?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
mm: hugetlb: fix hugepage memory leak caused by wrong reserve count
[linux/fpc-iii.git] / drivers / gpu / drm / amd / amdgpu / gfx_v7_0.c
blob72793f93e2fcc9989e5f936ec1d43ba0593ac69e
1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23 #include <linux/firmware.h>
24 #include "drmP.h"
25 #include "amdgpu.h"
26 #include "amdgpu_ih.h"
27 #include "amdgpu_gfx.h"
28 #include "cikd.h"
29 #include "cik.h"
30 #include "atom.h"
31 #include "amdgpu_ucode.h"
32 #include "clearstate_ci.h"
33
34 #include "uvd/uvd_4_2_d.h"
35
36 #include "dce/dce_8_0_d.h"
37 #include "dce/dce_8_0_sh_mask.h"
38
39 #include "bif/bif_4_1_d.h"
40 #include "bif/bif_4_1_sh_mask.h"
41
42 #include "gca/gfx_7_0_d.h"
43 #include "gca/gfx_7_2_enum.h"
44 #include "gca/gfx_7_2_sh_mask.h"
45
46 #include "gmc/gmc_7_0_d.h"
47 #include "gmc/gmc_7_0_sh_mask.h"
48
49 #include "oss/oss_2_0_d.h"
50 #include "oss/oss_2_0_sh_mask.h"
51
52 #define GFX7_NUM_GFX_RINGS 1
53 #define GFX7_NUM_COMPUTE_RINGS 8
54
55 static void gfx_v7_0_set_ring_funcs(struct amdgpu_device *adev);
56 static void gfx_v7_0_set_irq_funcs(struct amdgpu_device *adev);
57 static void gfx_v7_0_set_gds_init(struct amdgpu_device *adev);
58 int gfx_v7_0_get_cu_info(struct amdgpu_device *, struct amdgpu_cu_info *);
59
60 MODULE_FIRMWARE("radeon/bonaire_pfp.bin");
61 MODULE_FIRMWARE("radeon/bonaire_me.bin");
62 MODULE_FIRMWARE("radeon/bonaire_ce.bin");
63 MODULE_FIRMWARE("radeon/bonaire_rlc.bin");
64 MODULE_FIRMWARE("radeon/bonaire_mec.bin");
65
66 MODULE_FIRMWARE("radeon/hawaii_pfp.bin");
67 MODULE_FIRMWARE("radeon/hawaii_me.bin");
68 MODULE_FIRMWARE("radeon/hawaii_ce.bin");
69 MODULE_FIRMWARE("radeon/hawaii_rlc.bin");
70 MODULE_FIRMWARE("radeon/hawaii_mec.bin");
71
72 MODULE_FIRMWARE("radeon/kaveri_pfp.bin");
73 MODULE_FIRMWARE("radeon/kaveri_me.bin");
74 MODULE_FIRMWARE("radeon/kaveri_ce.bin");
75 MODULE_FIRMWARE("radeon/kaveri_rlc.bin");
76 MODULE_FIRMWARE("radeon/kaveri_mec.bin");
77 MODULE_FIRMWARE("radeon/kaveri_mec2.bin");
78
79 MODULE_FIRMWARE("radeon/kabini_pfp.bin");
80 MODULE_FIRMWARE("radeon/kabini_me.bin");
81 MODULE_FIRMWARE("radeon/kabini_ce.bin");
82 MODULE_FIRMWARE("radeon/kabini_rlc.bin");
83 MODULE_FIRMWARE("radeon/kabini_mec.bin");
84
85 MODULE_FIRMWARE("radeon/mullins_pfp.bin");
86 MODULE_FIRMWARE("radeon/mullins_me.bin");
87 MODULE_FIRMWARE("radeon/mullins_ce.bin");
88 MODULE_FIRMWARE("radeon/mullins_rlc.bin");
89 MODULE_FIRMWARE("radeon/mullins_mec.bin");
90
91 static const struct amdgpu_gds_reg_offset amdgpu_gds_reg_offset[] =
92 {
93 {mmGDS_VMID0_BASE, mmGDS_VMID0_SIZE, mmGDS_GWS_VMID0, mmGDS_OA_VMID0},
94 {mmGDS_VMID1_BASE, mmGDS_VMID1_SIZE, mmGDS_GWS_VMID1, mmGDS_OA_VMID1},
95 {mmGDS_VMID2_BASE, mmGDS_VMID2_SIZE, mmGDS_GWS_VMID2, mmGDS_OA_VMID2},
96 {mmGDS_VMID3_BASE, mmGDS_VMID3_SIZE, mmGDS_GWS_VMID3, mmGDS_OA_VMID3},
97 {mmGDS_VMID4_BASE, mmGDS_VMID4_SIZE, mmGDS_GWS_VMID4, mmGDS_OA_VMID4},
98 {mmGDS_VMID5_BASE, mmGDS_VMID5_SIZE, mmGDS_GWS_VMID5, mmGDS_OA_VMID5},
99 {mmGDS_VMID6_BASE, mmGDS_VMID6_SIZE, mmGDS_GWS_VMID6, mmGDS_OA_VMID6},
100 {mmGDS_VMID7_BASE, mmGDS_VMID7_SIZE, mmGDS_GWS_VMID7, mmGDS_OA_VMID7},
101 {mmGDS_VMID8_BASE, mmGDS_VMID8_SIZE, mmGDS_GWS_VMID8, mmGDS_OA_VMID8},
102 {mmGDS_VMID9_BASE, mmGDS_VMID9_SIZE, mmGDS_GWS_VMID9, mmGDS_OA_VMID9},
103 {mmGDS_VMID10_BASE, mmGDS_VMID10_SIZE, mmGDS_GWS_VMID10, mmGDS_OA_VMID10},
104 {mmGDS_VMID11_BASE, mmGDS_VMID11_SIZE, mmGDS_GWS_VMID11, mmGDS_OA_VMID11},
105 {mmGDS_VMID12_BASE, mmGDS_VMID12_SIZE, mmGDS_GWS_VMID12, mmGDS_OA_VMID12},
106 {mmGDS_VMID13_BASE, mmGDS_VMID13_SIZE, mmGDS_GWS_VMID13, mmGDS_OA_VMID13},
107 {mmGDS_VMID14_BASE, mmGDS_VMID14_SIZE, mmGDS_GWS_VMID14, mmGDS_OA_VMID14},
108 {mmGDS_VMID15_BASE, mmGDS_VMID15_SIZE, mmGDS_GWS_VMID15, mmGDS_OA_VMID15}
109 };
110
111 static const u32 spectre_rlc_save_restore_register_list[] =
112 {
113 (0x0e00 << 16) | (0xc12c >> 2),
114 0x00000000,
115 (0x0e00 << 16) | (0xc140 >> 2),
116 0x00000000,
117 (0x0e00 << 16) | (0xc150 >> 2),
118 0x00000000,
119 (0x0e00 << 16) | (0xc15c >> 2),
120 0x00000000,
121 (0x0e00 << 16) | (0xc168 >> 2),
122 0x00000000,
123 (0x0e00 << 16) | (0xc170 >> 2),
124 0x00000000,
125 (0x0e00 << 16) | (0xc178 >> 2),
126 0x00000000,
127 (0x0e00 << 16) | (0xc204 >> 2),
128 0x00000000,
129 (0x0e00 << 16) | (0xc2b4 >> 2),
130 0x00000000,
131 (0x0e00 << 16) | (0xc2b8 >> 2),
132 0x00000000,
133 (0x0e00 << 16) | (0xc2bc >> 2),
134 0x00000000,
135 (0x0e00 << 16) | (0xc2c0 >> 2),
136 0x00000000,
137 (0x0e00 << 16) | (0x8228 >> 2),
138 0x00000000,
139 (0x0e00 << 16) | (0x829c >> 2),
140 0x00000000,
141 (0x0e00 << 16) | (0x869c >> 2),
142 0x00000000,
143 (0x0600 << 16) | (0x98f4 >> 2),
144 0x00000000,
145 (0x0e00 << 16) | (0x98f8 >> 2),
146 0x00000000,
147 (0x0e00 << 16) | (0x9900 >> 2),
148 0x00000000,
149 (0x0e00 << 16) | (0xc260 >> 2),
150 0x00000000,
151 (0x0e00 << 16) | (0x90e8 >> 2),
152 0x00000000,
153 (0x0e00 << 16) | (0x3c000 >> 2),
154 0x00000000,
155 (0x0e00 << 16) | (0x3c00c >> 2),
156 0x00000000,
157 (0x0e00 << 16) | (0x8c1c >> 2),
158 0x00000000,
159 (0x0e00 << 16) | (0x9700 >> 2),
160 0x00000000,
161 (0x0e00 << 16) | (0xcd20 >> 2),
162 0x00000000,
163 (0x4e00 << 16) | (0xcd20 >> 2),
164 0x00000000,
165 (0x5e00 << 16) | (0xcd20 >> 2),
166 0x00000000,
167 (0x6e00 << 16) | (0xcd20 >> 2),
168 0x00000000,
169 (0x7e00 << 16) | (0xcd20 >> 2),
170 0x00000000,
171 (0x8e00 << 16) | (0xcd20 >> 2),
172 0x00000000,
173 (0x9e00 << 16) | (0xcd20 >> 2),
174 0x00000000,
175 (0xae00 << 16) | (0xcd20 >> 2),
176 0x00000000,
177 (0xbe00 << 16) | (0xcd20 >> 2),
178 0x00000000,
179 (0x0e00 << 16) | (0x89bc >> 2),
180 0x00000000,
181 (0x0e00 << 16) | (0x8900 >> 2),
182 0x00000000,
183 0x3,
184 (0x0e00 << 16) | (0xc130 >> 2),
185 0x00000000,
186 (0x0e00 << 16) | (0xc134 >> 2),
187 0x00000000,
188 (0x0e00 << 16) | (0xc1fc >> 2),
189 0x00000000,
190 (0x0e00 << 16) | (0xc208 >> 2),
191 0x00000000,
192 (0x0e00 << 16) | (0xc264 >> 2),
193 0x00000000,
194 (0x0e00 << 16) | (0xc268 >> 2),
195 0x00000000,
196 (0x0e00 << 16) | (0xc26c >> 2),
197 0x00000000,
198 (0x0e00 << 16) | (0xc270 >> 2),
199 0x00000000,
200 (0x0e00 << 16) | (0xc274 >> 2),
201 0x00000000,
202 (0x0e00 << 16) | (0xc278 >> 2),
203 0x00000000,
204 (0x0e00 << 16) | (0xc27c >> 2),
205 0x00000000,
206 (0x0e00 << 16) | (0xc280 >> 2),
207 0x00000000,
208 (0x0e00 << 16) | (0xc284 >> 2),
209 0x00000000,
210 (0x0e00 << 16) | (0xc288 >> 2),
211 0x00000000,
212 (0x0e00 << 16) | (0xc28c >> 2),
213 0x00000000,
214 (0x0e00 << 16) | (0xc290 >> 2),
215 0x00000000,
216 (0x0e00 << 16) | (0xc294 >> 2),
217 0x00000000,
218 (0x0e00 << 16) | (0xc298 >> 2),
219 0x00000000,
220 (0x0e00 << 16) | (0xc29c >> 2),
221 0x00000000,
222 (0x0e00 << 16) | (0xc2a0 >> 2),
223 0x00000000,
224 (0x0e00 << 16) | (0xc2a4 >> 2),
225 0x00000000,
226 (0x0e00 << 16) | (0xc2a8 >> 2),
227 0x00000000,
228 (0x0e00 << 16) | (0xc2ac >> 2),
229 0x00000000,
230 (0x0e00 << 16) | (0xc2b0 >> 2),
231 0x00000000,
232 (0x0e00 << 16) | (0x301d0 >> 2),
233 0x00000000,
234 (0x0e00 << 16) | (0x30238 >> 2),
235 0x00000000,
236 (0x0e00 << 16) | (0x30250 >> 2),
237 0x00000000,
238 (0x0e00 << 16) | (0x30254 >> 2),
239 0x00000000,
240 (0x0e00 << 16) | (0x30258 >> 2),
241 0x00000000,
242 (0x0e00 << 16) | (0x3025c >> 2),
243 0x00000000,
244 (0x4e00 << 16) | (0xc900 >> 2),
245 0x00000000,
246 (0x5e00 << 16) | (0xc900 >> 2),
247 0x00000000,
248 (0x6e00 << 16) | (0xc900 >> 2),
249 0x00000000,
250 (0x7e00 << 16) | (0xc900 >> 2),
251 0x00000000,
252 (0x8e00 << 16) | (0xc900 >> 2),
253 0x00000000,
254 (0x9e00 << 16) | (0xc900 >> 2),
255 0x00000000,
256 (0xae00 << 16) | (0xc900 >> 2),
257 0x00000000,
258 (0xbe00 << 16) | (0xc900 >> 2),
259 0x00000000,
260 (0x4e00 << 16) | (0xc904 >> 2),
261 0x00000000,
262 (0x5e00 << 16) | (0xc904 >> 2),
263 0x00000000,
264 (0x6e00 << 16) | (0xc904 >> 2),
265 0x00000000,
266 (0x7e00 << 16) | (0xc904 >> 2),
267 0x00000000,
268 (0x8e00 << 16) | (0xc904 >> 2),
269 0x00000000,
270 (0x9e00 << 16) | (0xc904 >> 2),
271 0x00000000,
272 (0xae00 << 16) | (0xc904 >> 2),
273 0x00000000,
274 (0xbe00 << 16) | (0xc904 >> 2),
275 0x00000000,
276 (0x4e00 << 16) | (0xc908 >> 2),
277 0x00000000,
278 (0x5e00 << 16) | (0xc908 >> 2),
279 0x00000000,
280 (0x6e00 << 16) | (0xc908 >> 2),
281 0x00000000,
282 (0x7e00 << 16) | (0xc908 >> 2),
283 0x00000000,
284 (0x8e00 << 16) | (0xc908 >> 2),
285 0x00000000,
286 (0x9e00 << 16) | (0xc908 >> 2),
287 0x00000000,
288 (0xae00 << 16) | (0xc908 >> 2),
289 0x00000000,
290 (0xbe00 << 16) | (0xc908 >> 2),
291 0x00000000,
292 (0x4e00 << 16) | (0xc90c >> 2),
293 0x00000000,
294 (0x5e00 << 16) | (0xc90c >> 2),
295 0x00000000,
296 (0x6e00 << 16) | (0xc90c >> 2),
297 0x00000000,
298 (0x7e00 << 16) | (0xc90c >> 2),
299 0x00000000,
300 (0x8e00 << 16) | (0xc90c >> 2),
301 0x00000000,
302 (0x9e00 << 16) | (0xc90c >> 2),
303 0x00000000,
304 (0xae00 << 16) | (0xc90c >> 2),
305 0x00000000,
306 (0xbe00 << 16) | (0xc90c >> 2),
307 0x00000000,
308 (0x4e00 << 16) | (0xc910 >> 2),
309 0x00000000,
310 (0x5e00 << 16) | (0xc910 >> 2),
311 0x00000000,
312 (0x6e00 << 16) | (0xc910 >> 2),
313 0x00000000,
314 (0x7e00 << 16) | (0xc910 >> 2),
315 0x00000000,
316 (0x8e00 << 16) | (0xc910 >> 2),
317 0x00000000,
318 (0x9e00 << 16) | (0xc910 >> 2),
319 0x00000000,
320 (0xae00 << 16) | (0xc910 >> 2),
321 0x00000000,
322 (0xbe00 << 16) | (0xc910 >> 2),
323 0x00000000,
324 (0x0e00 << 16) | (0xc99c >> 2),
325 0x00000000,
326 (0x0e00 << 16) | (0x9834 >> 2),
327 0x00000000,
328 (0x0000 << 16) | (0x30f00 >> 2),
329 0x00000000,
330 (0x0001 << 16) | (0x30f00 >> 2),
331 0x00000000,
332 (0x0000 << 16) | (0x30f04 >> 2),
333 0x00000000,
334 (0x0001 << 16) | (0x30f04 >> 2),
335 0x00000000,
336 (0x0000 << 16) | (0x30f08 >> 2),
337 0x00000000,
338 (0x0001 << 16) | (0x30f08 >> 2),
339 0x00000000,
340 (0x0000 << 16) | (0x30f0c >> 2),
341 0x00000000,
342 (0x0001 << 16) | (0x30f0c >> 2),
343 0x00000000,
344 (0x0600 << 16) | (0x9b7c >> 2),
345 0x00000000,
346 (0x0e00 << 16) | (0x8a14 >> 2),
347 0x00000000,
348 (0x0e00 << 16) | (0x8a18 >> 2),
349 0x00000000,
350 (0x0600 << 16) | (0x30a00 >> 2),
351 0x00000000,
352 (0x0e00 << 16) | (0x8bf0 >> 2),
353 0x00000000,
354 (0x0e00 << 16) | (0x8bcc >> 2),
355 0x00000000,
356 (0x0e00 << 16) | (0x8b24 >> 2),
357 0x00000000,
358 (0x0e00 << 16) | (0x30a04 >> 2),
359 0x00000000,
360 (0x0600 << 16) | (0x30a10 >> 2),
361 0x00000000,
362 (0x0600 << 16) | (0x30a14 >> 2),
363 0x00000000,
364 (0x0600 << 16) | (0x30a18 >> 2),
365 0x00000000,
366 (0x0600 << 16) | (0x30a2c >> 2),
367 0x00000000,
368 (0x0e00 << 16) | (0xc700 >> 2),
369 0x00000000,
370 (0x0e00 << 16) | (0xc704 >> 2),
371 0x00000000,
372 (0x0e00 << 16) | (0xc708 >> 2),
373 0x00000000,
374 (0x0e00 << 16) | (0xc768 >> 2),
375 0x00000000,
376 (0x0400 << 16) | (0xc770 >> 2),
377 0x00000000,
378 (0x0400 << 16) | (0xc774 >> 2),
379 0x00000000,
380 (0x0400 << 16) | (0xc778 >> 2),
381 0x00000000,
382 (0x0400 << 16) | (0xc77c >> 2),
383 0x00000000,
384 (0x0400 << 16) | (0xc780 >> 2),
385 0x00000000,
386 (0x0400 << 16) | (0xc784 >> 2),
387 0x00000000,
388 (0x0400 << 16) | (0xc788 >> 2),
389 0x00000000,
390 (0x0400 << 16) | (0xc78c >> 2),
391 0x00000000,
392 (0x0400 << 16) | (0xc798 >> 2),
393 0x00000000,
394 (0x0400 << 16) | (0xc79c >> 2),
395 0x00000000,
396 (0x0400 << 16) | (0xc7a0 >> 2),
397 0x00000000,
398 (0x0400 << 16) | (0xc7a4 >> 2),
399 0x00000000,
400 (0x0400 << 16) | (0xc7a8 >> 2),
401 0x00000000,
402 (0x0400 << 16) | (0xc7ac >> 2),
403 0x00000000,
404 (0x0400 << 16) | (0xc7b0 >> 2),
405 0x00000000,
406 (0x0400 << 16) | (0xc7b4 >> 2),
407 0x00000000,
408 (0x0e00 << 16) | (0x9100 >> 2),
409 0x00000000,
410 (0x0e00 << 16) | (0x3c010 >> 2),
411 0x00000000,
412 (0x0e00 << 16) | (0x92a8 >> 2),
413 0x00000000,
414 (0x0e00 << 16) | (0x92ac >> 2),
415 0x00000000,
416 (0x0e00 << 16) | (0x92b4 >> 2),
417 0x00000000,
418 (0x0e00 << 16) | (0x92b8 >> 2),
419 0x00000000,
420 (0x0e00 << 16) | (0x92bc >> 2),
421 0x00000000,
422 (0x0e00 << 16) | (0x92c0 >> 2),
423 0x00000000,
424 (0x0e00 << 16) | (0x92c4 >> 2),
425 0x00000000,
426 (0x0e00 << 16) | (0x92c8 >> 2),
427 0x00000000,
428 (0x0e00 << 16) | (0x92cc >> 2),
429 0x00000000,
430 (0x0e00 << 16) | (0x92d0 >> 2),
431 0x00000000,
432 (0x0e00 << 16) | (0x8c00 >> 2),
433 0x00000000,
434 (0x0e00 << 16) | (0x8c04 >> 2),
435 0x00000000,
436 (0x0e00 << 16) | (0x8c20 >> 2),
437 0x00000000,
438 (0x0e00 << 16) | (0x8c38 >> 2),
439 0x00000000,
440 (0x0e00 << 16) | (0x8c3c >> 2),
441 0x00000000,
442 (0x0e00 << 16) | (0xae00 >> 2),
443 0x00000000,
444 (0x0e00 << 16) | (0x9604 >> 2),
445 0x00000000,
446 (0x0e00 << 16) | (0xac08 >> 2),
447 0x00000000,
448 (0x0e00 << 16) | (0xac0c >> 2),
449 0x00000000,
450 (0x0e00 << 16) | (0xac10 >> 2),
451 0x00000000,
452 (0x0e00 << 16) | (0xac14 >> 2),
453 0x00000000,
454 (0x0e00 << 16) | (0xac58 >> 2),
455 0x00000000,
456 (0x0e00 << 16) | (0xac68 >> 2),
457 0x00000000,
458 (0x0e00 << 16) | (0xac6c >> 2),
459 0x00000000,
460 (0x0e00 << 16) | (0xac70 >> 2),
461 0x00000000,
462 (0x0e00 << 16) | (0xac74 >> 2),
463 0x00000000,
464 (0x0e00 << 16) | (0xac78 >> 2),
465 0x00000000,
466 (0x0e00 << 16) | (0xac7c >> 2),
467 0x00000000,
468 (0x0e00 << 16) | (0xac80 >> 2),
469 0x00000000,
470 (0x0e00 << 16) | (0xac84 >> 2),
471 0x00000000,
472 (0x0e00 << 16) | (0xac88 >> 2),
473 0x00000000,
474 (0x0e00 << 16) | (0xac8c >> 2),
475 0x00000000,
476 (0x0e00 << 16) | (0x970c >> 2),
477 0x00000000,
478 (0x0e00 << 16) | (0x9714 >> 2),
479 0x00000000,
480 (0x0e00 << 16) | (0x9718 >> 2),
481 0x00000000,
482 (0x0e00 << 16) | (0x971c >> 2),
483 0x00000000,
484 (0x0e00 << 16) | (0x31068 >> 2),
485 0x00000000,
486 (0x4e00 << 16) | (0x31068 >> 2),
487 0x00000000,
488 (0x5e00 << 16) | (0x31068 >> 2),
489 0x00000000,
490 (0x6e00 << 16) | (0x31068 >> 2),
491 0x00000000,
492 (0x7e00 << 16) | (0x31068 >> 2),
