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_1x16c4__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_1x16c4__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 <= 1);
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
43 do {
44 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
45 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52
53 size_t k = kc;
54
55 while (k >= 16 * sizeof(int8_t)) {
56 const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
57 const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
58 const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
59 const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
60
61 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77
78 const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
79 const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
80
81 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
82 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
84 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
85 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
86 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
88 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
89 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
90 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
92 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
93 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
94 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
96 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
97 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
98 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
100 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
101 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
102 const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
104 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
105 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
106 const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107 vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
108 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
109 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
110 const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
112 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
113 const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
114 const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
115
116 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
117 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
119 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
120 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
121 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
123 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
124 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
125 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
127 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
128 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
129 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
131 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
132 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
133 const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
135 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
136 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
137 const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
139 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
140 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
141 const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142 vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
143 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
144 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
145 const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146 vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
147 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
148
149 k -= 16 * sizeof(int8_t);
150 }
151
152 if (k >= 8 * sizeof(int8_t)) {
153 const int32x2_t va00 = vld1_dup_s32((const void*)a0);
154 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
155
156 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172
173 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
174
175 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
176 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
177 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
178 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
179 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
180 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
181 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
182 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
183 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
184 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
185 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
186 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
187 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
188 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
189 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
190 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
191 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
192
193 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
194 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
195 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
196 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
197 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
198 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
199 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
200 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
201 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
202 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
203 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
204 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
205 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
206 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
207 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
208 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
209
210 k -= 8 * sizeof(int8_t);
211 }
212
213 if XNN_UNLIKELY(k != 0) {
214 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
215
216 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224
225 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
226 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
227 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
228 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
229 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
230 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
231 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
232 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
233 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
234 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
235 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
236 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
237 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
238 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
239 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
240 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
241 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
242 }
243
244 #if XNN_ARCH_ARM64
245 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
246 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
247 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
248 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
249 #else
250 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
251 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
252 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
253 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
254 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
255 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
256 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
257 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
258 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
259 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
260 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
261 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
262 #endif
263
264 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
265 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
266 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
267
268 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
269 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
270 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
271 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
272
273 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
274 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
275 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
276 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
277
278 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
279 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
280 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
281 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
282
283 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
284 #if XNN_ARCH_ARM64
285 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
286 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
287
288 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
289 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
290
291 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
292 #else
293 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
294 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
295
296 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
297 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
298
299 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
300 #endif
301
302 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
303 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
304
305 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
306 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
307
308 if (nc >= 16) {
309 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
310
311 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
312
313 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
314
315 nc -= 16;
316 } else {
317 // Final case where not all of the 16 columns fit in the destination.
318 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
319 if (nc & 8) {
320 vst1_s8(c0, vout0x01234567); c0 += 8;
321 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
322 }
323 if (nc & 4) {
324 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
325 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
326 }
327 if (nc & 2) {
328 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
329 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
330 }
331 if (nc & 1) {
332 vst1_lane_s8(c0, vout0x01234567, 0);
333 }
334
335 nc = 0;
336 }
337 } while (nc != 0);
338 }
339