24 #ifdef ARM_COMPUTE_ENABLE_SVE
27 #include "../../asmlib.hpp"
31 void sve_interleaved_fp32_mmla_8x3VL(
const float *Apanel,
const float *Bpanel,
float *Cpanel,
int ablocks,
int bblocks,
int K) {
32 const float *a_ptr = Apanel;
33 float *c_ptr = Cpanel;
36 const long loops_count = (
K / 2) - 1;
37 const long tails_count =
K % 2;
39 for (
int yb=0; yb<ablocks; yb++) {
40 const float *a_ptr0 = a_ptr;
41 const float *b_ptr = Bpanel;
43 for (
int xb=0; xb<bblocks; xb++) {
45 long loops = loops_count;
46 long tails = tails_count;
54 "ld1rqw z0.s, p0/z, [%[a_ptr]]\n"
56 "ld1w z4.s, p0/z, [%[b_ptr]]\n"
58 "ld1rqw z1.s, p0/z, [%[a_ptr], #0x10]\n"
60 "ld1w z5.s, p0/z, [%[b_ptr], #1, MUL VL]\n"
62 "ld1rqw z2.s, p0/z, [%[a_ptr], #0x20]\n"
64 "ld1w z6.s, p0/z, [%[b_ptr], #2, MUL VL]\n"
66 "add %[a_ptr], %[a_ptr], #0x40\n"
68 "addvl %[b_ptr], %[b_ptr], #4\n"
84 ".inst 0x64a4e408 // fmmla z8.s, z0.s, z4.s\n"
85 "ld1w z7.s, p0/z, [%[b_ptr], #-1, MUL VL]\n"
86 ".inst 0x64a4e42e // fmmla z14.s, z1.s, z4.s\n"
87 "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n"
88 ".inst 0x64a4e454 // fmmla z20.s, z2.s, z4.s\n"
89 "subs %[loops], %[loops], #0x1\n"
90 ".inst 0x64a5e409 // fmmla z9.s, z0.s, z5.s\n"
91 ".inst 0x64a4e47a // fmmla z26.s, z3.s, z4.s\n"
92 "ld1w z4.s, p0/z, [%[b_ptr]]\n"
93 ".inst 0x64a5e42f // fmmla z15.s, z1.s, z5.s\n"
94 ".inst 0x64a5e455 // fmmla z21.s, z2.s, z5.s\n"
95 ".inst 0x64a5e47b // fmmla z27.s, z3.s, z5.s\n"
96 "ld1w z5.s, p0/z, [%[b_ptr], #1, MUL VL]\n"
97 ".inst 0x64a6e40a // fmmla z10.s, z0.s, z6.s\n"
98 ".inst 0x64a6e430 // fmmla z16.s, z1.s, z6.s\n"
99 ".inst 0x64a6e456 // fmmla z22.s, z2.s, z6.s\n"
100 ".inst 0x64a6e47c // fmmla z28.s, z3.s, z6.s\n"
101 "ld1w z6.s, p0/z, [%[b_ptr], #2, MUL VL]\n"
102 ".inst 0x64a7e40b // fmmla z11.s, z0.s, z7.s\n"
103 ".inst 0x64a7e431 // fmmla z17.s, z1.s, z7.s\n"
104 ".inst 0x64a7e457 // fmmla z23.s, z2.s, z7.s\n"
105 ".inst 0x64a7e47d // fmmla z29.s, z3.s, z7.s\n"
106 "ld1w z7.s, p0/z, [%[b_ptr], #3, MUL VL]\n"
107 ".inst 0x64a4e40c // fmmla z12.s, z0.s, z4.s\n"
108 ".inst 0x64a4e432 // fmmla z18.s, z1.s, z4.s\n"
109 ".inst 0x64a4e458 // fmmla z24.s, z2.s, z4.s\n"
110 ".inst 0x64a4e47e // fmmla z30.s, z3.s, z4.s\n"
