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_4x8c2__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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r(
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 <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (4 * 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 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 do {
58 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60 int32x4_t vacc1x0123 = vacc0x0123;
61 int32x4_t vacc1x4567 = vacc0x4567;
62 int32x4_t vacc2x0123 = vacc0x0123;
63 int32x4_t vacc2x4567 = vacc0x4567;
64 int32x4_t vacc3x0123 = vacc0x0123;
65 int32x4_t vacc3x4567 = vacc0x4567;
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const int8_t* restrict a1 = a[1];
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76 }
77 const int8_t* restrict a2 = a[2];
78 if XNN_UNPREDICTABLE(a2 != zero) {
79 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
80 }
81 const int8_t* restrict a3 = a[3];
82 if XNN_UNPREDICTABLE(a3 != zero) {
83 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
84 }
85 a += 4;
86
87 size_t k = kc;
88
89 while (k >= 16 * sizeof(int8_t)) {
90 const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
91 const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
92 const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
93 const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
94 const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
95 const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
96 const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
97 const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
98 const int16x4x2_t va20x0 = vld2_dup_s16((const void*)a2);
99 const int16x4x2_t va21x0 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
100 const int16x4x2_t va20x1 = vld2_dup_s16((const void*)a2);
101 const int16x4x2_t va21x1 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
102 const int16x4x2_t va30x0 = vld2_dup_s16((const void*)a3);
103 const int16x4x2_t va31x0 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
104 const int16x4x2_t va30x1 = vld2_dup_s16((const void*)a3);
105 const int16x4x2_t va31x1 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
106
107 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115
116 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
117 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
118 const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
119 const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
120 const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0.val[0]);
121 const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1.val[0]);
122 const int8x8_t va3c0x0 = vreinterpret_s8_s16(va30x0.val[0]);
123 const int8x8_t va3c0x1 = vreinterpret_s8_s16(va30x1.val[0]);
124
125 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
126 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
127 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
128 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3c0x0);
129 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
131 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
132 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
133 vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3c0x1);
134 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
135 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
136 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
137 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
138 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
139 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
140 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
141 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0);
142 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
143 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
144 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
145 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
146 vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3c0x1);
147 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
148 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
149 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
150 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
151 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
152 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
153 const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
154 const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
155 const int8x8_t va2c1x0 = vreinterpret_s8_s16(va20x0.val[1]);
156 const int8x8_t va2c1x1 = vreinterpret_s8_s16(va20x1.val[1]);
157 const int8x8_t va3c1x0 = vreinterpret_s8_s16(va30x0.val[1]);
158 const int8x8_t va3c1x1 = vreinterpret_s8_s16(va30x1.val[1]);
159
160 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
161 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
162 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
163 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0);
164 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
166 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
167 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
168 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1);
169 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
170 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
171 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
172 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
173 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
174 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
175 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
176 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3c1x0);
177 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
179 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
180 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
181 vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3c1x1);
182 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
183 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
184 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
185 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
186 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
187 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
188 const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
189 const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
190 const int8x8_t va2c2x0 = vreinterpret_s8_s16(va21x0.val[0]);
191 const int8x8_t va2c2x1 = vreinterpret_s8_s16(va21x1.val[0]);
192 const int8x8_t va3c2x0 = vreinterpret_s8_s16(va31x0.val[0]);
193 const int8x8_t va3c2x1 = vreinterpret_s8_s16(va31x1.val[0]);
194
195 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
196 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
197 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
198 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3c2x0);
199 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
201 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
202 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
203 vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3c2x1);
204 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
205 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
206 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
207 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
208 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
209 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
210 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
211 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0);
212 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
214 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
215 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
216 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1);
217 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
218 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
219 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
220 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
221 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
222 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
223 const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
224 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
225 const int8x8_t va2c3x0 = vreinterpret_s8_s16(va21x0.val[1]);
226 const int8x8_t va2c3x1 = vreinterpret_s8_s16(va21x1.val[1]);
227 const int8x8_t va3c3x0 = vreinterpret_s8_s16(va31x0.val[1]);
228 const int8x8_t va3c3x1 = vreinterpret_s8_s16(va31x1.val[1]);
229
230 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
231 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
232 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
233 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3c3x0);
234 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
236 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
237 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
238 vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3c3x1);
239 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
240 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
241 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
242 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
243 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
244 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
245 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
246 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0);
247 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
249 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
250 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
251 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1);
252 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
253 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
254 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
255 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
256
257 k -= 16 * sizeof(int8_t);
258 }
259
260 if (k >= 8 * sizeof(int8_t)) {
261 const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
262 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
263 const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
264 const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
265 const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
266 const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
267 const int16x4x2_t va30 = vld2_dup_s16((const void*)a3);
268 const int16x4x2_t va31 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
269
270 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
271 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
272 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
273 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
274 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
275 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
278
279 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
280 const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
281 const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
282 const int8x8_t va3c0 = vreinterpret_s8_s16(va30.val[0]);
283
284 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
285 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
286 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
287 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
288 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
289 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
290 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
291 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
292 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
293 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
294 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
295 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
296 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
297 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
298 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
299 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
300 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
301 const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
302 const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
303 const int8x8_t va3c1 = vreinterpret_s8_s16(va30.val[1]);
304
