1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 do {
47 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 a += 1;
63
64 size_t k = kc;
65
66 while (k >= 16 * sizeof(int8_t)) {
67 const int32x2x2_t va0x0 = vld2_dup_s32((const void*)a0); a0 += 8;
68 const int32x2x2_t va0x1 = vld2_dup_s32((const void*)a0); a0 += 8;
69
70 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86
87 const int8x8_t va0c0x0 = vreinterpret_s8_s32(va0x0.val[0]);
88 const int8x8_t va0c0x1 = vreinterpret_s8_s32(va0x1.val[0]);
89
90 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
91 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
93 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
94 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
95 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
97 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
98 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
99 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
101 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
102 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
103 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
105 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
106 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
107 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
109 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
110 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
111 const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
113 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
114 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
115 const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
117 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
118 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
119 const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
121 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
122 const int8x8_t va0c1x0 = vreinterpret_s8_s32(va0x0.val[1]);
123 const int8x8_t va0c1x1 = vreinterpret_s8_s32(va0x1.val[1]);
124
125 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
126 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
128 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
129 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
130 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
132 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
133 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
134 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
136 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
137 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
138 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
140 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
141 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
142 const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
143 vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
144 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
145 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
146 const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
148 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
149 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
150 const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
152 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
153 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
154 const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155 vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
156 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
157
158 k -= 16 * sizeof(int8_t);
159 }
160
161 if (k >= 8 * sizeof(int8_t)) {
162 const int32x2x2_t va0 = vld2_dup_s32((const void*)a0); a0 += 8;
163
164 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180
181 const int8x8_t va0c0 = vreinterpret_s8_s32(va0.val[0]);
182
183 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
184 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
185 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
186 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
187 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
188 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
189 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
190 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
191 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
192 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
193 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
194 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
195 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
196 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
197 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
198 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
199 const int8x8_t va0c1 = vreinterpret_s8_s32(va0.val[1]);
200
201 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
202 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
203 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
204 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
205 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
206 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
207 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
208 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
209 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
210 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
211 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
212 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
213 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
214 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
215 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
216 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
217
218 k -= 8 * sizeof(int8_t);
219 }
220
221 if XNN_UNLIKELY(k != 0) {
222 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
223
224 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232
233 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
234 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
235 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
236 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
237 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
238 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
239 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
240 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
241 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
242 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
243 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
244 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
245 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
246 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
247 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
248 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
249 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
250 }
251 p -= 1 * sizeof(void*);
252 } while (p != 0);
253
254 #if XNN_ARCH_ARM64
255 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
256 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
257 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
258 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
259 #else
260 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
261 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
262 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
263 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
264 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
265 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
266 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
267 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
268 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
269 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
270 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
271 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
272 #endif
273
274 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
275 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
276 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
277
278 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
279 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
280 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
281 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
282
283 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
284 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
285 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
286 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
287
288 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
289 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
290 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
291 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
292
293 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
294 #if XNN_ARCH_ARM64
295 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
296 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
297
298 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
299 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
300
301 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
302 #else
303 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
304 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
305
306 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
307 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
308
309 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
310 #endif
311
312 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
313 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
314
315 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
316 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
317
318 if (nc >= 16) {
319 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
320
321 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
322
323 a = (const int8_t**restrict) ((uintptr_t) a - ks);
324
325 nc -= 16;
326 } else {
327 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
328 if (nc & 8) {
329 vst1_s8(c0, vout0x01234567); c0 += 8;
330 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
331 }
332 if (nc & 4) {
333 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
334 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
335 }
336 if (nc & 2) {
337 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
338 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
339 }
340 if (nc & 1) {
341 vst1_lane_s8(c0, vout0x01234567, 0);
342 }
343
344 nc = 0;
345 }
346 } while (nc != 0);
347 }
348