Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x8-minmax-fp32-neon-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/neon-mlal-lane.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/igemm.h"
17 #include "xnnpack/prefetch.h"
18
19
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x8__neon_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55 do {
56 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
57 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
58 int32x4_t vacc1x0123 = vacc0x0123;
59 int32x4_t vacc1x4567 = vacc0x4567;
60 int32x4_t vacc2x0123 = vacc0x0123;
61 int32x4_t vacc2x4567 = vacc0x4567;
62
63 size_t p = ks;
64 do {
65 const int8_t* restrict a0 = a[0];
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
68 }
69 const int8_t* restrict a1 = a[1];
70 if XNN_UNPREDICTABLE(a1 != zero) {
71 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
72 }
73 const int8_t* restrict a2 = a[2];
74 if XNN_UNPREDICTABLE(a2 != zero) {
75 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
76 }
77 a += 3;
78
79 size_t k = kc;
80 while (k >= 8 * sizeof(int8_t)) {
81 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
82 const int16x8_t vxa0 = vmovl_s8(va0);
83 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
84 const int16x8_t vxa1 = vmovl_s8(va1);
85 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
86 const int16x8_t vxa2 = vmovl_s8(va2);
87
88 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
90
91 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
92 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
93 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
94 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
95 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
96 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
97 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
98 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
99
100 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
101 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
102 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
103 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
104 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
105 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
106 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
107 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
108
109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
110 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
111 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
112 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
113 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
114 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
115 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
116 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
117
118 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
119 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
120 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
121 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
122 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
123 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
124
125 xnn_prefetch_to_l1((const int8_t*) w + 448);
126
127 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
128 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
129
130 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
131 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
132 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
133 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
134 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
135 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
136 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
137 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
138
139 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
140 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
141 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
142 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
143 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
144 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
145 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
146 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
147
148 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
149 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
150 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
151 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
152 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
153 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
154 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
155 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
156
157 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
158 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
159 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
160 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
161 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
162 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
163
164 k -= 8 * sizeof(int8_t);
165 }
166 if XNN_UNLIKELY(k != 0) {
167 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
168 const int16x8_t vxa0 = vmovl_s8(va0);
169 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
170 const int16x8_t vxa1 = vmovl_s8(va1);
171 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
172 const int16x8_t vxa2 = vmovl_s8(va2);
173
174 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
175 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
176
177 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
178 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
179 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
180 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
181 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
182 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
183
184 if (k >= 2 * sizeof(int8_t)) {
185 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
186 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
187
188 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
189 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
190 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
191 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
192 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
193 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
194
195 if (k > 2 * sizeof(int8_t)) {
196 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
197 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
198
199 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
200 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
201 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
202 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
203 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
204 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
205
206 if (k >= 4 * sizeof(int8_t)) {
207 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
208 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
209
210 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
211 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
212 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
213 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
214 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
215 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
216
217 if (k > 4 * sizeof(int8_t)) {
218 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
220
221 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
222 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
223 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
224 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
225 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
226 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
227
228 if (k >= 6 * sizeof(int8_t)) {
229 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
230 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
231
232 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
233 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
234 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
235 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
236 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
237 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
238
239 if (k > 6 * sizeof(int8_t)) {
240 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
241 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
242
243 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
244 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
245 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
246 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
247 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
248 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
249 }
250 }
251 }
252 }
253 }
254 }
255 }
256 p -= 3 * sizeof(void*);
257 } while (p != 0);
258
259 // Post-accumulation work
260 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
261 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
262 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
263 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
264 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
265 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
266
267 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
268 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
269 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
270 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
271 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
272 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
273 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
274 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
275
276 const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
277 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
278 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
279 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
280 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
281 vacc2x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x0123, vmagic_bias));
282 vacc2x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x4567, vmagic_bias));
283
284 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
285 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
286 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
287 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
288 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
289 vacc2x0123 = vqsubq_s32(vacc2x0123, vmagic_bias_less_output_zero_point);
290 vacc2x4567 = vqsubq_s32(vacc2x4567, vmagic_bias_less_output_zero_point);
291
292 #if XNN_ARCH_ARM64
293 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
294 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
295 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
296
297
298 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
299 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
300 #else
301 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
302 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
303 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
304
305
306 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
307 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
308 #endif
309
310 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
311 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
312 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
313
314 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
315 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
316 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
317
318 if (nc >= 8) {
319 vst1_s8(c2 + 0, vout2x01234567);
320 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
321 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
322
323 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
324 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
325 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
326
327 a = (const int8_t**restrict) ((uintptr_t) a - ks);
328
329 nc -= 8;
330 } else {
331 if (nc & 4) {
332 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
333 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
334 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
335 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
336 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
337 }
338 if (nc & 2) {
339 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
340 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
341 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
342 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
343 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
344 }
345 if (nc & 1) {
346 vst1_lane_s8(c2, vout2x01234567, 0);
347 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
348 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
349 }
350
351 nc = 0;
352 }
353 } while (nc != 0);
354 }