1 /*
2 * Copyright (c) 2017-2018 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #ifdef __aarch64__
25
26 #include <arm_neon.h>
27
28 #include "../../asmlib.hpp"
29
30 namespace arm_gemm {
31
a64_gemm_u16_asimd_8x12(const uint16_t * Apanel,const uint16_t * Bpanel,uint32_t * Cpanel,int ablocks,int bblocks,int K)32 void a64_gemm_u16_asimd_8x12(const uint16_t *Apanel, const uint16_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, int K)
33 {
34 const uint16_t *a_ptr = Apanel;
35 uint32_t *c_ptr = Cpanel;
36
37 for (int yb = 0; yb < ablocks; yb++)
38 {
39 const uint16_t *a_ptr0 = a_ptr;
40 const uint16_t *b_ptr = Bpanel;
41
42 for (int xb = 0; xb < bblocks; xb++)
43 {
44 a_ptr = a_ptr0;
45 const bool odd_k = K & 0x1;
46 int k = (K+1)/2 - 1;
47
48 register uint16x8_t aa asm("v0");
49 register uint16x8_t ab asm("v1");
50 register uint16x8_t b0 asm("v2");
51 register uint16x8_t b1 asm("v3");
52 register uint16x8_t b2 asm("v4");
53
54 __asm __volatile (
55 "ldr %d[aa], [%x[a_ptr]]\n" // Load A[A].lower
56 "movi v5.4s, #0\n"
57 "ldr x20, [%x[a_ptr], #0x08]\n" // Load A[A].upper
58 "movi v6.4s, #0\n"
59 "ldr %d[b0], [%x[b_ptr]]\n" // Load B[0].lower
60 "ins %[aa].d[1], x20\n" // Merge A[A].lower and upper
61 "movi v7.4s, #0\n"
62 ASM_PREFETCH("[%[a_ptr], #64]")
63 "movi v8.4s, #0\n"
64 "ldr x20, [%x[b_ptr], #0x08]\n" // Load B[0].upper
65 "movi v9.4s, #0\n"
66 ASM_PREFETCH("[%[b_ptr], #64]")
67 "movi v10.4s, #0\n"
68 "ldr %d[b1], [%x[b_ptr], #0x10]\n" // Load B[1].lower
69 "ins %[b0].d[1], x20\n" // Merge B[0].lower and upper
70 "movi v11.4s, #0\n"
71 ASM_PREFETCH("[%[a_ptr], #96]")
72 "movi v12.4s, #0\n"
73 "movi v13.4s, #0\n"
74 ASM_PREFETCH("[%[b_ptr], #96]")
75 "movi v14.4s, #0\n"
76 "movi v15.4s, #0\n"
77 ASM_PREFETCH("[%[a_ptr], #128]")
78 "movi v16.4s, #0\n"
79 "movi v17.4s, #0\n"
80 ASM_PREFETCH("[%[b_ptr], #128]")
81 "movi v18.4s, #0\n"
82 "movi v19.4s, #0\n"
83 ASM_PREFETCH("[%[a_ptr], #160]")
84 "movi v20.4s, #0\n"
85 "movi v21.4s, #0\n"
86 ASM_PREFETCH("[%[b_ptr], #160]")
87 "movi v22.4s, #0\n"
88 "movi v23.4s, #0\n"
89 ASM_PREFETCH("[%[a_ptr], #192]")
90 "movi v24.4s, #0\n"
91 "add %x[a_ptr], %x[a_ptr], #0x10\n"
92 "movi v25.4s, #0\n"
93 ASM_PREFETCH("[%[b_ptr], #192]")
94 "movi v26.4s, #0\n"
95 "add %x[b_ptr], %x[b_ptr], #0x18\n"
96 "movi v27.4s, #0\n"
97 "movi v28.4s, #0\n"
98
99 "cbz %x[k], 2f\n" // Skip the loop if doing zero iterations.
