       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
287       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
288       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
289       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
290       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
291       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
292       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
293       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
294       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
295       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
296       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
297       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
298       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
299       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
300       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
301       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
302       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
303       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
304       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
305       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
306       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
307       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
308       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
309 
310       k -= 8 * sizeof(int8_t);
311     }
312 
313     if XNN_UNLIKELY(k != 0) {
314       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
315       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
316 
317       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
318       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
319       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
322       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
323       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
325 
326       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
327       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
328       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
329       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
330       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
331       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
332       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
333       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
334       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
335       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
336       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
337       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
338       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
339       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
340       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
341       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
342       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
343       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
344       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
345       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
346       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
347       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
348       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
349       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
350       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
351       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
352       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
353       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
354       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
355       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
356       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
357       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
358       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
359       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
360     }
361 
362 #if XNN_ARCH_ARM64
363     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
364     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
365     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
366     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
367     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
368     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
369     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
370     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
371 #else
372     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
373     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
374     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
375     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
376     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
377     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
378     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
379     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
380     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
381     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
382     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
383     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
384     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
385     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
386     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
387     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
388     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
389     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
390     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
391     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
392     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
393     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
394     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
395     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
396 #endif
397 
398     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
399     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
400     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
401 
402     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
403     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
404     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
405     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
406     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
407     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
408     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
409     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
410 
411     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
412     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
413     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
414     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
415     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
416     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
417     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
418     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
419 
420     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
421     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
422     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
423     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
424     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
425     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
426     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
427     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
428 
429     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
430 #if XNN_ARCH_ARM64
431     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
432     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
433     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
434     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
435 
436     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
437     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
438     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
439     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
440 
441     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
442     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
443 #else
444     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
445     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
446     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
447     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
448 
449     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
450     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
451     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
452     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
453 
454     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
455     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
456 #endif
457 
458     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
459     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
460     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
461 
462     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
463     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
464     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
465 
466     if (nc >= 16) {
467       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
468       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
469 
470       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
471       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
472 
473       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
474       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
475 
476       nc -= 16;
477     } else {
478       // Final case where not all of the 16 columns fit in the destination.
479       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
480       if (nc & 8) {
481         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
482         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
483         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
484       }
485       if (nc & 4) {
486         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
487         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
488         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
489       }
490       if (nc & 2) {
491         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
492         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
493         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
494       }
495       if (nc & 1) {
496         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
497         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
498       }
499 
500       nc = 0;
501     }
502   } while (nc != 0);
503 }
504