111 "ld1w z4.s, p0/z, [%[b_ptr], #4, MUL VL]\n"
112 ".inst 0x64a5e40d // fmmla z13.s, z0.s, z5.s\n"
113 "ld1rqw z0.s, p0/z, [%[a_ptr]]\n"
114 ".inst 0x64a5e433 // fmmla z19.s, z1.s, z5.s\n"
115 "ld1rqw z1.s, p0/z, [%[a_ptr], #0x10]\n"
116 ".inst 0x64a5e459 // fmmla z25.s, z2.s, z5.s\n"
117 "ld1rqw z2.s, p0/z, [%[a_ptr], #0x20]\n"
118 ".inst 0x64a5e47f // fmmla z31.s, z3.s, z5.s\n"
119 "ld1w z5.s, p0/z, [%[b_ptr], #5, MUL VL]\n"
120 ".inst 0x64a6e408 // fmmla z8.s, z0.s, z6.s\n"
121 "ld1rqw z3.s, p0/z, [%[a_ptr], #0x30]\n"
122 ".inst 0x64a6e42e // fmmla z14.s, z1.s, z6.s\n"
123 "add %[a_ptr], %[a_ptr], #0x80\n"
124 ".inst 0x64a6e454 // fmmla z20.s, z2.s, z6.s\n"
125 "addvl %[b_ptr], %[b_ptr], #12\n"
126 ".inst 0x64a6e47a // fmmla z26.s, z3.s, z6.s\n"
127 ".inst 0x64a7e409 // fmmla z9.s, z0.s, z7.s\n"
128 ".inst 0x64a7e42f // fmmla z15.s, z1.s, z7.s\n"
129 "ld1w z6.s, p0/z, [%[b_ptr], #-6, MUL VL]\n"
130 ".inst 0x64a7e455 // fmmla z21.s, z2.s, z7.s\n"
131 ".inst 0x64a7e47b // fmmla z27.s, z3.s, z7.s\n"
132 "ld1w z7.s, p0/z, [%[b_ptr], #-5, MUL VL]\n"
133 ".inst 0x64a4e40a // fmmla z10.s, z0.s, z4.s\n"
134 ".inst 0x64a4e430 // fmmla z16.s, z1.s, z4.s\n"
135 ".inst 0x64a4e456 // fmmla z22.s, z2.s, z4.s\n"
136 ".inst 0x64a4e47c // fmmla z28.s, z3.s, z4.s\n"
137 "ld1w z4.s, p0/z, [%[b_ptr], #-4, MUL VL]\n"
138 ".inst 0x64a5e40b // fmmla z11.s, z0.s, z5.s\n"
139 ".inst 0x64a5e431 // fmmla z17.s, z1.s, z5.s\n"
140 ".inst 0x64a5e457 // fmmla z23.s, z2.s, z5.s\n"
141 ".inst 0x64a5e47d // fmmla z29.s, z3.s, z5.s\n"
142 "ld1w z5.s, p0/z, [%[b_ptr], #-3, MUL VL]\n"
143 ".inst 0x64a6e40c // fmmla z12.s, z0.s, z6.s\n"
144 ".inst 0x64a6e432 // fmmla z18.s, z1.s, z6.s\n"
145 ".inst 0x64a6e458 // fmmla z24.s, z2.s, z6.s\n"
146 ".inst 0x64a6e47e // fmmla z30.s, z3.s, z6.s\n"
147 "ld1w z6.s, p0/z, [%[b_ptr], #-2, MUL VL]\n"
148 ".inst 0x64a7e40d // fmmla z13.s, z0.s, z7.s\n"
149 "ld1rqw z0.s, p0/z, [%[a_ptr], #-0x40]\n"
150 ".inst 0x64a7e433 // fmmla z19.s, z1.s, z7.s\n"
151 "ld1rqw z1.s, p0/z, [%[a_ptr], #-0x30]\n"
152 ".inst 0x64a7e459 // fmmla z25.s, z2.s, z7.s\n"
153 "ld1rqw z2.s, p0/z, [%[a_ptr], #-0x20]\n"
154 ".inst 0x64a7e47f // fmmla z31.s, z3.s, z7.s\n"
