1 // Auto-generated file. Do not edit!
2 // Template: src/f16-igemm/neonfp16arith-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/igemm.h>
16
17
xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const void ** restrict a,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const void * zero,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const void** restrict a,
24 const void* restrict w,
25 void* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const void* zero,
30 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 8);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(__fp16) == 0);
37 assert(ks != 0);
38 assert(ks % (8 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(__fp16) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 __fp16* c0 = (__fp16*) c;
45 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 __fp16* c4 = (__fp16*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 __fp16* c5 = (__fp16*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 6) {
63 c5 = c4;
64 }
65 __fp16* c6 = (__fp16*) ((uintptr_t) c5 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 6) {
67 c6 = c5;
68 }
69 __fp16* c7 = (__fp16*) ((uintptr_t) c6 + cm_stride);
70 if XNN_UNPREDICTABLE(mr != 8) {
71 c7 = c6;
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 float16x8_t vacc6x01234567 = vacc0x01234567;
88 float16x8_t vacc6x89ABCDEF = vacc0x89ABCDEF;
89 float16x8_t vacc7x01234567 = vacc0x01234567;
90 float16x8_t vacc7x89ABCDEF = vacc0x89ABCDEF;
91
92 size_t p = ks;
93 do {
94 const __fp16* restrict a0 = (const __fp16*) a[0];
95 assert(a0 != NULL);
96 if XNN_UNPREDICTABLE(a0 != zero) {
97 a0 = (const __fp16*) ((uintptr_t) a0 + a_offset);
98 }
99 const __fp16* restrict a1 = (const __fp16*) a[1];
100 assert(a1 != NULL);
101 if XNN_UNPREDICTABLE(a1 != zero) {
102 a1 = (const __fp16*) ((uintptr_t) a1 + a_offset);
103 }
104 const __fp16* restrict a2 = (const __fp16*) a[2];
105 assert(a2 != NULL);
106 if XNN_UNPREDICTABLE(a2 != zero) {
107 a2 = (const __fp16*) ((uintptr_t) a2 + a_offset);
108 }
109 const __fp16* restrict a3 = (const __fp16*) a[3];
110 assert(a3 != NULL);
111 if XNN_UNPREDICTABLE(a3 != zero) {
112 a3 = (const __fp16*) ((uintptr_t) a3 + a_offset);
113 }
114 const __fp16* restrict a4 = (const __fp16*) a[4];
115 assert(a4 != NULL);
116 if XNN_UNPREDICTABLE(a4 != zero) {
117 a4 = (const __fp16*) ((uintptr_t) a4 + a_offset);
118 }
119 const __fp16* restrict a5 = (const __fp16*) a[5];
120 assert(a5 != NULL);
121 if XNN_UNPREDICTABLE(a5 != zero) {
122 a5 = (const __fp16*) ((uintptr_t) a5 + a_offset);
123 }
124 const __fp16* restrict a6 = (const __fp16*) a[6];
125 assert(a6 != NULL);
126 if XNN_UNPREDICTABLE(a6 != zero) {
127 a6 = (const __fp16*) ((uintptr_t) a6 + a_offset);
128 }
129 const __fp16* restrict a7 = (const __fp16*) a[7];
130 assert(a7 != NULL);
131 if XNN_UNPREDICTABLE(a7 != zero) {
132 a7 = (const __fp16*) ((uintptr_t) a7 + a_offset);
133 }
134 a += 8;
135
136 size_t k = kc;
137 for (; k >= 4 * sizeof(__fp16); k -= 4 * sizeof(__fp16)) {
138 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
139 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
140 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
141 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
142 const float16x4_t va4 = vld1_f16(a4); a4 += 4;
143 const float16x4_t va5 = vld1_f16(a5); a5 += 4;
144 const float16x4_t va6 = vld1_f16(a6); a6 += 4;
145 const float16x4_t va7 = vld1_f16(a7); a7 += 4;
146
147 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
148 const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
149
150 #if XNN_ARCH_ARM64
151 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
152 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
153 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
154 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
155 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c0, va4, 0);
156 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c0, va5, 0);
157 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c0, va6, 0);
158 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c0, va7, 0);
159 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
160 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
161 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
162 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
163 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc0, va4, 0);
164 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc0, va5, 0);
165 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc0, va6, 0);
166 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc0, va7, 0);
167 #else
168 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
169 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
170 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
171 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
172 const float16x8_t va4c0 = vdupq_lane_f16(va4, 0);
173 const float16x8_t va5c0 = vdupq_lane_f16(va5, 0);
174 const float16x8_t va6c0 = vdupq_lane_f16(va6, 0);
175 const float16x8_t va7c0 = vdupq_lane_f16(va7, 0);
176
177 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
178 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
179 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
180 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
181 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c0, vb01234567c0);
182 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c0, vb01234567c0);
183 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c0, vb01234567c0);
184 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c0, vb01234567c0);
185 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