493 0x00000000,
494 (0x8e00 << 16) | (0x31068 >> 2),
495 0x00000000,
496 (0x9e00 << 16) | (0x31068 >> 2),
497 0x00000000,
498 (0xae00 << 16) | (0x31068 >> 2),
499 0x00000000,
500 (0xbe00 << 16) | (0x31068 >> 2),
501 0x00000000,
502 (0x0e00 << 16) | (0xcd10 >> 2),
503 0x00000000,
504 (0x0e00 << 16) | (0xcd14 >> 2),
505 0x00000000,
506 (0x0e00 << 16) | (0x88b0 >> 2),
507 0x00000000,
508 (0x0e00 << 16) | (0x88b4 >> 2),
509 0x00000000,
510 (0x0e00 << 16) | (0x88b8 >> 2),
511 0x00000000,
512 (0x0e00 << 16) | (0x88bc >> 2),
513 0x00000000,
514 (0x0400 << 16) | (0x89c0 >> 2),
515 0x00000000,
516 (0x0e00 << 16) | (0x88c4 >> 2),
517 0x00000000,
518 (0x0e00 << 16) | (0x88c8 >> 2),
519 0x00000000,
520 (0x0e00 << 16) | (0x88d0 >> 2),
521 0x00000000,
522 (0x0e00 << 16) | (0x88d4 >> 2),
523 0x00000000,
524 (0x0e00 << 16) | (0x88d8 >> 2),
525 0x00000000,
526 (0x0e00 << 16) | (0x8980 >> 2),
527 0x00000000,
528 (0x0e00 << 16) | (0x30938 >> 2),
529 0x00000000,
530 (0x0e00 << 16) | (0x3093c >> 2),
531 0x00000000,
532 (0x0e00 << 16) | (0x30940 >> 2),
533 0x00000000,
534 (0x0e00 << 16) | (0x89a0 >> 2),
535 0x00000000,
536 (0x0e00 << 16) | (0x30900 >> 2),
537 0x00000000,
538 (0x0e00 << 16) | (0x30904 >> 2),
539 0x00000000,
540 (0x0e00 << 16) | (0x89b4 >> 2),
541 0x00000000,
542 (0x0e00 << 16) | (0x3c210 >> 2),
543 0x00000000,
544 (0x0e00 << 16) | (0x3c214 >> 2),
545 0x00000000,
546 (0x0e00 << 16) | (0x3c218 >> 2),
547 0x00000000,
548 (0x0e00 << 16) | (0x8904 >> 2),
549 0x00000000,
550 0x5,
551 (0x0e00 << 16) | (0x8c28 >> 2),
552 (0x0e00 << 16) | (0x8c2c >> 2),
553 (0x0e00 << 16) | (0x8c30 >> 2),
554 (0x0e00 << 16) | (0x8c34 >> 2),
555 (0x0e00 << 16) | (0x9600 >> 2),
556 };
557
558 static const u32 kalindi_rlc_save_restore_register_list[] =
559 {
560 (0x0e00 << 16) | (0xc12c >> 2),
561 0x00000000,
562 (0x0e00 << 16) | (0xc140 >> 2),
563 0x00000000,
564 (0x0e00 << 16) | (0xc150 >> 2),
565 0x00000000,
566 (0x0e00 << 16) | (0xc15c >> 2),
567 0x00000000,
568 (0x0e00 << 16) | (0xc168 >> 2),
569 0x00000000,
570 (0x0e00 << 16) | (0xc170 >> 2),
571 0x00000000,
572 (0x0e00 << 16) | (0xc204 >> 2),
573 0x00000000,
574 (0x0e00 << 16) | (0xc2b4 >> 2),
575 0x00000000,
576 (0x0e00 << 16) | (0xc2b8 >> 2),
577 0x00000000,
578 (0x0e00 << 16) | (0xc2bc >> 2),
579 0x00000000,
580 (0x0e00 << 16) | (0xc2c0 >> 2),
581 0x00000000,
582 (0x0e00 << 16) | (0x8228 >> 2),
583 0x00000000,
584 (0x0e00 << 16) | (0x829c >> 2),
585 0x00000000,
586 (0x0e00 << 16) | (0x869c >> 2),
587 0x00000000,
588 (0x0600 << 16) | (0x98f4 >> 2),
589 0x00000000,
590 (0x0e00 << 16) | (0x98f8 >> 2),
591 0x00000000,
592 (0x0e00 << 16) | (0x9900 >> 2),
593 0x00000000,
594 (0x0e00 << 16) | (0xc260 >> 2),
595 0x00000000,
596 (0x0e00 << 16) | (0x90e8 >> 2),
597 0x00000000,
598 (0x0e00 << 16) | (0x3c000 >> 2),
599 0x00000000,
600 (0x0e00 << 16) | (0x3c00c >> 2),
601 0x00000000,
602 (0x0e00 << 16) | (0x8c1c >> 2),
603 0x00000000,
604 (0x0e00 << 16) | (0x9700 >> 2),
605 0x00000000,
606 (0x0e00 << 16) | (0xcd20 >> 2),
607 0x00000000,
608 (0x4e00 << 16) | (0xcd20 >> 2),
609 0x00000000,
610 (0x5e00 << 16) | (0xcd20 >> 2),
611 0x00000000,
612 (0x6e00 << 16) | (0xcd20 >> 2),
613 0x00000000,
614 (0x7e00 << 16) | (0xcd20 >> 2),
615 0x00000000,
616 (0x0e00 << 16) | (0x89bc >> 2),
617 0x00000000,
618 (0x0e00 << 16) | (0x8900 >> 2),
619 0x00000000,
620 0x3,
621 (0x0e00 << 16) | (0xc130 >> 2),
622 0x00000000,
623 (0x0e00 << 16) | (0xc134 >> 2),
624 0x00000000,
625 (0x0e00 << 16) | (0xc1fc >> 2),
626 0x00000000,
627 (0x0e00 << 16) | (0xc208 >> 2),
628 0x00000000,
629 (0x0e00 << 16) | (0xc264 >> 2),
630 0x00000000,
631 (0x0e00 << 16) | (0xc268 >> 2),
632 0x00000000,
633 (0x0e00 << 16) | (0xc26c >> 2),
634 0x00000000,
635 (0x0e00 << 16) | (0xc270 >> 2),
636 0x00000000,
637 (0x0e00 << 16) | (0xc274 >> 2),
638 0x00000000,
639 (0x0e00 << 16) | (0xc28c >> 2),
640 0x00000000,
641 (0x0e00 << 16) | (0xc290 >> 2),
642 0x00000000,
643 (0x0e00 << 16) | (0xc294 >> 2),
644 0x00000000,
645 (0x0e00 << 16) | (0xc298 >> 2),
646 0x00000000,
647 (0x0e00 << 16) | (0xc2a0 >> 2),
648 0x00000000,
649 (0x0e00 << 16) | (0xc2a4 >> 2),
650 0x00000000,
651 (0x0e00 << 16) | (0xc2a8 >> 2),
652 0x00000000,
653 (0x0e00 << 16) | (0xc2ac >> 2),
654 0x00000000,
655 (0x0e00 << 16) | (0x301d0 >> 2),
656 0x00000000,
657 (0x0e00 << 16) | (0x30238 >> 2),
658 0x00000000,
659 (0x0e00 << 16) | (0x30250 >> 2),
660 0x00000000,
661 (0x0e00 << 16) | (0x30254 >> 2),
662 0x00000000,
663 (0x0e00 << 16) | (0x30258 >> 2),
664 0x00000000,
665 (0x0e00 << 16) | (0x3025c >> 2),
666 0x00000000,
667 (0x4e00 << 16) | (0xc900 >> 2),
668 0x00000000,
669 (0x5e00 << 16) | (0xc900 >> 2),
670 0x00000000,
671 (0x6e00 << 16) | (0xc900 >> 2),
672 0x00000000,
673 (0x7e00 << 16) | (0xc900 >> 2),
674 0x00000000,
675 (0x4e00 << 16) | (0xc904 >> 2),
676 0x00000000,
677 (0x5e00 << 16) | (0xc904 >> 2),
678 0x00000000,
679 (0x6e00 << 16) | (0xc904 >> 2),
680 0x00000000,
681 (0x7e00 << 16) | (0xc904 >> 2),
682 0x00000000,
683 (0x4e00 << 16) | (0xc908 >> 2),
684 0x00000000,
685 (0x5e00 << 16) | (0xc908 >> 2),
686 0x00000000,
687 (0x6e00 << 16) | (0xc908 >> 2),
688 0x00000000,
689 (0x7e00 << 16) | (0xc908 >> 2),
690 0x00000000,
691 (0x4e00 << 16) | (0xc90c >> 2),
692 0x00000000,
693 (0x5e00 << 16) | (0xc90c >> 2),
694 0x00000000,
695 (0x6e00 << 16) | (0xc90c >> 2),
696 0x00000000,
697 (0x7e00 << 16) | (0xc90c >> 2),
698 0x00000000,
699 (0x4e00 << 16) | (0xc910 >> 2),
700 0x00000000,
701 (0x5e00 << 16) | (0xc910 >> 2),
702 0x00000000,
703 (0x6e00 << 16) | (0xc910 >> 2),
704 0x00000000,
705 (0x7e00 << 16) | (0xc910 >> 2),
706 0x00000000,
707 (0x0e00 << 16) | (0xc99c >> 2),
708 0x00000000,
709 (0x0e00 << 16) | (0x9834 >> 2),
710 0x00000000,
711 (0x0000 << 16) | (0x30f00 >> 2),
712 0x00000000,
713 (0x0000 << 16) | (0x30f04 >> 2),
714 0x00000000,
715 (0x0000 << 16) | (0x30f08 >> 2),
716 0x00000000,
717 (0x0000 << 16) | (0x30f0c >> 2),
718 0x00000000,
719 (0x0600 << 16) | (0x9b7c >> 2),
720 0x00000000,
721 (0x0e00 << 16) | (0x8a14 >> 2),
722 0x00000000,
723 (0x0e00 << 16) | (0x8a18 >> 2),
724 0x00000000,
725 (0x0600 << 16) | (0x30a00 >> 2),
726 0x00000000,
727 (0x0e00 << 16) | (0x8bf0 >> 2),
728 0x00000000,
729 (0x0e00 << 16) | (0x8bcc >> 2),
730 0x00000000,
731 (0x0e00 << 16) | (0x8b24 >> 2),
732 0x00000000,
733 (0x0e00 << 16) | (0x30a04 >> 2),
734 0x00000000,
735 (0x0600 << 16) | (0x30a10 >> 2),
736 0x00000000,
737 (0x0600 << 16) | (0x30a14 >> 2),
738 0x00000000,
739 (0x0600 << 16) | (0x30a18 >> 2),
740 0x00000000,
741 (0x0600 << 16) | (0x30a2c >> 2),
742 0x00000000,
743 (0x0e00 << 16) | (0xc700 >> 2),
744 0x00000000,
745 (0x0e00 << 16) | (0xc704 >> 2),
746 0x00000000,
747 (0x0e00 << 16) | (0xc708 >> 2),
748 0x00000000,
749 (0x0e00 << 16) | (0xc768 >> 2),
750 0x00000000,
751 (0x0400 << 16) | (0xc770 >> 2),
752 0x00000000,
753 (0x0400 << 16) | (0xc774 >> 2),
754 0x00000000,
755 (0x0400 << 16) | (0xc798 >> 2),
756 0x00000000,
757 (0x0400 << 16) | (0xc79c >> 2),
758 0x00000000,
759 (0x0e00 << 16) | (0x9100 >> 2),
760 0x00000000,
761 (0x0e00 << 16) | (0x3c010 >> 2),
762 0x00000000,
763 (0x0e00 << 16) | (0x8c00 >> 2),
764 0x00000000,
765 (0x0e00 << 16) | (0x8c04 >> 2),
766 0x00000000,
767 (0x0e00 << 16) | (0x8c20 >> 2),
768 0x00000000,
769 (0x0e00 << 16) | (0x8c38 >> 2),
770 0x00000000,
771 (0x0e00 << 16) | (0x8c3c >> 2),
772 0x00000000,
773 (0x0e00 << 16) | (0xae00 >> 2),
774 0x00000000,
775 (0x0e00 << 16) | (0x9604 >> 2),
776 0x00000000,
777 (0x0e00 << 16) | (0xac08 >> 2),
778 0x00000000,
779 (0x0e00 << 16) | (0xac0c >> 2),
780 0x00000000,
781 (0x0e00 << 16) | (0xac10 >> 2),
782 0x00000000,
783 (0x0e00 << 16) | (0xac14 >> 2),
784 0x00000000,
785 (0x0e00 << 16) | (0xac58 >> 2),
786 0x00000000,
787 (0x0e00 << 16) | (0xac68 >> 2),
788 0x00000000,
789 (0x0e00 << 16) | (0xac6c >> 2),
790 0x00000000,
791 (0x0e00 << 16) | (0xac70 >> 2),
792 0x00000000,
793 (0x0e00 << 16) | (0xac74 >> 2),
794 0x00000000,
795 (0x0e00 << 16) | (0xac78 >> 2),
796 0x00000000,
797 (0x0e00 << 16) | (0xac7c >> 2),
798 0x00000000,
799 (0x0e00 << 16) | (0xac80 >> 2),
800 0x00000000,
801 (0x0e00 << 16) | (0xac84 >> 2),
802 0x00000000,
803 (0x0e00 << 16) | (0xac88 >> 2),
804 0x00000000,
805 (0x0e00 << 16) | (0xac8c >> 2),
806 0x00000000,
807 (0x0e00 << 16) | (0x970c >> 2),
808 0x00000000,
809 (0x0e00 << 16) | (0x9714 >> 2),
810 0x00000000,
811 (0x0e00 << 16) | (0x9718 >> 2),
812 0x00000000,
813 (0x0e00 << 16) | (0x971c >> 2),
814 0x00000000,
815 (0x0e00 << 16) | (0x31068 >> 2),
816 0x00000000,
817 (0x4e00 << 16) | (0x31068 >> 2),
818 0x00000000,
819 (0x5e00 << 16) | (0x31068 >> 2),
820 0x00000000,
821 (0x6e00 << 16) | (0x31068 >> 2),
822 0x00000000,
823 (0x7e00 << 16) | (0x31068 >> 2),
824 0x00000000,
825 (0x0e00 << 16) | (0xcd10 >> 2),
826 0x00000000,
827 (0x0e00 << 16) | (0xcd14 >> 2),
828 0x00000000,
829 (0x0e00 << 16) | (0x88b0 >> 2),
830 0x00000000,
831 (0x0e00 << 16) | (0x88b4 >> 2),
832 0x00000000,
833 (0x0e00 << 16) | (0x88b8 >> 2),
834 0x00000000,
835 (0x0e00 << 16) | (0x88bc >> 2),
836 0x00000000,
837 (0x0400 << 16) | (0x89c0 >> 2),
838 0x00000000,
839 (0x0e00 << 16) | (0x88c4 >> 2),
840 0x00000000,
841 (0x0e00 << 16) | (0x88c8 >> 2),
842 0x00000000,
843 (0x0e00 << 16) | (0x88d0 >> 2),
844 0x00000000,
845 (0x0e00 << 16) | (0x88d4 >> 2),
846 0x00000000,
847 (0x0e00 << 16) | (0x88d8 >> 2),
848 0x00000000,
849 (0x0e00 << 16) | (0x8980 >> 2),
850 0x00000000,
851 (0x0e00 << 16) | (0x30938 >> 2),
852 0x00000000,
853 (0x0e00 << 16) | (0x3093c >> 2),
854 0x00000000,
855 (0x0e00 << 16) | (0x30940 >> 2),
856 0x00000000,
857 (0x0e00 << 16) | (0x89a0 >> 2),
858 0x00000000,
859 (0x0e00 << 16) | (0x30900 >> 2),
860 0x00000000,
861 (0x0e00 << 16) | (0x30904 >> 2),
862 0x00000000,
863 (0x0e00 << 16) | (0x89b4 >> 2),
864 0x00000000,
865 (0x0e00 << 16) | (0x3e1fc >> 2),
866 0x00000000,
867 (0x0e00 << 16) | (0x3c210 >> 2),
868 0x00000000,
869 (0x0e00 << 16) | (0x3c214 >> 2),
870 0x00000000,
871 (0x0e00 << 16) | (0x3c218 >> 2),
872 0x00000000,
873 (0x0e00 << 16) | (0x8904 >> 2),
874 0x00000000,
875 0x5,
876 (0x0e00 << 16) | (0x8c28 >> 2),
877 (0x0e00 << 16) | (0x8c2c >> 2),
878 (0x0e00 << 16) | (0x8c30 >> 2),
879 (0x0e00 << 16) | (0x8c34 >> 2),
880 (0x0e00 << 16) | (0x9600 >> 2),
881 };
882
883 static u32 gfx_v7_0_get_csb_size(struct amdgpu_device *adev);
884 static void gfx_v7_0_get_csb_buffer(struct amdgpu_device *adev, volatile u32 *buffer);
885 static void gfx_v7_0_init_cp_pg_table(struct amdgpu_device *adev);
886 static void gfx_v7_0_init_pg(struct amdgpu_device *adev);
887
888 /*
889 * Core functions
890 */
891 /**
892 * gfx_v7_0_init_microcode - load ucode images from disk
893 *
894 * @adev: amdgpu_device pointer
895 *
896 * Use the firmware interface to load the ucode images into
897 * the driver (not loaded into hw).
898 * Returns 0 on success, error on failure.
899 */
900 static int gfx_v7_0_init_microcode(struct amdgpu_device *adev)
901 {
902 const char *chip_name;
903 char fw_name[30];
904 int err;
905
906 DRM_DEBUG("\n");
907
908 switch (adev->asic_type) {
909 case CHIP_BONAIRE:
910 chip_name = "bonaire";
911 break;
912 case CHIP_HAWAII:
913 chip_name = "hawaii";
914 break;
915 case CHIP_KAVERI:
916 chip_name = "kaveri";
917 break;
918 case CHIP_KABINI:
919 chip_name = "kabini";
920 break;
921 case CHIP_MULLINS:
922 chip_name = "mullins";
923 break;
924 default: BUG();
925 }
926
927 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
928 err = request_firmware(&adev->gfx.pfp_fw, fw_name, adev->dev);
929 if (err)
930 goto out;
931 err = amdgpu_ucode_validate(adev->gfx.pfp_fw);
932 if (err)
933 goto out;
934
935 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
936 err = request_firmware(&adev->gfx.me_fw, fw_name, adev->dev);
937 if (err)
938 goto out;
939 err = amdgpu_ucode_validate(adev->gfx.me_fw);
940 if (err)
941 goto out;
942
943 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
944 err = request_firmware(&adev->gfx.ce_fw, fw_name, adev->dev);
945 if (err)
946 goto out;
947 err = amdgpu_ucode_validate(adev->gfx.ce_fw);
948 if (err)
949 goto out;
950
951 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mec.bin", chip_name);
952 err = request_firmware(&adev->gfx.mec_fw, fw_name, adev->dev);
953 if (err)
954 goto out;
955 err = amdgpu_ucode_validate(adev->gfx.mec_fw);
956 if (err)
957 goto out;
958
959 if (adev->asic_type == CHIP_KAVERI) {
960 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mec2.bin", chip_name);
961 err = request_firmware(&adev->gfx.mec2_fw, fw_name, adev->dev);
962 if (err)
963 goto out;
964 err = amdgpu_ucode_validate(adev->gfx.mec2_fw);
965 if (err)
966 goto out;
967 }
968
969 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", chip_name);
970 err = request_firmware(&adev->gfx.rlc_fw, fw_name, adev->dev);
971 if (err)
972 goto out;
973 err = amdgpu_ucode_validate(adev->gfx.rlc_fw);
974
975 out:
976 if (err) {
977 printk(KERN_ERR
978 "gfx7: Failed to load firmware \"%s\"\n",
979 fw_name);
980 release_firmware(adev->gfx.pfp_fw);
981 adev->gfx.pfp_fw = NULL;
982 release_firmware(adev->gfx.me_fw);
983 adev->gfx.me_fw = NULL;
984 release_firmware(adev->gfx.ce_fw);
985 adev->gfx.ce_fw = NULL;
986 release_firmware(adev->gfx.mec_fw);
987 adev->gfx.mec_fw = NULL;
988 release_firmware(adev->gfx.mec2_fw);
989 adev->gfx.mec2_fw = NULL;
990 release_firmware(adev->gfx.rlc_fw);
991 adev->gfx.rlc_fw = NULL;
992 }
993 return err;
994 }
995
996 /**
997 * gfx_v7_0_tiling_mode_table_init - init the hw tiling table
998 *
999 * @adev: amdgpu_device pointer
1000 *
1001 * Starting with SI, the tiling setup is done globally in a
1002 * set of 32 tiling modes. Rather than selecting each set of
1003 * parameters per surface as on older asics, we just select
1004 * which index in the tiling table we want to use, and the
1005 * surface uses those parameters (CIK).