158 ".inst 0x64a4e408 // fmmla z8.s, z0.s, z4.s\n"
159 "ld1w z7.s, p0/z, [%[b_ptr], #-1, MUL VL]\n"
160 ".inst 0x64a4e42e // fmmla z14.s, z1.s, z4.s\n"
161 "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n"
162 ".inst 0x64a4e454 // fmmla z20.s, z2.s, z4.s\n"
163 ".inst 0x64a5e409 // fmmla z9.s, z0.s, z5.s\n"
164 ".inst 0x64a5e42f // fmmla z15.s, z1.s, z5.s\n"
165 ".inst 0x64a4e47a // fmmla z26.s, z3.s, z4.s\n"
166 "ld1w z4.s, p0/z, [%[b_ptr]]\n"
167 ".inst 0x64a5e455 // fmmla z21.s, z2.s, z5.s\n"
168 ".inst 0x64a5e47b // fmmla z27.s, z3.s, z5.s\n"
169 "ld1w z5.s, p0/z, [%[b_ptr], #1, MUL VL]\n"
170 ".inst 0x64a6e40a // fmmla z10.s, z0.s, z6.s\n"
171 ".inst 0x64a6e430 // fmmla z16.s, z1.s, z6.s\n"
172 ".inst 0x64a6e456 // fmmla z22.s, z2.s, z6.s\n"
173 ".inst 0x64a6e47c // fmmla z28.s, z3.s, z6.s\n"
174 "ld1w z6.s, p0/z, [%[b_ptr], #2, MUL VL]\n"
175 ".inst 0x64a7e40b // fmmla z11.s, z0.s, z7.s\n"
176 ".inst 0x64a7e431 // fmmla z17.s, z1.s, z7.s\n"
177 ".inst 0x64a7e457 // fmmla z23.s, z2.s, z7.s\n"
178 ".inst 0x64a7e47d // fmmla z29.s, z3.s, z7.s\n"
179 "ld1w z7.s, p0/z, [%[b_ptr], #3, MUL VL]\n"
180 ".inst 0x64a4e40c // fmmla z12.s, z0.s, z4.s\n"
181 ".inst 0x64a4e432 // fmmla z18.s, z1.s, z4.s\n"
182 ".inst 0x64a4e458 // fmmla z24.s, z2.s, z4.s\n"
183 ".inst 0x64a4e47e // fmmla z30.s, z3.s, z4.s\n"
184 "ld1w z4.s, p0/z, [%[b_ptr], #4, MUL VL]\n"
185 ".inst 0x64a5e40d // fmmla z13.s, z0.s, z5.s\n"
186 "ld1rqw z0.s, p0/z, [%[a_ptr]]\n"
187 ".inst 0x64a5e433 // fmmla z19.s, z1.s, z5.s\n"
188 "ld1rqw z1.s, p0/z, [%[a_ptr], #0x10]\n"
189 ".inst 0x64a5e459 // fmmla z25.s, z2.s, z5.s\n"
190 "ld1rqw z2.s, p0/z, [%[a_ptr], #0x20]\n"
191 ".inst 0x64a5e47f // fmmla z31.s, z3.s, z5.s\n"
192 "ld1w z5.s, p0/z, [%[b_ptr], #5, MUL VL]\n"
193 ".inst 0x64a6e408 // fmmla z8.s, z0.s, z6.s\n"
194 "ld1rqw z3.s, p0/z, [%[a_ptr], #0x30]\n"
195 ".inst 0x64a6e42e // fmmla z14.s, z1.s, z6.s\n"
196 "add %[a_ptr], %[a_ptr], #0x80\n"
197 ".inst 0x64a6e454 // fmmla z20.s, z2.s, z6.s\n"
198 "addvl %[b_ptr], %[b_ptr], #14\n"
199 ".inst 0x64a6e47a // fmmla z26.s, z3.s, z6.s\n"
200 ".inst 0x64a7e409 // fmmla z9.s, z0.s, z7.s\n"
201 ".inst 0x64a7e42f // fmmla z15.s, z1.s, z7.s\n"