186 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c0, vb89ABCDEFc0);
187 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c0, vb89ABCDEFc0);
188 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c0, vb89ABCDEFc0);
189 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c0, vb89ABCDEFc0);
190 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c0, vb89ABCDEFc0);
191 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c0, vb89ABCDEFc0);
192 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c0, vb89ABCDEFc0);
193 #endif
194 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
195 const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
196
197 #if XNN_ARCH_ARM64
198 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
199 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
200 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
201 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
202 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c1, va4, 1);
203 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c1, va5, 1);
204 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c1, va6, 1);
205 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c1, va7, 1);
206 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
207 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
208 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
209 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
210 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc1, va4, 1);
211 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc1, va5, 1);
212 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc1, va6, 1);
213 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc1, va7, 1);
214 #else
215 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
216 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
217 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
218 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
219 const float16x8_t va4c1 = vdupq_lane_f16(va4, 1);
220 const float16x8_t va5c1 = vdupq_lane_f16(va5, 1);
221 const float16x8_t va6c1 = vdupq_lane_f16(va6, 1);
222 const float16x8_t va7c1 = vdupq_lane_f16(va7, 1);
223
224 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
225 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
226 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
227 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
228 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c1, vb01234567c1);
229 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c1, vb01234567c1);
230 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c1, vb01234567c1);
231 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c1, vb01234567c1);
232 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
233 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c1, vb89ABCDEFc1);
234 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c1, vb89ABCDEFc1);
235 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c1, vb89ABCDEFc1);
236 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c1, vb89ABCDEFc1);
237 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c1, vb89ABCDEFc1);
238 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c1, vb89ABCDEFc1);
239 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c1, vb89ABCDEFc1);
240 #endif
241 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
242 const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
243
244 #if XNN_ARCH_ARM64
245 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
246 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
247 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
248 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
249 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c2, va4, 2);
250 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c2, va5, 2);
251 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c2, va6, 2);
252 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c2, va7, 2);
253 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
254 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
255 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
256 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
257 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc2, va4, 2);
258 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc2, va5, 2);
259 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc2, va6, 2);
260 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc2, va7, 2);
261 #else
262 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
263 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
264 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
265 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
266 const float16x8_t va4c2 = vdupq_lane_f16(va4, 2);
267 const float16x8_t va5c2 = vdupq_lane_f16(va5, 2);
268 const float16x8_t va6c2 = vdupq_lane_f16(va6, 2);
269 const float16x8_t va7c2 = vdupq_lane_f16(va7, 2);
270
271 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
272 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
273 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
274 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
275 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c2, vb01234567c2);
276 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c2, vb01234567c2);
277 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c2, vb01234567c2);
278 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c2, vb01234567c2);
279 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
280 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c2, vb89ABCDEFc2);
