26 #if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
30 #include "../../asmlib.hpp"
44 void a64_hgemm_asimd_8x24(
const __fp16 *Apanel,
const __fp16 *Bpanel, __fp16 *Cpanel,
int ablocks,
int bblocks,
int K) {
45 const __fp16 *a_ptr = Apanel;
46 __fp16 *c_ptr = Cpanel;
48 for (
int yb=0; yb<ablocks; yb++) {
49 const __fp16 *a_ptr0 = a_ptr;
50 const __fp16 *b_ptr = Bpanel;
52 for (
int xb=0; xb<bblocks; xb++) {
57 int k = ((
K+1)/2) - 1;
59 register float16x8_t a0
asm(
"v0");
60 register float16x8_t a0a
asm(
"v1");
61 register float16x8_t b0
asm(
"v2");
62 register float16x8_t b1
asm(
"v3");
63 register float16x8_t b2
asm(
"v4");
64 register float16x8_t b0a
asm(
"v5");
65 register float16x8_t b1a
asm(
"v6");
66 register float16x8_t b2a
asm(
"v7");
71 "ldr %q[a0], [%[a_ptr]]\n"
73 "ldr %q[b0], [%[b_ptr]]\n"
75 "ldr %q[b1], [%[b_ptr], #16]\n"
77 "ldr %q[b2], [%[b_ptr], #32]\n"
79 "ldr %q[b0a], [%[b_ptr], #48]\n"
81 "ldr %q[b1a], [%[b_ptr], #64]\n"
100 "movi v26.8h, #0x0\n"
101 "movi v27.8h, #0x0\n"
102 "movi v28.8h, #0x0\n"
103 "movi v29.8h, #0x0\n"
104 "movi v30.8h, #0x0\n"
105 "movi v31.8h, #0x0\n"
111 "fmla v8.8h , %[b0].8h, %[a0].h[0]\n"
112 "fmla v9.8h , %[b0].8h, %[a0].h[1]\n"
113 "ldr %q[a0a], [%[a_ptr], #16]\n"
114 "fmla v10.8h, %[b0].8h, %[a0].h[2]\n"
115 "fmla v11.8h, %[b0].8h, %[a0].h[3]\n"
116 "ldr %q[b2a], [%[b_ptr], #80]\n"
117 "fmla v12.8h, %[b0].8h, %[a0].h[4]\n"
118 "fmla v13.8h, %[b0].8h, %[a0].h[5]\n"
119 "fmla v14.8h, %[b0].8h, %[a0].h[6]\n"
120 "fmla v15.8h, %[b0].8h, %[a0].h[7]\n"
121 "ldr %q[b0], [%[b_ptr], #96]\n"
123 "fmla v16.8h, %[b1].8h, %[a0].h[0]\n"
124 "fmla v17.8h, %[b1].8h, %[a0].h[1]\n"
126 "fmla v18.8h, %[b1].8h, %[a0].h[2]\n"
127 "fmla v19.8h, %[b1].8h, %[a0].h[3]\n"
128 "add %[b_ptr], %[b_ptr], #96\n"
129 "fmla v20.8h, %[b1].8h, %[a0].h[4]\n"
130 "fmla v21.8h, %[b1].8h, %[a0].h[5]\n"
131 "fmla v22.8h, %[b1].8h, %[a0].h[6]\n"
132 "fmla v23.8h, %[b1].8h, %[a0].h[7]\n"
133 "ldr %q[b1], [%[b_ptr], #16]\n"
135 "fmla v24.8h, %[b2].8h, %[a0].h[0]\n"
136 "fmla v25.8h, %[b2].8h, %[a0].h[1]\n"
138 "fmla v26.8h, %[b2].8h, %[a0].h[2]\n"
139 "fmla v27.8h, %[b2].8h, %[a0].h[3]\n"
140 "fmla v28.8h, %[b2].8h, %[a0].h[4]\n"
141 "fmla v29.8h, %[b2].8h, %[a0].h[5]\n"
