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
17
xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const uint8_t* a0 = a;
40 uint8_t* c0 = c;
41
42 const uint8x8_t vb_zero_point = vld1_dup_u8(¶ms->fp32_neon.kernel_zero_point[0]);
43 do {
44 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
45 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
46 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
47 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
48
49 size_t k = kc;
50 while (k >= 8 * sizeof(uint8_t)) {
51 const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
52 const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
53
54 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
55 const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
56
57 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
58 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
59 const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
60 const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
61
62 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
63 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
64 const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
65 const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
66
67 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
68 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
69 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
70 const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
71
72 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
73 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
74 const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
75 const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
76
77 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
78 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
79 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
80 const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
81
82 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
83 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
84 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
85 const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
86
87 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
88 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
89 const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
90 const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
91
92 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
93 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
94
95
96 const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
97 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
98
99 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
100 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
101 const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
102 const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
103
104 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
105 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
106 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
107 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
108
109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
110 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
111 const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
112 const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
113
114 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
115 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
116 const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
117 const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
118
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
121 const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
122 const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
123
124 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
125 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
126 const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
127 const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
128
129 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
130 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
131 const uint8x8_t vb89ABCDEFc7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
132 const int16x8_t vxb89ABCDEFc7 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc7, vb_zero_point));
133
134 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
135 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
136
137 k -= 8 * sizeof(uint8_t);
138 }
139 if XNN_UNLIKELY(k != 0) {
140 const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
141 const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
142
143 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
144 const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
145 const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
146 const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
147
148 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
149 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
150 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
151 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
152
153 if (k >= 2 * sizeof(uint8_t)) {
154 const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
155 const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
156 const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
157 const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
158
159 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
160 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
161 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
162 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
163
164 if (k > 2 * sizeof(uint8_t)) {
165 const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
166 const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
167 const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
168 const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
169
170 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
171 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
172 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
173 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
174
175 if (k >= 4 * sizeof(uint8_t)) {
176 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
177 const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
178 const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
179 const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
180
181 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
182 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
183 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
184 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
185
186 if (k > 4 * sizeof(uint8_t)) {
187 const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
188 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
189 const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
190 const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
191
192 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
193 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
194 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
195 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
196
197 if (k >= 6 * sizeof(uint8_t)) {
198 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
199 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
200 const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
201 const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
202
203 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
204 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
205 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
206 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
207
208 if (k > 6 * sizeof(uint8_t)) {
209 const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
210 const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
211 const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
212 const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
213
214 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
215 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
216 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
217 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
218 }
219 }
220 }
221 }
222 }
223 }
224 }
225
226 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
227 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
228 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
229 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
230
231 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
232 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
233 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
234 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
235 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
236
237 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
238 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
239 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
240 vacc0x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x89AB, vmagic_bias));
241 vacc0xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc0xCDEF, vmagic_bias));
242
243 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
244 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
245 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
246 vacc0x89AB = vqsubq_s32(vacc0x89AB, vmagic_bias_less_output_zero_point);
247 vacc0xCDEF = vqsubq_s32(vacc0xCDEF, vmagic_bias_less_output_zero_point);
248
249 #if XNN_ARCH_ARM64
250 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
251 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
252
253
254 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
255 #else
256 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
257 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
258
259
260 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
261 #endif
262
263 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neon.output_min);
264 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
265
266 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neon.output_max);
267 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
268
269 if (nc >= 16) {
270 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
271
272 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
273
274 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
275
276 nc -= 16;
277 } else {
278 uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
279 if (nc & 8) {
280 vst1_u8(c0, vout0x01234567); c0 += 8;
281 vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
282 }
283 if (nc & 4) {
284 vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
285 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
286 }
287 if (nc & 2) {
288 vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
289 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
290 }
291 if (nc & 1) {
292 vst1_lane_u8(c0, vout0x01234567, 0);
293 }
294
295 nc = 0;
296 }
297 } while (nc != 0);
298 }
299