1006 */
1007 static void gfx_v7_0_tiling_mode_table_init(struct amdgpu_device *adev)
1008 {
1009 const u32 num_tile_mode_states = 32;
1010 const u32 num_secondary_tile_mode_states = 16;
1011 u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
1012
1013 switch (adev->gfx.config.mem_row_size_in_kb) {
1014 case 1:
1015 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
1016 break;
1017 case 2:
1018 default:
1019 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
1020 break;
1021 case 4:
1022 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
1023 break;
1024 }
1025
1026 switch (adev->asic_type) {
1027 case CHIP_BONAIRE:
1028 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
1029 switch (reg_offset) {
1030 case 0:
1031 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1032 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1033 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1034 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1035 break;
1036 case 1:
1037 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1038 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1039 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
1040 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1041 break;
1042 case 2:
1043 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1044 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1045 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1046 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1047 break;
1048 case 3:
1049 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1050 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1051 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1052 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1053 break;
1054 case 4:
1055 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1056 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1057 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1058 TILE_SPLIT(split_equal_to_row_size));
1059 break;
1060 case 5:
1061 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1062 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1063 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1064 break;
1065 case 6:
1066 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1067 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1068 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1069 TILE_SPLIT(split_equal_to_row_size));
1070 break;
1071 case 7:
1072 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1073 break;
1074
1075 case 8:
1076 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
1077 PIPE_CONFIG(ADDR_SURF_P4_16x16));
1078 break;
1079 case 9:
1080 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1081 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1082 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
1083 break;
1084 case 10:
1085 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1086 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1087 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1088 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1089 break;
1090 case 11:
1091 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1092 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1093 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1094 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1095 break;
1096 case 12:
1097 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1098 break;
1099 case 13:
1100 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1101 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1102 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
1103 break;
1104 case 14:
1105 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1106 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1107 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1108 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1109 break;
1110 case 15:
1111 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_THIN1) |
1112 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1113 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1114 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1115 break;
1116 case 16:
1117 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1118 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1119 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1120 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1121 break;
1122 case 17:
1123 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1124 break;
1125 case 18:
1126 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1127 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1128 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1129 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1130 break;
1131 case 19:
1132 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1133 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1134 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
1135 break;
1136 case 20:
1137 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THICK) |
1138 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1139 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1140 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1141 break;
1142 case 21:
1143 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_THICK) |
1144 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1145 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1146 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1147 break;
1148 case 22:
1149 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THICK) |
1150 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1151 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1152 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1153 break;
1154 case 23:
1155 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1156 break;
1157 case 24:
1158 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THICK) |
1159 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1160 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1161 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1162 break;
1163 case 25:
1164 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_XTHICK) |
1165 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1166 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1167 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1168 break;
1169 case 26:
1170 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_XTHICK) |
1171 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1172 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1173 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1174 break;
1175 case 27:
1176 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1177 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1178 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
1179 break;
1180 case 28:
1181 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1182 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1183 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1184 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1185 break;
1186 case 29:
1187 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1188 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1189 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1190 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1191 break;
1192 case 30:
1193 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1194 break;
1195 default:
1196 gb_tile_moden = 0;
1197 break;
1198 }
1199 adev->gfx.config.tile_mode_array[reg_offset] = gb_tile_moden;
1200 WREG32(mmGB_TILE_MODE0 + reg_offset, gb_tile_moden);
1201 }
1202 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
1203 switch (reg_offset) {
1204 case 0:
1205 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1206 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1207 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1208 NUM_BANKS(ADDR_SURF_16_BANK));
1209 break;
1210 case 1:
1211 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1212 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1213 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1214 NUM_BANKS(ADDR_SURF_16_BANK));
1215 break;
1216 case 2:
1217 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1218 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1219 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1220 NUM_BANKS(ADDR_SURF_16_BANK));
1221 break;
1222 case 3:
1223 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1224 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1225 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1226 NUM_BANKS(ADDR_SURF_16_BANK));
1227 break;
1228 case 4:
1229 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1230 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1231 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1232 NUM_BANKS(ADDR_SURF_16_BANK));
1233 break;
1234 case 5:
1235 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1236 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1237 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1238 NUM_BANKS(ADDR_SURF_8_BANK));
1239 break;
1240 case 6:
1241 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1242 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1243 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1244 NUM_BANKS(ADDR_SURF_4_BANK));
1245 break;
1246 case 8:
1247 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
1248 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
1249 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1250 NUM_BANKS(ADDR_SURF_16_BANK));
1251 break;
1252 case 9:
1253 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
1254 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1255 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1256 NUM_BANKS(ADDR_SURF_16_BANK));
1257 break;
1258 case 10:
1259 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1260 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1261 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1262 NUM_BANKS(ADDR_SURF_16_BANK));
1263 break;
1264 case 11:
1265 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1266 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1267 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1268 NUM_BANKS(ADDR_SURF_16_BANK));
1269 break;
1270 case 12:
1271 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1272 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1273 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1274 NUM_BANKS(ADDR_SURF_16_BANK));
1275 break;
1276 case 13:
1277 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1278 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1279 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1280 NUM_BANKS(ADDR_SURF_8_BANK));
1281 break;
1282 case 14:
1283 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1284 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1285 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1286 NUM_BANKS(ADDR_SURF_4_BANK));
1287 break;
1288 default:
1289 gb_tile_moden = 0;
1290 break;
1291 }
1292 adev->gfx.config.macrotile_mode_array[reg_offset] = gb_tile_moden;
1293 WREG32(mmGB_MACROTILE_MODE0 + reg_offset, gb_tile_moden);
1294 }
1295 break;
1296 case CHIP_HAWAII:
1297 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
1298 switch (reg_offset) {
1299 case 0:
1300 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1301 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1302 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1303 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1304 break;
1305 case 1:
1306 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1307 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1308 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
1309 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1310 break;
1311 case 2:
1312 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1313 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1314 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1315 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1316 break;
1317 case 3:
1318 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1319 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1320 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1321 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1322 break;
1323 case 4:
1324 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1325 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1326 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1327 TILE_SPLIT(split_equal_to_row_size));
1328 break;
1329 case 5:
1330 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1331 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1332 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1333 TILE_SPLIT(split_equal_to_row_size));
1334 break;
1335 case 6:
1336 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1337 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1338 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1339 TILE_SPLIT(split_equal_to_row_size));
1340 break;
1341 case 7:
1342 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1343 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1344 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1345 TILE_SPLIT(split_equal_to_row_size));
1346 break;
1347
1348 case 8:
1349 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
1350 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16));
1351 break;
1352 case 9:
1353 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1354 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1355 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
1356 break;
1357 case 10:
1358 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1359 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1360 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1361 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1362 break;
1363 case 11:
1364 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1365 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1366 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1367 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1368 break;
1369 case 12:
1370 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
1371 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1372 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1373 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1374 break;
1375 case 13:
1376 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1377 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1378 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
1379 break;
1380 case 14:
1381 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1382 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1383 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1384 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1385 break;
1386 case 15:
1387 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_THIN1) |
1388 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1389 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1390 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1391 break;
1392 case 16:
1393 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1394 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1395 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1396 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1397 break;
1398 case 17:
1399 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1400 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1401 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1402 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1403 break;
1404 case 18:
1405 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1406 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1407 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1408 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1409 break;
1410 case 19:
1411 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1412 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1413 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING));
1414 break;
1415 case 20:
1416 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THICK) |
1417 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1418 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1419 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1420 break;
1421 case 21:
1422 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_THICK) |
1423 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1424 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1425 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1426 break;
1427 case 22:
1428 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THICK) |
1429 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1430 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1431 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1432 break;
1433 case 23:
1434 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THICK) |
1435 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1436 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1437 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1438 break;
1439 case 24:
1440 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THICK) |
1441 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1442 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1443 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1444 break;
1445 case 25:
1446 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_XTHICK) |
1447 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1448 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1449 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1450 break;
1451 case 26:
1452 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_XTHICK) |
1453 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1454 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1455 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1456 break;
1457 case 27:
1458 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1459 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1460 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
1461 break;
1462 case 28:
1463 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1464 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1465 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1466 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1467 break;
1468 case 29:
1469 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1470 PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
1471 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1472 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1473 break;
1474 case 30:
1475 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1476 PIPE_CONFIG(ADDR_SURF_P4_16x16) |
1477 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1478 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1479 break;
1480 default:
1481 gb_tile_moden = 0;
1482 break;
1483 }
1484 adev->gfx.config.tile_mode_array[reg_offset] = gb_tile_moden;
1485 WREG32(mmGB_TILE_MODE0 + reg_offset, gb_tile_moden);
1486 }
1487 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
1488 switch (reg_offset) {
1489 case 0:
1490 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1491 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1492 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1493 NUM_BANKS(ADDR_SURF_16_BANK));
1494 break;
1495 case 1:
1496 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1497 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1498 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1499 NUM_BANKS(ADDR_SURF_16_BANK));
1500 break;
1501 case 2:
1502 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1503 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1504 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1505 NUM_BANKS(ADDR_SURF_16_BANK));
1506 break;
1507 case 3:
1508 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1509 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1510 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1511 NUM_BANKS(ADDR_SURF_16_BANK));
1512 break;
1513 case 4:
1514 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1515 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1516 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1517 NUM_BANKS(ADDR_SURF_8_BANK));
1518 break;
1519 case 5:
1520 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1521 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1522 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1523 NUM_BANKS(ADDR_SURF_4_BANK));
1524 break;
1525 case 6:
1526 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1527 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1528 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1529 NUM_BANKS(ADDR_SURF_4_BANK));
1530 break;
1531 case 8:
1532 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1533 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1534 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1535 NUM_BANKS(ADDR_SURF_16_BANK));
1536 break;
1537 case 9:
1538 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1539 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1540 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1541 NUM_BANKS(ADDR_SURF_16_BANK));
1542 break;
1543 case 10:
1544 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1545 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1546 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1547 NUM_BANKS(ADDR_SURF_16_BANK));
1548 break;
1549 case 11:
1550 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1551 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1552 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1553 NUM_BANKS(ADDR_SURF_8_BANK));
1554 break;
1555 case 12:
1556 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1557 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1558 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1559 NUM_BANKS(ADDR_SURF_16_BANK));
1560 break;
1561 case 13:
1562 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1563 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1564 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1565 NUM_BANKS(ADDR_SURF_8_BANK));
1566 break;
1567 case 14:
1568 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1569 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1570 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1571 NUM_BANKS(ADDR_SURF_4_BANK));
1572 break;
1573 default:
1574 gb_tile_moden = 0;
1575 break;
1576 }
1577 adev->gfx.config.macrotile_mode_array[reg_offset] = gb_tile_moden;
1578 WREG32(mmGB_MACROTILE_MODE0 + reg_offset, gb_tile_moden);
1579 }
1580 break;
1581 case CHIP_KABINI:
1582 case CHIP_KAVERI:
1583 case CHIP_MULLINS:
1584 default:
1585 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
1586 switch (reg_offset) {
1587 case 0:
1588 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1589 PIPE_CONFIG(ADDR_SURF_P2) |
1590 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1591 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1592 break;
1593 case 1:
1594 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1595 PIPE_CONFIG(ADDR_SURF_P2) |
1596 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
1597 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1598 break;
1599 case 2:
1600 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1601 PIPE_CONFIG(ADDR_SURF_P2) |
1602 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1603 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1604 break;
1605 case 3:
1606 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1607 PIPE_CONFIG(ADDR_SURF_P2) |
1608 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1609 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1610 break;
1611 case 4:
1612 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1613 PIPE_CONFIG(ADDR_SURF_P2) |
1614 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1615 TILE_SPLIT(split_equal_to_row_size));
1616 break;
1617 case 5:
1618 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1619 PIPE_CONFIG(ADDR_SURF_P2) |
1620 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
1621 break;
1622 case 6:
1623 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1624 PIPE_CONFIG(ADDR_SURF_P2) |
1625 MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
1626 TILE_SPLIT(split_equal_to_row_size));
1627 break;
1628 case 7:
1629 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1630 break;
1631
1632 case 8:
1633 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
1634 PIPE_CONFIG(ADDR_SURF_P2));
1635 break;
1636 case 9:
1637 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1638 PIPE_CONFIG(ADDR_SURF_P2) |
1639 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
1640 break;
1641 case 10:
1642 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1643 PIPE_CONFIG(ADDR_SURF_P2) |
1644 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1645 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1646 break;
1647 case 11:
1648 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1649 PIPE_CONFIG(ADDR_SURF_P2) |
1650 MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
1651 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1652 break;
1653 case 12:
1654 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1655 break;
1656 case 13:
1657 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1658 PIPE_CONFIG(ADDR_SURF_P2) |
1659 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
1660 break;
1661 case 14:
1662 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1663 PIPE_CONFIG(ADDR_SURF_P2) |
1664 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1665 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1666 break;
1667 case 15:
1668 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_THIN1) |
1669 PIPE_CONFIG(ADDR_SURF_P2) |
1670 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1671 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1672 break;
1673 case 16:
1674 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1675 PIPE_CONFIG(ADDR_SURF_P2) |
1676 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1677 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1678 break;
1679 case 17:
1680 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1681 break;
1682 case 18:
1683 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1684 PIPE_CONFIG(ADDR_SURF_P2) |
1685 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1686 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1687 break;
1688 case 19:
1689 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1690 PIPE_CONFIG(ADDR_SURF_P2) |
1691 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING));
1692 break;
1693 case 20:
1694 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THICK) |
1695 PIPE_CONFIG(ADDR_SURF_P2) |
1696 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1697 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1698 break;
1699 case 21:
1700 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_THICK) |
1701 PIPE_CONFIG(ADDR_SURF_P2) |
1702 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1703 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1704 break;
1705 case 22:
1706 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THICK) |
1707 PIPE_CONFIG(ADDR_SURF_P2) |
1708 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1709 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1710 break;
1711 case 23:
1712 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1713 break;
1714 case 24:
1715 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THICK) |
1716 PIPE_CONFIG(ADDR_SURF_P2) |
1717 MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
1718 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1719 break;
1720 case 25:
1721 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_XTHICK) |
1722 PIPE_CONFIG(ADDR_SURF_P2) |
1723 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1724 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1725 break;
1726 case 26:
1727 gb_tile_moden = (ARRAY_MODE(ARRAY_3D_TILED_XTHICK) |
1728 PIPE_CONFIG(ADDR_SURF_P2) |
1729 MICRO_TILE_MODE_NEW(ADDR_SURF_THICK_MICRO_TILING) |
1730 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_1));
1731 break;
1732 case 27:
1733 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1734 PIPE_CONFIG(ADDR_SURF_P2) |
1735 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
1736 break;
1737 case 28:
1738 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1739 PIPE_CONFIG(ADDR_SURF_P2) |
1740 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1741 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
1742 break;
1743 case 29:
1744 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
1745 PIPE_CONFIG(ADDR_SURF_P2) |
1746 MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
1747 SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_8));
1748 break;
1749 case 30:
1750 gb_tile_moden = (TILE_SPLIT(split_equal_to_row_size));
1751 break;
1752 default:
1753 gb_tile_moden = 0;
1754 break;
1755 }
1756 adev->gfx.config.tile_mode_array[reg_offset] = gb_tile_moden;
1757 WREG32(mmGB_TILE_MODE0 + reg_offset, gb_tile_moden);
1758 }
1759 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
1760 switch (reg_offset) {
1761 case 0:
1762 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1763 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1764 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1765 NUM_BANKS(ADDR_SURF_8_BANK));
1766 break;
1767 case 1:
1768 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1769 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1770 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1771 NUM_BANKS(ADDR_SURF_8_BANK));
1772 break;
1773 case 2:
1774 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1775 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1776 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1777 NUM_BANKS(ADDR_SURF_8_BANK));
1778 break;
1779 case 3:
1780 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1781 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1782 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1783 NUM_BANKS(ADDR_SURF_8_BANK));
1784 break;
1785 case 4:
1786 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1787 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1788 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1789 NUM_BANKS(ADDR_SURF_8_BANK));
1790 break;
1791 case 5:
1792 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1793 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1794 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1795 NUM_BANKS(ADDR_SURF_8_BANK));
1796 break;
1797 case 6:
1798 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1799 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1800 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1801 NUM_BANKS(ADDR_SURF_8_BANK));
1802 break;
1803 case 8:
1804 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_4) |
1805 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
1806 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1807 NUM_BANKS(ADDR_SURF_16_BANK));
1808 break;
1809 case 9:
1810 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_4) |
1811 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1812 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1813 NUM_BANKS(ADDR_SURF_16_BANK));
1814 break;
1815 case 10:
1816 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
1817 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1818 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1819 NUM_BANKS(ADDR_SURF_16_BANK));
1820 break;
1821 case 11:
1822 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
1823 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1824 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1825 NUM_BANKS(ADDR_SURF_16_BANK));
1826 break;
1827 case 12:
1828 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1829 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1830 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1831 NUM_BANKS(ADDR_SURF_16_BANK));
1832 break;
1833 case 13:
1834 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1835 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1836 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
1837 NUM_BANKS(ADDR_SURF_16_BANK));
1838 break;
1839 case 14:
1840 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1841 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1842 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1843 NUM_BANKS(ADDR_SURF_8_BANK));
1844 break;
1845 default:
1846 gb_tile_moden = 0;
1847 break;
1848 }
1849 adev->gfx.config.macrotile_mode_array[reg_offset] = gb_tile_moden;
1850 WREG32(mmGB_MACROTILE_MODE0 + reg_offset, gb_tile_moden);
1851 }
1852 break;
1853 }
1854 }
1855
1856 /**
1857 * gfx_v7_0_select_se_sh - select which SE, SH to address
1858 *
1859 * @adev: amdgpu_device pointer
1860 * @se_num: shader engine to address
1861 * @sh_num: sh block to address
1862 *
1863 * Select which SE, SH combinations to address. Certain
1864 * registers are instanced per SE or SH. 0xffffffff means
1865 * broadcast to all SEs or SHs (CIK).
1866 */
1867 void gfx_v7_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, u32 sh_num)
1868 {
1869 u32 data = GRBM_GFX_INDEX__INSTANCE_BROADCAST_WRITES_MASK;
1870
1871 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
1872 data |= GRBM_GFX_INDEX__SH_BROADCAST_WRITES_MASK |
1873 GRBM_GFX_INDEX__SE_BROADCAST_WRITES_MASK;
1874 else if (se_num == 0xffffffff)
1875 data |= GRBM_GFX_INDEX__SE_BROADCAST_WRITES_MASK |
1876 (sh_num << GRBM_GFX_INDEX__SH_INDEX__SHIFT);
1877 else if (sh_num == 0xffffffff)
1878 data |= GRBM_GFX_INDEX__SH_BROADCAST_WRITES_MASK |
1879 (se_num << GRBM_GFX_INDEX__SE_INDEX__SHIFT);
1880 else
1881 data |= (sh_num << GRBM_GFX_INDEX__SH_INDEX__SHIFT) |
1882 (se_num << GRBM_GFX_INDEX__SE_INDEX__SHIFT);
1883 WREG32(mmGRBM_GFX_INDEX, data);
1884 }
1885
1886 /**
1887 * gfx_v7_0_create_bitmask - create a bitmask
1888 *
1889 * @bit_width: length of the mask
1890 *
1891 * create a variable length bit mask (CIK).
1892 * Returns the bitmask.
1893 */
1894 static u32 gfx_v7_0_create_bitmask(u32 bit_width)
1895 {
1896 u32 i, mask = 0;
1897
1898 for (i = 0; i < bit_width; i++) {
1899 mask <<= 1;
1900 mask |= 1;
1901 }
1902 return mask;
1903 }
1904
1905 /**
1906 * gfx_v7_0_get_rb_disabled - computes the mask of disabled RBs
1907 *
1908 * @adev: amdgpu_device pointer
1909 * @max_rb_num: max RBs (render backends) for the asic
1910 * @se_num: number of SEs (shader engines) for the asic
1911 * @sh_per_se: number of SH blocks per SE for the asic
1912 *
1913 * Calculates the bitmask of disabled RBs (CIK).
1914 * Returns the disabled RB bitmask.
1915 */
1916 static u32 gfx_v7_0_get_rb_disabled(struct amdgpu_device *adev,
1917 u32 max_rb_num_per_se,
1918 u32 sh_per_se)
1919 {
1920 u32 data, mask;
1921
1922 data = RREG32(mmCC_RB_BACKEND_DISABLE);
1923 if (data & 1)
1924 data &= CC_RB_BACKEND_DISABLE__BACKEND_DISABLE_MASK;
1925 else
1926 data = 0;
1927
1928 data |= RREG32(mmGC_USER_RB_BACKEND_DISABLE);
1929
1930 data >>= GC_USER_RB_BACKEND_DISABLE__BACKEND_DISABLE__SHIFT;
1931
1932 mask = gfx_v7_0_create_bitmask(max_rb_num_per_se / sh_per_se);
1933
1934 return data & mask;
1935 }
1936
1937 /**
1938 * gfx_v7_0_setup_rb - setup the RBs on the asic
1939 *
1940 * @adev: amdgpu_device pointer
1941 * @se_num: number of SEs (shader engines) for the asic
1942 * @sh_per_se: number of SH blocks per SE for the asic
1943 * @max_rb_num: max RBs (render backends) for the asic
1944 *
1945 * Configures per-SE/SH RB registers (CIK).
1946 */
1947 static void gfx_v7_0_setup_rb(struct amdgpu_device *adev,
1948 u32 se_num, u32 sh_per_se,
1949 u32 max_rb_num_per_se)
1950 {
1951 int i, j;
1952 u32 data, mask;
1953 u32 disabled_rbs = 0;
1954 u32 enabled_rbs = 0;
1955
1956 mutex_lock(&adev->grbm_idx_mutex);
1957 for (i = 0; i < se_num; i++) {
1958 for (j = 0; j < sh_per_se; j++) {
1959 gfx_v7_0_select_se_sh(adev, i, j);
1960 data = gfx_v7_0_get_rb_disabled(adev, max_rb_num_per_se, sh_per_se);
1961 if (adev->asic_type == CHIP_HAWAII)
1962 disabled_rbs |= data << ((i * sh_per_se + j) * HAWAII_RB_BITMAP_WIDTH_PER_SH);
1963 else
1964 disabled_rbs |= data << ((i * sh_per_se + j) * CIK_RB_BITMAP_WIDTH_PER_SH);
1965 }
1966 }
1967 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
1968 mutex_unlock(&adev->grbm_idx_mutex);
1969
1970 mask = 1;
1971 for (i = 0; i < max_rb_num_per_se * se_num; i++) {
1972 if (!(disabled_rbs & mask))
1973 enabled_rbs |= mask;
1974 mask <<= 1;
1975 }
1976
1977 adev->gfx.config.backend_enable_mask = enabled_rbs;
1978
1979 mutex_lock(&adev->grbm_idx_mutex);
1980 for (i = 0; i < se_num; i++) {
1981 gfx_v7_0_select_se_sh(adev, i, 0xffffffff);
1982 data = 0;
1983 for (j = 0; j < sh_per_se; j++) {
1984 switch (enabled_rbs & 3) {
1985 case 0:
1986 if (j == 0)
1987 data |= (RASTER_CONFIG_RB_MAP_3 <<
1988 PA_SC_RASTER_CONFIG__PKR_MAP__SHIFT);
1989 else
1990 data |= (RASTER_CONFIG_RB_MAP_0 <<
1991 PA_SC_RASTER_CONFIG__PKR_MAP__SHIFT);
1992 break;
1993 case 1:
1994 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
1995 break;
1996 case 2:
1997 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
1998 break;
1999 case 3:
2000 default:
2001 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
2002 break;
2003 }
2004 enabled_rbs >>= 2;
2005 }
2006 WREG32(mmPA_SC_RASTER_CONFIG, data);
2007 }
2008 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
2009 mutex_unlock(&adev->grbm_idx_mutex);
2010 }
2011
2012 /**
2013 * gmc_v7_0_init_compute_vmid - gart enable
2014 *
2015 * @rdev: amdgpu_device pointer
2016 *
2017 * Initialize compute vmid sh_mem registers
2018 *
2019 */
2020 #define DEFAULT_SH_MEM_BASES (0x6000)
2021 #define FIRST_COMPUTE_VMID (8)
2022 #define LAST_COMPUTE_VMID (16)
2023 static void gmc_v7_0_init_compute_vmid(struct amdgpu_device *adev)
2024 {
2025 int i;
2026 uint32_t sh_mem_config;
2027 uint32_t sh_mem_bases;
2028
2029 /*
2030 * Configure apertures:
2031 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB)
2032 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB)
2033 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB)
2034 */
2035 sh_mem_bases = DEFAULT_SH_MEM_BASES | (DEFAULT_SH_MEM_BASES << 16);
2036 sh_mem_config = SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
2037 SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT;
2038 sh_mem_config |= MTYPE_NONCACHED << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT;
2039 mutex_lock(&adev->srbm_mutex);
2040 for (i = FIRST_COMPUTE_VMID; i < LAST_COMPUTE_VMID; i++) {
2041 cik_srbm_select(adev, 0, 0, 0, i);
2042 /* CP and shaders */
2043 WREG32(mmSH_MEM_CONFIG, sh_mem_config);
2044 WREG32(mmSH_MEM_APE1_BASE, 1);
2045 WREG32(mmSH_MEM_APE1_LIMIT, 0);
2046 WREG32(mmSH_MEM_BASES, sh_mem_bases);
2047 }
2048 cik_srbm_select(adev, 0, 0, 0, 0);
2049 mutex_unlock(&adev->srbm_mutex);
2050 }
2051
2052 /**
2053 * gfx_v7_0_gpu_init - setup the 3D engine
2054 *
2055 * @adev: amdgpu_device pointer
2056 *
2057 * Configures the 3D engine and tiling configuration
2058 * registers so that the 3D engine is usable.