142 "fmla v30.8h, %[b2].8h, %[a0].h[6]\n"
143 "fmla v31.8h, %[b2].8h, %[a0].h[7]\n"
144 "ldr %q[a0], [%[a_ptr], #32]\n"
146 "fmla v8.8h , %[b0a].8h, %[a0a].h[0]\n"
147 "fmla v9.8h , %[b0a].8h, %[a0a].h[1]\n"
148 "ldr %q[b2], [%[b_ptr], #32]\n"
149 "fmla v10.8h, %[b0a].8h, %[a0a].h[2]\n"
150 "fmla v11.8h, %[b0a].8h, %[a0a].h[3]\n"
151 "fmla v12.8h, %[b0a].8h, %[a0a].h[4]\n"
152 "fmla v13.8h, %[b0a].8h, %[a0a].h[5]\n"
153 "fmla v14.8h, %[b0a].8h, %[a0a].h[6]\n"
154 "fmla v15.8h, %[b0a].8h, %[a0a].h[7]\n"
155 "ldr %q[b0a], [%[b_ptr], #48]\n"
157 "fmla v16.8h, %[b1a].8h, %[a0a].h[0]\n"
158 "fmla v17.8h, %[b1a].8h, %[a0a].h[1]\n"
160 "fmla v18.8h, %[b1a].8h, %[a0a].h[2]\n"
161 "fmla v19.8h, %[b1a].8h, %[a0a].h[3]\n"
162 "fmla v20.8h, %[b1a].8h, %[a0a].h[4]\n"
163 "fmla v21.8h, %[b1a].8h, %[a0a].h[5]\n"
164 "fmla v22.8h, %[b1a].8h, %[a0a].h[6]\n"
165 "fmla v23.8h, %[b1a].8h, %[a0a].h[7]\n"
166 "ldr %q[b1a], [%[b_ptr], #64]\n"
168 "fmla v24.8h, %[b2a].8h, %[a0a].h[0]\n"
169 "fmla v25.8h, %[b2a].8h, %[a0a].h[1]\n"
170 "add %[a_ptr], %[a_ptr], #32\n"
171 "fmla v26.8h, %[b2a].8h, %[a0a].h[2]\n"
172 "fmla v27.8h, %[b2a].8h, %[a0a].h[3]\n"
173 "fmla v28.8h, %[b2a].8h, %[a0a].h[4]\n"
174 "fmla v29.8h, %[b2a].8h, %[a0a].h[5]\n"
175 "subs %w[k], %w[k], #1\n"
176 "fmla v30.8h, %[b2a].8h, %[a0a].h[6]\n"
177 "fmla v31.8h, %[b2a].8h, %[a0a].h[7]\n"
183 "cbnz %w[oddk], 2f\n"
186 "fmla v8.8h , %[b0].8h, %[a0].h[0]\n"
187 "fmla v9.8h , %[b0].8h, %[a0].h[1]\n"
188 "ldr %q[a0a], [%[a_ptr], #16]\n"
189 "fmla v10.8h, %[b0].8h, %[a0].h[2]\n"
190 "fmla v11.8h, %[b0].8h, %[a0].h[3]\n"
191 "ldr %q[b2a], [%[b_ptr], #80]\n"
192 "fmla v12.8h, %[b0].8h, %[a0].h[4]\n"
193 "fmla v13.8h, %[b0].8h, %[a0].h[5]\n"
194 "fmla v14.8h, %[b0].8h, %[a0].h[6]\n"
195 "fmla v15.8h, %[b0].8h, %[a0].h[7]\n"
197 "fmla v16.8h, %[b1].8h, %[a0].h[0]\n"
198 "fmla v17.8h, %[b1].8h, %[a0].h[1]\n"
199 "add %[b_ptr], %[b_ptr], #96\n"
200 "fmla v18.8h, %[b1].8h, %[a0].h[2]\n"
201 "fmla v19.8h, %[b1].8h, %[a0].h[3]\n"
202 "fmla v20.8h, %[b1].8h, %[a0].h[4]\n"
203 "fmla v21.8h, %[b1].8h, %[a0].h[5]\n"
204 "add %[a_ptr], %[a_ptr], #32\n"
205 "fmla v22.8h, %[b1].8h, %[a0].h[6]\n"
206 "fmla v23.8h, %[b1].8h, %[a0].h[7]\n"
208 "fmla v24.8h, %[b2].8h, %[a0].h[0]\n"
