Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x16-minmax-neonfp16arith-mlal-lane.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/math.h"
17
18
19 void xnn_qd8_f16_qb4w_gemm_minmax_ukernel_1x16__neonfp16arith_mlal_lane(
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 xnn_float16* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const struct xnn_f16_qb4w_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 <= 1);
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 size_t bl = params->scalar.blocksize;
42 assert(bl <= round_up_po2(kc, 2));
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47
48 const int8x8_t vmask = vmov_n_s8(INT8_C(0xF0));
49 kc = round_up_po2(kc, 2);
50 do {
51 float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
52 float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
53 float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
54 float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
55 const float32x4_t vzp0 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[0].zero_point));
56 float32x4_t vout0x0123 = vmulq_f32(vksum0123, vzp0);
57 float32x4_t vout0x4567 = vmulq_f32(vksum4567, vzp0);
58 float32x4_t vout0x89AB = vmulq_f32(vksum89AB, vzp0);
59 float32x4_t vout0xCDEF = vmulq_f32(vksumCDEF, vzp0);
60
61 for (size_t kb=0; kb < kc; kb += bl) {
62 int32x4_t vacc0x0123 = vdupq_n_s32(0);
63 int32x4_t vacc0x4567 = vdupq_n_s32(0);
64 int32x4_t vacc0x89AB = vdupq_n_s32(0);
65 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
66
67 size_t k = bl;
68 while (k >= 8 * sizeof(int8_t)) {
69 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
70 const int16x8_t vxa0 = vmovl_s8(va0);
71
72 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const uint8_t*) w + 8;
73 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const uint8_t*) w + 8;
74 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const uint8_t*) w + 8;
75 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const uint8_t*) w + 8;
76 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const uint8_t*) w + 8;
77 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const uint8_t*) w + 8;
78 const int8x8_t vb01234567c67 = vld1_s8(w); w = (const uint8_t*) w + 8;
79 const int8x8_t vb89ABCDEFc67 = vld1_s8(w); w = (const uint8_t*) w + 8;
80 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
81 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
82 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
83 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
84 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
85 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
86 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
87 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
88 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
89 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
90 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
91 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
92 const int8x8_t vb01234567c6 = vshl_n_s8(vb01234567c67, 4);
93 const int8x8_t vb01234567c7 = vand_s8(vb01234567c67, vmask);
94 const int8x8_t vb89ABCDEFc6 = vshl_n_s8(vb89ABCDEFc67, 4);
95 const int8x8_t vb89ABCDEFc7 = vand_s8(vb89ABCDEFc67, vmask);
96
97 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
98 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
99 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
100 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
101 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
102 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
103 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
104 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
105 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
106 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
107 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
108 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
109 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
110 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
111 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
112 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
113
114
115 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
116 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
117 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
118 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
121 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
122 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
123 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
124 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
125 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
126 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
127 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
128 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
129 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
130 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
131
132 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
133 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
134 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
135 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
136 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
137 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
138 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
139 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
140 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
141 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
142 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
143 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
144 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
145 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
146 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
147 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
148
149 k -= 8 * sizeof(int8_t);
150 }
151 if XNN_UNLIKELY(k != 0) {
152 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
153 const int16x8_t vxa0 = vmovl_s8(va0);
154
155 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
158 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
159 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
160 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
161
162 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
163 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
164
165 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
166 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
167 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
168 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
169
170 if (k >= 2 * sizeof(int8_t)) {
171 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
172 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
173
174 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
175 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
176 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
177 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
178
179 if (k > 2 * sizeof(int8_t)) {
180 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const int8_t*) w + 8;
182 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
183 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
184 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
185 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
186 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
187 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
188
189 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
190 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
191 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
192 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
193
194 if (k >= 4 * sizeof(int8_t)) {
195 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
196 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
197
198 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
199 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
200 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
201 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
202
203 if (k > 4 * sizeof(int8_t)) {
204 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const int8_t*) w + 8;
205 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const int8_t*) w + 8;
206 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
207 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
208 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
209 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
210 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
211 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
212
213 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
214 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
215 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
216 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
217
218 if (k >= 6 * sizeof(int8_t)) {
219 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
220 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
221
222 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
223 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
224 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
225 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
226
227 }
228 }
229 }
230 }
231 }
232 }
233
234 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
235 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
236 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
237 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
238
239 const float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
240 const float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
241 const float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
242 const float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
243
244 #if XNN_ARCH_ARM64
245 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
246 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
247 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
248 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
249 #else
250 vout0x0123 = vmlaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
251 vout0x4567 = vmlaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
252 vout0x89AB = vmlaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
253 vout0xCDEF = vmlaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
254 #endif
255 }
256 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
257
258 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
259
260 #if XNN_ARCH_ARM64
261 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vinput_scale0);
262 #else
263 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vinput_scale0);
264 #endif
265 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
266
267 #if XNN_ARCH_ARM64
268 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vinput_scale0);
269 #else
270 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vinput_scale0);
271 #endif
272 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
273
274 #if XNN_ARCH_ARM64
275 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vinput_scale0);
276 #else
277 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vinput_scale0);
278 #endif
279 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
280
281 #if XNN_ARCH_ARM64
282 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vinput_scale0);
283 #else
284 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vinput_scale0);
285 #endif
286
287 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
288 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
289 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
290 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
291 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
292 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
293 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
294 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
295
296 if XNN_LIKELY(nc >= 16) {
297 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
298 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
299
300 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
301
302 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
303
304 nc -= 16;
305 } else {
306 if (nc & 8) {
307 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
308 vfp16out0x01234567 = vfp16out0x89ABCDEF;
309 }
310 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
311 if (nc & 4) {
312 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
313 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
314 }
315 if (nc & 2) {
316 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
317 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
318 }
319 if (nc & 1) {
320 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
321 }
322 nc = 0;
323 }
324 } while (nc != 0);
325 }