2059 */
2060 static void gfx_v7_0_gpu_init(struct amdgpu_device *adev)
2061 {
2062 u32 gb_addr_config;
2063 u32 mc_shared_chmap, mc_arb_ramcfg;
2064 u32 dimm00_addr_map, dimm01_addr_map, dimm10_addr_map, dimm11_addr_map;
2065 u32 sh_mem_cfg;
2066 u32 tmp;
2067 int i;
2068
2069 switch (adev->asic_type) {
2070 case CHIP_BONAIRE:
2071 adev->gfx.config.max_shader_engines = 2;
2072 adev->gfx.config.max_tile_pipes = 4;
2073 adev->gfx.config.max_cu_per_sh = 7;
2074 adev->gfx.config.max_sh_per_se = 1;
2075 adev->gfx.config.max_backends_per_se = 2;
2076 adev->gfx.config.max_texture_channel_caches = 4;
2077 adev->gfx.config.max_gprs = 256;
2078 adev->gfx.config.max_gs_threads = 32;
2079 adev->gfx.config.max_hw_contexts = 8;
2080
2081 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
2082 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
2083 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
2084 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
2085 gb_addr_config = BONAIRE_GB_ADDR_CONFIG_GOLDEN;
2086 break;
2087 case CHIP_HAWAII:
2088 adev->gfx.config.max_shader_engines = 4;
2089 adev->gfx.config.max_tile_pipes = 16;
2090 adev->gfx.config.max_cu_per_sh = 11;
2091 adev->gfx.config.max_sh_per_se = 1;
2092 adev->gfx.config.max_backends_per_se = 4;
2093 adev->gfx.config.max_texture_channel_caches = 16;
2094 adev->gfx.config.max_gprs = 256;
2095 adev->gfx.config.max_gs_threads = 32;
2096 adev->gfx.config.max_hw_contexts = 8;
2097
2098 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
2099 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
2100 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
2101 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
2102 gb_addr_config = HAWAII_GB_ADDR_CONFIG_GOLDEN;
2103 break;
2104 case CHIP_KAVERI:
2105 adev->gfx.config.max_shader_engines = 1;
2106 adev->gfx.config.max_tile_pipes = 4;
2107 if ((adev->pdev->device == 0x1304) ||
2108 (adev->pdev->device == 0x1305) ||
2109 (adev->pdev->device == 0x130C) ||
2110 (adev->pdev->device == 0x130F) ||
2111 (adev->pdev->device == 0x1310) ||
2112 (adev->pdev->device == 0x1311) ||
2113 (adev->pdev->device == 0x131C)) {
2114 adev->gfx.config.max_cu_per_sh = 8;
2115 adev->gfx.config.max_backends_per_se = 2;
2116 } else if ((adev->pdev->device == 0x1309) ||
2117 (adev->pdev->device == 0x130A) ||
2118 (adev->pdev->device == 0x130D) ||
2119 (adev->pdev->device == 0x1313) ||
2120 (adev->pdev->device == 0x131D)) {
2121 adev->gfx.config.max_cu_per_sh = 6;
2122 adev->gfx.config.max_backends_per_se = 2;
2123 } else if ((adev->pdev->device == 0x1306) ||
2124 (adev->pdev->device == 0x1307) ||
2125 (adev->pdev->device == 0x130B) ||
2126 (adev->pdev->device == 0x130E) ||
2127 (adev->pdev->device == 0x1315) ||
2128 (adev->pdev->device == 0x131B)) {
2129 adev->gfx.config.max_cu_per_sh = 4;
2130 adev->gfx.config.max_backends_per_se = 1;
2131 } else {
2132 adev->gfx.config.max_cu_per_sh = 3;
2133 adev->gfx.config.max_backends_per_se = 1;
2134 }
2135 adev->gfx.config.max_sh_per_se = 1;
2136 adev->gfx.config.max_texture_channel_caches = 4;
2137 adev->gfx.config.max_gprs = 256;
2138 adev->gfx.config.max_gs_threads = 16;
2139 adev->gfx.config.max_hw_contexts = 8;
2140
2141 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
2142 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
2143 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
2144 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
2145 gb_addr_config = BONAIRE_GB_ADDR_CONFIG_GOLDEN;
2146 break;
2147 case CHIP_KABINI:
2148 case CHIP_MULLINS:
2149 default:
2150 adev->gfx.config.max_shader_engines = 1;
2151 adev->gfx.config.max_tile_pipes = 2;
2152 adev->gfx.config.max_cu_per_sh = 2;
2153 adev->gfx.config.max_sh_per_se = 1;
2154 adev->gfx.config.max_backends_per_se = 1;
2155 adev->gfx.config.max_texture_channel_caches = 2;
2156 adev->gfx.config.max_gprs = 256;
2157 adev->gfx.config.max_gs_threads = 16;
2158 adev->gfx.config.max_hw_contexts = 8;
2159
2160 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
2161 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
2162 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
2163 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
2164 gb_addr_config = BONAIRE_GB_ADDR_CONFIG_GOLDEN;
2165 break;
2166 }
2167
2168 WREG32(mmGRBM_CNTL, (0xff << GRBM_CNTL__READ_TIMEOUT__SHIFT));
2169
2170 mc_shared_chmap = RREG32(mmMC_SHARED_CHMAP);
2171 adev->gfx.config.mc_arb_ramcfg = RREG32(mmMC_ARB_RAMCFG);
2172 mc_arb_ramcfg = adev->gfx.config.mc_arb_ramcfg;
2173
2174 adev->gfx.config.num_tile_pipes = adev->gfx.config.max_tile_pipes;
2175 adev->gfx.config.mem_max_burst_length_bytes = 256;
2176 if (adev->flags & AMD_IS_APU) {
2177 /* Get memory bank mapping mode. */
2178 tmp = RREG32(mmMC_FUS_DRAM0_BANK_ADDR_MAPPING);
2179 dimm00_addr_map = REG_GET_FIELD(tmp, MC_FUS_DRAM0_BANK_ADDR_MAPPING, DIMM0ADDRMAP);
2180 dimm01_addr_map = REG_GET_FIELD(tmp, MC_FUS_DRAM0_BANK_ADDR_MAPPING, DIMM1ADDRMAP);
2181
2182 tmp = RREG32(mmMC_FUS_DRAM1_BANK_ADDR_MAPPING);
2183 dimm10_addr_map = REG_GET_FIELD(tmp, MC_FUS_DRAM1_BANK_ADDR_MAPPING, DIMM0ADDRMAP);
2184 dimm11_addr_map = REG_GET_FIELD(tmp, MC_FUS_DRAM1_BANK_ADDR_MAPPING, DIMM1ADDRMAP);
2185
2186 /* Validate settings in case only one DIMM installed. */
2187 if ((dimm00_addr_map == 0) || (dimm00_addr_map == 3) || (dimm00_addr_map == 4) || (dimm00_addr_map > 12))
2188 dimm00_addr_map = 0;
2189 if ((dimm01_addr_map == 0) || (dimm01_addr_map == 3) || (dimm01_addr_map == 4) || (dimm01_addr_map > 12))
2190 dimm01_addr_map = 0;
2191 if ((dimm10_addr_map == 0) || (dimm10_addr_map == 3) || (dimm10_addr_map == 4) || (dimm10_addr_map > 12))
2192 dimm10_addr_map = 0;
2193 if ((dimm11_addr_map == 0) || (dimm11_addr_map == 3) || (dimm11_addr_map == 4) || (dimm11_addr_map > 12))
2194 dimm11_addr_map = 0;
2195
2196 /* If DIMM Addr map is 8GB, ROW size should be 2KB. Otherwise 1KB. */
2197 /* If ROW size(DIMM1) != ROW size(DMIMM0), ROW size should be larger one. */
2198 if ((dimm00_addr_map == 11) || (dimm01_addr_map == 11) || (dimm10_addr_map == 11) || (dimm11_addr_map == 11))
2199 adev->gfx.config.mem_row_size_in_kb = 2;
2200 else
2201 adev->gfx.config.mem_row_size_in_kb = 1;
2202 } else {
2203 tmp = (mc_arb_ramcfg & MC_ARB_RAMCFG__NOOFCOLS_MASK) >> MC_ARB_RAMCFG__NOOFCOLS__SHIFT;
2204 adev->gfx.config.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
2205 if (adev->gfx.config.mem_row_size_in_kb > 4)
2206 adev->gfx.config.mem_row_size_in_kb = 4;
2207 }
2208 /* XXX use MC settings? */
2209 adev->gfx.config.shader_engine_tile_size = 32;
2210 adev->gfx.config.num_gpus = 1;
2211 adev->gfx.config.multi_gpu_tile_size = 64;
2212
2213 /* fix up row size */
2214 gb_addr_config &= ~GB_ADDR_CONFIG__ROW_SIZE_MASK;
2215 switch (adev->gfx.config.mem_row_size_in_kb) {
2216 case 1:
2217 default:
2218 gb_addr_config |= (0 << GB_ADDR_CONFIG__ROW_SIZE__SHIFT);
2219 break;
2220 case 2:
2221 gb_addr_config |= (1 << GB_ADDR_CONFIG__ROW_SIZE__SHIFT);
2222 break;
2223 case 4:
2224 gb_addr_config |= (2 << GB_ADDR_CONFIG__ROW_SIZE__SHIFT);
2225 break;
2226 }
2227 adev->gfx.config.gb_addr_config = gb_addr_config;
2228
2229 WREG32(mmGB_ADDR_CONFIG, gb_addr_config);
2230 WREG32(mmHDP_ADDR_CONFIG, gb_addr_config);
2231 WREG32(mmDMIF_ADDR_CALC, gb_addr_config);
2232 WREG32(mmSDMA0_TILING_CONFIG + SDMA0_REGISTER_OFFSET, gb_addr_config & 0x70);
2233 WREG32(mmSDMA0_TILING_CONFIG + SDMA1_REGISTER_OFFSET, gb_addr_config & 0x70);
2234 WREG32(mmUVD_UDEC_ADDR_CONFIG, gb_addr_config);
2235 WREG32(mmUVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
2236 WREG32(mmUVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
2237
2238 gfx_v7_0_tiling_mode_table_init(adev);
2239
2240 gfx_v7_0_setup_rb(adev, adev->gfx.config.max_shader_engines,
2241 adev->gfx.config.max_sh_per_se,
2242 adev->gfx.config.max_backends_per_se);
2243
2244 /* set HW defaults for 3D engine */
2245 WREG32(mmCP_MEQ_THRESHOLDS,
2246 (0x30 << CP_MEQ_THRESHOLDS__MEQ1_START__SHIFT) |
2247 (0x60 << CP_MEQ_THRESHOLDS__MEQ2_START__SHIFT));
2248
2249 mutex_lock(&adev->grbm_idx_mutex);
2250 /*
2251 * making sure that the following register writes will be broadcasted
2252 * to all the shaders
2253 */
2254 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
2255
2256 /* XXX SH_MEM regs */
2257 /* where to put LDS, scratch, GPUVM in FSA64 space */
2258 sh_mem_cfg = REG_SET_FIELD(0, SH_MEM_CONFIG, ALIGNMENT_MODE,
2259 SH_MEM_ALIGNMENT_MODE_UNALIGNED);
2260
2261 mutex_lock(&adev->srbm_mutex);
2262 for (i = 0; i < 16; i++) {
2263 cik_srbm_select(adev, 0, 0, 0, i);
2264 /* CP and shaders */
2265 WREG32(mmSH_MEM_CONFIG, sh_mem_cfg);
2266 WREG32(mmSH_MEM_APE1_BASE, 1);
2267 WREG32(mmSH_MEM_APE1_LIMIT, 0);
2268 WREG32(mmSH_MEM_BASES, 0);
2269 }
2270 cik_srbm_select(adev, 0, 0, 0, 0);
2271 mutex_unlock(&adev->srbm_mutex);
2272
2273 gmc_v7_0_init_compute_vmid(adev);
2274
2275 WREG32(mmSX_DEBUG_1, 0x20);
2276
2277 WREG32(mmTA_CNTL_AUX, 0x00010000);
2278
2279 tmp = RREG32(mmSPI_CONFIG_CNTL);
2280 tmp |= 0x03000000;
2281 WREG32(mmSPI_CONFIG_CNTL, tmp);
2282
2283 WREG32(mmSQ_CONFIG, 1);
2284
2285 WREG32(mmDB_DEBUG, 0);
2286
2287 tmp = RREG32(mmDB_DEBUG2) & ~0xf00fffff;
2288 tmp |= 0x00000400;
2289 WREG32(mmDB_DEBUG2, tmp);
2290
2291 tmp = RREG32(mmDB_DEBUG3) & ~0x0002021c;
2292 tmp |= 0x00020200;
2293 WREG32(mmDB_DEBUG3, tmp);
2294
2295 tmp = RREG32(mmCB_HW_CONTROL) & ~0x00010000;
2296 tmp |= 0x00018208;
2297 WREG32(mmCB_HW_CONTROL, tmp);
2298
2299 WREG32(mmSPI_CONFIG_CNTL_1, (4 << SPI_CONFIG_CNTL_1__VTX_DONE_DELAY__SHIFT));
2300
2301 WREG32(mmPA_SC_FIFO_SIZE,
2302 ((adev->gfx.config.sc_prim_fifo_size_frontend << PA_SC_FIFO_SIZE__SC_FRONTEND_PRIM_FIFO_SIZE__SHIFT) |
2303 (adev->gfx.config.sc_prim_fifo_size_backend << PA_SC_FIFO_SIZE__SC_BACKEND_PRIM_FIFO_SIZE__SHIFT) |
2304 (adev->gfx.config.sc_hiz_tile_fifo_size << PA_SC_FIFO_SIZE__SC_HIZ_TILE_FIFO_SIZE__SHIFT) |
2305 (adev->gfx.config.sc_earlyz_tile_fifo_size << PA_SC_FIFO_SIZE__SC_EARLYZ_TILE_FIFO_SIZE__SHIFT)));
2306
2307 WREG32(mmVGT_NUM_INSTANCES, 1);
2308
2309 WREG32(mmCP_PERFMON_CNTL, 0);
2310
2311 WREG32(mmSQ_CONFIG, 0);
2312
2313 WREG32(mmPA_SC_FORCE_EOV_MAX_CNTS,
2314 ((4095 << PA_SC_FORCE_EOV_MAX_CNTS__FORCE_EOV_MAX_CLK_CNT__SHIFT) |
2315 (255 << PA_SC_FORCE_EOV_MAX_CNTS__FORCE_EOV_MAX_REZ_CNT__SHIFT)));
2316
2317 WREG32(mmVGT_CACHE_INVALIDATION,
2318 (VC_AND_TC << VGT_CACHE_INVALIDATION__CACHE_INVALIDATION__SHIFT) |
2319 (ES_AND_GS_AUTO << VGT_CACHE_INVALIDATION__AUTO_INVLD_EN__SHIFT));
2320
2321 WREG32(mmVGT_GS_VERTEX_REUSE, 16);
2322 WREG32(mmPA_SC_LINE_STIPPLE_STATE, 0);
2323
2324 WREG32(mmPA_CL_ENHANCE, PA_CL_ENHANCE__CLIP_VTX_REORDER_ENA_MASK |
2325 (3 << PA_CL_ENHANCE__NUM_CLIP_SEQ__SHIFT));
2326 WREG32(mmPA_SC_ENHANCE, PA_SC_ENHANCE__ENABLE_PA_SC_OUT_OF_ORDER_MASK);
2327 mutex_unlock(&adev->grbm_idx_mutex);
2328
2329 udelay(50);
2330 }
2331
2332 /*
2333 * GPU scratch registers helpers function.
2334 */
2335 /**
2336 * gfx_v7_0_scratch_init - setup driver info for CP scratch regs
2337 *
2338 * @adev: amdgpu_device pointer
2339 *
2340 * Set up the number and offset of the CP scratch registers.
2341 * NOTE: use of CP scratch registers is a legacy inferface and
2342 * is not used by default on newer asics (r6xx+). On newer asics,
2343 * memory buffers are used for fences rather than scratch regs.
2344 */
2345 static void gfx_v7_0_scratch_init(struct amdgpu_device *adev)
2346 {
2347 int i;
2348
2349 adev->gfx.scratch.num_reg = 7;
2350 adev->gfx.scratch.reg_base = mmSCRATCH_REG0;
2351 for (i = 0; i < adev->gfx.scratch.num_reg; i++) {
2352 adev->gfx.scratch.free[i] = true;
2353 adev->gfx.scratch.reg[i] = adev->gfx.scratch.reg_base + i;
2354 }
2355 }
2356
2357 /**
2358 * gfx_v7_0_ring_test_ring - basic gfx ring test
2359 *
2360 * @adev: amdgpu_device pointer
2361 * @ring: amdgpu_ring structure holding ring information
2362 *
2363 * Allocate a scratch register and write to it using the gfx ring (CIK).
2364 * Provides a basic gfx ring test to verify that the ring is working.
2365 * Used by gfx_v7_0_cp_gfx_resume();
2366 * Returns 0 on success, error on failure.
2367 */
2368 static int gfx_v7_0_ring_test_ring(struct amdgpu_ring *ring)
2369 {
2370 struct amdgpu_device *adev = ring->adev;
2371 uint32_t scratch;
2372 uint32_t tmp = 0;
2373 unsigned i;
2374 int r;
2375
2376 r = amdgpu_gfx_scratch_get(adev, &scratch);
2377 if (r) {
2378 DRM_ERROR("amdgpu: cp failed to get scratch reg (%d).\n", r);
2379 return r;
2380 }
2381 WREG32(scratch, 0xCAFEDEAD);
2382 r = amdgpu_ring_lock(ring, 3);
2383 if (r) {
2384 DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r);
2385 amdgpu_gfx_scratch_free(adev, scratch);
2386 return r;
2387 }
2388 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
2389 amdgpu_ring_write(ring, (scratch - PACKET3_SET_UCONFIG_REG_START));
2390 amdgpu_ring_write(ring, 0xDEADBEEF);
2391 amdgpu_ring_unlock_commit(ring);
2392
2393 for (i = 0; i < adev->usec_timeout; i++) {
2394 tmp = RREG32(scratch);
2395 if (tmp == 0xDEADBEEF)
2396 break;
2397 DRM_UDELAY(1);
2398 }
2399 if (i < adev->usec_timeout) {
2400 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
2401 } else {
2402 DRM_ERROR("amdgpu: ring %d test failed (scratch(0x%04X)=0x%08X)\n",
2403 ring->idx, scratch, tmp);
2404 r = -EINVAL;
2405 }
2406 amdgpu_gfx_scratch_free(adev, scratch);
2407 return r;
2408 }
2409
2410 /**
2411 * gfx_v7_0_ring_emit_hdp - emit an hdp flush on the cp
2412 *
2413 * @adev: amdgpu_device pointer
2414 * @ridx: amdgpu ring index
2415 *
2416 * Emits an hdp flush on the cp.
2417 */
2418 static void gfx_v7_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
2419 {
2420 u32 ref_and_mask;
2421 int usepfp = ring->type == AMDGPU_RING_TYPE_COMPUTE ? 0 : 1;
2422
2423 if (ring->type == AMDGPU_RING_TYPE_COMPUTE) {
2424 switch (ring->me) {
2425 case 1:
2426 ref_and_mask = GPU_HDP_FLUSH_DONE__CP2_MASK << ring->pipe;
2427 break;
2428 case 2:
2429 ref_and_mask = GPU_HDP_FLUSH_DONE__CP6_MASK << ring->pipe;
2430 break;
2431 default:
2432 return;
2433 }
2434 } else {
2435 ref_and_mask = GPU_HDP_FLUSH_DONE__CP0_MASK;
2436 }
2437
2438 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
2439 amdgpu_ring_write(ring, (WAIT_REG_MEM_OPERATION(1) | /* write, wait, write */
2440 WAIT_REG_MEM_FUNCTION(3) | /* == */
2441 WAIT_REG_MEM_ENGINE(usepfp))); /* pfp or me */
2442 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ);
2443 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE);
2444 amdgpu_ring_write(ring, ref_and_mask);
2445 amdgpu_ring_write(ring, ref_and_mask);
2446 amdgpu_ring_write(ring, 0x20); /* poll interval */
2447 }
2448
2449 /**
2450 * gfx_v7_0_ring_emit_fence_gfx - emit a fence on the gfx ring
2451 *
2452 * @adev: amdgpu_device pointer
2453 * @fence: amdgpu fence object
2454 *
2455 * Emits a fence sequnce number on the gfx ring and flushes
2456 * GPU caches.
2457 */
2458 static void gfx_v7_0_ring_emit_fence_gfx(struct amdgpu_ring *ring, u64 addr,
2459 u64 seq, unsigned flags)
2460 {
2461 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
2462 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
2463 /* Workaround for cache flush problems. First send a dummy EOP
2464 * event down the pipe with seq one below.
2465 */
2466 amdgpu_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
2467 amdgpu_ring_write(ring, (EOP_TCL1_ACTION_EN |
2468 EOP_TC_ACTION_EN |
2469 EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
2470 EVENT_INDEX(5)));
2471 amdgpu_ring_write(ring, addr & 0xfffffffc);
2472 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xffff) |
2473 DATA_SEL(1) | INT_SEL(0));
2474 amdgpu_ring_write(ring, lower_32_bits(seq - 1));
2475 amdgpu_ring_write(ring, upper_32_bits(seq - 1));
2476
2477 /* Then send the real EOP event down the pipe. */
2478 amdgpu_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
2479 amdgpu_ring_write(ring, (EOP_TCL1_ACTION_EN |
2480 EOP_TC_ACTION_EN |
2481 EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
2482 EVENT_INDEX(5)));
2483 amdgpu_ring_write(ring, addr & 0xfffffffc);
2484 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xffff) |
2485 DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0));
2486 amdgpu_ring_write(ring, lower_32_bits(seq));
2487 amdgpu_ring_write(ring, upper_32_bits(seq));
2488 }
2489
2490 /**
2491 * gfx_v7_0_ring_emit_fence_compute - emit a fence on the compute ring
2492 *
2493 * @adev: amdgpu_device pointer
2494 * @fence: amdgpu fence object
2495 *
2496 * Emits a fence sequnce number on the compute ring and flushes
2497 * GPU caches.
2498 */
2499 static void gfx_v7_0_ring_emit_fence_compute(struct amdgpu_ring *ring,
2500 u64 addr, u64 seq,
2501 unsigned flags)
2502 {
2503 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
2504 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
2505
2506 /* RELEASE_MEM - flush caches, send int */
2507 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 5));
2508 amdgpu_ring_write(ring, (EOP_TCL1_ACTION_EN |
2509 EOP_TC_ACTION_EN |
2510 EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
2511 EVENT_INDEX(5)));
2512 amdgpu_ring_write(ring, DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0));
2513 amdgpu_ring_write(ring, addr & 0xfffffffc);
2514 amdgpu_ring_write(ring, upper_32_bits(addr));
2515 amdgpu_ring_write(ring, lower_32_bits(seq));
2516 amdgpu_ring_write(ring, upper_32_bits(seq));
2517 }
2518
2519 /**
2520 * gfx_v7_0_ring_emit_semaphore - emit a semaphore on the CP ring
2521 *
2522 * @ring: amdgpu ring buffer object
2523 * @semaphore: amdgpu semaphore object
2524 * @emit_wait: Is this a sempahore wait?
2525 *
2526 * Emits a semaphore signal/wait packet to the CP ring and prevents the PFP
2527 * from running ahead of semaphore waits.
