Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x8-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_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 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 <= 3);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (3 * 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 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56
57 do {
58 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
60 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
61 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
62 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
63 int32x4_t vacc1x0123 = vacc0x0123;
64 int32x4_t vacc1x4567 = vacc0x4567;
65 int32x4_t vacc2x0123 = vacc0x0123;
66 int32x4_t vacc2x4567 = vacc0x4567;
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 const int8_t* restrict a2 = a[2];
83 if XNN_UNPREDICTABLE(a2 != zero) {
84 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
85 } else {
86 a2 = zero_data;
87 }
88 a += 3;
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
99 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
100 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
101
102 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
103 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
104 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
105 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
106 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
107 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
108 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
110
111 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
112 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
113 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
114 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
115 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
116 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
117 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
118 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
119
120 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
121 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
122 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
123 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
124 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
125 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
126 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
128
129 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
130 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
131 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
132 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
133 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
134 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
135
136 xnn_prefetch_to_l1((const int8_t*) w + 448);
137
138 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
139 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
140
141 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
142 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
143 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
144 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
145 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
146 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
147 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
148 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
149
150 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
151 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
152 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
153 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
154 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
155 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
156 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
158
159 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
160 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
161 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
162 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
163 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
164 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
165 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
166 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
167
168 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
169 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
170 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
171 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
172 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
173 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
174
175 k -= 8 * sizeof(int8_t);
176 }
177 if XNN_UNLIKELY(k != 0) {
178 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
179 const int16x8_t vxa0 = vmovl_s8(va0);
180 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
181 const int16x8_t vxa1 = vmovl_s8(va1);
182 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
183 const int16x8_t vxa2 = vmovl_s8(va2);
184
185 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
186 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
187
188 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
189 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
190 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
191 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
192 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
193 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
194
195 if (k >= 2 * sizeof(int8_t)) {
196 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
197 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
198
199 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
200 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
201 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
202 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
203 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
204 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
205
206 if (k > 2 * sizeof(int8_t)) {
207 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
208 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
209
210 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
211 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
212 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
213 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
214 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
215 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
216
217 if (k >= 4 * sizeof(int8_t)) {
218 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
220
221 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
222 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
223 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
224 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
225 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
226 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
227
228 if (k > 4 * sizeof(int8_t)) {
229 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
230 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
231
232 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
233 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
234 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
235 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
236 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
237 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
238
239 if (k >= 6 * sizeof(int8_t)) {
240 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
241 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
242
243 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
244 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
245 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
246 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
247 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
248 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
249
250 if (k > 6 * sizeof(int8_t)) {
251 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
252 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
253
254 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
255 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
256 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
257 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
258 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
259 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
260 }
261 }
262 }
263 }
264 }
265 }
266 }
267 p -= 3 * sizeof(void*);
268 } while (p != 0);
269
270 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
271 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
272 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
273 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
274 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
275 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
276
277 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
278 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
279 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
280 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
281 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
282 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
283 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
284
285 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
286 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
287
288 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
289 #if XNN_ARCH_ARM64
290 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
291 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
292 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
293 #else
294 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
295 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
296 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
297 #endif
298 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
299 #if XNN_ARCH_ARM64
300 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
301 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
302 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
303 #else
304 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
305 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
306 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
307 #endif
308
309 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
310 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
311 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
312 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
313 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
314 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
315 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
316
317 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
318 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
319 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
320 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
321 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
322 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
323 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
324
325 if XNN_LIKELY(nc >= 8) {
326 vst1q_f32(&c2[0], vout2x0123);
327 vst1q_f32(&c2[4], vout2x4567);
328 vst1q_f32(&c1[0], vout1x0123);
329 vst1q_f32(&c1[4], vout1x4567);
330 vst1q_f32(&c0[0], vout0x0123);
331 vst1q_f32(&c0[4], vout0x4567);
332
333 a = (const int8_t**restrict) ((uintptr_t) a - ks);
334
335 c0 = (float*) ((uintptr_t) c0 + cn_stride);
336 c1 = (float*) ((uintptr_t) c1 + cn_stride);
337 c2 = (float*) ((uintptr_t) c2 + cn_stride);
338
339 nc -= 8;
340 } else {
341 if (nc & 4) {
342 vst1q_f32(c2, vout2x0123); c2 += 4;
343 vout2x0123 = vout2x4567;
344 vst1q_f32(c1, vout1x0123); c1 += 4;
345 vout1x0123 = vout1x4567;
346 vst1q_f32(c0, vout0x0123); c0 += 4;
347 vout0x0123 = vout0x4567;
348 }
349 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
350 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
351 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
352 if (nc & 2) {
353 vst1_f32(c2, vout2x01); c2 += 2;
354 vst1_f32(c1, vout1x01); c1 += 2;
355 vst1_f32(c0, vout0x01); c0 += 2;
356 vout2x01 = vget_high_f32(vout2x0123);
357 vout1x01 = vget_high_f32(vout1x0123);
358 vout0x01 = vget_high_f32(vout0x0123);
359 }
360 if (nc & 1) {
361 vst1_lane_f32(c2, vout2x01, 0);
362 vst1_lane_f32(c1, vout1x01, 0);
363 vst1_lane_f32(c0, vout0x01, 0);
364 }
365 nc = 0;
366 }
367 } while (nc != 0);
368 }