1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gemm/neonfp16arith-ld64.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
17 #include <xnnpack/gemm.h>
18
19
xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const void* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 void* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 6);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(__fp16) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const __fp16* a0 = (const __fp16*) a;
42 __fp16* c0 = (__fp16*) c;
43 const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
44 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
50 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
56 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const __fp16* a4 = (const __fp16*) ((uintptr_t) a3 + a_stride);
62 __fp16* c4 = (__fp16*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const __fp16* a5 = (const __fp16*) ((uintptr_t) a4 + a_stride);
68 __fp16* c5 = (__fp16*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr != 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73
74 do {
75 float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
76 float16x8_t vacc0x89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
77 float16x8_t vacc1x01234567 = vacc0x01234567;
78 float16x8_t vacc1x89ABCDEF = vacc0x89ABCDEF;
79 float16x8_t vacc2x01234567 = vacc0x01234567;
80 float16x8_t vacc2x89ABCDEF = vacc0x89ABCDEF;
81 float16x8_t vacc3x01234567 = vacc0x01234567;
82 float16x8_t vacc3x89ABCDEF = vacc0x89ABCDEF;
83 float16x8_t vacc4x01234567 = vacc0x01234567;
84 float16x8_t vacc4x89ABCDEF = vacc0x89ABCDEF;
85 float16x8_t vacc5x01234567 = vacc0x01234567;
86 float16x8_t vacc5x89ABCDEF = vacc0x89ABCDEF;
87
88 size_t k = kc;
89 while (k >= 4 * sizeof(__fp16)) {
90 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
91 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
92 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
93 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
94 const float16x4_t va4 = vld1_f16(a4); a4 += 4;
95 const float16x4_t va5 = vld1_f16(a5); a5 += 4;
96
97 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
98 const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
99
100 #if XNN_ARCH_ARM64
101 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
102 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
103 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
104 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
105 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c0, va4, 0);
106 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c0, va5, 0);
107 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
108 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
109 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
110 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
111 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc0, va4, 0);
112 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc0, va5, 0);
113 #else
114 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
115 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
116 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
117 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
118 const float16x8_t va4c0 = vdupq_lane_f16(va4, 0);
119 const float16x8_t va5c0 = vdupq_lane_f16(va5, 0);
120
121 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
122 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
123 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
124 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
125 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c0, vb01234567c0);
126 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c0, vb01234567c0);
127 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
128 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c0, vb89ABCDEFc0);
129 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c0, vb89ABCDEFc0);
130 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c0, vb89ABCDEFc0);
131 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c0, vb89ABCDEFc0);
132 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c0, vb89ABCDEFc0);
133 #endif
134 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
135 const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
136
137 #if XNN_ARCH_ARM64
138 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
139 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
140 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
141 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
142 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c1, va4, 1);
143 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c1, va5, 1);
144 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
145 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
146 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
147 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
148 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc1, va4, 1);
149 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc1, va5, 1);
150 #else
151 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
152 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
153 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
154 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
155 const float16x8_t va4c1 = vdupq_lane_f16(va4, 1);
156 const float16x8_t va5c1 = vdupq_lane_f16(va5, 1);
157
158 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
159 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
160 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
161 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
162 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c1, vb01234567c1);
163 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c1, vb01234567c1);
164 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
165 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c1, vb89ABCDEFc1);
166 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c1, vb89ABCDEFc1);
167 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c1, vb89ABCDEFc1);
168 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c1, vb89ABCDEFc1);
169 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c1, vb89ABCDEFc1);
170 #endif
171 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
172 const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
173
174 #if XNN_ARCH_ARM64
175 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
176 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
177 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
178 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
179 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c2, va4, 2);
180 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c2, va5, 2);
181 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
182 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
183 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
184 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
185 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc2, va4, 2);
186 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc2, va5, 2);
187 #else
188 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
189 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
190 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
191 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
192 const float16x8_t va4c2 = vdupq_lane_f16(va4, 2);
193 const float16x8_t va5c2 = vdupq_lane_f16(va5, 2);
194
195 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
196 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
197 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
198 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
199 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c2, vb01234567c2);
200 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c2, vb01234567c2);
201 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
202 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c2, vb89ABCDEFc2);
203 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c2, vb89ABCDEFc2);
204 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c2, vb89ABCDEFc2);
205 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c2, vb89ABCDEFc2);
206 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c2, vb89ABCDEFc2);
207 #endif
208 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
209 const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