2528 */
2529 static bool gfx_v7_0_ring_emit_semaphore(struct amdgpu_ring *ring,
2530 struct amdgpu_semaphore *semaphore,
2531 bool emit_wait)
2532 {
2533 uint64_t addr = semaphore->gpu_addr;
2534 unsigned sel = emit_wait ? PACKET3_SEM_SEL_WAIT : PACKET3_SEM_SEL_SIGNAL;
2535
2536 amdgpu_ring_write(ring, PACKET3(PACKET3_MEM_SEMAPHORE, 1));
2537 amdgpu_ring_write(ring, addr & 0xffffffff);
2538 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xffff) | sel);
2539
2540 if (emit_wait && (ring->type == AMDGPU_RING_TYPE_GFX)) {
2541 /* Prevent the PFP from running ahead of the semaphore wait */
2542 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
2543 amdgpu_ring_write(ring, 0x0);
2544 }
2545
2546 return true;
2547 }
2548
2549 /*
2550 * IB stuff
2551 */
2552 /**
2553 * gfx_v7_0_ring_emit_ib - emit an IB (Indirect Buffer) on the ring
2554 *
2555 * @ring: amdgpu_ring structure holding ring information
2556 * @ib: amdgpu indirect buffer object
2557 *
2558 * Emits an DE (drawing engine) or CE (constant engine) IB
2559 * on the gfx ring. IBs are usually generated by userspace
2560 * acceleration drivers and submitted to the kernel for
2561 * sheduling on the ring. This function schedules the IB
2562 * on the gfx ring for execution by the GPU.
2563 */
2564 static void gfx_v7_0_ring_emit_ib_gfx(struct amdgpu_ring *ring,
2565 struct amdgpu_ib *ib)
2566 {
2567 bool need_ctx_switch = ring->current_ctx != ib->ctx;
2568 u32 header, control = 0;
2569 u32 next_rptr = ring->wptr + 5;
2570
2571 /* drop the CE preamble IB for the same context */
2572 if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) && !need_ctx_switch)
2573 return;
2574
2575 if (need_ctx_switch)
2576 next_rptr += 2;
2577
2578 next_rptr += 4;
2579 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
2580 amdgpu_ring_write(ring, WRITE_DATA_DST_SEL(5) | WR_CONFIRM);
2581 amdgpu_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
2582 amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
2583 amdgpu_ring_write(ring, next_rptr);
2584
2585 /* insert SWITCH_BUFFER packet before first IB in the ring frame */
2586 if (need_ctx_switch) {
2587 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
2588 amdgpu_ring_write(ring, 0);
2589 }
2590
2591 if (ib->flags & AMDGPU_IB_FLAG_CE)
2592 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
2593 else
2594 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
2595
2596 control |= ib->length_dw |
2597 (ib->vm ? (ib->vm->ids[ring->idx].id << 24) : 0);
2598
2599 amdgpu_ring_write(ring, header);
2600 amdgpu_ring_write(ring,
2601 #ifdef __BIG_ENDIAN
2602 (2 << 0) |
2603 #endif
2604 (ib->gpu_addr & 0xFFFFFFFC));
2605 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
2606 amdgpu_ring_write(ring, control);
2607 }
2608
2609 static void gfx_v7_0_ring_emit_ib_compute(struct amdgpu_ring *ring,
2610 struct amdgpu_ib *ib)
2611 {
2612 u32 header, control = 0;
2613 u32 next_rptr = ring->wptr + 5;
2614
2615 control |= INDIRECT_BUFFER_VALID;
2616 next_rptr += 4;
2617 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
2618 amdgpu_ring_write(ring, WRITE_DATA_DST_SEL(5) | WR_CONFIRM);
2619 amdgpu_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
2620 amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
2621 amdgpu_ring_write(ring, next_rptr);
2622
2623 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
2624
2625 control |= ib->length_dw |
2626 (ib->vm ? (ib->vm->ids[ring->idx].id << 24) : 0);
2627
2628 amdgpu_ring_write(ring, header);
2629 amdgpu_ring_write(ring,
2630 #ifdef __BIG_ENDIAN
2631 (2 << 0) |
2632 #endif
2633 (ib->gpu_addr & 0xFFFFFFFC));
2634 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
2635 amdgpu_ring_write(ring, control);
2636 }
2637
2638 /**
2639 * gfx_v7_0_ring_test_ib - basic ring IB test
2640 *
2641 * @ring: amdgpu_ring structure holding ring information
2642 *
2643 * Allocate an IB and execute it on the gfx ring (CIK).
2644 * Provides a basic gfx ring test to verify that IBs are working.
2645 * Returns 0 on success, error on failure.
2646 */
2647 static int gfx_v7_0_ring_test_ib(struct amdgpu_ring *ring)
2648 {
2649 struct amdgpu_device *adev = ring->adev;
2650 struct amdgpu_ib ib;
2651 struct fence *f = NULL;
2652 uint32_t scratch;
2653 uint32_t tmp = 0;
2654 unsigned i;
2655 int r;
2656
2657 r = amdgpu_gfx_scratch_get(adev, &scratch);
2658 if (r) {
2659 DRM_ERROR("amdgpu: failed to get scratch reg (%d).\n", r);
2660 return r;
2661 }
2662 WREG32(scratch, 0xCAFEDEAD);
2663 memset(&ib, 0, sizeof(ib));
2664 r = amdgpu_ib_get(ring, NULL, 256, &ib);
2665 if (r) {
2666 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
2667 goto err1;
2668 }
2669 ib.ptr[0] = PACKET3(PACKET3_SET_UCONFIG_REG, 1);
2670 ib.ptr[1] = ((scratch - PACKET3_SET_UCONFIG_REG_START));
2671 ib.ptr[2] = 0xDEADBEEF;
2672 ib.length_dw = 3;
2673
2674 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, &ib, 1, NULL,
2675 AMDGPU_FENCE_OWNER_UNDEFINED,
2676 &f);
2677 if (r)
2678 goto err2;
2679
2680 r = fence_wait(f, false);
2681 if (r) {
2682 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
2683 goto err2;
2684 }
2685 for (i = 0; i < adev->usec_timeout; i++) {
2686 tmp = RREG32(scratch);
2687 if (tmp == 0xDEADBEEF)
2688 break;
2689 DRM_UDELAY(1);
2690 }
2691 if (i < adev->usec_timeout) {
2692 DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
2693 ring->idx, i);
2694 goto err2;
2695 } else {
2696 DRM_ERROR("amdgpu: ib test failed (scratch(0x%04X)=0x%08X)\n",
2697 scratch, tmp);
2698 r = -EINVAL;
2699 }
2700
2701 err2:
2702 fence_put(f);
2703 amdgpu_ib_free(adev, &ib);
2704 err1:
2705 amdgpu_gfx_scratch_free(adev, scratch);
2706 return r;
2707 }
2708
2709 /*
2710 * CP.
2711 * On CIK, gfx and compute now have independant command processors.
2712 *
2713 * GFX
2714 * Gfx consists of a single ring and can process both gfx jobs and
2715 * compute jobs. The gfx CP consists of three microengines (ME):
2716 * PFP - Pre-Fetch Parser
2717 * ME - Micro Engine
2718 * CE - Constant Engine
2719 * The PFP and ME make up what is considered the Drawing Engine (DE).
2720 * The CE is an asynchronous engine used for updating buffer desciptors
2721 * used by the DE so that they can be loaded into cache in parallel
2722 * while the DE is processing state update packets.
2723 *
2724 * Compute
2725 * The compute CP consists of two microengines (ME):
2726 * MEC1 - Compute MicroEngine 1
2727 * MEC2 - Compute MicroEngine 2
2728 * Each MEC supports 4 compute pipes and each pipe supports 8 queues.
2729 * The queues are exposed to userspace and are programmed directly
2730 * by the compute runtime.
2731 */
2732 /**
2733 * gfx_v7_0_cp_gfx_enable - enable/disable the gfx CP MEs
2734 *
2735 * @adev: amdgpu_device pointer
2736 * @enable: enable or disable the MEs
2737 *
2738 * Halts or unhalts the gfx MEs.
2739 */
2740 static void gfx_v7_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable)
2741 {
2742 int i;
2743
2744 if (enable) {
2745 WREG32(mmCP_ME_CNTL, 0);
2746 } else {
2747 WREG32(mmCP_ME_CNTL, (CP_ME_CNTL__ME_HALT_MASK | CP_ME_CNTL__PFP_HALT_MASK | CP_ME_CNTL__CE_HALT_MASK));
2748 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
2749 adev->gfx.gfx_ring[i].ready = false;
2750 }
2751 udelay(50);
2752 }
2753
2754 /**
2755 * gfx_v7_0_cp_gfx_load_microcode - load the gfx CP ME ucode
2756 *
2757 * @adev: amdgpu_device pointer
2758 *
2759 * Loads the gfx PFP, ME, and CE ucode.
2760 * Returns 0 for success, -EINVAL if the ucode is not available.
2761 */
2762 static int gfx_v7_0_cp_gfx_load_microcode(struct amdgpu_device *adev)
2763 {
2764 const struct gfx_firmware_header_v1_0 *pfp_hdr;
2765 const struct gfx_firmware_header_v1_0 *ce_hdr;
2766 const struct gfx_firmware_header_v1_0 *me_hdr;
2767 const __le32 *fw_data;
2768 unsigned i, fw_size;
2769
2770 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw || !adev->gfx.ce_fw)
2771 return -EINVAL;
2772
2773 pfp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
2774 ce_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
2775 me_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
2776
2777 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header);
2778 amdgpu_ucode_print_gfx_hdr(&ce_hdr->header);
2779 amdgpu_ucode_print_gfx_hdr(&me_hdr->header);
2780 adev->gfx.pfp_fw_version = le32_to_cpu(pfp_hdr->header.ucode_version);
2781 adev->gfx.ce_fw_version = le32_to_cpu(ce_hdr->header.ucode_version);
2782 adev->gfx.me_fw_version = le32_to_cpu(me_hdr->header.ucode_version);
2783 adev->gfx.me_feature_version = le32_to_cpu(me_hdr->ucode_feature_version);
2784 adev->gfx.ce_feature_version = le32_to_cpu(ce_hdr->ucode_feature_version);
2785 adev->gfx.pfp_feature_version = le32_to_cpu(pfp_hdr->ucode_feature_version);
2786
2787 gfx_v7_0_cp_gfx_enable(adev, false);
2788
2789 /* PFP */
2790 fw_data = (const __le32 *)
2791 (adev->gfx.pfp_fw->data +
2792 le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
2793 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
2794 WREG32(mmCP_PFP_UCODE_ADDR, 0);
2795 for (i = 0; i < fw_size; i++)
2796 WREG32(mmCP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
2797 WREG32(mmCP_PFP_UCODE_ADDR, adev->gfx.pfp_fw_version);
2798
2799 /* CE */
2800 fw_data = (const __le32 *)
2801 (adev->gfx.ce_fw->data +
2802 le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
2803 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
2804 WREG32(mmCP_CE_UCODE_ADDR, 0);
2805 for (i = 0; i < fw_size; i++)
2806 WREG32(mmCP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
2807 WREG32(mmCP_CE_UCODE_ADDR, adev->gfx.ce_fw_version);
2808
2809 /* ME */
2810 fw_data = (const __le32 *)
2811 (adev->gfx.me_fw->data +
2812 le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
2813 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
2814 WREG32(mmCP_ME_RAM_WADDR, 0);
2815 for (i = 0; i < fw_size; i++)
2816 WREG32(mmCP_ME_RAM_DATA, le32_to_cpup(fw_data++));
2817 WREG32(mmCP_ME_RAM_WADDR, adev->gfx.me_fw_version);
2818
2819 return 0;
2820 }
2821
2822 /**
2823 * gfx_v7_0_cp_gfx_start - start the gfx ring
2824 *
2825 * @adev: amdgpu_device pointer
2826 *
2827 * Enables the ring and loads the clear state context and other
2828 * packets required to init the ring.
2829 * Returns 0 for success, error for failure.
2830 */
2831 static int gfx_v7_0_cp_gfx_start(struct amdgpu_device *adev)
2832 {
2833 struct amdgpu_ring *ring = &adev->gfx.gfx_ring[0];
2834 const struct cs_section_def *sect = NULL;
2835 const struct cs_extent_def *ext = NULL;
2836 int r, i;
2837
2838 /* init the CP */
2839 WREG32(mmCP_MAX_CONTEXT, adev->gfx.config.max_hw_contexts - 1);
2840 WREG32(mmCP_ENDIAN_SWAP, 0);
2841 WREG32(mmCP_DEVICE_ID, 1);
2842
2843 gfx_v7_0_cp_gfx_enable(adev, true);
2844
2845 r = amdgpu_ring_lock(ring, gfx_v7_0_get_csb_size(adev) + 8);
2846 if (r) {
2847 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r);
2848 return r;
2849 }
2850
2851 /* init the CE partitions. CE only used for gfx on CIK */
2852 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
2853 amdgpu_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
2854 amdgpu_ring_write(ring, 0x8000);
2855 amdgpu_ring_write(ring, 0x8000);
2856
2857 /* clear state buffer */
2858 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
2859 amdgpu_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
2860
2861 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
2862 amdgpu_ring_write(ring, 0x80000000);
2863 amdgpu_ring_write(ring, 0x80000000);
2864
2865 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
2866 for (ext = sect->section; ext->extent != NULL; ++ext) {
2867 if (sect->id == SECT_CONTEXT) {
2868 amdgpu_ring_write(ring,
2869 PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
2870 amdgpu_ring_write(ring, ext->reg_index - PACKET3_SET_CONTEXT_REG_START);
2871 for (i = 0; i < ext->reg_count; i++)
2872 amdgpu_ring_write(ring, ext->extent[i]);
2873 }
2874 }
2875 }
2876
2877 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
2878 amdgpu_ring_write(ring, mmPA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
2879 switch (adev->asic_type) {
2880 case CHIP_BONAIRE:
2881 amdgpu_ring_write(ring, 0x16000012);
2882 amdgpu_ring_write(ring, 0x00000000);
2883 break;
2884 case CHIP_KAVERI:
2885 amdgpu_ring_write(ring, 0x00000000); /* XXX */
2886 amdgpu_ring_write(ring, 0x00000000);
2887 break;
2888 case CHIP_KABINI:
2889 case CHIP_MULLINS:
2890 amdgpu_ring_write(ring, 0x00000000); /* XXX */
2891 amdgpu_ring_write(ring, 0x00000000);
2892 break;
2893 case CHIP_HAWAII:
2894 amdgpu_ring_write(ring, 0x3a00161a);
2895 amdgpu_ring_write(ring, 0x0000002e);
2896 break;
2897 default:
2898 amdgpu_ring_write(ring, 0x00000000);
2899 amdgpu_ring_write(ring, 0x00000000);
2900 break;
2901 }
2902
2903 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
2904 amdgpu_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
2905
2906 amdgpu_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
2907 amdgpu_ring_write(ring, 0);
2908
2909 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
2910 amdgpu_ring_write(ring, 0x00000316);
2911 amdgpu_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
2912 amdgpu_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
2913
2914 amdgpu_ring_unlock_commit(ring);
2915
2916 return 0;
2917 }
2918
2919 /**
2920 * gfx_v7_0_cp_gfx_resume - setup the gfx ring buffer registers
2921 *
2922 * @adev: amdgpu_device pointer
2923 *
2924 * Program the location and size of the gfx ring buffer
2925 * and test it to make sure it's working.
2926 * Returns 0 for success, error for failure.
2927 */
2928 static int gfx_v7_0_cp_gfx_resume(struct amdgpu_device *adev)
2929 {
2930 struct amdgpu_ring *ring;
2931 u32 tmp;
2932 u32 rb_bufsz;
2933 u64 rb_addr, rptr_addr;
2934 int r;
2935
2936 WREG32(mmCP_SEM_WAIT_TIMER, 0x0);
2937 if (adev->asic_type != CHIP_HAWAII)
2938 WREG32(mmCP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
2939
2940 /* Set the write pointer delay */
2941 WREG32(mmCP_RB_WPTR_DELAY, 0);
2942
2943 /* set the RB to use vmid 0 */
2944 WREG32(mmCP_RB_VMID, 0);
2945
2946 WREG32(mmSCRATCH_ADDR, 0);
2947
2948 /* ring 0 - compute and gfx */
2949 /* Set ring buffer size */
2950 ring = &adev->gfx.gfx_ring[0];
2951 rb_bufsz = order_base_2(ring->ring_size / 8);
2952 tmp = (order_base_2(AMDGPU_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2953 #ifdef __BIG_ENDIAN
2954 tmp |= 2 << CP_RB0_CNTL__BUF_SWAP__SHIFT;
2955 #endif
2956 WREG32(mmCP_RB0_CNTL, tmp);
2957
2958 /* Initialize the ring buffer's read and write pointers */
2959 WREG32(mmCP_RB0_CNTL, tmp | CP_RB0_CNTL__RB_RPTR_WR_ENA_MASK);
2960 ring->wptr = 0;
2961 WREG32(mmCP_RB0_WPTR, ring->wptr);
2962
2963 /* set the wb address wether it's enabled or not */
2964 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
2965 WREG32(mmCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr));
2966 WREG32(mmCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 0xFF);
2967
2968 /* scratch register shadowing is no longer supported */
2969 WREG32(mmSCRATCH_UMSK, 0);
2970
2971 mdelay(1);
2972 WREG32(mmCP_RB0_CNTL, tmp);
2973
2974 rb_addr = ring->gpu_addr >> 8;
2975 WREG32(mmCP_RB0_BASE, rb_addr);
2976 WREG32(mmCP_RB0_BASE_HI, upper_32_bits(rb_addr));
2977
2978 /* start the ring */
2979 gfx_v7_0_cp_gfx_start(adev);
2980 ring->ready = true;
2981 r = amdgpu_ring_test_ring(ring);
2982 if (r) {
2983 ring->ready = false;
2984 return r;
2985 }
2986
2987 return 0;
2988 }
2989
2990 static u32 gfx_v7_0_ring_get_rptr_gfx(struct amdgpu_ring *ring)
2991 {
2992 u32 rptr;
2993
2994 rptr = ring->adev->wb.wb[ring->rptr_offs];
2995
2996 return rptr;
2997 }
2998
2999 static u32 gfx_v7_0_ring_get_wptr_gfx(struct amdgpu_ring *ring)
3000 {
3001 struct amdgpu_device *adev = ring->adev;
3002 u32 wptr;
3003
3004 wptr = RREG32(mmCP_RB0_WPTR);
3005
3006 return wptr;
3007 }
3008
3009 static void gfx_v7_0_ring_set_wptr_gfx(struct amdgpu_ring *ring)
3010 {
3011 struct amdgpu_device *adev = ring->adev;
3012
3013 WREG32(mmCP_RB0_WPTR, ring->wptr);
3014 (void)RREG32(mmCP_RB0_WPTR);
3015 }
3016
3017 static u32 gfx_v7_0_ring_get_rptr_compute(struct amdgpu_ring *ring)
3018 {
3019 u32 rptr;
3020
3021 rptr = ring->adev->wb.wb[ring->rptr_offs];
3022
3023 return rptr;
3024 }
3025
3026 static u32 gfx_v7_0_ring_get_wptr_compute(struct amdgpu_ring *ring)
3027 {
3028 u32 wptr;
3029
3030 /* XXX check if swapping is necessary on BE */
3031 wptr = ring->adev->wb.wb[ring->wptr_offs];
3032
3033 return wptr;
3034 }
3035
3036 static void gfx_v7_0_ring_set_wptr_compute(struct amdgpu_ring *ring)
3037 {
3038 struct amdgpu_device *adev = ring->adev;
3039
3040 /* XXX check if swapping is necessary on BE */
3041 adev->wb.wb[ring->wptr_offs] = ring->wptr;
3042 WDOORBELL32(ring->doorbell_index, ring->wptr);
3043 }
3044
3045 /**
3046 * gfx_v7_0_cp_compute_enable - enable/disable the compute CP MEs
3047 *
3048 * @adev: amdgpu_device pointer
3049 * @enable: enable or disable the MEs
3050 *
3051 * Halts or unhalts the compute MEs.
3052 */
3053 static void gfx_v7_0_cp_compute_enable(struct amdgpu_device *adev, bool enable)
3054 {
3055 int i;
3056
3057 if (enable) {
3058 WREG32(mmCP_MEC_CNTL, 0);
3059 } else {
3060 WREG32(mmCP_MEC_CNTL, (CP_MEC_CNTL__MEC_ME1_HALT_MASK | CP_MEC_CNTL__MEC_ME2_HALT_MASK));
3061 for (i = 0; i < adev->gfx.num_compute_rings; i++)
3062 adev->gfx.compute_ring[i].ready = false;
3063 }
3064 udelay(50);
3065 }
3066
3067 /**
3068 * gfx_v7_0_cp_compute_load_microcode - load the compute CP ME ucode
3069 *
3070 * @adev: amdgpu_device pointer
3071 *
3072 * Loads the compute MEC1&2 ucode.
3073 * Returns 0 for success, -EINVAL if the ucode is not available.
3074 */
3075 static int gfx_v7_0_cp_compute_load_microcode(struct amdgpu_device *adev)
3076 {
3077 const struct gfx_firmware_header_v1_0 *mec_hdr;
3078 const __le32 *fw_data;
3079 unsigned i, fw_size;
3080
3081 if (!adev->gfx.mec_fw)
3082 return -EINVAL;
3083
3084 mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
3085 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header);
3086 adev->gfx.mec_fw_version = le32_to_cpu(mec_hdr->header.ucode_version);
3087 adev->gfx.mec_feature_version = le32_to_cpu(
3088 mec_hdr->ucode_feature_version);
3089
3090 gfx_v7_0_cp_compute_enable(adev, false);
3091
3092 /* MEC1 */
3093 fw_data = (const __le32 *)
3094 (adev->gfx.mec_fw->data +
3095 le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes));
3096 fw_size = le32_to_cpu(mec_hdr->header.ucode_size_bytes) / 4;
3097 WREG32(mmCP_MEC_ME1_UCODE_ADDR, 0);
3098 for (i = 0; i < fw_size; i++)
3099 WREG32(mmCP_MEC_ME1_UCODE_DATA, le32_to_cpup(fw_data++));
3100 WREG32(mmCP_MEC_ME1_UCODE_ADDR, 0);
3101
3102 if (adev->asic_type == CHIP_KAVERI) {
3103 const struct gfx_firmware_header_v1_0 *mec2_hdr;
3104
3105 if (!adev->gfx.mec2_fw)
3106 return -EINVAL;
3107
3108 mec2_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec2_fw->data;
3109 amdgpu_ucode_print_gfx_hdr(&mec2_hdr->header);
3110 adev->gfx.mec2_fw_version = le32_to_cpu(mec2_hdr->header.ucode_version);
3111 adev->gfx.mec2_feature_version = le32_to_cpu(
3112 mec2_hdr->ucode_feature_version);
3113
3114 /* MEC2 */
3115 fw_data = (const __le32 *)
3116 (adev->gfx.mec2_fw->data +
3117 le32_to_cpu(mec2_hdr->header.ucode_array_offset_bytes));
3118 fw_size = le32_to_cpu(mec2_hdr->header.ucode_size_bytes) / 4;
3119 WREG32(mmCP_MEC_ME2_UCODE_ADDR, 0);
3120 for (i = 0; i < fw_size; i++)
3121 WREG32(mmCP_MEC_ME2_UCODE_DATA, le32_to_cpup(fw_data++));
3122 WREG32(mmCP_MEC_ME2_UCODE_ADDR, 0);
3123 }
3124
3125 return 0;
3126 }
3127
3128 /**
3129 * gfx_v7_0_cp_compute_start - start the compute queues
3130 *
3131 * @adev: amdgpu_device pointer
3132 *
3133 * Enable the compute queues.
