28 #include "../../asmlib.hpp"
32 void a64_gemm_u8_8x12(
const uint8_t *Apanel,
const uint8_t *Bpanel, uint32_t *Cpanel,
int ablocks,
int bblocks,
int K) {
33 const uint8_t *a_ptr = Apanel;
34 uint32_t *c_ptr = Cpanel;
39 const int oddk = (W & 1);
40 const int init_value_k = ((W+1)/2) - 1;
41 for (
int yb=0; yb<ablocks; yb++) {
42 const uint8_t *a_ptr0 = a_ptr;
43 const uint8_t *b_ptr = Bpanel;
44 for (
int xb=0; xb<bblocks; xb++) {
47 register uint8x16_t a0
asm(
"v0");
48 register uint8x16_t a1
asm(
"v1");
49 register uint8x16_t b0
asm(
"v2");
50 register uint8x16_t b1
asm(
"v3");
51 register uint8x16_t b2
asm(
"v4");
52 register uint8x16_t a0a
asm(
"v5");
53 register uint8x16_t a1a
asm(
"v6");
57 "ldr %q[a0], [%[a_ptr]]\n"
59 "ldr %q[b0], [%[b_ptr]]\n"
61 "ldr %q[a1], [%[a_ptr], #16]\n"
63 "ldr %q[b1], [%[b_ptr], #16]\n"
100 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
101 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
103 "ldr %q[b2], [%[b_ptr], #32]\n"
104 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
105 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
106 "ldr %q[a0a], [%[a_ptr], #32]\n"
107 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
108 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
109 "ldr %q[a1a], [%[a_ptr], #48]\n"
110 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
111 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
112 "ldr %q[b0], [%[b_ptr], #48]\n"
114 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
115 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
117 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
118 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
119 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
120 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
121 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
122 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
123 "ldr %q[b1], [%[b_ptr], #64]\n"
125 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
126 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
128 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
129 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
130 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
131 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
132 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
133 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
134 "ldr %q[b2], [%[b_ptr], #80]\n"
136 ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
137 ".word 0x6fa5e049 // udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
138 "ldr %q[a0], [%[a_ptr], #64]\n"
139 ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
140 ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
141 ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
142 "ldr %q[a1], [%[a_ptr], #80]\n"
143 ".word 0x6fa6e04d // udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
144 ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
145 ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
146 "ldr %q[b0], [%[b_ptr], #96]\n"
148 ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
149 ".word 0x6fa5e071 // udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
151 ".word 0x6f85e872 // udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
152 ".word 0x6fa5e873 // udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
153 ".word 0x6f86e074 // udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
154 ".word 0x6fa6e075 // udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
155 ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
156 ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
157 "ldr %q[b1], [%[b_ptr], #112]\n"
159 ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
160 ".word 0x6fa5e099 // udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
161 "add %[a_ptr], %[a_ptr], #64\n"
162 ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
163 ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
164 "add %[b_ptr], %[b_ptr], #96\n"
165 ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
166 ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
167 "subs %w[k], %w[k], #1\n"
168 ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
169 ".word 0x6fa6e89f // udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
176 "cbnz %w[oddk], 2f\n"
179 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
180 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
181 "ldr %q[b2], [%[b_ptr], #32]\n"
182 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
183 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
184 "ldr %q[a0a], [%[a_ptr], #32]\n"
185 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
186 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
187 "ldr %q[a1a], [%[a_ptr], #48]\n"
188 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
189 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
190 "ldr %q[b0], [%[b_ptr], #48]\n"
192 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
193 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
194 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
195 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
196 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
197 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
198 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
199 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
200 "ldr %q[b1], [%[b_ptr], #64]\n"
202 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