202 "ld1w z6.s, p0/z, [%[b_ptr], #-8, MUL VL]\n"
203 ".inst 0x64a7e455 // fmmla z21.s, z2.s, z7.s\n"
204 ".inst 0x64a7e47b // fmmla z27.s, z3.s, z7.s\n"
205 "ld1w z7.s, p0/z, [%[b_ptr], #-7, MUL VL]\n"
206 ".inst 0x64a4e40a // fmmla z10.s, z0.s, z4.s\n"
207 ".inst 0x64a4e430 // fmmla z16.s, z1.s, z4.s\n"
208 ".inst 0x64a4e456 // fmmla z22.s, z2.s, z4.s\n"
209 ".inst 0x64a4e47c // fmmla z28.s, z3.s, z4.s\n"
210 "ld1w z4.s, p0/z, [%[b_ptr], #-6, MUL VL]\n"
211 ".inst 0x64a5e40b // fmmla z11.s, z0.s, z5.s\n"
212 ".inst 0x64a5e431 // fmmla z17.s, z1.s, z5.s\n"
213 ".inst 0x64a5e457 // fmmla z23.s, z2.s, z5.s\n"
214 ".inst 0x64a5e47d // fmmla z29.s, z3.s, z5.s\n"
215 "ld1w z5.s, p0/z, [%[b_ptr], #-5, MUL VL]\n"
216 ".inst 0x64a6e40c // fmmla z12.s, z0.s, z6.s\n"
217 ".inst 0x64a6e432 // fmmla z18.s, z1.s, z6.s\n"
218 ".inst 0x64a6e458 // fmmla z24.s, z2.s, z6.s\n"
219 ".inst 0x64a6e47e // fmmla z30.s, z3.s, z6.s\n"
220 "ld1w z6.s, p0/z, [%[b_ptr], #-4, MUL VL]\n"
221 ".inst 0x64a7e40d // fmmla z13.s, z0.s, z7.s\n"
222 "ld1rqw z0.s, p0/z, [%[a_ptr], #-0x40]\n"
223 ".inst 0x64a7e433 // fmmla z19.s, z1.s, z7.s\n"
224 "ld1rqw z1.s, p0/z, [%[a_ptr], #-0x30]\n"
225 ".inst 0x64a7e459 // fmmla z25.s, z2.s, z7.s\n"
226 "ld1rqw z2.s, p0/z, [%[a_ptr], #-0x20]\n"
227 ".inst 0x64a7e47f // fmmla z31.s, z3.s, z7.s\n"
228 "ld1w z7.s, p0/z, [%[b_ptr], #-3, MUL VL]\n"
229 ".inst 0x64a4e408 // fmmla z8.s, z0.s, z4.s\n"
230 "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n"
231 ".inst 0x64a4e42e // fmmla z14.s, z1.s, z4.s\n"
232 ".inst 0x64a4e454 // fmmla z20.s, z2.s, z4.s\n"
233 ".inst 0x64a5e409 // fmmla z9.s, z0.s, z5.s\n"
234 ".inst 0x64a4e47a // fmmla z26.s, z3.s, z4.s\n"
235 "ld1w z4.s, p0/z, [%[b_ptr], #-2, MUL VL]\n"
236 ".inst 0x64a5e42f // fmmla z15.s, z1.s, z5.s\n"
237 ".inst 0x64a5e455 // fmmla z21.s, z2.s, z5.s\n"
238 ".inst 0x64a5e47b // fmmla z27.s, z3.s, z5.s\n"
239 "ld1w z5.s, p0/z, [%[b_ptr], #-1, MUL VL]\n"
240 ".inst 0x64a6e40a // fmmla z10.s, z0.s, z6.s\n"
241 ".inst 0x64a6e430 // fmmla z16.s, z1.s, z6.s\n"
242 ".inst 0x64a6e456 // fmmla z22.s, z2.s, z6.s\n"
243 ".inst 0x64a6e47c // fmmla z28.s, z3.s, z6.s\n"
244 "uzp1 z6.d, z14.d, z15.d\n"