209 "fmla v25.8h, %[b2].8h, %[a0].h[1]\n"
210 "fmla v26.8h, %[b2].8h, %[a0].h[2]\n"
211 "fmla v27.8h, %[b2].8h, %[a0].h[3]\n"
212 "fmla v28.8h, %[b2].8h, %[a0].h[4]\n"
213 "fmla v29.8h, %[b2].8h, %[a0].h[5]\n"
214 "fmla v30.8h, %[b2].8h, %[a0].h[6]\n"
215 "fmla v31.8h, %[b2].8h, %[a0].h[7]\n"
217 "fmla v8.8h , %[b0a].8h, %[a0a].h[0]\n"
218 "fmla v16.8h, %[b1a].8h, %[a0a].h[0]\n"
219 "str q8, [%[c_ptr]]\n"
220 "fmla v24.8h, %[b2a].8h, %[a0a].h[0]\n"
221 "str q16, [%[c_ptr], #16]\n"
223 "fmla v9.8h , %[b0a].8h, %[a0a].h[1]\n"
224 "str q24, [%[c_ptr], #32]\n"
225 "fmla v17.8h, %[b1a].8h, %[a0a].h[1]\n"
226 "str q9, [%[c_ptr], #48]\n"
227 "fmla v25.8h, %[b2a].8h, %[a0a].h[1]\n"
228 "str q17, [%[c_ptr], #64]\n"
230 "fmla v10.8h, %[b0a].8h, %[a0a].h[2]\n"
231 "str q25, [%[c_ptr], #80]\n"
232 "fmla v18.8h, %[b1a].8h, %[a0a].h[2]\n"
233 "str q10, [%[c_ptr], #96]\n"
234 "fmla v26.8h, %[b2a].8h, %[a0a].h[2]\n"
235 "str q18, [%[c_ptr], #112]\n"
237 "fmla v11.8h, %[b0a].8h, %[a0a].h[3]\n"
238 "str q26, [%[c_ptr], #128]\n"
239 "fmla v19.8h, %[b1a].8h, %[a0a].h[3]\n"
240 "str q11, [%[c_ptr], #144]\n"
241 "fmla v27.8h, %[b2a].8h, %[a0a].h[3]\n"
242 "str q19, [%[c_ptr], #160]\n"
244 "fmla v12.8h, %[b0a].8h, %[a0a].h[4]\n"
245 "str q27, [%[c_ptr], #176]\n"
246 "fmla v20.8h, %[b1a].8h, %[a0a].h[4]\n"
247 "str q12, [%[c_ptr], #192]\n"
248 "fmla v28.8h, %[b2a].8h, %[a0a].h[4]\n"
249 "str q20, [%[c_ptr], #208]\n"
251 "fmla v13.8h, %[b0a].8h, %[a0a].h[5]\n"
252 "str q28, [%[c_ptr], #224]\n"
253 "fmla v21.8h, %[b1a].8h, %[a0a].h[5]\n"
254 "str q13, [%[c_ptr], #240]\n"
255 "fmla v29.8h, %[b2a].8h, %[a0a].h[5]\n"
256 "str q21, [%[c_ptr], #256]\n"
258 "fmla v14.8h, %[b0a].8h, %[a0a].h[6]\n"
259 "str q29, [%[c_ptr], #272]\n"
260 "fmla v22.8h, %[b1a].8h, %[a0a].h[6]\n"
261 "str q14, [%[c_ptr], #288]\n"
262 "fmla v30.8h, %[b2a].8h, %[a0a].h[6]\n"
263 "str q22, [%[c_ptr], #304]\n"
265 "fmla v15.8h, %[b0a].8h, %[a0a].h[7]\n"
266 "str q30, [%[c_ptr], #320]\n"
267 "fmla v23.8h, %[b1a].8h, %[a0a].h[7]\n"
268 "str q15, [%[c_ptr], #336]\n"
269 "fmla v31.8h, %[b2a].8h, %[a0a].h[7]\n"
274 "fmla v8.8h , %[b0].8h, %[a0].h[0]\n"
275 "add %[b_ptr], %[b_ptr], #48\n"
276 "fmla v16.8h, %[b1].8h, %[a0].h[0]\n"