203 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
204 "add %[a_ptr], %[a_ptr], #64\n"
205 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
206 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
207 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
208 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
209 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
210 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
211 "ldr %q[b2], [%[b_ptr], #80]\n"
213 ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
215 ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
216 "add %[b_ptr], %[b_ptr], #96\n"
217 ".word 0x6fa5e049 // udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
218 "str q8, [%[c_ptr], #0]\n"
219 ".word 0x6fa5e071 // udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
220 "str q16, [%[c_ptr], #16]\n"
221 ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
222 "str q24, [%[c_ptr], #32]\n"
224 ".word 0x6fa5e099 // udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
225 "str q9, [%[c_ptr], #48]\n"
226 ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
227 "str q17, [%[c_ptr], #64]\n"
228 ".word 0x6f85e872 // udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
229 "str q25, [%[c_ptr], #80]\n"
230 ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
231 "str q10, [%[c_ptr], #96]\n"
233 ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
234 "str q18, [%[c_ptr], #112]\n"
235 ".word 0x6fa5e873 // udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
236 "str q26, [%[c_ptr], #128]\n"
237 ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
238 "str q11, [%[c_ptr], #144]\n"
240 ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
241 "str q19, [%[c_ptr], #160]\n"
242 ".word 0x6f86e074 // udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
243 "str q27, [%[c_ptr], #176]\n"
244 ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
245 "str q12, [%[c_ptr], #192]\n"
247 ".word 0x6fa6e04d // udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
248 "str q20, [%[c_ptr], #208]\n"
249 ".word 0x6fa6e075 // udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
250 "str q28, [%[c_ptr], #224]\n"
251 ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
252 "str q13, [%[c_ptr], #240]\n"
254 ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
255 "str q21, [%[c_ptr], #256]\n"
256 ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
257 "str q29, [%[c_ptr], #272]\n"
258 ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
259 "str q14, [%[c_ptr], #288]\n"
261 ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
262 "str q22, [%[c_ptr], #304]\n"
263 ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
264 "str q30, [%[c_ptr], #320]\n"
265 ".word 0x6fa6e89f // udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
266 "str q15, [%[c_ptr], #336]\n"
272 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
273 "ldr %q[b2], [%[b_ptr], #32]\n"
274 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
275 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
276 "str q8, [%[c_ptr], #0]\n"
277 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
278 "str q16, [%[c_ptr], #16]\n"
279 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
280 "add %[b_ptr], %[b_ptr], #48\n"
281 "add %[a_ptr], %[a_ptr], #32\n"
282 "str q24, [%[c_ptr], #32]\n"
283 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
284 "str q9, [%[c_ptr], #48]\n"
286 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
287 "str q17, [%[c_ptr], #64]\n"
288 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
289 "str q25, [%[c_ptr], #80]\n"
290 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
291 "str q10, [%[c_ptr], #96]\n"
293 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
294 "str q18, [%[c_ptr], #112]\n"
295 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
296 "str q26, [%[c_ptr], #128]\n"
297 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
298 "str q11, [%[c_ptr], #144]\n"
300 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
301 "str q19, [%[c_ptr], #160]\n"
302 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
303 "str q27, [%[c_ptr], #176]\n"
304 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
305 "str q12, [%[c_ptr], #192]\n"
307 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
308 "str q20, [%[c_ptr], #208]\n"
309 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
310 "str q28, [%[c_ptr], #224]\n"
311 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
312 "str q13, [%[c_ptr], #240]\n"
314 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
315 "str q21, [%[c_ptr], #256]\n"
316 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
317 "str q29, [%[c_ptr], #272]\n"
318 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
319 "str q14, [%[c_ptr], #288]\n"
321 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
322 "str q22, [%[c_ptr], #304]\n"
323 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
324 "str q30, [%[c_ptr], #320]\n"
325 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
326 "str q15, [%[c_ptr], #336]\n"
331 "str q23, [%[c_ptr], #352]\n"
332 "str q31, [%[c_ptr], #368]\n"
333 "add %[c_ptr], %[c_ptr], #384\n"
336 [a_ptr]
"+r" (a_ptr), [b_ptr]
"+r" (b_ptr), [c_ptr]
"+r" (c_ptr),
337 [a0]
"+w" (a0), [a1]
"+w" (a1), [a0a]
"+w" (a0a), [a1a]
"+w" (a1a),
338 [b0]
"+w" (b0), [b1]
"+w" (b1), [b2]
"+w" (b2), [k]
"+r" (k)
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"