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_rndnu_ukernel_4x8__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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59
60 do {
61 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
63 int32x4_t vacc1x0123 = vacc0x0123;
64 int32x4_t vacc1x4567 = vacc0x4567;
65 int32x4_t vacc2x0123 = vacc0x0123;
66 int32x4_t vacc2x4567 = vacc0x4567;
67 int32x4_t vacc3x0123 = vacc0x0123;
68 int32x4_t vacc3x4567 = vacc0x4567;
69
70 size_t k = kc;
71 while (k >= 8 * sizeof(int8_t)) {
72 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
73 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
74 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
75 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
76
77 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78
79 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
80 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
81 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
82 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
83 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
84 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
85 const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
86 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
87 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
88 const int16x8_t vprod3x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va3, 0));
89 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c0));
90 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));
91 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92
93 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
94 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
95 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
96 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
97 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
98 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
99 const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
100 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
101 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
102 const int16x8_t vprod3x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va3, 1));
103 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c1));
104 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));
105 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106
107 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
108 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
109 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
110 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
111 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
112 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
113 const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
114 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
115 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
116 const int16x8_t vprod3x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va3, 2));
117 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c2));
118 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));
119 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120
121 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
122 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
123 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
124 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
125 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
126 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
127 const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
128 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
129 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
130 const int16x8_t vprod3x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va3, 3));
131 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c3));
132 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));
133 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134
135 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
136 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
137 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
138 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
139 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
140 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
141 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
142 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
143 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
144 const int16x8_t vprod3x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va3, 4));
145 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c4));
146 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));
147 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148
149 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
150 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
151 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
152 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
153 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
154 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
155 const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
156 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
157 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
158 const int16x8_t vprod3x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va3, 5));
159 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c5));
160 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));
161 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162
163 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
164 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
165 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
166 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
167 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
168 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
169 const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
170 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
171 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
172 const int16x8_t vprod3x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va3, 6));
173 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c6));
174 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));
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 int16x8_t vprod2x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va2, 7));
184 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c7));
185 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c7));
186 const int16x8_t vprod3x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va3, 7));
187 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c7));
188 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c7));
189
190 k -= 8 * sizeof(int8_t);
191 }
192 if XNN_UNLIKELY(k != 0) {
193 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
194 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
195 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
196 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
197
198 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199
200 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
201 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
202 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
203 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
204 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
205 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
206 const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
207 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
208 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
209 const int16x8_t vprod3x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va3, 0));
210 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c0));
211 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c0));
212
213 if (k >= 2 * sizeof(int8_t)) {
214 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215
216 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
217 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
218 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
219 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
220 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
221 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
222 const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
223 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
224 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
225 const int16x8_t vprod3x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va3, 1));
226 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c1));
227 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c1));
228
229 if (k > 2 * sizeof(int8_t)) {
230 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231
232 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
233 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
234 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
235 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
236 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
237 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
238 const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
239 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
240 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
241 const int16x8_t vprod3x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va3, 2));
242 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c2));
243 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c2));
244
245 if (k >= 4 * sizeof(int8_t)) {
246 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247
248 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
249 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
250 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
251 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
252 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
253 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
254 const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
255 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
256 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
257 const int16x8_t vprod3x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va3, 3));
258 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c3));
259 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c3));
260
261 if (k > 4 * sizeof(int8_t)) {
262 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
263
264 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
265 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
266 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
267 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
268 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
269 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
270 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
271 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
272 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
273 const int16x8_t vprod3x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va3, 4));
274 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c4));
275 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c4));
276
277 if (k >= 6 * sizeof(int8_t)) {
278 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
279
280 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
281 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
282 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
283 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
284 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
285 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
286 const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
287 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
288 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
289 const int16x8_t vprod3x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va3, 5));
290 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c5));
291 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c5));
292
293 if (k > 6 * sizeof(int8_t)) {
294 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
295
296 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
297 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
298 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
299 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
300 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
301 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
302 const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
303 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
304 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
305 const int16x8_t vprod3x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va3, 6));
306 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c6));
307 vacc3x4567 = vaddw_s16(vacc3x4567, vget_high_s16(vprod3x01234567c6));
308 }
309 }
310 }
311 }
312 }
313 }
314 }
315
316 // Post-accumulation work
317 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
318 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
319 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
320
321 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
322 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
323 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
324 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
325 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
326 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
327 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
328 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
329
330 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
331 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
332 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
333 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
334 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
335 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
336 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
337 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
338
339 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
340 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
341 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
342 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
343 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
344 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
345 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
346 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
347
348 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
349 #if XNN_ARCH_ARM64
350 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
351 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
352 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
353 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
354
355 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
356 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
357 #else
358 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
359 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
360 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
361 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
362
363 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
364 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
365 #endif
366 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
367 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
368
369 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
370 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
371
372 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
373 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
374
375 if (nc >= 8) {
376 // Main case where there the 8 columns fit in the destination.
377 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
378 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
379 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
380 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
381
382 // Advance to the next 8 columns.
383 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
384 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
385 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
386 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
387
388 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
389 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
390 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
391 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
392
393 nc -= 8;
394 } else {
395 // Final case where not all of the 8 columns fit in the destination.
396 if (nc & 4) {
397 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
398 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
399 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
400 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
401 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
402 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
403 }
404 if (nc & 2) {
405 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
406 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
407 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
408 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
409 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
410 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
411 }
412 if (nc & 1) {
413 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
414 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
415 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
416 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
417 }
418
419 nc = 0;
420 }
421 } while (nc != 0);
422 }
423