1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-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
xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (1 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
43 int8_t* c0 = c;
44
45 do {
46 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
49 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50
51 size_t p = ks;
52 do {
53 const int8_t* restrict a0 = a[0];
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60
61 while (k >= 16 * sizeof(int8_t)) {
62 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
63 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
64
65 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81
82 const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0));
83 const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0));
84
85 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
86 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
88 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
89 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
90 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
92 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
93 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
94 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
96 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
97 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
98 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
100 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
101 const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1));
102 const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1));
103
104 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
105 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
107 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
108 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
109 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
111 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
112 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
113 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
115 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
116 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
117 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
119 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
120 const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2));
121 const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2));
122
123 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
124 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
126 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
127 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
128 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
130 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
131 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
132 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
134 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
135 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
136 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
138 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
139 const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3));
140 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3));
141
142 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
143 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
145 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
146 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
147 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
149 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
150 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
151 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
153 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
154 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
155 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
157 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
158
159 k -= 16 * sizeof(int8_t);
160 }
161
162 if (k >= 8 * sizeof(int8_t)) {
163 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
164
165 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181
182 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
183
184 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
185 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
186 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
187 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
188 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
189 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
190 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
191 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
192 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
193
194 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
195 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
196 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
197 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
198 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
199 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
200 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
201 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
202 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
203
204 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
205 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
206 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
207 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
208 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
209 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
210 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
211 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
212 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
213
214 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
215 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
216 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
217 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
218 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
219 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
220 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
221 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
222
223 k -= 8 * sizeof(int8_t);
224 }
225
226 if XNN_UNLIKELY(k != 0) {
227 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
228
229 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233
234 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
235 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
236 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
237 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
238 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
239 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
240 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
241 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
242 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
243
244 if (k > 2 * sizeof(int8_t)) {
245 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249
250 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
251 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
252 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
253 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
254 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
255 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
256 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
257 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
258 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
259
260 if (k > 4 * sizeof(int8_t)) {
261 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
262 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
263 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
264 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
265
266 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
267 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
268 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
269 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
270 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
271 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
272 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
273 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
274 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
275 }
276 }
277 }
278 p -= 1 * sizeof(void*);
279 } while (p != 0);
280
281 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
282 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
283 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
284
285 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
286 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
287 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
288 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
289
290 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
291 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
292 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
293 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
294
295 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
296 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
297 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
298 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
299
300 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
301 #if XNN_ARCH_ARM64
302 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
303 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
304
305 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
306 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
307
308 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
309 #else
310 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
311 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
312
313 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
314 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
315
316 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
317 #endif
318
319 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
320 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
321
322 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
323 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
324
325 if (nc >= 16) {
326 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
327
328 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
329
330 a = (const int8_t**restrict) ((uintptr_t) a - ks);
331
332 nc -= 16;
333 } else {
334 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
335 if (nc & 8) {
336 vst1_s8(c0, vout0x01234567); c0 += 8;
337 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
338 }
339 if (nc & 4) {
340 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
341 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
342 }
343 if (nc & 2) {
344 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
345 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
346 }
347 if (nc & 1) {
348 vst1_lane_s8(c0, vout0x01234567, 0);
349 }
350
351 nc = 0;
352 }
353 } while (nc != 0);
354 }
355