Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8-minmax-neon-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/prefetch.h"
17
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x8__neon_mlal_lane_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
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 vzp01 = vld1q_s32(&quantization_params[0].zero_point);
60 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vzp01), 0);
61 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vzp01), 0);
62 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vzp01), 0);
63 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vzp01), 0);
64 const int32x4_t vzp2 = vld1q_dup_s32(&quantization_params[2].zero_point);
65 int32x4_t vacc2x0123 = vmulq_s32(vksum0123, vzp2);
66 int32x4_t vacc2x4567 = vmulq_s32(vksum4567, vzp2);
67
68 size_t k = kc;
69 while (k >= 8 * sizeof(int8_t)) {
70 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
71 const int16x8_t vxa0 = vmovl_s8(va0);
72 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
73 const int16x8_t vxa1 = vmovl_s8(va1);
74 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
75 const int16x8_t vxa2 = vmovl_s8(va2);
76
77 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
85
86 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
87 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
88 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
89 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
90 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
91 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
92 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
93 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
94
95 xnn_prefetch_to_l1((const int8_t*) w + 448);
96
97 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
98 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
99 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
100 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
101 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
102 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
103 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
104 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
105 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
106 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
107 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
108 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
110 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
111 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
112 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
113 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
114 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
115 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
116 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
117 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
118 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
119 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
120 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
121
122 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
123 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
124 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
125 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
126 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
127 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
128 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
129 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
130 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
131 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
132 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
133 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
134 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
135 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
136 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
137 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
138 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
139 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
140 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
141 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
142 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
143 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
144 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
145 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
146
147 k -= 8 * sizeof(int8_t);
148 }
149 if XNN_UNLIKELY(k != 0) {
150 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
151 const int16x8_t vxa0 = vmovl_s8(va0);
152 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
153 const int16x8_t vxa1 = vmovl_s8(va1);
154 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
155 const int16x8_t vxa2 = vmovl_s8(va2);
156
157 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
158
159 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
160
161 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
162 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
163 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
164 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
165 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
166 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
167
168 if (k >= 2 * sizeof(int8_t)) {
169 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
171
172 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
173 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
174 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
175 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
176 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
177 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
178
179 if (k > 2 * sizeof(int8_t)) {
180 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
182
183 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
184 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
185 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
186 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
187 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
188 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
189
190 if (k >= 4 * sizeof(int8_t)) {
191 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
192 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
193
194 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
195 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
196 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
197 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
198 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
199 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
200
201 if (k > 4 * sizeof(int8_t)) {
202 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
203 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
204
205 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
206 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
207 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
208 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
209 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
210 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
211
212 if (k >= 6 * sizeof(int8_t)) {
213 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
214 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
215
216 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
217 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
218 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
219 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
220 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
221 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
222
223 if (k > 6 * sizeof(int8_t)) {
224 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
225 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
226
227 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
228 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
229 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
230 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
231 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
232 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
233 }
234 }
235 }
236 }
237 }
238 }
239 }
240
241 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
242 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
243 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
244 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
245 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
246 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
247
248 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
249 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
250 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
251 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
252 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
253 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
254 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
255 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
256
257 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
258 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
259
260 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
261 #if XNN_ARCH_ARM64
262 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
263 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
264 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
265 #else
266 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
267 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
268 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
269 #endif
270 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
271 #if XNN_ARCH_ARM64
272 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
273 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
274 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
275 #else
276 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
277 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
278 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
279 #endif
280
281 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
282 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
283 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
284 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
285 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
286 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
287 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
288
289 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
290 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
291 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
292 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
293 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
294 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
295 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
296
297 if XNN_LIKELY(nc >= 8) {
298 vst1q_f32(&c0[0], vout0x0123);
299 vst1q_f32(&c0[4], vout0x4567);
300 vst1q_f32(&c1[0], vout1x0123);
301 vst1q_f32(&c1[4], vout1x4567);
302 vst1q_f32(&c2[0], vout2x0123);
303 vst1q_f32(&c2[4], vout2x4567);
304
305 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
306 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
307 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
308
309 c0 = (float*) ((uintptr_t) c0 + cn_stride);
310 c1 = (float*) ((uintptr_t) c1 + cn_stride);
311 c2 = (float*) ((uintptr_t) c2 + cn_stride);
312
313 nc -= 8;
314 } else {
315 if (nc & 4) {
316 vst1q_f32(c0, vout0x0123); c0 += 4;
317 vout0x0123 = vout0x4567;
318 vst1q_f32(c1, vout1x0123); c1 += 4;
319 vout1x0123 = vout1x4567;
320 vst1q_f32(c2, vout2x0123); c2 += 4;
321 vout2x0123 = vout2x4567;
322 }
323 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
324 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
325 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
326 if (nc & 2) {
327 vst1_f32(c0, vout0x01); c0 += 2;
328 vst1_f32(c1, vout1x01); c1 += 2;
329 vst1_f32(c2, vout2x01); c2 += 2;
330 vout0x01 = vget_high_f32(vout0x0123);
331 vout1x01 = vget_high_f32(vout1x0123);
332 vout2x01 = vget_high_f32(vout2x0123);
333 }
334 if (nc & 1) {
335 vst1_lane_f32(c0, vout0x01, 0);
336 vst1_lane_f32(c1, vout1x01, 0);
337 vst1_lane_f32(c2, vout2x01, 0);
338 }
339 nc = 0;
340 }
341 } while (nc != 0);
342 }