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_gemminc_minmax_ukernel_8x16__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 void * restrict acc,const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemminc_minmax_ukernel_8x16__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 void*restrict acc,
31 const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
32 {
33 assert(mr != 0);
34 assert(mr <= 8);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(__fp16) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41 assert(acc != NULL);
42
43 const __fp16* a0 = (const __fp16*) a;
44 __fp16* c0 = (__fp16*) c;
45 const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
46 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
52 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
58 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr < 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63 const __fp16* a4 = (const __fp16*) ((uintptr_t) a3 + a_stride);
64 __fp16* c4 = (__fp16*) ((uintptr_t) c3 + cm_stride);
65 if XNN_UNPREDICTABLE(mr <= 4) {
66 a4 = a3;
67 c4 = c3;
68 }
69 const __fp16* a5 = (const __fp16*) ((uintptr_t) a4 + a_stride);
70 __fp16* c5 = (__fp16*) ((uintptr_t) c4 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 6) {
72 a5 = a4;
73 c5 = c4;
74 }
75 const __fp16* a6 = (const __fp16*) ((uintptr_t) a5 + a_stride);
76 __fp16* c6 = (__fp16*) ((uintptr_t) c5 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 6) {
78 a6 = a5;
79 c6 = c5;
80 }
81 const __fp16* a7 = (const __fp16*) ((uintptr_t) a6 + a_stride);
82 __fp16* c7 = (__fp16*) ((uintptr_t) c6 + cm_stride);
83 if XNN_UNPREDICTABLE(mr != 8) {
84 a7 = a6;
85 c7 = c6;
86 }
87
88 do {
89 float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
90 float16x8_t vacc0x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
91 float16x8_t vacc1x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
92 float16x8_t vacc1x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
93 float16x8_t vacc2x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
94 float16x8_t vacc2x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
95 float16x8_t vacc3x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
96 float16x8_t vacc3x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
97 float16x8_t vacc4x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
98 float16x8_t vacc4x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
99 float16x8_t vacc5x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
100 float16x8_t vacc5x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
101 float16x8_t vacc6x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
102 float16x8_t vacc6x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
103 float16x8_t vacc7x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
104 float16x8_t vacc7x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
105
106 size_t k = kc;
107 while (k >= 4 * sizeof(__fp16)) {
108 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
109 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
110 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
111 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
112 const float16x4_t va4 = vld1_f16(a4); a4 += 4;
113 const float16x4_t va5 = vld1_f16(a5); a5 += 4;
114 const float16x4_t va6 = vld1_f16(a6); a6 += 4;
115 const float16x4_t va7 = vld1_f16(a7); a7 += 4;
116
117 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
118 const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
119
120 #if XNN_ARCH_ARM64
121 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
122 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
123 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
124 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
125 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c0, va4, 0);
126 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c0, va5, 0);
127 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c0, va6, 0);
128 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c0, va7, 0);
129 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
130 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
131 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
132 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
133 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc0, va4, 0);
134 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc0, va5, 0);
135 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc0, va6, 0);
136 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc0, va7, 0);
137 #else
138 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
139 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
140 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
141 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
142 const float16x8_t va4c0 = vdupq_lane_f16(va4, 0);
143 const float16x8_t va5c0 = vdupq_lane_f16(va5, 0);
144 const float16x8_t va6c0 = vdupq_lane_f16(va6, 0);
145 const float16x8_t va7c0 = vdupq_lane_f16(va7, 0);
146
147 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
148 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
149 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
150 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
151 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c0, vb01234567c0);
152 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c0, vb01234567c0);
153 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c0, vb01234567c0);
154 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c0, vb01234567c0);
155 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
156 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c0, vb89ABCDEFc0);
157 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c0, vb89ABCDEFc0);
158 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c0, vb89ABCDEFc0);
159 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c0, vb89ABCDEFc0);
160 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c0, vb89ABCDEFc0);
