Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x8-minmax-fp32-neonv8-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/intrinsics-polyfill.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x8__neonv8_mlal_lane_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 3);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (3 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 do {
57 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
58 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
59 int32x4_t vacc1x0123 = vacc0x0123;
60 int32x4_t vacc1x4567 = vacc0x4567;
61 int32x4_t vacc2x0123 = vacc0x0123;
62 int32x4_t vacc2x4567 = vacc0x4567;
63
64 size_t p = ks;
65 do {
66 const int8_t* restrict a0 = a[0];
67 if XNN_UNPREDICTABLE(a0 != zero) {
68 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
69 }
70 const int8_t* restrict a1 = a[1];
71 if XNN_UNPREDICTABLE(a1 != zero) {
72 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
73 }
74 const int8_t* restrict a2 = a[2];
75 if XNN_UNPREDICTABLE(a2 != zero) {
76 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
77 }
78 a += 3;
79
80 size_t k = kc;
81 while (k >= 8 * sizeof(int8_t)) {
82 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
83 const int16x8_t vxa0 = vmovl_s8(va0);
84 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
85 const int16x8_t vxa1 = vmovl_s8(va1);
86 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
87 const int16x8_t vxa2 = vmovl_s8(va2);
88
89 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
91
92 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
93 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
94 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
95 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
96 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
97 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
98 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
100
101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
103 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
104 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
105 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
106 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
107 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
109
110 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
111 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
112 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
113 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
114 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
115 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
116 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
118
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
121 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
122 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
123 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
124 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
125
126 xnn_prefetch_to_l1((const int8_t*) w + 448);
127
128 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
130
131 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
132 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
133 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
134 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
135 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
136 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
137 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
138 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
139
140 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
141 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
142 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
143 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
144 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
145 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
146 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
148
149 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
150 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
151 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
152 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
153 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
154 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
155 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
157
158 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
159 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
160 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
161 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
162 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
163 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
164
165 k -= 8 * sizeof(int8_t);
166 }
167 if XNN_UNLIKELY(k != 0) {
168 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
169 const int16x8_t vxa0 = vmovl_s8(va0);
170 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
171 const int16x8_t vxa1 = vmovl_s8(va1);
172 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
173 const int16x8_t vxa2 = vmovl_s8(va2);
174
175 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
176 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
177
178 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
179 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
180 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
181 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
182 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
183 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
184
185 if (k >= 2 * sizeof(int8_t)) {
186 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
187 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
188
189 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
190 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
191 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
192 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
193 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
194 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
195
196 if (k > 2 * sizeof(int8_t)) {
197 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
198 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
199
200 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
201 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
202 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
203 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
204 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
205 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
206
207 if (k >= 4 * sizeof(int8_t)) {
208 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
209 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
210
211 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
212 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
213 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
214 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
215 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
216 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
217
218 if (k > 4 * sizeof(int8_t)) {
219 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
220 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
221
222 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
223 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
224 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
225 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
226 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
227 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
228
229 if (k >= 6 * sizeof(int8_t)) {
230 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
231 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
232
233 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
234 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
235 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
236 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
237 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
238 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
239
240 if (k > 6 * sizeof(int8_t)) {
241 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
242 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
243
244 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
245 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
246 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
247 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
248 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
249 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
250 }
251 }
252 }
253 }
254 }
255 }
256 }
257 p -= 3 * sizeof(void*);
258 } while (p != 0);
259
260 // Post-accumulation work
261 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
262 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
263 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
264 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
265 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
266 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
267
268 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
269 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
270 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
271 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
272 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
273 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
274 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
275 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
276
277 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
278 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
279 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
280 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
281 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
282 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
283
284 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
285 #if XNN_ARCH_ARM64
286 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
287 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
288 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
289
290 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
291 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
292 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
293
294 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
295 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
296 #else
297 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
298 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
299 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
300
301 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
302 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
303 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
304
305 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
306 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
307 #endif
308
309 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
310 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
311 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
312
313 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
314 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
315 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
316
317 if (nc >= 8) {
318 vst1_s8(c2 + 0, vout2x01234567);
319 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
320 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
321
322 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
323 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
324 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
325
326 a = (const int8_t**restrict) ((uintptr_t) a - ks);
327
328 nc -= 8;
329 } else {
330 if (nc & 4) {
331 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
332 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
333 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
334 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
335 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
336 }
337 if (nc & 2) {
338 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
339 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
340 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
341 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
342 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
343 }
344 if (nc & 1) {
345 vst1_lane_s8(c2, vout2x01234567, 0);
346 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
347 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
348 }
349
350 nc = 0;
351 }
352 } while (nc != 0);
353 }