100
101 "1:\n" // Main loop
102 // First unroll
103 "umlal v5.4s, %[b0].4h, %[aa].h[0]\n"
104 "ldr x20, [%x[b_ptr]]\n" // Load B[1].upper
105 "umlal v6.4s, %[b0].4h, %[aa].h[1]\n"
106 "umlal v7.4s, %[b0].4h, %[aa].h[2]\n"
107 "ldr %d[ab], [%x[a_ptr]]\n" // Load A[B].lower
108 "ins %[b1].d[1], x20\n" // Merge B[1].lower and .upper
109 "umlal v8.4s, %[b0].4h, %[aa].h[3]\n"
110 "umlal v9.4s, %[b0].4h, %[aa].h[4]\n"
111 "ldr x20, [%x[a_ptr], #0x8]\n" // Load A[B].upper
112 "umlal v10.4s, %[b0].4h, %[aa].h[5]\n"
113 "umlal v11.4s, %[b0].4h, %[aa].h[6]\n"
114 "ldr %d[b2], [%x[b_ptr], #0x8]\n" // Load B[2].lower
115 "ins %[ab].d[1], x20\n" // Merge A[B].lower and .upper
116 "umlal v12.4s, %[b0].4h, %[aa].h[7]\n"
117 "umlal2 v13.4s, %[b0].8h, %[aa].h[0]\n"
118 "ldr x20, [%x[b_ptr], #0x10]\n" // Load B[2].upper
119 "umlal2 v14.4s, %[b0].8h, %[aa].h[1]\n"
120 "umlal2 v15.4s, %[b0].8h, %[aa].h[2]\n"
121 "umlal2 v16.4s, %[b0].8h, %[aa].h[3]\n"
122 "umlal2 v17.4s, %[b0].8h, %[aa].h[4]\n"
123 "umlal2 v18.4s, %[b0].8h, %[aa].h[5]\n"
124 "umlal2 v19.4s, %[b0].8h, %[aa].h[6]\n"
125 "umlal2 v20.4s, %[b0].8h, %[aa].h[7]\n"
126 "ldr %d[b0], [%x[b_ptr], #0x18]\n" // Load B[0].lower
127 "ins %[b2].d[1], x20\n" // Merge B[2].lower and .upper
128 "umlal v21.4s, %[b1].4h, %[aa].h[0]\n"
129 "umlal v22.4s, %[b1].4h, %[aa].h[1]\n"
130 "ldr x20, [%x[b_ptr], #0x20]\n" // Load B[0].upper
131 "umlal v23.4s, %[b1].4h, %[aa].h[2]\n"
132 "umlal v24.4s, %[b1].4h, %[aa].h[3]\n"
133 "umlal v25.4s, %[b1].4h, %[aa].h[4]\n"
134 "umlal v26.4s, %[b1].4h, %[aa].h[5]\n"
135 "umlal v27.4s, %[b1].4h, %[aa].h[6]\n"
136 "umlal v28.4s, %[b1].4h, %[aa].h[7]\n"
137
138 // Second unroll
139 "umlal2 v5.4s, %[b1].8h, %[ab].h[0]\n"
140 "ldr %d[aa], [%x[a_ptr], #0x10]\n" // Load A[A].lower
141 "ins %[b0].d[1], x20\n" // Merge B[0].lower and .upper
142 "umlal2 v6.4s, %[b1].8h, %[ab].h[1]\n"
143 "umlal2 v7.4s, %[b1].8h, %[ab].h[2]\n"
144 "ldr x20, [%x[a_ptr], #0x18]\n" // Load A[A].upper
145 "umlal2 v8.4s, %[b1].8h, %[ab].h[3]\n"
146 "umlal2 v9.4s, %[b1].8h, %[ab].h[4]\n"
147 "umlal2 v10.4s, %[b1].8h, %[ab].h[5]\n"
148 "umlal2 v11.4s, %[b1].8h, %[ab].h[6]\n"
149 "add %x[a_ptr], %x[a_ptr], #0x20\n"
150 "umlal2 v12.4s, %[b1].8h, %[ab].h[7]\n"
151 "umlal v13.4s, %[b2].4h, %[ab].h[0]\n"
