1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_1x16c2__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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47
48 size_t k = kc;
49
50 while (k >= 16 * sizeof(int8_t)) {
51 const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
52 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
53 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
54 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
55 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
56 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
57 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
58 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
59 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75
76 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
77 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
78
79 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
80 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
82 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
83 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
84 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
86 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
87 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
88 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
90 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
91 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
92 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
94 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
95 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
96 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
97
98 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
99 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
101 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
102 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
103 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
105 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
106 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
107 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
109 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
110 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
111 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
113 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
114 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
115 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
116
117 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
118 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
120 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
121 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
122 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
124 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
125 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
126 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
128 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
129 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
130 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
132 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
133 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
134 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
135
136 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
137 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
139 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
140 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
141 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
144 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
145 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
147 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
148 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
149 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
151 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
152
153 k -= 16 * sizeof(int8_t);
154 }
155
156 if (k >= 8 * sizeof(int8_t)) {
157 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
158 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
159 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
160 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
161
162 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178
179 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
180
181 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
182 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
183 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
184 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
185 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
186 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
187 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
188 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
189 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
190
191 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
192 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
193 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
194 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
195 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
196 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
197 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
198 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
199 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
200
201 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
202 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
203 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
204 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
205 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
206 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
207 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
208 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
209 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
210
211 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
212 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
213 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
214 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
215 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
216 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
217 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
218 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
219
220 k -= 8 * sizeof(int8_t);
221 }
222
223 if XNN_UNLIKELY(k != 0) {
224 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
225
226 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230
231 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
232 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
233 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
234 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
235 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
236 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
237 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
238 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
239 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
240
241 if (k > 2 * sizeof(int8_t)) {
242 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246
247 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
248 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
249 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
250 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
251 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
252 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
253 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
254 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
255 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
256
257 if (k > 4 * sizeof(int8_t)) {
258 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
259 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
260 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
261 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
262
263 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
264 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
265 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
266 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
267 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
268 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
269 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
270 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
271 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
272 }
273 }
274 }
275
276 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
277 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
278 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
279
280 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
281 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
282 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
283 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
284
285 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
286 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
287 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
288 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
289
290 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
291 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
292 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
293 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
294
295 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
296 #if XNN_ARCH_ARM64
297 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
298 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
299
300 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
301 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
302
303 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
304 #else
305 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
306 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
307
308 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
309 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
310
311 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
312 #endif
313
314 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
315 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
316
317 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
318 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
319
320 if (nc >= 16) {
321 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
322
323 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
324
325 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
326
327 nc -= 16;
328 } else {
329 // Final case where not all of the 16 columns fit in the destination.
330 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
331 if (nc & 8) {
332 vst1_s8(c0, vout0x01234567); c0 += 8;
333 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
334 }
335 if (nc & 4) {
336 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
337 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
338 }
339 if (nc & 2) {
340 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
341 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
342 }
343 if (nc & 1) {
344 vst1_lane_s8(c0, vout0x01234567, 0);
345 }
346
347 nc = 0;
348 }
349 } while (nc != 0);
350 }
351