305 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
306 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
307 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
308 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
309 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
310 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
311 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
312 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
313 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
314 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
315 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
316 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
317 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
318 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
319 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
320 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
321 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
322 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
323 const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
324 const int8x8_t va3c2 = vreinterpret_s8_s16(va31.val[0]);
325
326 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
327 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
328 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
329 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
330 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
331 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
332 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
333 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
334 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
335 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
336 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
337 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
338 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
339 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
340 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
341 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
342 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
343 const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
344 const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
345 const int8x8_t va3c3 = vreinterpret_s8_s16(va31.val[1]);
346
347 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
348 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
349 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
350 const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
351 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
352 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
353 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
354 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
355 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
356 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
357 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
358 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
359 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
360 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
361 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
362 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
363
364 k -= 8 * sizeof(int8_t);
365 }
366
367 if XNN_UNLIKELY(k != 0) {
368 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
369 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
370 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
371 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
372
373 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
374 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
375
376 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
377 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
378 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
379 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
380 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
381 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
382 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
383 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
384 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
385 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
386 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
387 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
388 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
389 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
390 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
391 const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
392 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
393 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
394 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
395 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
396
397 if (k > 2 * sizeof(int8_t)) {
398 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
399 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
400
401 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
402 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
403 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
404 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
405 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
406 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
407 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
408 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
409 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
410 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
411 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
412 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
413 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
414 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
415 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
416 const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
417 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
418 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
419 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
420 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
421
422 if (k > 4 * sizeof(int8_t)) {
423 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
424 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
425
426 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
427 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
428 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
429 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
430 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
431 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
432 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
433 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
434 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
435 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
436 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
437 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
438 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
439 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
440 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
441 const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
442 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
443 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
444 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
445 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
446 }
447 }
448 }
449 p -= 4 * sizeof(void*);
450 } while (p != 0);
451
452 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
453 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
454 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
455
456 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
457 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
458 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
459 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
460 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
461 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
462 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
463 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
464
465 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
466 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
467 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
468 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
469 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
470 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
471 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
472 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
473
474 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
475 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
476 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
477 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
478 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
479 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
480 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
481 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
482
483 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
484 #if XNN_ARCH_ARM64
485 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
486 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
487 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
488 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
489
490 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
491 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
492 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
493 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
494
495 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
496 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
497 #else
498 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
499 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
500 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
501 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
502
503 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
504 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
505 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
506 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
507
508 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
509 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
510 #endif
511
512 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
513 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
514 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
515
516 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
517 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
518 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
519
520 if (nc >= 8) {
521 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
522 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
523 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
524 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
525
526 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
527 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
528 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
529 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
530
531 a = (const int8_t**restrict) ((uintptr_t) a - ks);
532
533 nc -= 8;
534 } else {
535 if (nc & 4) {
536 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
537 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
538 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
539 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
540 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
541 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
542 }
543 if (nc & 2) {
544 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
545 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
546 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
547 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
548 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
549 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
550 }
551 if (nc & 1) {
552 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
553 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
554 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
555 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
556 }
557
558 nc = 0;
559 }
560 } while (nc != 0);
561 }
562