152 ASM_PREFETCH("[%[b_ptr], #320]")
153 "umlal v14.4s, %[b2].4h, %[ab].h[1]\n"
154 "umlal v15.4s, %[b2].4h, %[ab].h[2]\n"
155 ASM_PREFETCH("[%[a_ptr], #320]")
156 "umlal v16.4s, %[b2].4h, %[ab].h[3]\n"
157 "umlal v17.4s, %[b2].4h, %[ab].h[4]\n"
158 ASM_PREFETCH("[%[b_ptr], #448]")
159 "umlal v18.4s, %[b2].4h, %[ab].h[5]\n"
160 "umlal v19.4s, %[b2].4h, %[ab].h[6]\n"
161 "umlal v20.4s, %[b2].4h, %[ab].h[7]\n"
162 "umlal2 v21.4s, %[b2].8h, %[ab].h[0]\n"
163 "umlal2 v22.4s, %[b2].8h, %[ab].h[1]\n"
164 "subs %x[k], %x[k], #0x1\n"
165 "umlal2 v23.4s, %[b2].8h, %[ab].h[2]\n"
166 "umlal2 v24.4s, %[b2].8h, %[ab].h[3]\n"
167 "ldr %d[b1], [%x[b_ptr], #0x28]\n" // Load B[1].lower
168 "ins %[aa].d[1], x20\n" // Merge A[A].lower and .upper
169 "umlal2 v25.4s, %[b2].8h, %[ab].h[4]\n"
170 "umlal2 v26.4s, %[b2].8h, %[ab].h[5]\n"
171 "add %x[b_ptr], %x[b_ptr], #0x30\n"
172 "umlal2 v27.4s, %[b2].8h, %[ab].h[6]\n"
173 "umlal2 v28.4s, %[b2].8h, %[ab].h[7]\n"
174 "bne 1b\n"
175
176 "2:\n" // Even tail
177 "cbnz %x[odd_k], 3f\n"
178
179 "umlal v5.4s, %[b0].4h, %[aa].h[0]\n"
180 "ldr x20, [%x[b_ptr]]\n" // Load B[1].upper
181 "umlal v6.4s, %[b0].4h, %[aa].h[1]\n"
182 "umlal v7.4s, %[b0].4h, %[aa].h[2]\n"
183 "ldr %d[ab], [%x[a_ptr]]\n" // Load A[B].lower
184 "ins %[b1].d[1], x20\n" // Merge B[1].lower and .upper
185 "umlal v8.4s, %[b0].4h, %[aa].h[3]\n"
186 "umlal v9.4s, %[b0].4h, %[aa].h[4]\n"
187 "ldr x20, [%x[a_ptr], #0x8]\n" // Load A[B].upper
188 "umlal v10.4s, %[b0].4h, %[aa].h[5]\n"
189 "umlal v11.4s, %[b0].4h, %[aa].h[6]\n"
190 "ldr %d[b2], [%x[b_ptr], #0x8]\n" // Load B[2].lower
191 "ins %[ab].d[1], x20\n" // Merge A[B].lower and .upper
192 "umlal v12.4s, %[b0].4h, %[aa].h[7]\n"
193 "umlal2 v13.4s, %[b0].8h, %[aa].h[0]\n"
194 "ldr x20, [%x[b_ptr], #0x10]\n" // Load B[2].upper
195 "umlal2 v14.4s, %[b0].8h, %[aa].h[1]\n"
196 "umlal2 v15.4s, %[b0].8h, %[aa].h[2]\n"
197 "umlal2 v16.4s, %[b0].8h, %[aa].h[3]\n"
198 "add %[a_ptr], %[a_ptr], #0x10\n"
199 "umlal2 v17.4s, %[b0].8h, %[aa].h[4]\n"
200 "add %[b_ptr], %[b_ptr], #0x18\n"
201 "umlal2 v18.4s, %[b0].8h, %[aa].h[5]\n"
202 "umlal2 v19.4s, %[b0].8h, %[aa].h[6]\n"
203 "umlal2 v20.4s, %[b0].8h, %[aa].h[7]\n"
204 "ins %[b2].d[1], x20\n" // Merge B[2].lower and .upper