281 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c2, vb89ABCDEFc2);
282 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c2, vb89ABCDEFc2);
283 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c2, vb89ABCDEFc2);
284 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c2, vb89ABCDEFc2);
285 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c2, vb89ABCDEFc2);
286 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c2, vb89ABCDEFc2);
287 #endif
288 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
289 const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
290
291 #if XNN_ARCH_ARM64
292 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
293 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
294 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
295 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
296 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c3, va4, 3);
297 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c3, va5, 3);
298 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c3, va6, 3);
299 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c3, va7, 3);
300 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
301 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
302 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
303 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
304 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc3, va4, 3);
305 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc3, va5, 3);
306 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc3, va6, 3);
307 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc3, va7, 3);
308 #else
309 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
310 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
311 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
312 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
313 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3);
314 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3);
315 const float16x8_t va6c3 = vdupq_lane_f16(va6, 3);
316 const float16x8_t va7c3 = vdupq_lane_f16(va7, 3);
317
318 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
319 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
320 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
321 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
322 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3);
323 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3);
324 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c3, vb01234567c3);
325 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c3, vb01234567c3);
326 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
327 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c3, vb89ABCDEFc3);
328 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c3, vb89ABCDEFc3);
329 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c3, vb89ABCDEFc3);
330 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3);
331 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3);
332 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c3, vb89ABCDEFc3);
333 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c3, vb89ABCDEFc3);
334 #endif
335 }
336 if XNN_UNLIKELY(k != 0) {
337 do {
338 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
339 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
340 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
341 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
342 const float16x8_t va4 = vld1q_dup_f16(a4); a4 += 1;
343 const float16x8_t va5 = vld1q_dup_f16(a5); a5 += 1;
344 const float16x8_t va6 = vld1q_dup_f16(a6); a6 += 1;
345 const float16x8_t va7 = vld1q_dup_f16(a7); a7 += 1;
346
347 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
348 const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
349
350 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
351 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
352 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
353 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
354 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4, vb01234567);
355 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5, vb01234567);
356 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6, vb01234567);
357 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7, vb01234567);
358 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
359 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1, vb89ABCDEF);
360 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2, vb89ABCDEF);
361 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);
362 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4, vb89ABCDEF);
363 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5, vb89ABCDEF);
364 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6, vb89ABCDEF);
365 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7, vb89ABCDEF);
366
367 k -= sizeof(__fp16);
368 } while (k != 0);
369 }
370 p -= 8 * sizeof(void*);
371 } while (p != 0);
372
373
374 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
375 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
376 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
377 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
378 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
379 vacc4x01234567 = vminq_f16(vacc4x01234567, vmax);
380 vacc5x01234567 = vminq_f16(vacc5x01234567, vmax);
381 vacc6x01234567 = vminq_f16(vacc6x01234567, vmax);
382 vacc7x01234567 = vminq_f16(vacc7x01234567, vmax);
383 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
384 vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
385 vacc2x89ABCDEF = vminq_f16(vacc2x89ABCDEF, vmax);