3134 * Returns 0 for success, error for failure.
3135 */
3136 static int gfx_v7_0_cp_compute_start(struct amdgpu_device *adev)
3137 {
3138 gfx_v7_0_cp_compute_enable(adev, true);
3139
3140 return 0;
3141 }
3142
3143 /**
3144 * gfx_v7_0_cp_compute_fini - stop the compute queues
3145 *
3146 * @adev: amdgpu_device pointer
3147 *
3148 * Stop the compute queues and tear down the driver queue
3149 * info.
3150 */
3151 static void gfx_v7_0_cp_compute_fini(struct amdgpu_device *adev)
3152 {
3153 int i, r;
3154
3155 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
3156 struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
3157
3158 if (ring->mqd_obj) {
3159 r = amdgpu_bo_reserve(ring->mqd_obj, false);
3160 if (unlikely(r != 0))
3161 dev_warn(adev->dev, "(%d) reserve MQD bo failed\n", r);
3162
3163 amdgpu_bo_unpin(ring->mqd_obj);
3164 amdgpu_bo_unreserve(ring->mqd_obj);
3165
3166 amdgpu_bo_unref(&ring->mqd_obj);
3167 ring->mqd_obj = NULL;
3168 }
3169 }
3170 }
3171
3172 static void gfx_v7_0_mec_fini(struct amdgpu_device *adev)
3173 {
3174 int r;
3175
3176 if (adev->gfx.mec.hpd_eop_obj) {
3177 r = amdgpu_bo_reserve(adev->gfx.mec.hpd_eop_obj, false);
3178 if (unlikely(r != 0))
3179 dev_warn(adev->dev, "(%d) reserve HPD EOP bo failed\n", r);
3180 amdgpu_bo_unpin(adev->gfx.mec.hpd_eop_obj);
3181 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
3182
3183 amdgpu_bo_unref(&adev->gfx.mec.hpd_eop_obj);
3184 adev->gfx.mec.hpd_eop_obj = NULL;
3185 }
3186 }
3187
3188 #define MEC_HPD_SIZE 2048
3189
3190 static int gfx_v7_0_mec_init(struct amdgpu_device *adev)
3191 {
3192 int r;
3193 u32 *hpd;
3194
3195 /*
3196 * KV: 2 MEC, 4 Pipes/MEC, 8 Queues/Pipe - 64 Queues total
3197 * CI/KB: 1 MEC, 4 Pipes/MEC, 8 Queues/Pipe - 32 Queues total
3198 * Nonetheless, we assign only 1 pipe because all other pipes will
3199 * be handled by KFD
3200 */
3201 adev->gfx.mec.num_mec = 1;
3202 adev->gfx.mec.num_pipe = 1;
3203 adev->gfx.mec.num_queue = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe * 8;
3204
3205 if (adev->gfx.mec.hpd_eop_obj == NULL) {
3206 r = amdgpu_bo_create(adev,
3207 adev->gfx.mec.num_mec *adev->gfx.mec.num_pipe * MEC_HPD_SIZE * 2,
3208 PAGE_SIZE, true,
3209 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL,
3210 &adev->gfx.mec.hpd_eop_obj);
3211 if (r) {
3212 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
3213 return r;
3214 }
3215 }
3216
3217 r = amdgpu_bo_reserve(adev->gfx.mec.hpd_eop_obj, false);
3218 if (unlikely(r != 0)) {
3219 gfx_v7_0_mec_fini(adev);
3220 return r;
3221 }
3222 r = amdgpu_bo_pin(adev->gfx.mec.hpd_eop_obj, AMDGPU_GEM_DOMAIN_GTT,
3223 &adev->gfx.mec.hpd_eop_gpu_addr);
3224 if (r) {
3225 dev_warn(adev->dev, "(%d) pin HDP EOP bo failed\n", r);
3226 gfx_v7_0_mec_fini(adev);
3227 return r;
3228 }
3229 r = amdgpu_bo_kmap(adev->gfx.mec.hpd_eop_obj, (void **)&hpd);
3230 if (r) {
3231 dev_warn(adev->dev, "(%d) map HDP EOP bo failed\n", r);
3232 gfx_v7_0_mec_fini(adev);
3233 return r;
3234 }
3235
3236 /* clear memory. Not sure if this is required or not */
3237 memset(hpd, 0, adev->gfx.mec.num_mec *adev->gfx.mec.num_pipe * MEC_HPD_SIZE * 2);
3238
3239 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
3240 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
3241
3242 return 0;
3243 }
3244
3245 struct hqd_registers
3246 {
3247 u32 cp_mqd_base_addr;
3248 u32 cp_mqd_base_addr_hi;
3249 u32 cp_hqd_active;
3250 u32 cp_hqd_vmid;
3251 u32 cp_hqd_persistent_state;
3252 u32 cp_hqd_pipe_priority;
3253 u32 cp_hqd_queue_priority;
3254 u32 cp_hqd_quantum;
3255 u32 cp_hqd_pq_base;
3256 u32 cp_hqd_pq_base_hi;
3257 u32 cp_hqd_pq_rptr;
3258 u32 cp_hqd_pq_rptr_report_addr;
3259 u32 cp_hqd_pq_rptr_report_addr_hi;
3260 u32 cp_hqd_pq_wptr_poll_addr;
3261 u32 cp_hqd_pq_wptr_poll_addr_hi;
3262 u32 cp_hqd_pq_doorbell_control;
3263 u32 cp_hqd_pq_wptr;
3264 u32 cp_hqd_pq_control;
3265 u32 cp_hqd_ib_base_addr;
3266 u32 cp_hqd_ib_base_addr_hi;
3267 u32 cp_hqd_ib_rptr;
3268 u32 cp_hqd_ib_control;
3269 u32 cp_hqd_iq_timer;
3270 u32 cp_hqd_iq_rptr;
3271 u32 cp_hqd_dequeue_request;
3272 u32 cp_hqd_dma_offload;
3273 u32 cp_hqd_sema_cmd;
3274 u32 cp_hqd_msg_type;
3275 u32 cp_hqd_atomic0_preop_lo;
3276 u32 cp_hqd_atomic0_preop_hi;
3277 u32 cp_hqd_atomic1_preop_lo;
3278 u32 cp_hqd_atomic1_preop_hi;
3279 u32 cp_hqd_hq_scheduler0;
3280 u32 cp_hqd_hq_scheduler1;
3281 u32 cp_mqd_control;
3282 };
3283
3284 struct bonaire_mqd
3285 {
3286 u32 header;
3287 u32 dispatch_initiator;
3288 u32 dimensions[3];
3289 u32 start_idx[3];
3290 u32 num_threads[3];
3291 u32 pipeline_stat_enable;
3292 u32 perf_counter_enable;
3293 u32 pgm[2];
3294 u32 tba[2];
3295 u32 tma[2];
3296 u32 pgm_rsrc[2];
3297 u32 vmid;
3298 u32 resource_limits;
3299 u32 static_thread_mgmt01[2];
3300 u32 tmp_ring_size;
3301 u32 static_thread_mgmt23[2];
3302 u32 restart[3];
3303 u32 thread_trace_enable;
3304 u32 reserved1;
3305 u32 user_data[16];
3306 u32 vgtcs_invoke_count[2];
3307 struct hqd_registers queue_state;
3308 u32 dequeue_cntr;
3309 u32 interrupt_queue[64];
3310 };
3311
3312 /**
3313 * gfx_v7_0_cp_compute_resume - setup the compute queue registers
3314 *
3315 * @adev: amdgpu_device pointer
3316 *
3317 * Program the compute queues and test them to make sure they
3318 * are working.
3319 * Returns 0 for success, error for failure.
3320 */
3321 static int gfx_v7_0_cp_compute_resume(struct amdgpu_device *adev)
3322 {
3323 int r, i, j;
3324 u32 tmp;
3325 bool use_doorbell = true;
3326 u64 hqd_gpu_addr;
3327 u64 mqd_gpu_addr;
3328 u64 eop_gpu_addr;
3329 u64 wb_gpu_addr;
3330 u32 *buf;
3331 struct bonaire_mqd *mqd;
3332
3333 r = gfx_v7_0_cp_compute_start(adev);
3334 if (r)
3335 return r;
3336
3337 /* fix up chicken bits */
3338 tmp = RREG32(mmCP_CPF_DEBUG);
3339 tmp |= (1 << 23);
3340 WREG32(mmCP_CPF_DEBUG, tmp);
3341
3342 /* init the pipes */
3343 mutex_lock(&adev->srbm_mutex);
3344 for (i = 0; i < (adev->gfx.mec.num_pipe * adev->gfx.mec.num_mec); i++) {
3345 int me = (i < 4) ? 1 : 2;
3346 int pipe = (i < 4) ? i : (i - 4);
3347
3348 eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr + (i * MEC_HPD_SIZE * 2);
3349
3350 cik_srbm_select(adev, me, pipe, 0, 0);
3351
3352 /* write the EOP addr */
3353 WREG32(mmCP_HPD_EOP_BASE_ADDR, eop_gpu_addr >> 8);
3354 WREG32(mmCP_HPD_EOP_BASE_ADDR_HI, upper_32_bits(eop_gpu_addr) >> 8);
3355
3356 /* set the VMID assigned */
3357 WREG32(mmCP_HPD_EOP_VMID, 0);
3358
3359 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
3360 tmp = RREG32(mmCP_HPD_EOP_CONTROL);
3361 tmp &= ~CP_HPD_EOP_CONTROL__EOP_SIZE_MASK;
3362 tmp |= order_base_2(MEC_HPD_SIZE / 8);
3363 WREG32(mmCP_HPD_EOP_CONTROL, tmp);
3364 }
3365 cik_srbm_select(adev, 0, 0, 0, 0);
3366 mutex_unlock(&adev->srbm_mutex);
3367
3368 /* init the queues. Just two for now. */
3369 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
3370 struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
3371
3372 if (ring->mqd_obj == NULL) {
3373 r = amdgpu_bo_create(adev,
3374 sizeof(struct bonaire_mqd),
3375 PAGE_SIZE, true,
3376 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL,
3377 &ring->mqd_obj);
3378 if (r) {
3379 dev_warn(adev->dev, "(%d) create MQD bo failed\n", r);
3380 return r;
3381 }
3382 }
3383
3384 r = amdgpu_bo_reserve(ring->mqd_obj, false);
3385 if (unlikely(r != 0)) {
3386 gfx_v7_0_cp_compute_fini(adev);
3387 return r;
3388 }
3389 r = amdgpu_bo_pin(ring->mqd_obj, AMDGPU_GEM_DOMAIN_GTT,
3390 &mqd_gpu_addr);
3391 if (r) {
3392 dev_warn(adev->dev, "(%d) pin MQD bo failed\n", r);
3393 gfx_v7_0_cp_compute_fini(adev);
3394 return r;
3395 }
3396 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&buf);
3397 if (r) {
3398 dev_warn(adev->dev, "(%d) map MQD bo failed\n", r);
3399 gfx_v7_0_cp_compute_fini(adev);
3400 return r;
3401 }
3402
3403 /* init the mqd struct */
3404 memset(buf, 0, sizeof(struct bonaire_mqd));
3405
3406 mqd = (struct bonaire_mqd *)buf;
3407 mqd->header = 0xC0310800;
3408 mqd->static_thread_mgmt01[0] = 0xffffffff;
3409 mqd->static_thread_mgmt01[1] = 0xffffffff;
3410 mqd->static_thread_mgmt23[0] = 0xffffffff;
3411 mqd->static_thread_mgmt23[1] = 0xffffffff;
3412
3413 mutex_lock(&adev->srbm_mutex);
3414 cik_srbm_select(adev, ring->me,
3415 ring->pipe,
3416 ring->queue, 0);
3417
3418 /* disable wptr polling */
3419 tmp = RREG32(mmCP_PQ_WPTR_POLL_CNTL);
3420 tmp &= ~CP_PQ_WPTR_POLL_CNTL__EN_MASK;
3421 WREG32(mmCP_PQ_WPTR_POLL_CNTL, tmp);
3422
3423 /* enable doorbell? */
3424 mqd->queue_state.cp_hqd_pq_doorbell_control =
3425 RREG32(mmCP_HQD_PQ_DOORBELL_CONTROL);
3426 if (use_doorbell)
3427 mqd->queue_state.cp_hqd_pq_doorbell_control |= CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_EN_MASK;
3428 else
3429 mqd->queue_state.cp_hqd_pq_doorbell_control &= ~CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_EN_MASK;
3430 WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL,
3431 mqd->queue_state.cp_hqd_pq_doorbell_control);
3432
3433 /* disable the queue if it's active */
3434 mqd->queue_state.cp_hqd_dequeue_request = 0;
3435 mqd->queue_state.cp_hqd_pq_rptr = 0;
3436 mqd->queue_state.cp_hqd_pq_wptr= 0;
3437 if (RREG32(mmCP_HQD_ACTIVE) & 1) {
3438 WREG32(mmCP_HQD_DEQUEUE_REQUEST, 1);
3439 for (j = 0; j < adev->usec_timeout; j++) {
3440 if (!(RREG32(mmCP_HQD_ACTIVE) & 1))
3441 break;
3442 udelay(1);
3443 }
3444 WREG32(mmCP_HQD_DEQUEUE_REQUEST, mqd->queue_state.cp_hqd_dequeue_request);
3445 WREG32(mmCP_HQD_PQ_RPTR, mqd->queue_state.cp_hqd_pq_rptr);
3446 WREG32(mmCP_HQD_PQ_WPTR, mqd->queue_state.cp_hqd_pq_wptr);
3447 }
3448
3449 /* set the pointer to the MQD */
3450 mqd->queue_state.cp_mqd_base_addr = mqd_gpu_addr & 0xfffffffc;
3451 mqd->queue_state.cp_mqd_base_addr_hi = upper_32_bits(mqd_gpu_addr);
3452 WREG32(mmCP_MQD_BASE_ADDR, mqd->queue_state.cp_mqd_base_addr);
3453 WREG32(mmCP_MQD_BASE_ADDR_HI, mqd->queue_state.cp_mqd_base_addr_hi);
3454 /* set MQD vmid to 0 */
3455 mqd->queue_state.cp_mqd_control = RREG32(mmCP_MQD_CONTROL);
3456 mqd->queue_state.cp_mqd_control &= ~CP_MQD_CONTROL__VMID_MASK;
3457 WREG32(mmCP_MQD_CONTROL, mqd->queue_state.cp_mqd_control);
3458
3459 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
3460 hqd_gpu_addr = ring->gpu_addr >> 8;
3461 mqd->queue_state.cp_hqd_pq_base = hqd_gpu_addr;
3462 mqd->queue_state.cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
3463 WREG32(mmCP_HQD_PQ_BASE, mqd->queue_state.cp_hqd_pq_base);
3464 WREG32(mmCP_HQD_PQ_BASE_HI, mqd->queue_state.cp_hqd_pq_base_hi);
3465
3466 /* set up the HQD, this is similar to CP_RB0_CNTL */
3467 mqd->queue_state.cp_hqd_pq_control = RREG32(mmCP_HQD_PQ_CONTROL);
3468 mqd->queue_state.cp_hqd_pq_control &=
3469 ~(CP_HQD_PQ_CONTROL__QUEUE_SIZE_MASK |
3470 CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE_MASK);
3471
3472 mqd->queue_state.cp_hqd_pq_control |=
3473 order_base_2(ring->ring_size / 8);
3474 mqd->queue_state.cp_hqd_pq_control |=
3475 (order_base_2(AMDGPU_GPU_PAGE_SIZE/8) << 8);
3476 #ifdef __BIG_ENDIAN
3477 mqd->queue_state.cp_hqd_pq_control |=
3478 2 << CP_HQD_PQ_CONTROL__ENDIAN_SWAP__SHIFT;
3479 #endif
3480 mqd->queue_state.cp_hqd_pq_control &=
3481 ~(CP_HQD_PQ_CONTROL__UNORD_DISPATCH_MASK |
3482 CP_HQD_PQ_CONTROL__ROQ_PQ_IB_FLIP_MASK |
3483 CP_HQD_PQ_CONTROL__PQ_VOLATILE_MASK);
3484 mqd->queue_state.cp_hqd_pq_control |=
3485 CP_HQD_PQ_CONTROL__PRIV_STATE_MASK |
3486 CP_HQD_PQ_CONTROL__KMD_QUEUE_MASK; /* assuming kernel queue control */
3487 WREG32(mmCP_HQD_PQ_CONTROL, mqd->queue_state.cp_hqd_pq_control);
3488
3489 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
3490 wb_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
3491 mqd->queue_state.cp_hqd_pq_wptr_poll_addr = wb_gpu_addr & 0xfffffffc;
3492 mqd->queue_state.cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
3493 WREG32(mmCP_HQD_PQ_WPTR_POLL_ADDR, mqd->queue_state.cp_hqd_pq_wptr_poll_addr);
3494 WREG32(mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
3495 mqd->queue_state.cp_hqd_pq_wptr_poll_addr_hi);
3496
3497 /* set the wb address wether it's enabled or not */
3498 wb_gpu_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
3499 mqd->queue_state.cp_hqd_pq_rptr_report_addr = wb_gpu_addr & 0xfffffffc;
3500 mqd->queue_state.cp_hqd_pq_rptr_report_addr_hi =
3501 upper_32_bits(wb_gpu_addr) & 0xffff;
3502 WREG32(mmCP_HQD_PQ_RPTR_REPORT_ADDR,
3503 mqd->queue_state.cp_hqd_pq_rptr_report_addr);
3504 WREG32(mmCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
3505 mqd->queue_state.cp_hqd_pq_rptr_report_addr_hi);
3506
3507 /* enable the doorbell if requested */
3508 if (use_doorbell) {
3509 mqd->queue_state.cp_hqd_pq_doorbell_control =
3510 RREG32(mmCP_HQD_PQ_DOORBELL_CONTROL);
3511 mqd->queue_state.cp_hqd_pq_doorbell_control &=
3512 ~CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET_MASK;
3513 mqd->queue_state.cp_hqd_pq_doorbell_control |=
3514 (ring->doorbell_index <<
3515 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT);
3516 mqd->queue_state.cp_hqd_pq_doorbell_control |=
3517 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_EN_MASK;
3518 mqd->queue_state.cp_hqd_pq_doorbell_control &=
3519 ~(CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_SOURCE_MASK |
3520 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_HIT_MASK);
3521
3522 } else {
3523 mqd->queue_state.cp_hqd_pq_doorbell_control = 0;
3524 }
3525 WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL,
3526 mqd->queue_state.cp_hqd_pq_doorbell_control);
3527
3528 /* read and write pointers, similar to CP_RB0_WPTR/_RPTR */
3529 ring->wptr = 0;
3530 mqd->queue_state.cp_hqd_pq_wptr = ring->wptr;
3531 WREG32(mmCP_HQD_PQ_WPTR, mqd->queue_state.cp_hqd_pq_wptr);
3532 mqd->queue_state.cp_hqd_pq_rptr = RREG32(mmCP_HQD_PQ_RPTR);
3533
3534 /* set the vmid for the queue */
3535 mqd->queue_state.cp_hqd_vmid = 0;
3536 WREG32(mmCP_HQD_VMID, mqd->queue_state.cp_hqd_vmid);
3537
3538 /* activate the queue */
3539 mqd->queue_state.cp_hqd_active = 1;
3540 WREG32(mmCP_HQD_ACTIVE, mqd->queue_state.cp_hqd_active);
3541
3542 cik_srbm_select(adev, 0, 0, 0, 0);
3543 mutex_unlock(&adev->srbm_mutex);
3544
3545 amdgpu_bo_kunmap(ring->mqd_obj);
3546 amdgpu_bo_unreserve(ring->mqd_obj);
3547
3548 ring->ready = true;
3549 r = amdgpu_ring_test_ring(ring);
3550 if (r)
3551 ring->ready = false;
3552 }
3553
3554 return 0;
3555 }
3556
3557 static void gfx_v7_0_cp_enable(struct amdgpu_device *adev, bool enable)
3558 {
3559 gfx_v7_0_cp_gfx_enable(adev, enable);
3560 gfx_v7_0_cp_compute_enable(adev, enable);
3561 }
3562
3563 static int gfx_v7_0_cp_load_microcode(struct amdgpu_device *adev)
3564 {
3565 int r;
3566
3567 r = gfx_v7_0_cp_gfx_load_microcode(adev);
3568 if (r)
3569 return r;
3570 r = gfx_v7_0_cp_compute_load_microcode(adev);
3571 if (r)
3572 return r;
3573
3574 return 0;
3575 }
3576
3577 static void gfx_v7_0_enable_gui_idle_interrupt(struct amdgpu_device *adev,
3578 bool enable)
3579 {
3580 u32 tmp = RREG32(mmCP_INT_CNTL_RING0);
3581
3582 if (enable)
3583 tmp |= (CP_INT_CNTL_RING0__CNTX_BUSY_INT_ENABLE_MASK |
3584 CP_INT_CNTL_RING0__CNTX_EMPTY_INT_ENABLE_MASK);
3585 else
3586 tmp &= ~(CP_INT_CNTL_RING0__CNTX_BUSY_INT_ENABLE_MASK |
3587 CP_INT_CNTL_RING0__CNTX_EMPTY_INT_ENABLE_MASK);
3588 WREG32(mmCP_INT_CNTL_RING0, tmp);
3589 }
3590
3591 static int gfx_v7_0_cp_resume(struct amdgpu_device *adev)
3592 {
3593 int r;
3594
3595 gfx_v7_0_enable_gui_idle_interrupt(adev, false);
3596
3597 r = gfx_v7_0_cp_load_microcode(adev);
3598 if (r)
3599 return r;
3600
3601 r = gfx_v7_0_cp_gfx_resume(adev);
3602 if (r)
3603 return r;
3604 r = gfx_v7_0_cp_compute_resume(adev);
3605 if (r)
3606 return r;
3607
3608 gfx_v7_0_enable_gui_idle_interrupt(adev, true);
3609
3610 return 0;
3611 }
3612
3613 /*
3614 * vm
3615 * VMID 0 is the physical GPU addresses as used by the kernel.
3616 * VMIDs 1-15 are used for userspace clients and are handled
3617 * by the amdgpu vm/hsa code.
3618 */
3619 /**
3620 * gfx_v7_0_ring_emit_vm_flush - cik vm flush using the CP
3621 *
3622 * @adev: amdgpu_device pointer
3623 *
3624 * Update the page table base and flush the VM TLB
3625 * using the CP (CIK).
