Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8-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_4x8__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 <= 4);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (4 * 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 do {
61 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
62 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
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 p = ks;
71 do {
72 const int8_t* restrict a0 = a[0];
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
75 }
76 const int8_t* restrict a1 = a[1];
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
79 }
80 const int8_t* restrict a2 = a[2];
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 const int8_t* restrict a3 = a[3];
85 if XNN_UNPREDICTABLE(a3 != zero) {
86 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
87 }
88 a += 4;
89
90 size_t k = kc;
91 while (k >= 8 * sizeof(int8_t)) {
92 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
93 const int16x8_t vxa0 = vmovl_s8(va0);
94 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
95 const int16x8_t vxa1 = vmovl_s8(va1);
96 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
97 const int16x8_t vxa2 = vmovl_s8(va2);
98 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
99 const int16x8_t vxa3 = vmovl_s8(va3);
100
101 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
102 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
103
104 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
105 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
106 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
107 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
108 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
109 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
110 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
111 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
112 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
113 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
114
115 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
116 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
117 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
118 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
119 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
120 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
121 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
122 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
123 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
124 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
125
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
128 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
129 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
130 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
131 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
132 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
133 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
134 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
135 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
136
137 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
138 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
139 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
140 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
141 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
142 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
143 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
144 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
145
146 xnn_prefetch_to_l1((const int8_t*) w + 448);
147
148 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
149 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
150
151 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
152 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
153 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
154 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
155 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
156 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
157 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
158 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
159 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
161
162 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
163 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
164 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
165 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
166 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
167 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
168 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
169 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
170 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
171 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
172
173 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
174 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
175 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
176 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
177 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
178 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
179 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
180 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
181 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
182 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
183
184 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
185 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
186 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
187 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
188 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
189 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
190 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
191 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
192
193 k -= 8 * sizeof(int8_t);
194 }
195 if XNN_UNLIKELY(k != 0) {
196 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
197 const int16x8_t vxa0 = vmovl_s8(va0);
198 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
199 const int16x8_t vxa1 = vmovl_s8(va1);
200 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
201 const int16x8_t vxa2 = vmovl_s8(va2);
202 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
203 const int16x8_t vxa3 = vmovl_s8(va3);
204
205 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
206 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
207
208 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
209 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
210 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
211 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
212 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
213 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
214 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
215 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
216
217 if (k >= 2 * sizeof(int8_t)) {
218 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
220
221 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
222 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
223 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
224 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
225 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
226 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
227 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
228 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
229
230 if (k > 2 * sizeof(int8_t)) {
231 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
232 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
233
234 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
235 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
236 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
237 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
238 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
239 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
240 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
241 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
242
243 if (k >= 4 * sizeof(int8_t)) {
244 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
245 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
246
247 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
248 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
249 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
250 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
251 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
252 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
253 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
254 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
255
256 if (k > 4 * sizeof(int8_t)) {
257 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
258 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
259
260 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
261 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
262 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
263 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
264 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
265 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
266 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
267 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
268
269 if (k >= 6 * sizeof(int8_t)) {
270 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
271 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
272
273 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
274 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
275 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
276 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
277 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
278 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
279 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
280 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
281
282 if (k > 6 * sizeof(int8_t)) {
283 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
284 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
285
286 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
287 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
288 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
289 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
290 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
291 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
292 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
293 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
294 }
295 }
296 }
297 }
298 }
299 }
300 }
301 p -= 4 * sizeof(void*);
302 } while (p != 0);
303
304 // Post-accumulation work
305 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
306 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
307 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
308 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
309 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
310 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
311 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
312 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
313
314 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
315 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
316 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
317 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
318 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
319 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
320 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
321 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
322 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
323 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
324
325 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
326 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
327 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
328 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
329 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
330 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
331 vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
332 vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
333
334 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
335 #if XNN_ARCH_ARM64
336 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
337 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
338 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
339 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
340
341 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
342 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
343 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
344 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
345
346 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
347 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
348 #else
349 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
350 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
351 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
352 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
353
354 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
355 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
356 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
357 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
358
359 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
360 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
361 #endif
362
363 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
364 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
365 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
366
367 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
368 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
369 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
370
371 if (nc >= 8) {
372 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
373 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
374 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
375 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
376
377 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
378 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
379 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
380 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
381
382 a = (const int8_t**restrict) ((uintptr_t) a - ks);
383
384 nc -= 8;
385 } else {
386 if (nc & 4) {
387 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
388 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
389 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
390 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
391 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
392 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
393 }
394 if (nc & 2) {
395 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
396 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
397 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
398 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
399 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
400 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
401 }
402 if (nc & 1) {
403 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
404 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
405 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
406 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
407 }
408
409 nc = 0;
410 }
411 } while (nc != 0);
412 }