161 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c0, vb89ABCDEFc0);
162 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c0, vb89ABCDEFc0);
163 #endif
164 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
165 const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
166
167 #if XNN_ARCH_ARM64
168 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
169 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
170 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
171 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
172 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c1, va4, 1);
173 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c1, va5, 1);
174 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c1, va6, 1);
175 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c1, va7, 1);
176 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
177 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
178 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
179 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
180 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc1, va4, 1);
181 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc1, va5, 1);
182 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc1, va6, 1);
183 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc1, va7, 1);
184 #else
185 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
186 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
187 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
188 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
189 const float16x8_t va4c1 = vdupq_lane_f16(va4, 1);
190 const float16x8_t va5c1 = vdupq_lane_f16(va5, 1);
191 const float16x8_t va6c1 = vdupq_lane_f16(va6, 1);
192 const float16x8_t va7c1 = vdupq_lane_f16(va7, 1);
193
194 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
195 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
196 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
197 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
198 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c1, vb01234567c1);
199 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c1, vb01234567c1);
200 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c1, vb01234567c1);
201 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c1, vb01234567c1);
202 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
203 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c1, vb89ABCDEFc1);
204 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c1, vb89ABCDEFc1);
205 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c1, vb89ABCDEFc1);
206 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c1, vb89ABCDEFc1);
207 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c1, vb89ABCDEFc1);
208 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c1, vb89ABCDEFc1);
209 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c1, vb89ABCDEFc1);
210 #endif
211 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
212 const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
213
214 #if XNN_ARCH_ARM64
215 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
216 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
217 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
218 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
219 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c2, va4, 2);
220 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c2, va5, 2);
221 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c2, va6, 2);
222 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c2, va7, 2);
223 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
224 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
225 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
226 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
227 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc2, va4, 2);
228 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc2, va5, 2);
229 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc2, va6, 2);
230 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc2, va7, 2);
231 #else
232 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
233 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
234 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
235 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
236 const float16x8_t va4c2 = vdupq_lane_f16(va4, 2);
237 const float16x8_t va5c2 = vdupq_lane_f16(va5, 2);
238 const float16x8_t va6c2 = vdupq_lane_f16(va6, 2);
239 const float16x8_t va7c2 = vdupq_lane_f16(va7, 2);
240
241 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
242 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
243 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
244 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
245 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c2, vb01234567c2);
246 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c2, vb01234567c2);
247 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c2, vb01234567c2);
248 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c2, vb01234567c2);
249 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
250 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c2, vb89ABCDEFc2);
251 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c2, vb89ABCDEFc2);
252 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c2, vb89ABCDEFc2);
253 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c2, vb89ABCDEFc2);
254 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c2, vb89ABCDEFc2);
255 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c2, vb89ABCDEFc2);
256 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c2, vb89ABCDEFc2);
257 #endif
258 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
259 const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
260
261 #if XNN_ARCH_ARM64
262 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
263 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