205 "umlal v21.4s, %[b1].4h, %[aa].h[0]\n"
206 "umlal v22.4s, %[b1].4h, %[aa].h[1]\n"
207 "umlal v23.4s, %[b1].4h, %[aa].h[2]\n"
208 "umlal v24.4s, %[b1].4h, %[aa].h[3]\n"
209 "umlal v25.4s, %[b1].4h, %[aa].h[4]\n"
210 "umlal v26.4s, %[b1].4h, %[aa].h[5]\n"
211 "umlal v27.4s, %[b1].4h, %[aa].h[6]\n"
212 "umlal v28.4s, %[b1].4h, %[aa].h[7]\n"
213
214 "umlal2 v5.4s, %[b1].8h, %[ab].h[0]\n"
215 "umlal v13.4s, %[b2].4h, %[ab].h[0]\n"
216 "umlal2 v21.4s, %[b2].8h, %[ab].h[0]\n"
217 "umlal2 v6.4s, %[b1].8h, %[ab].h[1]\n"
218 "umlal v14.4s, %[b2].4h, %[ab].h[1]\n"
219 "str q5, [%x[c_ptr]]\n"
220 "umlal2 v22.4s, %[b2].8h, %[ab].h[1]\n"
221 "str q13, [%x[c_ptr], #0x10]\n"
222 "umlal2 v7.4s, %[b1].8h, %[ab].h[2]\n"
223 "str q21, [%x[c_ptr], #0x20]\n"
224 "umlal v15.4s, %[b2].4h, %[ab].h[2]\n"
225 "str q6, [%x[c_ptr], #0x30]\n"
226 "umlal2 v23.4s, %[b2].8h, %[ab].h[2]\n"
227 "str q14, [%x[c_ptr], #0x40]\n"
228 "umlal2 v8.4s, %[b1].8h, %[ab].h[3]\n"
229 "str q22, [%x[c_ptr], #0x50]\n"
230 "umlal v16.4s, %[b2].4h, %[ab].h[3]\n"
231 "str q7, [%x[c_ptr], #0x60]\n"
232 "umlal2 v24.4s, %[b2].8h, %[ab].h[3]\n"
233 "str q15, [%x[c_ptr], #0x70]\n"
234 "umlal2 v9.4s, %[b1].8h, %[ab].h[4]\n"
235 "str q23, [%x[c_ptr], #0x80]\n"
236 "umlal v17.4s, %[b2].4h, %[ab].h[4]\n"
237 "str q8, [%x[c_ptr], #0x90]\n"
238 "umlal2 v25.4s, %[b2].8h, %[ab].h[4]\n"
239 "str q16, [%x[c_ptr], #0xa0]\n"
240 "umlal2 v10.4s, %[b1].8h, %[ab].h[5]\n"
241 "str q24, [%x[c_ptr], #0xb0]\n"
242 "umlal v18.4s, %[b2].4h, %[ab].h[5]\n"
243 "str q9, [%x[c_ptr], #0xc0]\n"
244 "umlal2 v26.4s, %[b2].8h, %[ab].h[5]\n"
245 "str q17, [%x[c_ptr], #0xd0]\n"
246 "umlal2 v11.4s, %[b1].8h, %[ab].h[6]\n"
247 "str q25, [%x[c_ptr], #0xe0]\n"
248 "umlal v19.4s, %[b2].4h, %[ab].h[6]\n"
249 "str q10, [%x[c_ptr], #0xf0]\n"
250 "umlal2 v27.4s, %[b2].8h, %[ab].h[6]\n"
251 "str q18, [%x[c_ptr], #0x100]\n"
252 "umlal2 v12.4s, %[b1].8h, %[ab].h[7]\n"
253 "str q26, [%x[c_ptr], #0x110]\n"
254 "umlal v20.4s, %[b2].4h, %[ab].h[7]\n"
255 "str q11, [%x[c_ptr], #0x120]\n"
256 "umlal2 v28.4s, %[b2].8h, %[ab].h[7]\n"
257 "str q19, [%x[c_ptr], #0x130]\n"
258 "b 4f\n" // Complete write out
259
260 "3:\n" // Odd tail
261 "umlal v5.4s, %[b0].4h, %[aa].h[0]\n"
262 "umlal2 v13.4s, %[b0].8h, %[aa].h[0]\n"
