Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8-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_4x8__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 <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 do {
60 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
61 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
62 int32x4_t vacc1x0123 = vacc0x0123;
63 int32x4_t vacc1x4567 = vacc0x4567;
64 int32x4_t vacc2x0123 = vacc0x0123;
65 int32x4_t vacc2x4567 = vacc0x4567;
66 int32x4_t vacc3x0123 = vacc0x0123;
67 int32x4_t vacc3x4567 = vacc0x4567;
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 const int8_t* restrict a1 = a[1];
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 if XNN_UNPREDICTABLE(a2 != zero) {
81 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
82 }
83 const int8_t* restrict a3 = a[3];
84 if XNN_UNPREDICTABLE(a3 != zero) {
85 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
86 }
87 a += 4;
88
89 size_t k = kc;
90 while (k >= 8 * sizeof(int8_t)) {
91 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
92 const int16x8_t vxa0 = vmovl_s8(va0);
93 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
94 const int16x8_t vxa1 = vmovl_s8(va1);
95 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
96 const int16x8_t vxa2 = vmovl_s8(va2);
97 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
98 const int16x8_t vxa3 = vmovl_s8(va3);
99
100 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
101 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
102
103 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
104 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
105 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
106 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
107 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
108 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
109 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
110 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
111 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
112 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
113
114 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
115 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
116 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
117 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
118 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
119 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
120 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
121 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
122 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
123 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
124
125 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
126 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
127 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
128 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
129 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
130 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
131 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
132 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
133 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
134 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
135
136 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
137 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
138 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
139 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
140 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
141 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
142 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
143 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
144
145 xnn_prefetch_to_l1((const int8_t*) w + 448);
146
147 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
148 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
149
150 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
151 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
152 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
153 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
154 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
155 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
156 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
157 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
158 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
160
161 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
162 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
163 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
164 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
165 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
166 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
167 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
168 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
169 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
171
172 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
173 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
174 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
175 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
176 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
177 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
178 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
179 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
180 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
182
183 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
184 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
185 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
186 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
187 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
188 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
189 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
190 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
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 int16x8_t vxa0 = vmovl_s8(va0);
197 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
198 const int16x8_t vxa1 = vmovl_s8(va1);
199 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
200 const int16x8_t vxa2 = vmovl_s8(va2);
201 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
202 const int16x8_t vxa3 = vmovl_s8(va3);
203
204 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
205 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
206
207 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
208 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
209 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
210 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
211 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
212 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
213 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
214 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
215
216 if (k >= 2 * sizeof(int8_t)) {
217 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
218 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
219
220 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
221 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
222 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
223 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
224 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
225 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
226 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
227 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
228
229 if (k > 2 * sizeof(int8_t)) {
230 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
231 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
232
233 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
234 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
235 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
236 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
237 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
238 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
239 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
240 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
241
242 if (k >= 4 * sizeof(int8_t)) {
243 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
244 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
245
246 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
247 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
248 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
249 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
250 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
251 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
252 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
253 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
254
255 if (k > 4 * sizeof(int8_t)) {
256 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
257 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
258
259 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
260 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
261 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
262 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
263 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
264 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
265 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
266 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
267
268 if (k >= 6 * sizeof(int8_t)) {
269 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
270 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
271
272 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
273 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
274 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
275 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
276 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
277 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
278 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
279 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
280
281 if (k > 6 * sizeof(int8_t)) {
282 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
283 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
284
285 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
286 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
287 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
288 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
289 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
290 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
291 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
292 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
293 }
294 }
295 }
296 }
297 }
298 }
299 }
300 p -= 4 * sizeof(void*);
301 } while (p != 0);
302
303 // Post-accumulation work
304 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
305 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
306 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
307 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
308 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
309 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
310 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
311 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
312
313 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
314 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
315 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
316 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
317 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
318 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
319 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
320 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
321 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
322 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
323
324 const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
325 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
326 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
327 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
328 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
329 vacc2x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x0123, vmagic_bias));
330 vacc2x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x4567, vmagic_bias));
331 vacc3x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x0123, vmagic_bias));
332 vacc3x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x4567, vmagic_bias));
333
334 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
335 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
336 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
337 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
338 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
339 vacc2x0123 = vqsubq_s32(vacc2x0123, vmagic_bias_less_output_zero_point);
340 vacc2x4567 = vqsubq_s32(vacc2x4567, vmagic_bias_less_output_zero_point);
341 vacc3x0123 = vqsubq_s32(vacc3x0123, vmagic_bias_less_output_zero_point);
342 vacc3x4567 = vqsubq_s32(vacc3x4567, vmagic_bias_less_output_zero_point);
343
344 #if XNN_ARCH_ARM64
345 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
346 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
347 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
348 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
349
350
351 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
352 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
353 #else
354 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
355 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
356 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
357 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
358
359
360 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
361 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
362 #endif
363
364 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
365 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
366 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
367
368 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
369 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
370 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
371
372 if (nc >= 8) {
373 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
374 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
375 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
376 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
377
378 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
379 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
380 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
381 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
382
383 a = (const int8_t**restrict) ((uintptr_t) a - ks);
384
385 nc -= 8;
386 } else {
387 if (nc & 4) {
388 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
389 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
390 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
391 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
392 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
393 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
394 }
395 if (nc & 2) {
396 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
397 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
398 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
399 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
400 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
401 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
402 }
403 if (nc & 1) {
404 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
405 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
406 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
407 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
408 }
409
410 nc = 0;
411 }
412 } while (nc != 0);
413 }