3626 */
3627 static void gfx_v7_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
3628 unsigned vm_id, uint64_t pd_addr)
3629 {
3630 int usepfp = (ring->type == AMDGPU_RING_TYPE_GFX);
3631 if (usepfp) {
3632 /* synce CE with ME to prevent CE fetch CEIB before context switch done */
3633 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3634 amdgpu_ring_write(ring, 0);
3635 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3636 amdgpu_ring_write(ring, 0);
3637 }
3638
3639 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3640 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) |
3641 WRITE_DATA_DST_SEL(0)));
3642 if (vm_id < 8) {
3643 amdgpu_ring_write(ring,
3644 (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
3645 } else {
3646 amdgpu_ring_write(ring,
3647 (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
3648 }
3649 amdgpu_ring_write(ring, 0);
3650 amdgpu_ring_write(ring, pd_addr >> 12);
3651
3652 /* bits 0-15 are the VM contexts0-15 */
3653 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3654 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3655 WRITE_DATA_DST_SEL(0)));
3656 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
3657 amdgpu_ring_write(ring, 0);
3658 amdgpu_ring_write(ring, 1 << vm_id);
3659
3660 /* wait for the invalidate to complete */
3661 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
3662 amdgpu_ring_write(ring, (WAIT_REG_MEM_OPERATION(0) | /* wait */
3663 WAIT_REG_MEM_FUNCTION(0) | /* always */
3664 WAIT_REG_MEM_ENGINE(0))); /* me */
3665 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
3666 amdgpu_ring_write(ring, 0);
3667 amdgpu_ring_write(ring, 0); /* ref */
3668 amdgpu_ring_write(ring, 0); /* mask */
3669 amdgpu_ring_write(ring, 0x20); /* poll interval */
3670
3671 /* compute doesn't have PFP */
3672 if (usepfp) {
3673 /* sync PFP to ME, otherwise we might get invalid PFP reads */
3674 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
3675 amdgpu_ring_write(ring, 0x0);
3676
3677 /* synce CE with ME to prevent CE fetch CEIB before context switch done */
3678 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3679 amdgpu_ring_write(ring, 0);
3680 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3681 amdgpu_ring_write(ring, 0);
3682 }
3683 }
3684
3685 /*
3686 * RLC
3687 * The RLC is a multi-purpose microengine that handles a
3688 * variety of functions.
3689 */
3690 static void gfx_v7_0_rlc_fini(struct amdgpu_device *adev)
3691 {
3692 int r;
3693
3694 /* save restore block */
3695 if (adev->gfx.rlc.save_restore_obj) {
3696 r = amdgpu_bo_reserve(adev->gfx.rlc.save_restore_obj, false);
3697 if (unlikely(r != 0))
3698 dev_warn(adev->dev, "(%d) reserve RLC sr bo failed\n", r);
3699 amdgpu_bo_unpin(adev->gfx.rlc.save_restore_obj);
3700 amdgpu_bo_unreserve(adev->gfx.rlc.save_restore_obj);
3701
3702 amdgpu_bo_unref(&adev->gfx.rlc.save_restore_obj);
3703 adev->gfx.rlc.save_restore_obj = NULL;
3704 }
3705
3706 /* clear state block */
3707 if (adev->gfx.rlc.clear_state_obj) {
3708 r = amdgpu_bo_reserve(adev->gfx.rlc.clear_state_obj, false);
3709 if (unlikely(r != 0))
3710 dev_warn(adev->dev, "(%d) reserve RLC c bo failed\n", r);
3711 amdgpu_bo_unpin(adev->gfx.rlc.clear_state_obj);
3712 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
3713
3714 amdgpu_bo_unref(&adev->gfx.rlc.clear_state_obj);
3715 adev->gfx.rlc.clear_state_obj = NULL;
3716 }
3717
3718 /* clear state block */
3719 if (adev->gfx.rlc.cp_table_obj) {
3720 r = amdgpu_bo_reserve(adev->gfx.rlc.cp_table_obj, false);
3721 if (unlikely(r != 0))
3722 dev_warn(adev->dev, "(%d) reserve RLC cp table bo failed\n", r);
3723 amdgpu_bo_unpin(adev->gfx.rlc.cp_table_obj);
3724 amdgpu_bo_unreserve(adev->gfx.rlc.cp_table_obj);
3725
3726 amdgpu_bo_unref(&adev->gfx.rlc.cp_table_obj);
3727 adev->gfx.rlc.cp_table_obj = NULL;
3728 }
3729 }
3730
3731 static int gfx_v7_0_rlc_init(struct amdgpu_device *adev)
3732 {
3733 const u32 *src_ptr;
3734 volatile u32 *dst_ptr;
3735 u32 dws, i;
3736 const struct cs_section_def *cs_data;
3737 int r;
3738
3739 /* allocate rlc buffers */
3740 if (adev->flags & AMD_IS_APU) {
3741 if (adev->asic_type == CHIP_KAVERI) {
3742 adev->gfx.rlc.reg_list = spectre_rlc_save_restore_register_list;
3743 adev->gfx.rlc.reg_list_size =
3744 (u32)ARRAY_SIZE(spectre_rlc_save_restore_register_list);
3745 } else {
3746 adev->gfx.rlc.reg_list = kalindi_rlc_save_restore_register_list;
3747 adev->gfx.rlc.reg_list_size =
3748 (u32)ARRAY_SIZE(kalindi_rlc_save_restore_register_list);
3749 }
3750 }
3751 adev->gfx.rlc.cs_data = ci_cs_data;
3752 adev->gfx.rlc.cp_table_size = CP_ME_TABLE_SIZE * 5 * 4;
3753
3754 src_ptr = adev->gfx.rlc.reg_list;
3755 dws = adev->gfx.rlc.reg_list_size;
3756 dws += (5 * 16) + 48 + 48 + 64;
3757
3758 cs_data = adev->gfx.rlc.cs_data;
3759
3760 if (src_ptr) {
3761 /* save restore block */
3762 if (adev->gfx.rlc.save_restore_obj == NULL) {
3763 r = amdgpu_bo_create(adev, dws * 4, PAGE_SIZE, true,
3764 AMDGPU_GEM_DOMAIN_VRAM,
3765 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
3766 NULL, NULL,
3767 &adev->gfx.rlc.save_restore_obj);
3768 if (r) {
3769 dev_warn(adev->dev, "(%d) create RLC sr bo failed\n", r);
3770 return r;
3771 }
3772 }
3773
3774 r = amdgpu_bo_reserve(adev->gfx.rlc.save_restore_obj, false);
3775 if (unlikely(r != 0)) {
3776 gfx_v7_0_rlc_fini(adev);
3777 return r;
3778 }
3779 r = amdgpu_bo_pin(adev->gfx.rlc.save_restore_obj, AMDGPU_GEM_DOMAIN_VRAM,
3780 &adev->gfx.rlc.save_restore_gpu_addr);
3781 if (r) {
3782 amdgpu_bo_unreserve(adev->gfx.rlc.save_restore_obj);
3783 dev_warn(adev->dev, "(%d) pin RLC sr bo failed\n", r);
3784 gfx_v7_0_rlc_fini(adev);
3785 return r;
3786 }
3787
3788 r = amdgpu_bo_kmap(adev->gfx.rlc.save_restore_obj, (void **)&adev->gfx.rlc.sr_ptr);
3789 if (r) {
3790 dev_warn(adev->dev, "(%d) map RLC sr bo failed\n", r);
3791 gfx_v7_0_rlc_fini(adev);
3792 return r;
3793 }
3794 /* write the sr buffer */
3795 dst_ptr = adev->gfx.rlc.sr_ptr;
3796 for (i = 0; i < adev->gfx.rlc.reg_list_size; i++)
3797 dst_ptr[i] = cpu_to_le32(src_ptr[i]);
3798 amdgpu_bo_kunmap(adev->gfx.rlc.save_restore_obj);
3799 amdgpu_bo_unreserve(adev->gfx.rlc.save_restore_obj);
3800 }
3801
3802 if (cs_data) {
3803 /* clear state block */
3804 adev->gfx.rlc.clear_state_size = dws = gfx_v7_0_get_csb_size(adev);
3805
3806 if (adev->gfx.rlc.clear_state_obj == NULL) {
3807 r = amdgpu_bo_create(adev, dws * 4, PAGE_SIZE, true,
3808 AMDGPU_GEM_DOMAIN_VRAM,
3809 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
3810 NULL, NULL,
3811 &adev->gfx.rlc.clear_state_obj);
3812 if (r) {
3813 dev_warn(adev->dev, "(%d) create RLC c bo failed\n", r);
3814 gfx_v7_0_rlc_fini(adev);
3815 return r;
3816 }
3817 }
3818 r = amdgpu_bo_reserve(adev->gfx.rlc.clear_state_obj, false);
3819 if (unlikely(r != 0)) {
3820 gfx_v7_0_rlc_fini(adev);
3821 return r;
3822 }
3823 r = amdgpu_bo_pin(adev->gfx.rlc.clear_state_obj, AMDGPU_GEM_DOMAIN_VRAM,
3824 &adev->gfx.rlc.clear_state_gpu_addr);
3825 if (r) {
3826 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
3827 dev_warn(adev->dev, "(%d) pin RLC c bo failed\n", r);
3828 gfx_v7_0_rlc_fini(adev);
3829 return r;
3830 }
3831
3832 r = amdgpu_bo_kmap(adev->gfx.rlc.clear_state_obj, (void **)&adev->gfx.rlc.cs_ptr);
3833 if (r) {
3834 dev_warn(adev->dev, "(%d) map RLC c bo failed\n", r);
3835 gfx_v7_0_rlc_fini(adev);
3836 return r;
3837 }
3838 /* set up the cs buffer */
3839 dst_ptr = adev->gfx.rlc.cs_ptr;
3840 gfx_v7_0_get_csb_buffer(adev, dst_ptr);
3841 amdgpu_bo_kunmap(adev->gfx.rlc.clear_state_obj);
3842 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
3843 }
3844
3845 if (adev->gfx.rlc.cp_table_size) {
3846 if (adev->gfx.rlc.cp_table_obj == NULL) {
3847 r = amdgpu_bo_create(adev, adev->gfx.rlc.cp_table_size, PAGE_SIZE, true,
3848 AMDGPU_GEM_DOMAIN_VRAM,
3849 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
3850 NULL, NULL,
3851 &adev->gfx.rlc.cp_table_obj);
3852 if (r) {
3853 dev_warn(adev->dev, "(%d) create RLC cp table bo failed\n", r);
3854 gfx_v7_0_rlc_fini(adev);
3855 return r;
3856 }
3857 }
3858
3859 r = amdgpu_bo_reserve(adev->gfx.rlc.cp_table_obj, false);
3860 if (unlikely(r != 0)) {
3861 dev_warn(adev->dev, "(%d) reserve RLC cp table bo failed\n", r);
3862 gfx_v7_0_rlc_fini(adev);
3863 return r;
3864 }
3865 r = amdgpu_bo_pin(adev->gfx.rlc.cp_table_obj, AMDGPU_GEM_DOMAIN_VRAM,
3866 &adev->gfx.rlc.cp_table_gpu_addr);
3867 if (r) {
3868 amdgpu_bo_unreserve(adev->gfx.rlc.cp_table_obj);
3869 dev_warn(adev->dev, "(%d) pin RLC cp_table bo failed\n", r);
3870 gfx_v7_0_rlc_fini(adev);
3871 return r;
3872 }
3873 r = amdgpu_bo_kmap(adev->gfx.rlc.cp_table_obj, (void **)&adev->gfx.rlc.cp_table_ptr);
3874 if (r) {
3875 dev_warn(adev->dev, "(%d) map RLC cp table bo failed\n", r);
3876 gfx_v7_0_rlc_fini(adev);
3877 return r;
3878 }
3879
3880 gfx_v7_0_init_cp_pg_table(adev);
3881
3882 amdgpu_bo_kunmap(adev->gfx.rlc.cp_table_obj);
3883 amdgpu_bo_unreserve(adev->gfx.rlc.cp_table_obj);
3884
3885 }
3886
3887 return 0;
3888 }
3889
3890 static void gfx_v7_0_enable_lbpw(struct amdgpu_device *adev, bool enable)
3891 {
3892 u32 tmp;
3893
3894 tmp = RREG32(mmRLC_LB_CNTL);
3895 if (enable)
3896 tmp |= RLC_LB_CNTL__LOAD_BALANCE_ENABLE_MASK;
3897 else
3898 tmp &= ~RLC_LB_CNTL__LOAD_BALANCE_ENABLE_MASK;
3899 WREG32(mmRLC_LB_CNTL, tmp);
3900 }
3901
3902 static void gfx_v7_0_wait_for_rlc_serdes(struct amdgpu_device *adev)
3903 {
3904 u32 i, j, k;
3905 u32 mask;
3906
3907 mutex_lock(&adev->grbm_idx_mutex);
3908 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
3909 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
3910 gfx_v7_0_select_se_sh(adev, i, j);
3911 for (k = 0; k < adev->usec_timeout; k++) {
3912 if (RREG32(mmRLC_SERDES_CU_MASTER_BUSY) == 0)
3913 break;
3914 udelay(1);
3915 }
3916 }
3917 }
3918 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
3919 mutex_unlock(&adev->grbm_idx_mutex);
3920
3921 mask = RLC_SERDES_NONCU_MASTER_BUSY__SE_MASTER_BUSY_MASK |
3922 RLC_SERDES_NONCU_MASTER_BUSY__GC_MASTER_BUSY_MASK |
3923 RLC_SERDES_NONCU_MASTER_BUSY__TC0_MASTER_BUSY_MASK |
3924 RLC_SERDES_NONCU_MASTER_BUSY__TC1_MASTER_BUSY_MASK;
3925 for (k = 0; k < adev->usec_timeout; k++) {
3926 if ((RREG32(mmRLC_SERDES_NONCU_MASTER_BUSY) & mask) == 0)
3927 break;
3928 udelay(1);
3929 }
3930 }
3931
3932 static void gfx_v7_0_update_rlc(struct amdgpu_device *adev, u32 rlc)
3933 {
3934 u32 tmp;
3935
3936 tmp = RREG32(mmRLC_CNTL);
3937 if (tmp != rlc)
3938 WREG32(mmRLC_CNTL, rlc);
3939 }
3940
3941 static u32 gfx_v7_0_halt_rlc(struct amdgpu_device *adev)
3942 {
3943 u32 data, orig;
3944
3945 orig = data = RREG32(mmRLC_CNTL);
3946
3947 if (data & RLC_CNTL__RLC_ENABLE_F32_MASK) {
3948 u32 i;
3949
3950 data &= ~RLC_CNTL__RLC_ENABLE_F32_MASK;
3951 WREG32(mmRLC_CNTL, data);
3952
3953 for (i = 0; i < adev->usec_timeout; i++) {
3954 if ((RREG32(mmRLC_GPM_STAT) & RLC_GPM_STAT__RLC_BUSY_MASK) == 0)
3955 break;
3956 udelay(1);
3957 }
3958
3959 gfx_v7_0_wait_for_rlc_serdes(adev);
3960 }
3961
3962 return orig;
3963 }
3964
3965 void gfx_v7_0_enter_rlc_safe_mode(struct amdgpu_device *adev)
3966 {
3967 u32 tmp, i, mask;
3968
3969 tmp = 0x1 | (1 << 1);
3970 WREG32(mmRLC_GPR_REG2, tmp);
3971
3972 mask = RLC_GPM_STAT__GFX_POWER_STATUS_MASK |
3973 RLC_GPM_STAT__GFX_CLOCK_STATUS_MASK;
3974 for (i = 0; i < adev->usec_timeout; i++) {
3975 if ((RREG32(mmRLC_GPM_STAT) & mask) == mask)
3976 break;
3977 udelay(1);
3978 }
3979
3980 for (i = 0; i < adev->usec_timeout; i++) {
3981 if ((RREG32(mmRLC_GPR_REG2) & 0x1) == 0)
3982 break;
3983 udelay(1);
3984 }
3985 }
3986
3987 void gfx_v7_0_exit_rlc_safe_mode(struct amdgpu_device *adev)
3988 {
3989 u32 tmp;
3990
3991 tmp = 0x1 | (0 << 1);
3992 WREG32(mmRLC_GPR_REG2, tmp);
3993 }
3994
3995 /**
3996 * gfx_v7_0_rlc_stop - stop the RLC ME
3997 *
3998 * @adev: amdgpu_device pointer
3999 *
4000 * Halt the RLC ME (MicroEngine) (CIK).
4001 */
4002 void gfx_v7_0_rlc_stop(struct amdgpu_device *adev)
4003 {
4004 WREG32(mmRLC_CNTL, 0);
4005
4006 gfx_v7_0_enable_gui_idle_interrupt(adev, false);
4007
4008 gfx_v7_0_wait_for_rlc_serdes(adev);
4009 }
4010
4011 /**
4012 * gfx_v7_0_rlc_start - start the RLC ME
4013 *
4014 * @adev: amdgpu_device pointer
4015 *
4016 * Unhalt the RLC ME (MicroEngine) (CIK).
4017 */
4018 static void gfx_v7_0_rlc_start(struct amdgpu_device *adev)
4019 {
4020 WREG32(mmRLC_CNTL, RLC_CNTL__RLC_ENABLE_F32_MASK);
4021
4022 gfx_v7_0_enable_gui_idle_interrupt(adev, true);
4023
4024 udelay(50);
4025 }
4026
4027 static void gfx_v7_0_rlc_reset(struct amdgpu_device *adev)
4028 {
4029 u32 tmp = RREG32(mmGRBM_SOFT_RESET);
4030
4031 tmp |= GRBM_SOFT_RESET__SOFT_RESET_RLC_MASK;
4032 WREG32(mmGRBM_SOFT_RESET, tmp);
4033 udelay(50);
4034 tmp &= ~GRBM_SOFT_RESET__SOFT_RESET_RLC_MASK;
4035 WREG32(mmGRBM_SOFT_RESET, tmp);
4036 udelay(50);
4037 }
4038
4039 /**
4040 * gfx_v7_0_rlc_resume - setup the RLC hw
4041 *
4042 * @adev: amdgpu_device pointer
4043 *
4044 * Initialize the RLC registers, load the ucode,
4045 * and start the RLC (CIK).
4046 * Returns 0 for success, -EINVAL if the ucode is not available.