263 "umlal v21.4s, %[b1].4h, %[aa].h[0]\n"
264 "umlal v6.4s, %[b0].4h, %[aa].h[1]\n"
265 "umlal2 v14.4s, %[b0].8h, %[aa].h[1]\n"
266 "umlal v22.4s, %[b1].4h, %[aa].h[1]\n"
267 "str q5, [%x[c_ptr]]\n"
268 "umlal v7.4s, %[b0].4h, %[aa].h[2]\n"
269 "str q13, [%x[c_ptr], #0x10]\n"
270 "umlal2 v15.4s, %[b0].8h, %[aa].h[2]\n"
271 "str q21, [%x[c_ptr], #0x20]\n"
272 "umlal v23.4s, %[b1].4h, %[aa].h[2]\n"
273 "str q6, [%x[c_ptr], #0x30]\n"
274 "umlal v8.4s, %[b0].4h, %[aa].h[3]\n"
275 "str q14, [%x[c_ptr], #0x40]\n"
276 "umlal2 v16.4s, %[b0].8h, %[aa].h[3]\n"
277 "str q22, [%x[c_ptr], #0x50]\n"
278 "umlal v24.4s, %[b1].4h, %[aa].h[3]\n"
279 "str q7, [%x[c_ptr], #0x60]\n"
280 "umlal v9.4s, %[b0].4h, %[aa].h[4]\n"
281 "str q15, [%x[c_ptr], #0x70]\n"
282 "umlal2 v17.4s, %[b0].8h, %[aa].h[4]\n"
283 "str q23, [%x[c_ptr], #0x80]\n"
284 "umlal v25.4s, %[b1].4h, %[aa].h[4]\n"
285 "str q8, [%x[c_ptr], #0x90]\n"
286 "umlal v10.4s, %[b0].4h, %[aa].h[5]\n"
287 "str q16, [%x[c_ptr], #0xa0]\n"
288 "umlal2 v18.4s, %[b0].8h, %[aa].h[5]\n"
289 "str q24, [%x[c_ptr], #0xb0]\n"
290 "umlal v26.4s, %[b1].4h, %[aa].h[5]\n"
291 "str q9, [%x[c_ptr], #0xc0]\n"
292 "umlal v11.4s, %[b0].4h, %[aa].h[6]\n"
293 "str q17, [%x[c_ptr], #0xd0]\n"
294 "umlal2 v19.4s, %[b0].8h, %[aa].h[6]\n"
295 "str q25, [%x[c_ptr], #0xe0]\n"
296 "umlal v27.4s, %[b1].4h, %[aa].h[6]\n"
297 "str q10, [%x[c_ptr], #0xf0]\n"
298 "umlal v12.4s, %[b0].4h, %[aa].h[7]\n"
299 "str q18, [%x[c_ptr], #0x100]\n"
300 "umlal2 v20.4s, %[b0].8h, %[aa].h[7]\n"
301 "str q26, [%x[c_ptr], #0x110]\n"
302 "umlal v28.4s, %[b1].4h, %[aa].h[7]\n"
303 "str q11, [%x[c_ptr], #0x120]\n"
304
305 "4:\n" // End of function
306 "str q19, [%x[c_ptr], #0x130]\n"
307 "str q27, [%x[c_ptr], #0x140]\n"
308 "str q12, [%x[c_ptr], #0x150]\n"
309 "str q20, [%x[c_ptr], #0x160]\n"
310 "str q28, [%x[c_ptr], #0x170]\n"
311 "add %x[c_ptr], %x[c_ptr], #0x180\n"
312 : [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr), [k] "+r" (k),
313 [aa] "+w" (aa), [ab] "+w" (ab), [b0] "+w" (b0), [b1] "+w" (b1), [b2] "+w" (b2)
314 : [odd_k] "r" (odd_k)
315 : "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "x20", "cc", "memory"
316 );
317 }
318 }
319 }
320
321 } // namespace arm_gemm
322
323 #endif
324