1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/neon-mull-addw-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/common.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_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)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup(
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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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 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 while (k >= 8 * sizeof(int8_t)) {
89 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
90 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
91 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
92 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
93
94 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95
96 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
97 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
98 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
99 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
100 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
101 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
102 const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
103 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
104 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
105 const int16x8_t vprod3x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va3, 0));
106 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c0));
107 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));
108 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109
110 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
111 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
112 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
113 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
114 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
115 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
116 const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
117 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
118 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
119 const int16x8_t vprod3x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va3, 1));
120 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c1));
121 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));
122 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123
124 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
125 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
126 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
127 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
128 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
129 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
130 const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
131 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
132 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
133 const int16x8_t vprod3x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va3, 2));
134 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c2));
135 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));
136 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137
138 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
139 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
140 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
141 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
142 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
143 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
144 const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
145 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
146 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
147 const int16x8_t vprod3x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va3, 3));
148 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c3));
149 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));
150 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151
152 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
153 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
154 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
155 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
156 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
157 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
158 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
159 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
160 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
161 const int16x8_t vprod3x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va3, 4));
162 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c4));
163 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));
164 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165
166 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
167 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
168 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
169 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
170 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
171 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
172 const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
173 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
174 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
175 const int16x8_t vprod3x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va3, 5));
176 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c5));
177 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));
178 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179
180 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
181 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
182 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
183 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
184 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
185 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
186 const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
187 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
188 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
189 const int16x8_t vprod3x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va3, 6));
190 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c6));
191 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));
192 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193
194 const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
195 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
196 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
197 const int16x8_t vprod1x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va1, 7));
198 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));
199 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c7));
200 const int16x8_t vprod2x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va2, 7));
201 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c7));
202 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c7));
203 const int16x8_t vprod3x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va3, 7));
204 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c7));
205 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c7));
206
207 k -= 8 * sizeof(int8_t);
208 }
209 if XNN_UNLIKELY(k != 0) {
210 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
211 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
212 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
213 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
214
215 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216
217 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
218 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
219 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
220 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
221 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
222 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
223 const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
224 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
225 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
226 const int16x8_t vprod3x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va3, 0));
227 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c0));
228 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));
229
230 if (k >= 2 * sizeof(int8_t)) {
231 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232
233 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
234 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
235 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
236 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
237 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
238 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
239 const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
240 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
241 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
242 const int16x8_t vprod3x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va3, 1));
243 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c1));
244 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));
245
246 if (k > 2 * sizeof(int8_t)) {
247 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248
249 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
250 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
251 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
252 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
253 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
254 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
255 const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
256 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
257 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
258 const int16x8_t vprod3x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va3, 2));
259 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c2));
260 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));
261
262 if (k >= 4 * sizeof(int8_t)) {
263 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
264
265 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
266 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
267 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
268 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
269 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
270 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
271 const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
272 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
273 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
274 const int16x8_t vprod3x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va3, 3));
275 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c3));
276 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));
277
278 if (k > 4 * sizeof(int8_t)) {
279 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
280
281 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
282 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
283 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
284 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
285 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
286 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
287 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
288 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
289 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
290 const int16x8_t vprod3x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va3, 4));
291 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c4));
292 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));
293
294 if (k >= 6 * sizeof(int8_t)) {
295 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
296
297 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
298 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
299 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
300 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
301 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
302 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
303 const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
304 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
305 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
306 const int16x8_t vprod3x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va3, 5));
307 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c5));
308 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));
309
310 if (k > 6 * sizeof(int8_t)) {
311 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
312
313 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
314 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
315 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
316 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
317 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
318 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
319 const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
320 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
321 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
322 const int16x8_t vprod3x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va3, 6));
323 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c6));
324 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));
325 }
326 }
327 }
328 }
329 }
330 }
331 }
332 p -= 4 * sizeof(void*);
333 } while (p != 0);
334
335 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
336 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
337 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
338
339 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
340 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
341 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
342 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
343 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
344 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
345 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
346 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
347
348 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
349 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
350 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
351 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
352 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
353 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
354 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
355 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
356
357 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
358 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
359 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
360 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
361 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
362 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
363 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
364 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
365
366 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
367 #if XNN_ARCH_ARM64
368 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
369 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
370 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
371 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
372
373 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
374 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
375 #else
376 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
377 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
378 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
379 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
380
381 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
382 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
383 #endif
384 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
385 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
386
387 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
388 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
389
390 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
391 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
392
393 if (nc >= 8) {
394 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
395 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
396 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
397 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
398
399 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
400 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
401 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
402 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
403
404 a = (const int8_t**restrict) ((uintptr_t) a - ks);
405
406 nc -= 8;
407 } else {
408 if (nc & 4) {
409 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
410 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
411 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
412 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
413 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
414 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
415 }
416 if (nc & 2) {
417 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
418 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
419 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
420 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
421 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
422 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
423 }
424 if (nc & 1) {
425 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
426 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
427 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
428 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
429 }
430
431 nc = 0;
432 }
433 } while (nc != 0);
434 }
435