264 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
265 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
266 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c3, va4, 3);
267 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c3, va5, 3);
268 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c3, va6, 3);
269 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c3, va7, 3);
270 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
271 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
272 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
273 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
274 vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc3, va4, 3);
275 vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc3, va5, 3);
276 vacc6x89ABCDEF = vfmaq_lane_f16(vacc6x89ABCDEF, vb89ABCDEFc3, va6, 3);
277 vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc3, va7, 3);
278 #else
279 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
280 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
281 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
282 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
283 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3);
284 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3);
285 const float16x8_t va6c3 = vdupq_lane_f16(va6, 3);
286 const float16x8_t va7c3 = vdupq_lane_f16(va7, 3);
287
288 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
289 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
290 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
291 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
292 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3);
293 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3);
294 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c3, vb01234567c3);
295 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c3, vb01234567c3);
296 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
297 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c3, vb89ABCDEFc3);
298 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c3, vb89ABCDEFc3);
299 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c3, vb89ABCDEFc3);
300 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3);
301 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3);
302 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6c3, vb89ABCDEFc3);
303 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7c3, vb89ABCDEFc3);
304 #endif
305
306 k -= 4 * sizeof(__fp16);
307 }
308 if XNN_UNLIKELY(k != 0) {
309 do {
310 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
311 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
312 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
313 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
314 const float16x8_t va4 = vld1q_dup_f16(a4); a4 += 1;
315 const float16x8_t va5 = vld1q_dup_f16(a5); a5 += 1;
316 const float16x8_t va6 = vld1q_dup_f16(a6); a6 += 1;
317 const float16x8_t va7 = vld1q_dup_f16(a7); a7 += 1;
318
319 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
320 const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
321
322 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
323 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
324 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
325 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
326 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4, vb01234567);
327 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5, vb01234567);
328 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6, vb01234567);
329 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7, vb01234567);
330 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
331 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1, vb89ABCDEF);
332 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2, vb89ABCDEF);
333 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);
334 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4, vb89ABCDEF);
335 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5, vb89ABCDEF);
336 vacc6x89ABCDEF = vfmaq_f16(vacc6x89ABCDEF, va6, vb89ABCDEF);
337 vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7, vb89ABCDEF);
338
339 k -= sizeof(__fp16);
340 } while (k != 0);
341 }
342
343 const float16x8_t vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
344 vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
345 vacc1x01234567 = vmulq_f16(vacc1x01234567, vscale);
346 vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);
347 vacc3x01234567 = vmulq_f16(vacc3x01234567, vscale);
348 vacc4x01234567 = vmulq_f16(vacc4x01234567, vscale);
349 vacc5x01234567 = vmulq_f16(vacc5x01234567, vscale);
350 vacc6x01234567 = vmulq_f16(vacc6x01234567, vscale);
351 vacc7x01234567 = vmulq_f16(vacc7x01234567, vscale);
352 vacc0x89ABCDEF = vmulq_f16(vacc0x89ABCDEF, vscale);
353 vacc1x89ABCDEF = vmulq_f16(vacc1x89ABCDEF, vscale);
354 vacc2x89ABCDEF = vmulq_f16(vacc2x89ABCDEF, vscale);
355 vacc3x89ABCDEF = vmulq_f16(vacc3x89ABCDEF, vscale);
356 vacc4x89ABCDEF = vmulq_f16(vacc4x89ABCDEF, vscale);
357 vacc5x89ABCDEF = vmulq_f16(vacc5x89ABCDEF, vscale);
358 vacc6x89ABCDEF = vmulq_f16(vacc6x89ABCDEF, vscale);
359 vacc7x89ABCDEF = vmulq_f16(vacc7x89ABCDEF, vscale);
360
361 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
362 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
363 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
364 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
365 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
366 vacc4x01234567 = vminq_f16(vacc4x01234567, vmax);
367 vacc5x01234567 = vminq_f16(vacc5x01234567, vmax);
368 vacc6x01234567 = vminq_f16(vacc6x01234567, vmax);
369 vacc7x01234567 = vminq_f16(vacc7x01234567, vmax);
370 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
371 vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
372 vacc2x89ABCDEF = vminq_f16(vacc2x89ABCDEF, vmax);
373 vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);
