Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x16-minmax-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_qd8_f32_qc8w_igemm_minmax_ukernel_2x16__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 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 2);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (2 * sizeof(void*)) == 0);
42 assert(a_offset % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr != 2) {
50 c1 = c0;
51 }
52
53 do {
54 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
55 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
58 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
59 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
60 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
61 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
62 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
63 int32x4_t vacc1x0123 = vacc0x0123;
64 int32x4_t vacc1x4567 = vacc0x4567;
65 int32x4_t vacc1x89AB = vacc0x89AB;
66 int32x4_t vacc1xCDEF = vacc0xCDEF;
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 } else {
74 a0 = zero_data;
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 } else {
80 a1 = zero_data;
81 }
82 a += 2;
83
84 size_t k = kc;
85 while (k >= 8 * sizeof(int8_t)) {
86 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
87 const int16x8_t vxa0 = vmovl_s8(va0);
88 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
89 const int16x8_t vxa1 = vmovl_s8(va1);
90
91 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
93
94 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
95 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
96 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
97 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
98 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
100
101 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
102 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
103 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
104 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
105 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
107
108 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
109 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
110 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
111 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
112 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
113 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
114
115 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
116 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
117 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
118 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
119 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
120 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
121
122 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
123 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
124 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
125 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
126 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
128
129 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
130 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
131 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
132 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 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 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
141 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
142
143 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
144 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
145 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
146 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
147
148 xnn_prefetch_to_l1((const int8_t*) w + 448);
149 xnn_prefetch_to_l1((const int8_t*) w + 512);
150
151 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
152 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
153
154 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
155 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
156 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
157 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
158 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
160
161 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
162 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
163 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
164 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
165 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
166 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
167
168 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
169 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
170 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
171 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
172 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
173 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
174
175 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
176 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
177 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
178 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
179 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
180 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
181
182 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
183 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
184 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
185 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
186 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
187 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
188
189 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
190 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
191 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
192 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
193 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
194 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
195
196 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
197 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
198 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
199 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
200 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
201 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
202
203 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
204 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
205 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
206 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
207
208 k -= 8 * sizeof(int8_t);
209 }
210 if XNN_UNLIKELY(k != 0) {
211 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
212 const int16x8_t vxa0 = vmovl_s8(va0);
213 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
214 const int16x8_t vxa1 = vmovl_s8(va1);
215
216 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
217 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
218 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
220
221 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
222 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
223 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
224 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
225 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
226 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
227 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
228 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
229
230 if (k >= 2 * sizeof(int8_t)) {
231 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
232 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
233 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
234 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
235
236 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
237 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
238 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
239 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
240 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
241 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
242 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
243 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
244
245 if (k > 2 * sizeof(int8_t)) {
246 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
247 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
248 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
249 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
250
251 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
252 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
253 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
254 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
255 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
256 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
257 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
258 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
259
260 if (k >= 4 * sizeof(int8_t)) {
261 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
262 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
263 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
264 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
265
266 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
267 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
268 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
269 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
270 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
271 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
272 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
273 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
274
275 if (k > 4 * sizeof(int8_t)) {
276 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
277 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
278 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
279 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
280
281 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
282 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
283 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
284 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
285 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
286 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
287 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
288 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
289
290 if (k >= 6 * sizeof(int8_t)) {
291 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
292 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
293 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
294 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
295
296 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
297 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
298 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
299 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
300 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
301 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
302 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
303 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
304
305 if (k > 6 * sizeof(int8_t)) {
306 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
307 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
308 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
309 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
310
311 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
312 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
313 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
314 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
315 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
316 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
317 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
318 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
319 }
320 }
321 }
322 }
323 }
324 }
325 }
326 p -= 2 * sizeof(void*);
327 } while (p != 0);
328
329 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
330 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
331 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
332 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
333 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
334 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
335 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
336 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
337
338 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
339 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
340 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
341 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
342 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
343 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
344 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
345 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
346 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
347
348 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
349 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
350 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
351 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
352
353 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
354 #if XNN_ARCH_ARM64
355 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
356 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
357 #else
358 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
359 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
360 #endif
361 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
362 #if XNN_ARCH_ARM64
363 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
364 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
365 #else
366 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
367 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
368 #endif
369 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
370 #if XNN_ARCH_ARM64
371 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
372 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
373 #else
374 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
375 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
376 #endif
377 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
378 #if XNN_ARCH_ARM64
379 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
380 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
381 #else
382 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
383 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
384 #endif
385
386 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
387 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
388 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
389 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
390 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
391 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
392 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
393 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
394 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
395
396 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
397 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
398 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
399 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
400 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
401 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
402 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
403 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
404 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
405
406 if XNN_LIKELY(nc >= 16) {
407 vst1q_f32(&c1[0], vout1x0123);
408 vst1q_f32(&c1[4], vout1x4567);
409 vst1q_f32(&c1[8], vout1x89AB);
410 vst1q_f32(&c1[12], vout1xCDEF);
411 vst1q_f32(&c0[0], vout0x0123);
412 vst1q_f32(&c0[4], vout0x4567);
413 vst1q_f32(&c0[8], vout0x89AB);
414 vst1q_f32(&c0[12], vout0xCDEF);
415
416 a = (const int8_t**restrict) ((uintptr_t) a - ks);
417
418 c0 = (float*) ((uintptr_t) c0 + cn_stride);
419 c1 = (float*) ((uintptr_t) c1 + cn_stride);
420
421 nc -= 16;
422 } else {
423 if (nc & 8) {
424 vst1q_f32(c1, vout1x0123); c1 += 4;
425 vout1x0123 = vout1x89AB;
426 vst1q_f32(c0, vout0x0123); c0 += 4;
427 vout0x0123 = vout0x89AB;
428 vst1q_f32(c1, vout1x4567); c1 += 4;
429 vout1x4567 = vout1xCDEF;
430 vst1q_f32(c0, vout0x4567); c0 += 4;
431 vout0x4567 = vout0xCDEF;
432 }
433 if (nc & 4) {
434 vst1q_f32(c1, vout1x0123); c1 += 4;
435 vout1x0123 = vout1x4567;
436 vst1q_f32(c0, vout0x0123); c0 += 4;
437 vout0x0123 = vout0x4567;
438 }
439 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
440 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
441 if (nc & 2) {
442 vst1_f32(c1, vout1x01); c1 += 2;
443 vst1_f32(c0, vout0x01); c0 += 2;
444 vout1x01 = vget_high_f32(vout1x0123);
445 vout0x01 = vget_high_f32(vout0x0123);
446 }
447 if (nc & 1) {
448 vst1_lane_f32(c1, vout1x01, 0);
449 vst1_lane_f32(c0, vout0x01, 0);
450 }
451 nc = 0;
452 }
453 } while (nc != 0);
454 }