210
211 #if XNN_ARCH_ARM64
212 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
213 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
214 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
215 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
216 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c3, va4, 3);
217 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c3, va5, 3);
218 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
219 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
220 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
221 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
222 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc3, va4, 3);
223 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc3, va5, 3);
224 #else
225 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
226 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
227 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
228 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
229 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3);
230 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3);
231
232 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
233 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
234 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
235 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
236 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3);
237 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3);
238 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
239 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c3, vb89ABCDEFc3);
240 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c3, vb89ABCDEFc3);
241 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c3, vb89ABCDEFc3);
242 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3);
243 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3);
244 #endif
245
246 k -= 4 * sizeof(__fp16);
247 }
248 if XNN_UNLIKELY(k != 0) {
249 do {
250 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
251 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
252 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
253 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
254 const float16x8_t va4 = vld1q_dup_f16(a4); a4 += 1;
255 const float16x8_t va5 = vld1q_dup_f16(a5); a5 += 1;
256
257 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
258 const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
259
260 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
261 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
262 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
263 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
264 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4, vb01234567);
265 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5, vb01234567);
266 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
267 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1, vb89ABCDEF);
268 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2, vb89ABCDEF);
269 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);
270 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4, vb89ABCDEF);
271 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5, vb89ABCDEF);
272
273 k -= sizeof(__fp16);
274 } while (k != 0);
275 }
276
277
278 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
279 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
280 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
281 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
282 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
283 vacc4x01234567 = vminq_f16(vacc4x01234567, vmax);
284 vacc5x01234567 = vminq_f16(vacc5x01234567, vmax);
285 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
286 vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
287 vacc2x89ABCDEF = vminq_f16(vacc2x89ABCDEF, vmax);
288 vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);
289 vacc4x89ABCDEF = vminq_f16(vacc4x89ABCDEF, vmax);
290 vacc5x89ABCDEF = vminq_f16(vacc5x89ABCDEF, vmax);
291
292 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
293 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
294 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
295 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
296 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
297 vacc4x01234567 = vmaxq_f16(vacc4x01234567, vmin);
298 vacc5x01234567 = vmaxq_f16(vacc5x01234567, vmin);
299 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
300 vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
301 vacc2x89ABCDEF = vmaxq_f16(vacc2x89ABCDEF, vmin);
302 vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);
303 vacc4x89ABCDEF = vmaxq_f16(vacc4x89ABCDEF, vmin);
304 vacc5x89ABCDEF = vmaxq_f16(vacc5x89ABCDEF, vmin);
305
306 if XNN_LIKELY(nc >= 16) {
307 vst1q_f16(c0, vacc0x01234567);
308 vst1q_f16(c0 + 8, vacc0x89ABCDEF);
309 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
310 vst1q_f16(c1, vacc1x01234567);
311 vst1q_f16(c1 + 8, vacc1x89ABCDEF);
312 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
313 vst1q_f16(c2, vacc2x01234567);
314 vst1q_f16(c2 + 8, vacc2x89ABCDEF);
315 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
316 vst1q_f16(c3, vacc3x01234567);
317 vst1q_f16(c3 + 8, vacc3x89ABCDEF);
318 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
319 vst1q_f16(c4, vacc4x01234567);
320 vst1q_f16(c4 + 8, vacc4x89ABCDEF);
321 c4 = (__fp16*) ((uintptr_t) c4 + cn_stride);
322 vst1q_f16(c5, vacc5x01234567);
323 vst1q_f16(c5 + 8, vacc5x89ABCDEF);
324 c5 = (__fp16*) ((uintptr_t) c5 + cn_stride);
325
326 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
327 a1 = (const __fp16*) ((uintptr_t) a1 - kc);
328 a2 = (const __fp16*) ((uintptr_t) a2 - kc);
329 a3 = (const __fp16*) ((uintptr_t) a3 - kc);
330 a4 = (const __fp16*) ((uintptr_t) a4 - kc);
331 a5 = (const __fp16*) ((uintptr_t) a5 - kc);
332
333 nc -= 16;
334 } else {
335 if (nc & 8) {
336 vst1q_f16(c0, vacc0x01234567); c0 += 8;
337 vst1q_f16(c1, vacc1x01234567); c1 += 8;
338 vst1q_f16(c2, vacc2x01234567); c2 += 8;
339 vst1q_f16(c3, vacc3x01234567); c3 += 8;
340 vst1q_f16(c4, vacc4x01234567); c4 += 8;
341 vst1q_f16(c5, vacc5x01234567); c5 += 8;
342
343 vacc0x01234567 = vacc0x89ABCDEF;
344 vacc1x01234567 = vacc1x89ABCDEF;
345 vacc2x01234567 = vacc2x89ABCDEF;
346 vacc3x01234567 = vacc3x89ABCDEF;
347 vacc4x01234567 = vacc4x89ABCDEF;
348 vacc5x01234567 = vacc5x89ABCDEF;
349 }
350 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
351 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
352 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
353 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
354 float16x4_t vacc4x0123 = vget_low_f16(vacc4x01234567);
355 float16x4_t vacc5x0123 = vget_low_f16(vacc5x01234567);
356 if (nc & 4) {
357 vst1_f16(c0, vacc0x0123); c0 += 4;
358 vst1_f16(c1, vacc1x0123); c1 += 4;
359 vst1_f16(c2, vacc2x0123); c2 += 4;
360 vst1_f16(c3, vacc3x0123); c3 += 4;
361 vst1_f16(c4, vacc4x0123); c4 += 4;
362 vst1_f16(c5, vacc5x0123); c5 += 4;
363
364 vacc0x0123 = vget_high_f16(vacc0x01234567);
365 vacc1x0123 = vget_high_f16(vacc1x01234567);
366 vacc2x0123 = vget_high_f16(vacc2x01234567);
367 vacc3x0123 = vget_high_f16(vacc3x01234567);
368 vacc4x0123 = vget_high_f16(vacc4x01234567);
369 vacc5x0123 = vget_high_f16(vacc5x01234567);
370 }
371 if (nc & 2) {
372 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
373 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
374 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
375 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
376 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vacc4x0123), 0); c4 += 2;
377 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vacc5x0123), 0); c5 += 2;
378
379 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
380 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
381 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
382 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
383 vacc4x0123 = vext_f16(vacc4x0123, vacc4x0123, 2);
384 vacc5x0123 = vext_f16(vacc5x0123, vacc5x0123, 2);
385 }
386 if (nc & 1) {
387 vst1_lane_f16(c0, vacc0x0123, 0);
388 vst1_lane_f16(c1, vacc1x0123, 0);
389 vst1_lane_f16(c2, vacc2x0123, 0);
390 vst1_lane_f16(c3, vacc3x0123, 0);
391 vst1_lane_f16(c4, vacc4x0123, 0);
392 vst1_lane_f16(c5, vacc5x0123, 0);
393 }
394
395 nc = 0;
396 }
397 } while (nc != 0);
398 }
399