245 ".inst 0x64a7e40b // fmmla z11.s, z0.s, z7.s\n"
246 ".inst 0x64a7e431 // fmmla z17.s, z1.s, z7.s\n"
247 ".inst 0x64a7e457 // fmmla z23.s, z2.s, z7.s\n"
248 ".inst 0x64a7e47d // fmmla z29.s, z3.s, z7.s\n"
249 ".inst 0x64a4e40c // fmmla z12.s, z0.s, z4.s\n"
250 "uzp1 z7.d, z16.d, z17.d\n"
251 ".inst 0x64a4e432 // fmmla z18.s, z1.s, z4.s\n"
252 ".inst 0x64a4e458 // fmmla z24.s, z2.s, z4.s\n"
253 ".inst 0x64a4e47e // fmmla z30.s, z3.s, z4.s\n"
254 "uzp2 z4.d, z10.d, z11.d\n"
255 ".inst 0x64a5e40d // fmmla z13.s, z0.s, z5.s\n"
256 "uzp1 z0.d, z8.d, z9.d\n"
257 ".inst 0x64a5e433 // fmmla z19.s, z1.s, z5.s\n"
258 "uzp1 z1.d, z10.d, z11.d\n"
259 ".inst 0x64a5e459 // fmmla z25.s, z2.s, z5.s\n"
260 "st1w z0.s, p0, [%[c_ptr]]\n"
261 "uzp1 z2.d, z12.d, z13.d\n"
262 "uzp1 z0.d, z18.d, z19.d\n"
263 ".inst 0x64a5e47f // fmmla z31.s, z3.s, z5.s\n"
264 "st1w z1.s, p0, [%[c_ptr], #1, MUL VL]\n"
265 "uzp2 z3.d, z8.d, z9.d\n"
266 "uzp2 z5.d, z12.d, z13.d\n"
267 "uzp2 z1.d, z14.d, z15.d\n"
268 "st1w z2.s, p0, [%[c_ptr], #2, MUL VL]\n"
271 ".inst 0x64a4e408 // fmmla z8.s, z0.s, z4.s\n"
272 "ld1w z7.s, p0/z, [%[b_ptr], #-1, MUL VL]\n"
273 ".inst 0x64a4e42e // fmmla z14.s, z1.s, z4.s\n"
274 "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n"
275 ".inst 0x64a4e454 // fmmla z20.s, z2.s, z4.s\n"
276 "add %[a_ptr], %[a_ptr], #0x40\n"
277 ".inst 0x64a5e409 // fmmla z9.s, z0.s, z5.s\n"
278 "addvl %[b_ptr], %[b_ptr], #8\n"
279 ".inst 0x64a4e47a // fmmla z26.s, z3.s, z4.s\n"
280 ".inst 0x64a5e42f // fmmla z15.s, z1.s, z5.s\n"
281 ".inst 0x64a5e455 // fmmla z21.s, z2.s, z5.s\n"
282 "ld1w z4.s, p0/z, [%[b_ptr], #-8, MUL VL]\n"
283 ".inst 0x64a5e47b // fmmla z27.s, z3.s, z5.s\n"
284 "ld1w z5.s, p0/z, [%[b_ptr], #-7, MUL VL]\n"
285 ".inst 0x64a6e40a // fmmla z10.s, z0.s, z6.s\n"
286 ".inst 0x64a6e430 // fmmla z16.s, z1.s, z6.s\n"
287 ".inst 0x64a6e456 // fmmla z22.s, z2.s, z6.s\n"
288 ".inst 0x64a6e47c // fmmla z28.s, z3.s, z6.s\n"
289 "ld1w z6.s, p0/z, [%[b_ptr], #-6, MUL VL]\n"
290 ".inst 0x64a7e40b // fmmla z11.s, z0.s, z7.s\n"
291 ".inst 0x64a7e431 // fmmla z17.s, z1.s, z7.s\n"
292 ".inst 0x64a7e457 // fmmla z23.s, z2.s, z7.s\n"
293 ".inst 0x64a7e47d // fmmla z29.s, z3.s, z7.s\n"
294 "ld1w z7.s, p0/z, [%[b_ptr], #-5, MUL VL]\n"