374 vacc4x89ABCDEF = vminq_f16(vacc4x89ABCDEF, vmax);
375 vacc5x89ABCDEF = vminq_f16(vacc5x89ABCDEF, vmax);
376 vacc6x89ABCDEF = vminq_f16(vacc6x89ABCDEF, vmax);
377 vacc7x89ABCDEF = vminq_f16(vacc7x89ABCDEF, vmax);
378
379 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
380 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
381 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
382 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
383 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
384 vacc4x01234567 = vmaxq_f16(vacc4x01234567, vmin);
385 vacc5x01234567 = vmaxq_f16(vacc5x01234567, vmin);
386 vacc6x01234567 = vmaxq_f16(vacc6x01234567, vmin);
387 vacc7x01234567 = vmaxq_f16(vacc7x01234567, vmin);
388 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
389 vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
390 vacc2x89ABCDEF = vmaxq_f16(vacc2x89ABCDEF, vmin);
391 vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);
392 vacc4x89ABCDEF = vmaxq_f16(vacc4x89ABCDEF, vmin);
393 vacc5x89ABCDEF = vmaxq_f16(vacc5x89ABCDEF, vmin);
394 vacc6x89ABCDEF = vmaxq_f16(vacc6x89ABCDEF, vmin);
395 vacc7x89ABCDEF = vmaxq_f16(vacc7x89ABCDEF, vmin);
396
397 if XNN_LIKELY(nc >= 16) {
398 vst1q_f16(c0, vacc0x01234567);
399 vst1q_f16(c0 + 8, vacc0x89ABCDEF);
400 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
401 vst1q_f16(c1, vacc1x01234567);
402 vst1q_f16(c1 + 8, vacc1x89ABCDEF);
403 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
404 vst1q_f16(c2, vacc2x01234567);
405 vst1q_f16(c2 + 8, vacc2x89ABCDEF);
406 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
407 vst1q_f16(c3, vacc3x01234567);
408 vst1q_f16(c3 + 8, vacc3x89ABCDEF);
409 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
410 vst1q_f16(c4, vacc4x01234567);
411 vst1q_f16(c4 + 8, vacc4x89ABCDEF);
412 c4 = (__fp16*) ((uintptr_t) c4 + cn_stride);
413 vst1q_f16(c5, vacc5x01234567);
414 vst1q_f16(c5 + 8, vacc5x89ABCDEF);
415 c5 = (__fp16*) ((uintptr_t) c5 + cn_stride);
416 vst1q_f16(c6, vacc6x01234567);
417 vst1q_f16(c6 + 8, vacc6x89ABCDEF);
418 c6 = (__fp16*) ((uintptr_t) c6 + cn_stride);
419 vst1q_f16(c7, vacc7x01234567);
420 vst1q_f16(c7 + 8, vacc7x89ABCDEF);
421 c7 = (__fp16*) ((uintptr_t) c7 + cn_stride);
422
423 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
424 a1 = (const __fp16*) ((uintptr_t) a1 - kc);
425 a2 = (const __fp16*) ((uintptr_t) a2 - kc);
426 a3 = (const __fp16*) ((uintptr_t) a3 - kc);
427 a4 = (const __fp16*) ((uintptr_t) a4 - kc);
428 a5 = (const __fp16*) ((uintptr_t) a5 - kc);
429 a6 = (const __fp16*) ((uintptr_t) a6 - kc);
430 a7 = (const __fp16*) ((uintptr_t) a7 - kc);
431
432 nc -= 16;
433 } else {
434 if (nc & 8) {
435 vst1q_f16(c0, vacc0x01234567); c0 += 8;
436 vst1q_f16(c1, vacc1x01234567); c1 += 8;
437 vst1q_f16(c2, vacc2x01234567); c2 += 8;
438 vst1q_f16(c3, vacc3x01234567); c3 += 8;
439 vst1q_f16(c4, vacc4x01234567); c4 += 8;
440 vst1q_f16(c5, vacc5x01234567); c5 += 8;
441 vst1q_f16(c6, vacc6x01234567); c6 += 8;
442 vst1q_f16(c7, vacc7x01234567); c7 += 8;
443
444 vacc0x01234567 = vacc0x89ABCDEF;
445 vacc1x01234567 = vacc1x89ABCDEF;
446 vacc2x01234567 = vacc2x89ABCDEF;
447 vacc3x01234567 = vacc3x89ABCDEF;
448 vacc4x01234567 = vacc4x89ABCDEF;
449 vacc5x01234567 = vacc5x89ABCDEF;
450 vacc6x01234567 = vacc6x89ABCDEF;
451 vacc7x01234567 = vacc7x89ABCDEF;
452 }
453 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
454 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
455 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
456 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
457 float16x4_t vacc4x0123 = vget_low_f16(vacc4x01234567);
458 float16x4_t vacc5x0123 = vget_low_f16(vacc5x01234567);
459 float16x4_t vacc6x0123 = vget_low_f16(vacc6x01234567);
460 float16x4_t vacc7x0123 = vget_low_f16(vacc7x01234567);
461 if (nc & 4) {
462 vst1_f16(c0, vacc0x0123); c0 += 4;
463 vst1_f16(c1, vacc1x0123); c1 += 4;
464 vst1_f16(c2, vacc2x0123); c2 += 4;
465 vst1_f16(c3, vacc3x0123); c3 += 4;
466 vst1_f16(c4, vacc4x0123); c4 += 4;
467 vst1_f16(c5, vacc5x0123); c5 += 4;
468 vst1_f16(c6, vacc6x0123); c6 += 4;
469 vst1_f16(c7, vacc7x0123); c7 += 4;
470
471 vacc0x0123 = vget_high_f16(vacc0x01234567);
472 vacc1x0123 = vget_high_f16(vacc1x01234567);
473 vacc2x0123 = vget_high_f16(vacc2x01234567);
474 vacc3x0123 = vget_high_f16(vacc3x01234567);
475 vacc4x0123 = vget_high_f16(vacc4x01234567);
476 vacc5x0123 = vget_high_f16(vacc5x01234567);
477 vacc6x0123 = vget_high_f16(vacc6x01234567);
478 vacc7x0123 = vget_high_f16(vacc7x01234567);
479 }
480 if (nc & 2) {
481 vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
482 vst1_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
483 vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
484 vst1_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
485 vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_f16(vacc4x0123), 0); c4 += 2;
486 vst1_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpret_u32_f16(vacc5x0123), 0); c5 += 2;
487 vst1_lane_u32(__builtin_assume_aligned(c6, 1), vreinterpret_u32_f16(vacc6x0123), 0); c6 += 2;
488 vst1_lane_u32(__builtin_assume_aligned(c7, 1), vreinterpret_u32_f16(vacc7x0123), 0); c7 += 2;
489
490 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
491 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
492 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
493 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
494 vacc4x0123 = vext_f16(vacc4x0123, vacc4x0123, 2);
495 vacc5x0123 = vext_f16(vacc5x0123, vacc5x0123, 2);
496 vacc6x0123 = vext_f16(vacc6x0123, vacc6x0123, 2);
497 vacc7x0123 = vext_f16(vacc7x0123, vacc7x0123, 2);
498 }
499 if (nc & 1) {
500 vst1_lane_f16(c0, vacc0x0123, 0);
501 vst1_lane_f16(c1, vacc1x0123, 0);
502 vst1_lane_f16(c2, vacc2x0123, 0);
503 vst1_lane_f16(c3, vacc3x0123, 0);
504 vst1_lane_f16(c4, vacc4x0123, 0);
505 vst1_lane_f16(c5, vacc5x0123, 0);
506 vst1_lane_f16(c6, vacc6x0123, 0);
507 vst1_lane_f16(c7, vacc7x0123, 0);
508 }
509
510 nc = 0;
511 }
512 } while (nc != 0);
513 }
514