xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x8c4-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_4x8c4__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_4x8c4__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 <= 4);
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   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56   if XNN_UNPREDICTABLE(mr != 4) {
57     a3 = a2;
58     c3 = c2;
59   }
60 
61   do {
62     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
64     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
65     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
66     int32x4_t vacc1x01 = vacc0x01;
67     int32x4_t vacc1x23 = vacc0x23;
68     int32x4_t vacc1x45 = vacc0x45;
69     int32x4_t vacc1x67 = vacc0x67;
70     int32x4_t vacc2x01 = vacc0x01;
71     int32x4_t vacc2x23 = vacc0x23;
72     int32x4_t vacc2x45 = vacc0x45;
73     int32x4_t vacc2x67 = vacc0x67;
74     int32x4_t vacc3x01 = vacc0x01;
75     int32x4_t vacc3x23 = vacc0x23;
76     int32x4_t vacc3x45 = vacc0x45;
77     int32x4_t vacc3x67 = vacc0x67;
78 
79     size_t k = kc;
80 
81     while (k >= 16 * sizeof(int8_t)) {
82       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
83       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
84       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
85       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
86       const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
87       const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
88       const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
89       const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
90 
91       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 
100       const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
101       const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
102       const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
103       const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
104       const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0));
105       const int8x8_t va2c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x1), 0));
106       const int8x8_t va3c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x0), 0));
107       const int8x8_t va3c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x1), 0));
108 
109       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
110       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
111       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
112       int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0);
113       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
115       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
116       vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
117       vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1);
118       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
119       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
120       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
121       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
122       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
123       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
124       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
125       int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0);
126       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
128       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
129       vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
130       vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1);
131       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
132       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
133       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
134       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
135       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
136       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
137       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
138       int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0);
139       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
141       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
142       vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
143       vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1);
144       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
145       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
146       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
147       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
148       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
149       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
150       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
151       int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0);
152       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
154       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
155       vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
156       vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1);
157       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
158       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
159       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
160       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
161       const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
162       const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
163       const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
164       const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
165       const int8x8_t va2c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 1));
166       const int8x8_t va2c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x1), 1));
167       const int8x8_t va3c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x0), 1));
168       const int8x8_t va3c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x1), 1));
169 
170       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
171       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
172       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
173       int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0);
174       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
176       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
177       vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
178       vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1);
179       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
180       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
181       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
182       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
183       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
184       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
185       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
186       int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0);
187       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
189       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
190       vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
191       vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1);
192       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
193       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
194       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
195       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
196       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
197       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
198       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
199       int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0);
200       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
201       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
202       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
203       vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
204       vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1);
205       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
206       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
207       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
208       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
209       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
210       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
211       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
212       int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0);
213       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
215       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
216       vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
217       vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1);
218       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
219       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
220       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
221       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
222 
223       k -= 16 * sizeof(int8_t);
224     }
225 
226     if (k >= 8 * sizeof(int8_t)) {
227       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
228       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
229       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
230       const int8x8_t va3 = vld1_s8(a3); a3 += 8;
231 
232       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235       const int8x8_t vb67c0 = 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 
241       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
242       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
243       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
244       const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
245 
246       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
247       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
248       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
249       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
250       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
251       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
252       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
253       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
254       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
255       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
256       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
257       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
258       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
259       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
260       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
261       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
262       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
263       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
264       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
265       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
266       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
267       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
268       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
269       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
270       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
271       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
272       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
273       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
274       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
275       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
276       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
277       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
278       const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
279       const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
280       const int8x8_t va2c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 1));
281       const int8x8_t va3c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 1));
282 
283       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
284       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
285       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
286       const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
287       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
288       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
289       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
290       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
291       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
292       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
293       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
294       const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
295       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
296       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
297       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
298       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
299       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
300       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
301       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
302       const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
303       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
304       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
305       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
306       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
307       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
308       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
309       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
310       const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
311       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
312       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
313       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
314       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
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       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
323       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
324 
325       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
326       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
328       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329 
330       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
331       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
332       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
333       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
334       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
335       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
336       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
337       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
338       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
339       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
340       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
341       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
342       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
343       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
344       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
345       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
346       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
347       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
348       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
349       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
350       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
351       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
352       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
353       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
354       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
355       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
356       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
357       const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
358       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
359       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
360       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
361       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
362       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
363       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
364       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
365       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
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 vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
372     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
373     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
374     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
375     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
376     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
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 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 vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
391     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
392     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
393     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
394     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
395     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
396     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
397     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
398     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
399     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
400     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
401     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
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     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
411     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
412     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
413     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
414     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
415     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
416 
417     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
418     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
419     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
420     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
421     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
422     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
423     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
424     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
425 
426     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
427     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
428     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
429     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
430     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
431     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
432     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
433     vacc3x4567 = vrshlq_s32(vacc3x4567, 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 vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
439     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
440     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
441 
442     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
443     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
444     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
445     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
446 
447     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
448     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
449 #else
450     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
451     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
452     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
453     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
454 
455     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
456     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
457     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
458     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
459 
460     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
461     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
462 #endif
463 
464     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
465     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
466     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
467 
468     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
469     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
470     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
471 
472     if (nc >= 8) {
473       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
474       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
475       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
476       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
477 
478       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
479       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
480       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
481       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
482 
483       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
484       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
485       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
486       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
487 
488       nc -= 8;
489     } else {
490       // Final case where not all of the 8 columns fit in the destination.
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         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
495         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
496         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
497         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
498       }
499       if (nc & 2) {
500         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
501         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
502         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
503         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
504         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
505         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
506       }
507       if (nc & 1) {
508         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
509         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
510         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
511         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
512       }
513 
514       nc = 0;
515     }
516   } while (nc != 0);
517 }
518