295 ".inst 0x64a4e40c // fmmla z12.s, z0.s, z4.s\n"
296 ".inst 0x64a4e432 // fmmla z18.s, z1.s, z4.s\n"
297 ".inst 0x64a4e458 // fmmla z24.s, z2.s, z4.s\n"
298 ".inst 0x64a4e47e // fmmla z30.s, z3.s, z4.s\n"
299 "ld1w z4.s, p0/z, [%[b_ptr], #-4, MUL VL]\n"
300 ".inst 0x64a5e40d // fmmla z13.s, z0.s, z5.s\n"
301 "ld1rqw z0.s, p0/z, [%[a_ptr], #-0x40]\n"
302 ".inst 0x64a5e433 // fmmla z19.s, z1.s, z5.s\n"
303 "ld1rqw z1.s, p0/z, [%[a_ptr], #-0x30]\n"
304 ".inst 0x64a5e459 // fmmla z25.s, z2.s, z5.s\n"
305 "ld1rqw z2.s, p0/z, [%[a_ptr], #-0x20]\n"
306 ".inst 0x64a5e47f // fmmla z31.s, z3.s, z5.s\n"
307 "ld1w z5.s, p0/z, [%[b_ptr], #-3, MUL VL]\n"
308 ".inst 0x64a6e408 // fmmla z8.s, z0.s, z6.s\n"
309 "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n"
310 ".inst 0x64a6e42e // fmmla z14.s, z1.s, z6.s\n"
311 ".inst 0x64a6e454 // fmmla z20.s, z2.s, z6.s\n"
312 ".inst 0x64a7e409 // fmmla z9.s, z0.s, z7.s\n"
313 ".inst 0x64a6e47a // fmmla z26.s, z3.s, z6.s\n"
314 "ld1w z6.s, p0/z, [%[b_ptr], #-2, MUL VL]\n"
315 ".inst 0x64a7e42f // fmmla z15.s, z1.s, z7.s\n"
316 ".inst 0x64a7e455 // fmmla z21.s, z2.s, z7.s\n"
317 ".inst 0x64a7e47b // fmmla z27.s, z3.s, z7.s\n"
318 "ld1w z7.s, p0/z, [%[b_ptr], #-1, MUL VL]\n"
319 ".inst 0x64a4e40a // fmmla z10.s, z0.s, z4.s\n"
320 ".inst 0x64a4e430 // fmmla z16.s, z1.s, z4.s\n"
321 ".inst 0x64a4e456 // fmmla z22.s, z2.s, z4.s\n"
322 ".inst 0x64a4e47c // fmmla z28.s, z3.s, z4.s\n"
323 ".inst 0x64a5e40b // fmmla z11.s, z0.s, z5.s\n"
324 ".inst 0x64a5e431 // fmmla z17.s, z1.s, z5.s\n"
325 ".inst 0x64a5e457 // fmmla z23.s, z2.s, z5.s\n"
326 ".inst 0x64a5e47d // fmmla z29.s, z3.s, z5.s\n"
327 "uzp2 z4.d, z10.d, z11.d\n"
328 ".inst 0x64a6e40c // fmmla z12.s, z0.s, z6.s\n"
329 ".inst 0x64a6e432 // fmmla z18.s, z1.s, z6.s\n"
330 ".inst 0x64a6e458 // fmmla z24.s, z2.s, z6.s\n"
331 ".inst 0x64a6e47e // fmmla z30.s, z3.s, z6.s\n"
332 "uzp1 z6.d, z14.d, z15.d\n"
333 ".inst 0x64a7e40d // fmmla z13.s, z0.s, z7.s\n"
334 "uzp1 z0.d, z8.d, z9.d\n"
335 ".inst 0x64a7e433 // fmmla z19.s, z1.s, z7.s\n"
336 "uzp1 z1.d, z10.d, z11.d\n"
337 "uzp2 z5.d, z12.d, z13.d\n"
338 "st1w z0.s, p0, [%[c_ptr]]\n"
339 ".inst 0x64a7e459 // fmmla z25.s, z2.s, z7.s\n"
