Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x8-minmax-neon-mlal-lane.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
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_3x8__neon_mlal_lane(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const int8_t* zero_data,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_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 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 do {
57 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
58 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
60 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
61 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
62 int32x4_t vacc1x0123 = vacc0x0123;
63 int32x4_t vacc1x4567 = vacc0x4567;
64 int32x4_t vacc2x0123 = vacc0x0123;
65 int32x4_t vacc2x4567 = vacc0x4567;
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 } else {
73 a0 = zero_data;
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 } else {
79 a1 = zero_data;
80 }
81 const int8_t* restrict a2 = a[2];
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
84 } else {
85 a2 = zero_data;
86 }
87 a += 3;
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
98 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
100
101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
103 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
104 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
105 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
106 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
107 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
109
110 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
111 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
112 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
113 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
114 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
115 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
116 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
118
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
121 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
122 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
123 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
124 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
125 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
126 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
127
128 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
129 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
130 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
131 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
132 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
133 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
134
135
136 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
137 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
138
139 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
140 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
141 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
142 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
143 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
144 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
145 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
146 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
147
148 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
149 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
150 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
151 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
152 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
153 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
154 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
155 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
156
157 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
158 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
159 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
160 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
161 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
162 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
163 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
164 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
165
166 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
167 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
168 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
169 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
170 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
171 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
172
173 k -= 8 * sizeof(int8_t);
174 }
175 if XNN_UNLIKELY(k != 0) {
176 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
177 const int16x8_t vxa0 = vmovl_s8(va0);
178 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
179 const int16x8_t vxa1 = vmovl_s8(va1);
180 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
181 const int16x8_t vxa2 = vmovl_s8(va2);
182
183 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
184 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
185
186 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
187 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
188 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
189 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
190 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
191 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
192
193 if (k >= 2 * sizeof(int8_t)) {
194 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
195 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
196
197 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
198 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
199 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
200 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
201 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
202 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
203
204 if (k > 2 * sizeof(int8_t)) {
205 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
206 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
207
208 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
209 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
210 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
211 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
212 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
213 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
214
215 if (k >= 4 * sizeof(int8_t)) {
216 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
217 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
218
219 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
220 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
221 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
222 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
223 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
224 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
225
226 if (k > 4 * sizeof(int8_t)) {
227 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
228 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
229
230 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
231 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
232 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
233 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
234 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
235 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
236
237 if (k >= 6 * sizeof(int8_t)) {
238 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
239 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
240
241 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
242 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
243 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
244 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
245 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
246 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
247
248 if (k > 6 * sizeof(int8_t)) {
249 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
250 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
251
252 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
253 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
254 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
255 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
256 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
257 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
258 }
259 }
260 }
261 }
262 }
263 }
264 }
265 p -= 3 * sizeof(void*);
266 } while (p != 0);
267
268 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
269 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
270 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
271 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
272 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
273 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
274
275 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
276 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
277 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
278 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
279 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
280 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
281 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
282
283 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
284 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
285
286 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
287 #if XNN_ARCH_ARM64
288 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
289 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
290 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
291 #else
292 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
293 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
294 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
295 #endif
296 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
297 #if XNN_ARCH_ARM64
298 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
299 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
300 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
301 #else
302 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
303 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
304 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
305 #endif
306
307 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
308 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
309 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
310 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
311 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
312 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
313 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
314
315 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
316 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
317 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
318 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
319 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
320 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
321 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
322
323 if XNN_LIKELY(nc >= 8) {
324 vst1q_f32(&c2[0], vout2x0123);
325 vst1q_f32(&c2[4], vout2x4567);
326 vst1q_f32(&c1[0], vout1x0123);
327 vst1q_f32(&c1[4], vout1x4567);
328 vst1q_f32(&c0[0], vout0x0123);
329 vst1q_f32(&c0[4], vout0x4567);
330
331 a = (const int8_t**restrict) ((uintptr_t) a - ks);
332
333 c0 = (float*) ((uintptr_t) c0 + cn_stride);
334 c1 = (float*) ((uintptr_t) c1 + cn_stride);
335 c2 = (float*) ((uintptr_t) c2 + cn_stride);
336
337 nc -= 8;
338 } else {
339 if (nc & 4) {
340 vst1q_f32(c2, vout2x0123); c2 += 4;
341 vout2x0123 = vout2x4567;
342 vst1q_f32(c1, vout1x0123); c1 += 4;
343 vout1x0123 = vout1x4567;
344 vst1q_f32(c0, vout0x0123); c0 += 4;
345 vout0x0123 = vout0x4567;
346 }
347 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
348 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
349 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
350 if (nc & 2) {
351 vst1_f32(c2, vout2x01); c2 += 2;
352 vst1_f32(c1, vout1x01); c1 += 2;
353 vst1_f32(c0, vout0x01); c0 += 2;
354 vout2x01 = vget_high_f32(vout2x0123);
355 vout1x01 = vget_high_f32(vout1x0123);
356 vout0x01 = vget_high_f32(vout0x0123);
357 }
358 if (nc & 1) {
359 vst1_lane_f32(c2, vout2x01, 0);
360 vst1_lane_f32(c1, vout1x01, 0);
361 vst1_lane_f32(c0, vout0x01, 0);
362 }
363 nc = 0;
364 }
365 } while (nc != 0);
366 }