1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mull-addw-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_gemm_minmax_ukernel_2x16__neon_mull_addw_dup(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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 do {
49 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53 int32x4_t vacc1x0123 = vacc0x0123;
54 int32x4_t vacc1x4567 = vacc0x4567;
55 int32x4_t vacc1x89AB = vacc0x89AB;
56 int32x4_t vacc1xCDEF = vacc0xCDEF;
57
58 size_t k = kc;
59 while (k >= 8 * sizeof(int8_t)) {
60 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
61 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
62
63 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64
65 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
66 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
67 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
68 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
69 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
70 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
71 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72
73 const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
74 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
75 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
76 const int16x8_t vprod1x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va1, 0));
77 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc0));
78 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));
79 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80
81 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
82 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
83 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
84 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
85 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
86 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
87 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88
89 const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
90 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
91 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
92 const int16x8_t vprod1x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va1, 1));
93 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc1));
94 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));
95 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96
97 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
98 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
99 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
100 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
101 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
102 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
103 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104
105 const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
106 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
107 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
108 const int16x8_t vprod1x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va1, 2));
109 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc2));
110 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));
111 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112
113 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
114 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
115 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
116 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
117 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
118 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
119 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120
121 const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
122 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
123 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
124 const int16x8_t vprod1x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va1, 3));
125 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc3));
126 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));
127 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128
129 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
130 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
131 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
132 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
133 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
134 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
135 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136
137 const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
138 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
139 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
140 const int16x8_t vprod1x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va1, 4));
141 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc4));
142 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));
143 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144
145 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
146 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
147 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
148 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
149 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
150 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
151 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152
153 const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
154 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
155 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
156 const int16x8_t vprod1x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va1, 5));
157 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc5));
158 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));
159 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160
161 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
162 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
163 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
164 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
165 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
166 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
167 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168
169 const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
170 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
171 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
172 const int16x8_t vprod1x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va1, 6));
173 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc6));
174 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));
175 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176
177 const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
178 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
179 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
180 const int16x8_t vprod1x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va1, 7));
181 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));
182 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c7));
183 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184
185 const int16x8_t vprod0x89ABCDEFc7 = vmull_s8(vb89ABCDEFc7, vdup_lane_s8(va0, 7));
186 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));
187 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc7));
188 const int16x8_t vprod1x89ABCDEFc7 = vmull_s8(vb89ABCDEFc7, vdup_lane_s8(va1, 7));
189 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc7));
190 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc7));
191
192 k -= 8 * sizeof(int8_t);
193 }
194 if XNN_UNLIKELY(k != 0) {
195 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
196 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
197
198 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200
201 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
202 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
203 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
204 const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
205 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
206 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
207 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
208 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
209 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
210 const int16x8_t vprod1x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va1, 0));
211 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc0));
212 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc0));
213
214 if (k >= 2 * sizeof(int8_t)) {
215 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217
218 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
219 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
220 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
221 const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
222 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
223 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
224 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
225 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
226 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
227 const int16x8_t vprod1x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va1, 1));
228 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc1));
229 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc1));
230
231 if (k > 2 * sizeof(int8_t)) {
232 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234
235 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
236 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
237 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
238 const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
239 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
240 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
241 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
242 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
243 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
244 const int16x8_t vprod1x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va1, 2));
245 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc2));
246 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc2));
247
248 if (k >= 4 * sizeof(int8_t)) {
249 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251
252 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
253 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
254 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
255 const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
256 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
257 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
258 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
259 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
260 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
261 const int16x8_t vprod1x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va1, 3));
262 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc3));
263 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc3));
264
265 if (k > 4 * sizeof(int8_t)) {
266 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268
269 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
270 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
271 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
272 const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
273 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
274 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
275 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
276 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
277 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
278 const int16x8_t vprod1x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va1, 4));
279 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc4));
280 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc4));
281
282 if (k >= 6 * sizeof(int8_t)) {
283 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
284 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
285
286 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
287 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
288 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
289 const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
290 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
291 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
292 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
293 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
294 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
295 const int16x8_t vprod1x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va1, 5));
296 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc5));
297 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc5));
298
299 if (k > 6 * sizeof(int8_t)) {
300 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
301 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
302
303 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
304 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
305 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
306 const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
307 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
308 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
309 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
310 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
311 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
312 const int16x8_t vprod1x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va1, 6));
313 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc6));
314 vacc1xCDEF = vaddw_s16(vacc1xCDEF, vget_high_s16(vprod1x89ABCDEFc6));
315 }
316 }
317 }
318 }
319 }
320 }
321 }
322 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier);
323 vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
324 vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
325 vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
326 vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
327 vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
328 vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
329 vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
330 vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
331
332 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
333 const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
334 vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
335 vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
336 vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
337 vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
338 vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
339 vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
340 vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
341 vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
342
343 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
344 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
345 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
346 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
347 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
348 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
349 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
350 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
351
352 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
353 #if XNN_ARCH_ARM64
354 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
355 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
356 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
357 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
358
359 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
360 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
361 #else
362 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
363 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
364 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
365 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
366
367 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
368 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
369 #endif
370 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
371 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
372
373 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
374 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
375
376 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
377 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
378
379 if (nc >= 16) {
380 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
381 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
382
383 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
384 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
385
386 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
387 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
388
389 nc -= 16;
390 } else {
391 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
392 if (nc & 8) {
393 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
394 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
395 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
396 }
397 if (nc & 4) {
398 vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
399 vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
400 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
401 }
402 if (nc & 2) {
403 vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
404 vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
405 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
406 }
407 if (nc & 1) {
408 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
409 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
410 }
411
412 nc = 0;
413 }
414 } while (nc != 0);
415 }
416