386 vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);
387 vacc4x89ABCDEF = vminq_f16(vacc4x89ABCDEF, vmax);
388 vacc5x89ABCDEF = vminq_f16(vacc5x89ABCDEF, vmax);
389 vacc6x89ABCDEF = vminq_f16(vacc6x89ABCDEF, vmax);
390 vacc7x89ABCDEF = vminq_f16(vacc7x89ABCDEF, vmax);
391
392 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
393 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
394 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
395 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
396 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
397 vacc4x01234567 = vmaxq_f16(vacc4x01234567, vmin);
398 vacc5x01234567 = vmaxq_f16(vacc5x01234567, vmin);
399 vacc6x01234567 = vmaxq_f16(vacc6x01234567, vmin);
400 vacc7x01234567 = vmaxq_f16(vacc7x01234567, vmin);
401 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
402 vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
403 vacc2x89ABCDEF = vmaxq_f16(vacc2x89ABCDEF, vmin);
404 vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);
405 vacc4x89ABCDEF = vmaxq_f16(vacc4x89ABCDEF, vmin);
406 vacc5x89ABCDEF = vmaxq_f16(vacc5x89ABCDEF, vmin);
407 vacc6x89ABCDEF = vmaxq_f16(vacc6x89ABCDEF, vmin);
408 vacc7x89ABCDEF = vmaxq_f16(vacc7x89ABCDEF, vmin);
409
410 if XNN_LIKELY(nc >= 16) {
411 vst1q_f16(c7, vacc7x01234567);
412 vst1q_f16(c7 + 8, vacc7x89ABCDEF);
413 c7 = (__fp16*) ((uintptr_t) c7 + cn_stride);
414 vst1q_f16(c6, vacc6x01234567);
415 vst1q_f16(c6 + 8, vacc6x89ABCDEF);
416 c6 = (__fp16*) ((uintptr_t) c6 + cn_stride);
417 vst1q_f16(c5, vacc5x01234567);
418 vst1q_f16(c5 + 8, vacc5x89ABCDEF);
419 c5 = (__fp16*) ((uintptr_t) c5 + cn_stride);
420 vst1q_f16(c4, vacc4x01234567);
421 vst1q_f16(c4 + 8, vacc4x89ABCDEF);
422 c4 = (__fp16*) ((uintptr_t) c4 + cn_stride);
423 vst1q_f16(c3, vacc3x01234567);
424 vst1q_f16(c3 + 8, vacc3x89ABCDEF);
425 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
426 vst1q_f16(c2, vacc2x01234567);
427 vst1q_f16(c2 + 8, vacc2x89ABCDEF);
428 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
429 vst1q_f16(c1, vacc1x01234567);
430 vst1q_f16(c1 + 8, vacc1x89ABCDEF);
431 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
432 vst1q_f16(c0, vacc0x01234567);
433 vst1q_f16(c0 + 8, vacc0x89ABCDEF);
434 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
435
436 a = (const void**restrict) ((uintptr_t) a - ks);
437 nc -= 16;
438 } else {
439 if (nc & 8) {
440 vst1q_f16(c7, vacc7x01234567); c7 += 8;
441 vst1q_f16(c6, vacc6x01234567); c6 += 8;
442 vst1q_f16(c5, vacc5x01234567); c5 += 8;
443 vst1q_f16(c4, vacc4x01234567); c4 += 8;
444 vst1q_f16(c3, vacc3x01234567); c3 += 8;
445 vst1q_f16(c2, vacc2x01234567); c2 += 8;
446 vst1q_f16(c1, vacc1x01234567); c1 += 8;
447 vst1q_f16(c0, vacc0x01234567); c0 += 8;
448
449 vacc7x01234567 = vacc7x89ABCDEF;
450 vacc6x01234567 = vacc6x89ABCDEF;
451 vacc5x01234567 = vacc5x89ABCDEF;
452 vacc4x01234567 = vacc4x89ABCDEF;
453 vacc3x01234567 = vacc3x89ABCDEF;
454 vacc2x01234567 = vacc2x89ABCDEF;
455 vacc1x01234567 = vacc1x89ABCDEF;
456 vacc0x01234567 = vacc0x89ABCDEF;
457 }
458 float16x4_t vacc7x0123 = vget_low_f16(vacc7x01234567);
459 float16x4_t vacc6x0123 = vget_low_f16(vacc6x01234567);
460 float16x4_t vacc5x0123 = vget_low_f16(vacc5x01234567);
461 float16x4_t vacc4x0123 = vget_low_f16(vacc4x01234567);
462 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
463 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
464 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
465 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
466 if (nc & 4) {
467 vst1_f16(c7, vacc7x0123); c7 += 4;
468 vst1_f16(c6, vacc6x0123); c6 += 4;
469 vst1_f16(c5, vacc5x0123); c5 += 4;
470 vst1_f16(c4, vacc4x0123); c4 += 4;
471 vst1_f16(c3, vacc3x0123); c3 += 4;
472 vst1_f16(c2, vacc2x0123); c2 += 4;
473 vst1_f16(c1, vacc1x0123); c1 += 4;
474 vst1_f16(c0, vacc0x0123); c0 += 4;
475
476 vacc7x0123 = vget_high_f16(vacc7x01234567);
477 vacc6x0123 = vget_high_f16(vacc6x01234567);
478 vacc5x0123 = vget_high_f16(vacc5x01234567);
479 vacc4x0123 = vget_high_f16(vacc4x01234567);
480 vacc3x0123 = vget_high_f16(vacc3x01234567);
481 vacc2x0123 = vget_high_f16(vacc2x01234567);
482 vacc1x0123 = vget_high_f16(vacc1x01234567);
483 vacc0x0123 = vget_high_f16(vacc0x01234567);
484 }
485 if (nc & 2) {
486 vst1_lane_u32((void*) c7, vreinterpret_u32_f16(vacc7x0123), 0); c7 += 2;
487 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vacc6x0123), 0); c6 += 2;
488 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vacc5x0123), 0); c5 += 2;
489 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vacc4x0123), 0); c4 += 2;
490 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
491 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
492 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
493 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
494
495 vacc7x0123 = vext_f16(vacc7x0123, vacc7x0123, 2);
496 vacc6x0123 = vext_f16(vacc6x0123, vacc6x0123, 2);
497 vacc5x0123 = vext_f16(vacc5x0123, vacc5x0123, 2);
498 vacc4x0123 = vext_f16(vacc4x0123, vacc4x0123, 2);
499 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
500 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
501 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
502 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
503 }
504 if (nc & 1) {
505 vst1_lane_f16(c7, vacc7x0123, 0);
506 vst1_lane_f16(c6, vacc6x0123, 0);
507 vst1_lane_f16(c5, vacc5x0123, 0);
508 vst1_lane_f16(c4, vacc4x0123, 0);
509 vst1_lane_f16(c3, vacc3x0123, 0);
510 vst1_lane_f16(c2, vacc2x0123, 0);
511 vst1_lane_f16(c1, vacc1x0123, 0);
512 vst1_lane_f16(c0, vacc0x0123, 0);
513 }
514
515 nc = 0;
516 }
517 } while (nc != 0);
518 }
519