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_ukernel_2x16__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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_2x16__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_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
54 int32x4_t vacc1x0123 = vacc0x0123;
55 int32x4_t vacc1x4567 = vacc0x4567;
56 int32x4_t vacc1x89AB = vacc0x89AB;
57 int32x4_t vacc1xCDEF = vacc0xCDEF;
58
59 size_t p = ks;
60 do {
61 const int8_t* restrict a0 = a[0];
62 if XNN_UNPREDICTABLE(a0 != zero) {
63 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
64 }
65 const int8_t* restrict a1 = a[1];
66 if XNN_UNPREDICTABLE(a1 != zero) {
67 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
68 }
69 a += 2;
70
71 size_t k = kc;
72 while (k >= 8 * sizeof(int8_t)) {
73 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
74 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
75
76 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77
78 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
79 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
80 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
81 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
82 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
83 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
84 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85
86 const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
87 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
88 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
89 const int16x8_t vprod1x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va1, 0));
90 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc0));
91 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));
92 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93
94 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
95 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
96 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
97 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
98 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
99 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
100 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101
102 const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
103 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
104 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
105 const int16x8_t vprod1x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va1, 1));
106 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc1));
107 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));
108 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109
110 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
111 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
112 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
113 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
114 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
115 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
116 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117
118 const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
119 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
120 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
121 const int16x8_t vprod1x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va1, 2));
122 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc2));
123 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));
124 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125
126 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
127 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
128 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
129 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
130 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
131 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
132 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133
134 const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
135 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
136 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
137 const int16x8_t vprod1x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va1, 3));
138 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc3));
139 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));
140 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141
142 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
143 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
144 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
145 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
146 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
147 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
148 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149
150 const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
151 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
152 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
153 const int16x8_t vprod1x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va1, 4));
154 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc4));
155 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));
156 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157
158 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
159 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
160 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
161 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
162 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
163 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
164 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165
166 const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
167 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
168 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
169 const int16x8_t vprod1x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va1, 5));
170 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc5));
171 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));
172 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173
174 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
175 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
176 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
177 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
178 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
179 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
180 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181
182 const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
183 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
184 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
185 const int16x8_t vprod1x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va1, 6));
186 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc6));
187 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));
188 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189
190 const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
191 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
192 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
193 const int16x8_t vprod1x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va1, 7));
194 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));
195 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c7));
196 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
197
198 const int16x8_t vprod0x89ABCDEFc7 = vmull_s8(vb89ABCDEFc7, vdup_lane_s8(va0, 7));
199 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));
200 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc7));
201 const int16x8_t vprod1x89ABCDEFc7 = vmull_s8(vb89ABCDEFc7, vdup_lane_s8(va1, 7));
202 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc7));
203 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc7));
204
205 k -= 8 * sizeof(int8_t);
206 }
207 if XNN_UNLIKELY(k != 0) {
208 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
209 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
210
211 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213
214 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
215 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
216 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
217 const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
218 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
219 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
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 vprod1x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va1, 0));
224 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc0));
225 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));
226
227 if (k >= 2 * sizeof(int8_t)) {
228 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230
231 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
232 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
233 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
234 const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
235 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
236 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
237 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
238 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
239 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
240 const int16x8_t vprod1x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va1, 1));
241 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc1));
242 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));
243
244 if (k > 2 * sizeof(int8_t)) {
245 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247
248 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
249 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
250 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
251 const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
252 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
253 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
254 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
255 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
256 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
257 const int16x8_t vprod1x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va1, 2));
258 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc2));
259 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));
260
261 if (k >= 4 * sizeof(int8_t)) {
262 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
263 const int8x8_t vb89ABCDEFc3 = 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 vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
269 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
270 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
271 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
272 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
273 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
274 const int16x8_t vprod1x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va1, 3));
275 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc3));
276 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));
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 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281
282 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
283 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
284 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
285 const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
286 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
287 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
288 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
289 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
290 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
291 const int16x8_t vprod1x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va1, 4));
292 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc4));
293 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));
294
295 if (k >= 6 * sizeof(int8_t)) {
296 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
297 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
298
299 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
300 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
301 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
302 const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
303 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
304 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
305 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
306 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
307 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
308 const int16x8_t vprod1x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va1, 5));
309 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc5));
310 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));
311
312 if (k > 6 * sizeof(int8_t)) {
313 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
314 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
315
316 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
317 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
318 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
319 const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
320 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
321 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
322 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
323 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
324 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
325 const int16x8_t vprod1x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va1, 6));
326 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc6));
327 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));
328 }
329 }
330 }
331 }
332 }
333 }
334 }
335 p -= 2 * sizeof(void*);
336 } while (p != 0);
337
338 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier);
339 vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
340 vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
341 vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
342 vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
343 vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
344 vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
345 vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
346 vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
347
348 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
349 const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
350 vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
351 vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
352 vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
353 vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
354 vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
355 vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
356 vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
357 vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
358
359 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
360 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
361 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
362 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
363 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
364 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
365 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
366 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
367
368 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
369 #if XNN_ARCH_ARM64
370 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
371 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
372 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
373 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
374
375 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
376 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
377 #else
378 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
379 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
380 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
381 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
382
383 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
384 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
385 #endif
386 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
387 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
388
389 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
390 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
391
392 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
393 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
394
395 if (nc >= 16) {
396 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
397 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
398
399 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
400 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
401
402 a = (const int8_t**restrict) ((uintptr_t) a - ks);
403
404 nc -= 16;
405 } else {
406 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
407 if (nc & 8) {
408 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
409 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
410 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
411 }
412 if (nc & 4) {
413 vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
414 vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
415 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
416 }
417 if (nc & 2) {
418 vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
419 vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
420 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
421 }
422 if (nc & 1) {
423 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
424 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
425 }
426
427 nc = 0;
428 }
429 } while (nc != 0);
430 }
431