340 "uzp1 z2.d, z12.d, z13.d\n"
341 "uzp1 z0.d, z18.d, z19.d\n"
342 "st1w z1.s, p0, [%[c_ptr], #1, MUL VL]\n"
343 "uzp2 z1.d, z14.d, z15.d\n"
344 ".inst 0x64a7e47f // fmmla z31.s, z3.s, z7.s\n"
345 "uzp2 z3.d, z8.d, z9.d\n"
346 "st1w z2.s, p0, [%[c_ptr], #2, MUL VL]\n"
347 "uzp1 z7.d, z16.d, z17.d\n"
349 "uzp2 z2.d, z16.d, z17.d\n"
350 "st1w z3.s, p0, [%[c_ptr], #3, MUL VL]\n"
351 "uzp2 z3.d, z18.d, z19.d\n"
352 "st1w z4.s, p0, [%[c_ptr], #4, MUL VL]\n"
353 "uzp1 z4.d, z20.d, z21.d\n"
354 "st1w z5.s, p0, [%[c_ptr], #5, MUL VL]\n"
355 "uzp1 z5.d, z22.d, z23.d\n"
356 "st1w z6.s, p0, [%[c_ptr], #6, MUL VL]\n"
357 "uzp1 z6.d, z24.d, z25.d\n"
358 "st1w z7.s, p0, [%[c_ptr], #7, MUL VL]\n"
359 "addvl %[c_ptr], %[c_ptr], #16\n"
360 "uzp2 z7.d, z20.d, z21.d\n"
361 "st1w z0.s, p0, [%[c_ptr], #-8, MUL VL]\n"
362 "uzp2 z0.d, z22.d, z23.d\n"
363 "st1w z1.s, p0, [%[c_ptr], #-7, MUL VL]\n"
364 "uzp2 z1.d, z24.d, z25.d\n"
365 "st1w z2.s, p0, [%[c_ptr], #-6, MUL VL]\n"
366 "uzp1 z2.d, z26.d, z27.d\n"
367 "st1w z3.s, p0, [%[c_ptr], #-5, MUL VL]\n"
368 "uzp1 z3.d, z28.d, z29.d\n"
369 "st1w z4.s, p0, [%[c_ptr], #-4, MUL VL]\n"
370 "uzp1 z4.d, z30.d, z31.d\n"
371 "st1w z5.s, p0, [%[c_ptr], #-3, MUL VL]\n"
372 "uzp2 z5.d, z26.d, z27.d\n"
373 "st1w z6.s, p0, [%[c_ptr], #-2, MUL VL]\n"
374 "uzp2 z6.d, z28.d, z29.d\n"
375 "st1w z7.s, p0, [%[c_ptr], #-1, MUL VL]\n"
376 "uzp2 z7.d, z30.d, z31.d\n"
377 "st1w z0.s, p0, [%[c_ptr]]\n"
378 "st1w z1.s, p0, [%[c_ptr], #1, MUL VL]\n"
379 "st1w z2.s, p0, [%[c_ptr], #2, MUL VL]\n"
380 "st1w z3.s, p0, [%[c_ptr], #3, MUL VL]\n"
381 "st1w z4.s, p0, [%[c_ptr], #4, MUL VL]\n"
382 "st1w z5.s, p0, [%[c_ptr], #5, MUL VL]\n"
383 "st1w z6.s, p0, [%[c_ptr], #6, MUL VL]\n"
384 "st1w z7.s, p0, [%[c_ptr], #7, MUL VL]\n"
385 "addvl %[c_ptr], %[c_ptr], #8\n"
386 : [a_ptr]
"+r" (a_ptr), [b_ptr]
"+r" (b_ptr), [c_ptr]
"+r" (c_ptr),
387 [loops]
"+r" (loops), [tails]
"+r" (tails)
389 :
"z0",
"z1",
"z2",
"z3",
"z4",
"z5",
"z6",
"z7",
"z8",
"z9",
"z10",
"z11",
"z12",
"z13",
"z14",
"z15",
"z16",
"z17",
"z18",
"z19",
"z20",
"z21",
"z22",
"z23",
"z24",
"z25",
"z26",
"z27",
"z28",
"z29",
"z30",
"z31",
"cc",
"memory"
397 #endif // __ARM_FEATURE_SVE