277 "add %[a_ptr], %[a_ptr], #16\n"
278 "str q8, [%[c_ptr]]\n"
279 "fmla v24.8h, %[b2].8h, %[a0].h[0]\n"
280 "str q16, [%[c_ptr], #16]\n"
282 "fmla v9.8h , %[b0].8h, %[a0].h[1]\n"
283 "str q24, [%[c_ptr], #32]\n"
284 "fmla v17.8h, %[b1].8h, %[a0].h[1]\n"
285 "str q9, [%[c_ptr], #48]\n"
286 "fmla v25.8h, %[b2].8h, %[a0].h[1]\n"
287 "str q17, [%[c_ptr], #64]\n"
289 "fmla v10.8h, %[b0].8h, %[a0].h[2]\n"
290 "str q25, [%[c_ptr], #80]\n"
291 "fmla v18.8h, %[b1].8h, %[a0].h[2]\n"
292 "str q10, [%[c_ptr], #96]\n"
293 "fmla v26.8h, %[b2].8h, %[a0].h[2]\n"
294 "str q18, [%[c_ptr], #112]\n"
296 "fmla v11.8h, %[b0].8h, %[a0].h[3]\n"
297 "str q26, [%[c_ptr], #128]\n"
298 "fmla v19.8h, %[b1].8h, %[a0].h[3]\n"
299 "str q11, [%[c_ptr], #144]\n"
300 "fmla v27.8h, %[b2].8h, %[a0].h[3]\n"
301 "str q19, [%[c_ptr], #160]\n"
303 "fmla v12.8h, %[b0].8h, %[a0].h[4]\n"
304 "str q27, [%[c_ptr], #176]\n"
305 "fmla v20.8h, %[b1].8h, %[a0].h[4]\n"
306 "str q12, [%[c_ptr], #192]\n"
307 "fmla v28.8h, %[b2].8h, %[a0].h[4]\n"
308 "str q20, [%[c_ptr], #208]\n"
310 "fmla v13.8h, %[b0].8h, %[a0].h[5]\n"
311 "str q28, [%[c_ptr], #224]\n"
312 "fmla v21.8h, %[b1].8h, %[a0].h[5]\n"
313 "str q13, [%[c_ptr], #240]\n"
314 "fmla v29.8h, %[b2].8h, %[a0].h[5]\n"
315 "str q21, [%[c_ptr], #256]\n"
317 "fmla v14.8h, %[b0].8h, %[a0].h[6]\n"
318 "str q29, [%[c_ptr], #272]\n"
319 "fmla v22.8h, %[b1].8h, %[a0].h[6]\n"
320 "str q14, [%[c_ptr], #288]\n"
321 "fmla v30.8h, %[b2].8h, %[a0].h[6]\n"
322 "str q22, [%[c_ptr], #304]\n"
324 "fmla v15.8h, %[b0].8h, %[a0].h[7]\n"
325 "str q30, [%[c_ptr], #320]\n"
326 "fmla v23.8h, %[b1].8h, %[a0].h[7]\n"
327 "str q15, [%[c_ptr], #336]\n"
328 "fmla v31.8h, %[b2].8h, %[a0].h[7]\n"
331 "str q23, [%[c_ptr], #352]\n"
332 "str q31, [%[c_ptr], #368]\n"
333 "add %[c_ptr], %[c_ptr], #384\n"
335 [a_ptr]
"+r" (a_ptr), [b_ptr]
"+r" (b_ptr), [c_ptr]
"+r" (c_ptr),
336 [a0]
"+w" (a0), [a0a]
"+w" (a0a),
337 [b0]
"+w" (b0), [b1]
"+w" (b1), [b2]
"+w" (b2), [k]
"+r" (k),
338 [b0a]
"+w" (b0a), [b1a]
"+w" (b1a), [b2a]
"+w" (b2a)
340 :
"x20",
"x21",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
341 "v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"cc"
349 #endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)