4047 */
4048 static int gfx_v7_0_rlc_resume(struct amdgpu_device *adev)
4049 {
4050 const struct rlc_firmware_header_v1_0 *hdr;
4051 const __le32 *fw_data;
4052 unsigned i, fw_size;
4053 u32 tmp;
4054
4055 if (!adev->gfx.rlc_fw)
4056 return -EINVAL;
4057
4058 hdr = (const struct rlc_firmware_header_v1_0 *)adev->gfx.rlc_fw->data;
4059 amdgpu_ucode_print_rlc_hdr(&hdr->header);
4060 adev->gfx.rlc_fw_version = le32_to_cpu(hdr->header.ucode_version);
4061 adev->gfx.rlc_feature_version = le32_to_cpu(
4062 hdr->ucode_feature_version);
4063
4064 gfx_v7_0_rlc_stop(adev);
4065
4066 /* disable CG */
4067 tmp = RREG32(mmRLC_CGCG_CGLS_CTRL) & 0xfffffffc;
4068 WREG32(mmRLC_CGCG_CGLS_CTRL, tmp);
4069
4070 gfx_v7_0_rlc_reset(adev);
4071
4072 gfx_v7_0_init_pg(adev);
4073
4074 WREG32(mmRLC_LB_CNTR_INIT, 0);
4075 WREG32(mmRLC_LB_CNTR_MAX, 0x00008000);
4076
4077 mutex_lock(&adev->grbm_idx_mutex);
4078 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
4079 WREG32(mmRLC_LB_INIT_CU_MASK, 0xffffffff);
4080 WREG32(mmRLC_LB_PARAMS, 0x00600408);
4081 WREG32(mmRLC_LB_CNTL, 0x80000004);
4082 mutex_unlock(&adev->grbm_idx_mutex);
4083
4084 WREG32(mmRLC_MC_CNTL, 0);
4085 WREG32(mmRLC_UCODE_CNTL, 0);
4086
4087 fw_data = (const __le32 *)
4088 (adev->gfx.rlc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
4089 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
4090 WREG32(mmRLC_GPM_UCODE_ADDR, 0);
4091 for (i = 0; i < fw_size; i++)
4092 WREG32(mmRLC_GPM_UCODE_DATA, le32_to_cpup(fw_data++));
4093 WREG32(mmRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version);
4094
4095 /* XXX - find out what chips support lbpw */
4096 gfx_v7_0_enable_lbpw(adev, false);
4097
4098 if (adev->asic_type == CHIP_BONAIRE)
4099 WREG32(mmRLC_DRIVER_CPDMA_STATUS, 0);
4100
4101 gfx_v7_0_rlc_start(adev);
4102
4103 return 0;
4104 }
4105
4106 static void gfx_v7_0_enable_cgcg(struct amdgpu_device *adev, bool enable)
4107 {
4108 u32 data, orig, tmp, tmp2;
4109
4110 orig = data = RREG32(mmRLC_CGCG_CGLS_CTRL);
4111
4112 if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_CGCG)) {
4113 gfx_v7_0_enable_gui_idle_interrupt(adev, true);
4114
4115 tmp = gfx_v7_0_halt_rlc(adev);
4116
4117 mutex_lock(&adev->grbm_idx_mutex);
4118 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
4119 WREG32(mmRLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
4120 WREG32(mmRLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
4121 tmp2 = RLC_SERDES_WR_CTRL__BPM_ADDR_MASK |
4122 RLC_SERDES_WR_CTRL__CGCG_OVERRIDE_0_MASK |
4123 RLC_SERDES_WR_CTRL__CGLS_ENABLE_MASK;
4124 WREG32(mmRLC_SERDES_WR_CTRL, tmp2);
4125 mutex_unlock(&adev->grbm_idx_mutex);
4126
4127 gfx_v7_0_update_rlc(adev, tmp);
4128
4129 data |= RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK | RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK;
4130 } else {
4131 gfx_v7_0_enable_gui_idle_interrupt(adev, false);
4132
4133 RREG32(mmCB_CGTT_SCLK_CTRL);
4134 RREG32(mmCB_CGTT_SCLK_CTRL);
4135 RREG32(mmCB_CGTT_SCLK_CTRL);
4136 RREG32(mmCB_CGTT_SCLK_CTRL);
4137
4138 data &= ~(RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK | RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK);
4139 }
4140
4141 if (orig != data)
4142 WREG32(mmRLC_CGCG_CGLS_CTRL, data);
4143
4144 }
4145
4146 static void gfx_v7_0_enable_mgcg(struct amdgpu_device *adev, bool enable)
4147 {
4148 u32 data, orig, tmp = 0;
4149
4150 if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_MGCG)) {
4151 if (adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_MGLS) {
4152 if (adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_CP_LS) {
4153 orig = data = RREG32(mmCP_MEM_SLP_CNTL);
4154 data |= CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
4155 if (orig != data)
4156 WREG32(mmCP_MEM_SLP_CNTL, data);
4157 }
4158 }
4159
4160 orig = data = RREG32(mmRLC_CGTT_MGCG_OVERRIDE);
4161 data |= 0x00000001;
4162 data &= 0xfffffffd;
4163 if (orig != data)
4164 WREG32(mmRLC_CGTT_MGCG_OVERRIDE, data);
4165
4166 tmp = gfx_v7_0_halt_rlc(adev);
4167
4168 mutex_lock(&adev->grbm_idx_mutex);
4169 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
4170 WREG32(mmRLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
4171 WREG32(mmRLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
4172 data = RLC_SERDES_WR_CTRL__BPM_ADDR_MASK |
4173 RLC_SERDES_WR_CTRL__MGCG_OVERRIDE_0_MASK;
4174 WREG32(mmRLC_SERDES_WR_CTRL, data);
4175 mutex_unlock(&adev->grbm_idx_mutex);
4176
4177 gfx_v7_0_update_rlc(adev, tmp);
4178
4179 if (adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_CGTS) {
4180 orig = data = RREG32(mmCGTS_SM_CTRL_REG);
4181 data &= ~CGTS_SM_CTRL_REG__SM_MODE_MASK;
4182 data |= (0x2 << CGTS_SM_CTRL_REG__SM_MODE__SHIFT);
4183 data |= CGTS_SM_CTRL_REG__SM_MODE_ENABLE_MASK;
4184 data &= ~CGTS_SM_CTRL_REG__OVERRIDE_MASK;
4185 if ((adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_MGLS) &&
4186 (adev->cg_flags & AMDGPU_CG_SUPPORT_GFX_CGTS_LS))
4187 data &= ~CGTS_SM_CTRL_REG__LS_OVERRIDE_MASK;
4188 data &= ~CGTS_SM_CTRL_REG__ON_MONITOR_ADD_MASK;
4189 data |= CGTS_SM_CTRL_REG__ON_MONITOR_ADD_EN_MASK;
4190 data |= (0x96 << CGTS_SM_CTRL_REG__ON_MONITOR_ADD__SHIFT);
4191 if (orig != data)
4192 WREG32(mmCGTS_SM_CTRL_REG, data);
4193 }
4194 } else {
4195 orig = data = RREG32(mmRLC_CGTT_MGCG_OVERRIDE);
4196 data |= 0x00000003;
4197 if (orig != data)
4198 WREG32(mmRLC_CGTT_MGCG_OVERRIDE, data);
4199
4200 data = RREG32(mmRLC_MEM_SLP_CNTL);
4201 if (data & RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK) {
4202 data &= ~RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK;
4203 WREG32(mmRLC_MEM_SLP_CNTL, data);
4204 }
4205
4206 data = RREG32(mmCP_MEM_SLP_CNTL);
4207 if (data & CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK) {
4208 data &= ~CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
4209 WREG32(mmCP_MEM_SLP_CNTL, data);
4210 }
4211
4212 orig = data = RREG32(mmCGTS_SM_CTRL_REG);
4213 data |= CGTS_SM_CTRL_REG__OVERRIDE_MASK | CGTS_SM_CTRL_REG__LS_OVERRIDE_MASK;
4214 if (orig != data)
4215 WREG32(mmCGTS_SM_CTRL_REG, data);
4216
4217 tmp = gfx_v7_0_halt_rlc(adev);
4218
4219 mutex_lock(&adev->grbm_idx_mutex);
4220 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
4221 WREG32(mmRLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
4222 WREG32(mmRLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
4223 data = RLC_SERDES_WR_CTRL__BPM_ADDR_MASK | RLC_SERDES_WR_CTRL__MGCG_OVERRIDE_1_MASK;
4224 WREG32(mmRLC_SERDES_WR_CTRL, data);
4225 mutex_unlock(&adev->grbm_idx_mutex);
4226
4227 gfx_v7_0_update_rlc(adev, tmp);
4228 }
4229 }
4230
4231 static void gfx_v7_0_update_cg(struct amdgpu_device *adev,
4232 bool enable)
4233 {
4234 gfx_v7_0_enable_gui_idle_interrupt(adev, false);
4235 /* order matters! */
4236 if (enable) {
4237 gfx_v7_0_enable_mgcg(adev, true);
4238 gfx_v7_0_enable_cgcg(adev, true);
4239 } else {
4240 gfx_v7_0_enable_cgcg(adev, false);
4241 gfx_v7_0_enable_mgcg(adev, false);
4242 }
4243 gfx_v7_0_enable_gui_idle_interrupt(adev, true);
4244 }
4245
4246 static void gfx_v7_0_enable_sclk_slowdown_on_pu(struct amdgpu_device *adev,
4247 bool enable)
4248 {
4249 u32 data, orig;
4250
4251 orig = data = RREG32(mmRLC_PG_CNTL);
4252 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_RLC_SMU_HS))
4253 data |= RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PU_ENABLE_MASK;
4254 else
4255 data &= ~RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PU_ENABLE_MASK;
4256 if (orig != data)
4257 WREG32(mmRLC_PG_CNTL, data);
4258 }
4259
4260 static void gfx_v7_0_enable_sclk_slowdown_on_pd(struct amdgpu_device *adev,
4261 bool enable)
4262 {
4263 u32 data, orig;
4264
4265 orig = data = RREG32(mmRLC_PG_CNTL);
4266 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_RLC_SMU_HS))
4267 data |= RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PD_ENABLE_MASK;
4268 else
4269 data &= ~RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PD_ENABLE_MASK;
4270 if (orig != data)
4271 WREG32(mmRLC_PG_CNTL, data);
4272 }
4273
4274 static void gfx_v7_0_enable_cp_pg(struct amdgpu_device *adev, bool enable)
4275 {
4276 u32 data, orig;
4277
4278 orig = data = RREG32(mmRLC_PG_CNTL);
4279 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_CP))
4280 data &= ~0x8000;
4281 else
4282 data |= 0x8000;
4283 if (orig != data)
4284 WREG32(mmRLC_PG_CNTL, data);
4285 }
4286
4287 static void gfx_v7_0_enable_gds_pg(struct amdgpu_device *adev, bool enable)
4288 {
4289 u32 data, orig;
4290
4291 orig = data = RREG32(mmRLC_PG_CNTL);
4292 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_GDS))
4293 data &= ~0x2000;
4294 else
4295 data |= 0x2000;
4296 if (orig != data)
4297 WREG32(mmRLC_PG_CNTL, data);
4298 }
4299
4300 static void gfx_v7_0_init_cp_pg_table(struct amdgpu_device *adev)
4301 {
4302 const __le32 *fw_data;
4303 volatile u32 *dst_ptr;
4304 int me, i, max_me = 4;
4305 u32 bo_offset = 0;
4306 u32 table_offset, table_size;
4307
4308 if (adev->asic_type == CHIP_KAVERI)
4309 max_me = 5;
4310
4311 if (adev->gfx.rlc.cp_table_ptr == NULL)
4312 return;
4313
4314 /* write the cp table buffer */
4315 dst_ptr = adev->gfx.rlc.cp_table_ptr;
4316 for (me = 0; me < max_me; me++) {
4317 if (me == 0) {
4318 const struct gfx_firmware_header_v1_0 *hdr =
4319 (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
4320 fw_data = (const __le32 *)
4321 (adev->gfx.ce_fw->data +
4322 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
4323 table_offset = le32_to_cpu(hdr->jt_offset);
4324 table_size = le32_to_cpu(hdr->jt_size);
4325 } else if (me == 1) {
4326 const struct gfx_firmware_header_v1_0 *hdr =
4327 (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
4328 fw_data = (const __le32 *)
4329 (adev->gfx.pfp_fw->data +
4330 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
4331 table_offset = le32_to_cpu(hdr->jt_offset);
4332 table_size = le32_to_cpu(hdr->jt_size);
4333 } else if (me == 2) {
4334 const struct gfx_firmware_header_v1_0 *hdr =
4335 (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
4336 fw_data = (const __le32 *)
4337 (adev->gfx.me_fw->data +
4338 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
4339 table_offset = le32_to_cpu(hdr->jt_offset);
4340 table_size = le32_to_cpu(hdr->jt_size);
4341 } else if (me == 3) {
4342 const struct gfx_firmware_header_v1_0 *hdr =
4343 (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
4344 fw_data = (const __le32 *)
4345 (adev->gfx.mec_fw->data +
4346 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
4347 table_offset = le32_to_cpu(hdr->jt_offset);
4348 table_size = le32_to_cpu(hdr->jt_size);
4349 } else {
4350 const struct gfx_firmware_header_v1_0 *hdr =
4351 (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec2_fw->data;
4352 fw_data = (const __le32 *)
4353 (adev->gfx.mec2_fw->data +
4354 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
4355 table_offset = le32_to_cpu(hdr->jt_offset);
4356 table_size = le32_to_cpu(hdr->jt_size);
4357 }
4358
4359 for (i = 0; i < table_size; i ++) {
4360 dst_ptr[bo_offset + i] =
4361 cpu_to_le32(le32_to_cpu(fw_data[table_offset + i]));
4362 }
4363
4364 bo_offset += table_size;
4365 }
4366 }
4367
4368 static void gfx_v7_0_enable_gfx_cgpg(struct amdgpu_device *adev,
4369 bool enable)
4370 {
4371 u32 data, orig;
4372
4373 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_GFX_PG)) {
4374 orig = data = RREG32(mmRLC_PG_CNTL);
4375 data |= RLC_PG_CNTL__GFX_POWER_GATING_ENABLE_MASK;
4376 if (orig != data)
4377 WREG32(mmRLC_PG_CNTL, data);
4378
4379 orig = data = RREG32(mmRLC_AUTO_PG_CTRL);
4380 data |= RLC_AUTO_PG_CTRL__AUTO_PG_EN_MASK;
4381 if (orig != data)
4382 WREG32(mmRLC_AUTO_PG_CTRL, data);
4383 } else {
4384 orig = data = RREG32(mmRLC_PG_CNTL);
4385 data &= ~RLC_PG_CNTL__GFX_POWER_GATING_ENABLE_MASK;
4386 if (orig != data)
4387 WREG32(mmRLC_PG_CNTL, data);
4388
4389 orig = data = RREG32(mmRLC_AUTO_PG_CTRL);
4390 data &= ~RLC_AUTO_PG_CTRL__AUTO_PG_EN_MASK;
4391 if (orig != data)
4392 WREG32(mmRLC_AUTO_PG_CTRL, data);
4393
4394 data = RREG32(mmDB_RENDER_CONTROL);
4395 }
4396 }
4397
4398 static u32 gfx_v7_0_get_cu_active_bitmap(struct amdgpu_device *adev,
4399 u32 se, u32 sh)
4400 {
4401 u32 mask = 0, tmp, tmp1;
4402 int i;
4403
4404 gfx_v7_0_select_se_sh(adev, se, sh);
4405 tmp = RREG32(mmCC_GC_SHADER_ARRAY_CONFIG);
4406 tmp1 = RREG32(mmGC_USER_SHADER_ARRAY_CONFIG);
4407 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
4408
4409 tmp &= 0xffff0000;
4410
4411 tmp |= tmp1;
4412 tmp >>= 16;
4413
4414 for (i = 0; i < adev->gfx.config.max_cu_per_sh; i ++) {
4415 mask <<= 1;
4416 mask |= 1;
4417 }
4418
4419 return (~tmp) & mask;
4420 }
4421
4422 static void gfx_v7_0_init_ao_cu_mask(struct amdgpu_device *adev)
4423 {
4424 uint32_t tmp, active_cu_number;
4425 struct amdgpu_cu_info cu_info;
4426
4427 gfx_v7_0_get_cu_info(adev, &cu_info);
4428 tmp = cu_info.ao_cu_mask;
4429 active_cu_number = cu_info.number;
4430
4431 WREG32(mmRLC_PG_ALWAYS_ON_CU_MASK, tmp);
4432
4433 tmp = RREG32(mmRLC_MAX_PG_CU);
4434 tmp &= ~RLC_MAX_PG_CU__MAX_POWERED_UP_CU_MASK;
4435 tmp |= (active_cu_number << RLC_MAX_PG_CU__MAX_POWERED_UP_CU__SHIFT);
4436 WREG32(mmRLC_MAX_PG_CU, tmp);
4437 }
4438
4439 static void gfx_v7_0_enable_gfx_static_mgpg(struct amdgpu_device *adev,
4440 bool enable)
4441 {
4442 u32 data, orig;
4443
4444 orig = data = RREG32(mmRLC_PG_CNTL);
4445 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_GFX_SMG))
4446 data |= RLC_PG_CNTL__STATIC_PER_CU_PG_ENABLE_MASK;
4447 else
4448 data &= ~RLC_PG_CNTL__STATIC_PER_CU_PG_ENABLE_MASK;
4449 if (orig != data)
4450 WREG32(mmRLC_PG_CNTL, data);
4451 }
4452
4453 static void gfx_v7_0_enable_gfx_dynamic_mgpg(struct amdgpu_device *adev,
4454 bool enable)
4455 {
4456 u32 data, orig;
4457
4458 orig = data = RREG32(mmRLC_PG_CNTL);
4459 if (enable && (adev->pg_flags & AMDGPU_PG_SUPPORT_GFX_DMG))
4460 data |= RLC_PG_CNTL__DYN_PER_CU_PG_ENABLE_MASK;
4461 else
4462 data &= ~RLC_PG_CNTL__DYN_PER_CU_PG_ENABLE_MASK;
4463 if (orig != data)
4464 WREG32(mmRLC_PG_CNTL, data);
4465 }
4466
4467 #define RLC_SAVE_AND_RESTORE_STARTING_OFFSET 0x90
4468 #define RLC_CLEAR_STATE_DESCRIPTOR_OFFSET 0x3D
4469
4470 static void gfx_v7_0_init_gfx_cgpg(struct amdgpu_device *adev)
4471 {
4472 u32 data, orig;
4473 u32 i;
4474
4475 if (adev->gfx.rlc.cs_data) {
4476 WREG32(mmRLC_GPM_SCRATCH_ADDR, RLC_CLEAR_STATE_DESCRIPTOR_OFFSET);
4477 WREG32(mmRLC_GPM_SCRATCH_DATA, upper_32_bits(adev->gfx.rlc.clear_state_gpu_addr));
4478 WREG32(mmRLC_GPM_SCRATCH_DATA, lower_32_bits(adev->gfx.rlc.clear_state_gpu_addr));
4479 WREG32(mmRLC_GPM_SCRATCH_DATA, adev->gfx.rlc.clear_state_size);
4480 } else {
4481 WREG32(mmRLC_GPM_SCRATCH_ADDR, RLC_CLEAR_STATE_DESCRIPTOR_OFFSET);
4482 for (i = 0; i < 3; i++)
4483 WREG32(mmRLC_GPM_SCRATCH_DATA, 0);
4484 }
4485 if (adev->gfx.rlc.reg_list) {
4486 WREG32(mmRLC_GPM_SCRATCH_ADDR, RLC_SAVE_AND_RESTORE_STARTING_OFFSET);
4487 for (i = 0; i < adev->gfx.rlc.reg_list_size; i++)
4488 WREG32(mmRLC_GPM_SCRATCH_DATA, adev->gfx.rlc.reg_list[i]);
4489 }
4490
4491 orig = data = RREG32(mmRLC_PG_CNTL);
4492 data |= RLC_PG_CNTL__GFX_POWER_GATING_SRC_MASK;
4493 if (orig != data)
4494 WREG32(mmRLC_PG_CNTL, data);
4495
4496 WREG32(mmRLC_SAVE_AND_RESTORE_BASE, adev->gfx.rlc.save_restore_gpu_addr >> 8);
4497 WREG32(mmRLC_JUMP_TABLE_RESTORE, adev->gfx.rlc.cp_table_gpu_addr >> 8);
4498
4499 data = RREG32(mmCP_RB_WPTR_POLL_CNTL);
4500 data &= ~CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT_MASK;
4501 data |= (0x60 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
4502 WREG32(mmCP_RB_WPTR_POLL_CNTL, data);
4503
4504 data = 0x10101010;
4505 WREG32(mmRLC_PG_DELAY, data);
4506
4507 data = RREG32(mmRLC_PG_DELAY_2);
4508 data &= ~0xff;
4509 data |= 0x3;
4510 WREG32(mmRLC_PG_DELAY_2, data);
4511
4512 data = RREG32(mmRLC_AUTO_PG_CTRL);
4513 data &= ~RLC_AUTO_PG_CTRL__GRBM_REG_SAVE_GFX_IDLE_THRESHOLD_MASK;
4514 data |= (0x700 << RLC_AUTO_PG_CTRL__GRBM_REG_SAVE_GFX_IDLE_THRESHOLD__SHIFT);
4515 WREG32(mmRLC_AUTO_PG_CTRL, data);
4516
4517 }
4518
4519 static void gfx_v7_0_update_gfx_pg(struct amdgpu_device *adev, bool enable)
4520 {
4521 gfx_v7_0_enable_gfx_cgpg(adev, enable);
4522 gfx_v7_0_enable_gfx_static_mgpg(adev, enable);
4523 gfx_v7_0_enable_gfx_dynamic_mgpg(adev, enable);
4524 }
4525
4526 static u32 gfx_v7_0_get_csb_size(struct amdgpu_device *adev)
4527 {
4528 u32 count = 0;
4529 const struct cs_section_def *sect = NULL;
4530 const struct cs_extent_def *ext = NULL;
4531
4532 if (adev->gfx.rlc.cs_data == NULL)
4533 return 0;
4534
4535 /* begin clear state */
4536 count += 2;
4537 /* context control state */
4538 count += 3;
4539
4540 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
4541 for (ext = sect->section; ext->extent != NULL; ++ext) {
4542 if (sect->id == SECT_CONTEXT)
4543 count += 2 + ext->reg_count;
4544 else
4545 return 0;
4546 }
4547 }
4548 /* pa_sc_raster_config/pa_sc_raster_config1 */
4549 count += 4;
4550 /* end clear state */
4551 count += 2;
4552 /* clear state */
4553 count += 2;
4554
4555 return count;
4556 }
4557
4558 static void gfx_v7_0_get_csb_buffer(struct amdgpu_device *adev,
4559 volatile u32 *buffer)
4560 {
4561 u32 count = 0, i;
4562 const struct cs_section_def *sect = NULL;
4563 const struct cs_extent_def *ext = NULL;
4564
4565 if (adev->gfx.rlc.cs_data == NULL)
4566 return;
4567 if (buffer == NULL)
4568 return;
4569
4570 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
4571 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
4572
4573 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
4574 buffer[count++] = cpu_to_le32(0x80000000);
4575 buffer[count++] = cpu_to_le32(0x80000000);
4576
4577 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
4578 for (ext = sect->section; ext->extent != NULL; ++ext) {
4579 if (sect->id == SECT_CONTEXT) {
4580 buffer[count++] =
4581 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
4582 buffer[count++] = cpu_to_le32(ext->reg_index - PACKET3_SET_CONTEXT_REG_START);
4583 for (i = 0; i < ext->reg_count; i++)
4584 buffer[count++] = cpu_to_le32(ext->extent[i]);
4585 } else {
4586 return;
4587 }
4588 }
4589 }
4590
4591 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 2));
4592 buffer[count++] = cpu_to_le32(mmPA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
4593 switch (adev->asic_type) {
4594 case CHIP_BONAIRE:
4595 buffer[count++] = cpu_to_le32(0x16000012);
4596 buffer[count++] = cpu_to_le32(0x00000000);
4597 break;
4598 case CHIP_KAVERI:
4599 buffer[count++] = cpu_to_le32(0x00000000); /* XXX */
4600 buffer[count++] = cpu_to_le32(0x00000000);
4601 break;
4602 case CHIP_KABINI:
4603 case CHIP_MULLINS:
4604 buffer[count++] = cpu_to_le32(0x00000000); /* XXX */
4605 buffer[count++] = cpu_to_le32(0x00000000);
4606 break;
4607 case CHIP_HAWAII:
4608 buffer[count++] = cpu_to_le32(0x3a00161a);
4609 buffer[count++] = cpu_to_le32(0x0000002e);
4610 break;
4611 default:
4612 buffer[count++] = cpu_to_le32(0x00000000);
4613 buffer[count++] = cpu_to_le32(0x00000000);
4614 break;
4615 }
4616
4617 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
4618 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
4619
4620 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
4621 buffer[count++] = cpu_to_le32(0);
4622 }
4623
4624 static void gfx_v7_0_init_pg(struct amdgpu_device *adev)
4625 {
4626 if (adev->pg_flags & (AMDGPU_PG_SUPPORT_GFX_PG |
4627 AMDGPU_PG_SUPPORT_GFX_SMG |
4628 AMDGPU_PG_SUPPORT_GFX_DMG |
4629 AMDGPU_PG_SUPPORT_CP |
4630 AMDGPU_PG_SUPPORT_GDS |
4631 AMDGPU_PG_SUPPORT_RLC_SMU_HS)) {
4632 gfx_v7_0_enable_sclk_slowdown_on_pu(adev, true);
4633 gfx_v7_0_enable_sclk_slowdown_on_pd(adev, true);
4634 if (adev->pg_flags & AMDGPU_PG_SUPPORT_GFX_PG) {
4635 gfx_v7_0_init_gfx_cgpg(adev);
4636 gfx_v7_0_enable_cp_pg(adev, true);
4637 gfx_v7_0_enable_gds_pg(adev, true);
4638 }
4639 gfx_v7_0_init_ao_cu_mask(adev);
4640 gfx_v7_0_update_gfx_pg(adev, true);
4641 }
4642 }
4643
4644 static void gfx_v7_0_fini_pg(struct amdgpu_device *adev)
4645 {
4646 if (adev->pg_flags & (AMDGPU_PG_SUPPORT_GFX_PG |
4647 AMDGPU_PG_SUPPORT_GFX_SMG |
4648 AMDGPU_PG_SUPPORT_GFX_DMG |
4649 AMDGPU_PG_SUPPORT_CP |
4650 AMDGPU_PG_SUPPORT_GDS |
4651 AMDGPU_PG_SUPPORT_RLC_SMU_HS)) {
4652 gfx_v7_0_update_gfx_pg(adev, false);
4653 if (adev->pg_flags & AMDGPU_PG_SUPPORT_GFX_PG) {
4654 gfx_v7_0_enable_cp_pg(adev, false);
4655 gfx_v7_0_enable_gds_pg(adev, false);
4656 }
4657 }
4658 }
4659
4660 /**
4661 * gfx_v7_0_get_gpu_clock_counter - return GPU clock counter snapshot
4662 *
4663 * @adev: amdgpu_device pointer
4664 *
4665 * Fetches a GPU clock counter snapshot (SI).
4666 * Returns the 64 bit clock counter snapshot.
4667 */
4668 uint64_t gfx_v7_0_get_gpu_clock_counter(struct amdgpu_device *adev)
4669 {
4670 uint64_t clock;
4671
4672 mutex_lock(&adev->gfx.gpu_clock_mutex);
4673 WREG32(mmRLC_CAPTURE_GPU_CLOCK_COUNT, 1);
4674 clock = (uint64_t)RREG32(mmRLC_GPU_CLOCK_COUNT_LSB) |
4675 ((uint64_t)RREG32(mmRLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
4676 mutex_unlock(&adev->gfx.gpu_clock_mutex);
4677 return clock;
4678 }
4679
4680 static void gfx_v7_0_ring_emit_gds_switch(struct amdgpu_ring *ring,
4681 uint32_t vmid,
4682 uint32_t gds_base, uint32_t gds_size,
4683 uint32_t gws_base, uint32_t gws_size,
4684 uint32_t oa_base, uint32_t oa_size)
4685 {
4686 gds_base = gds_base >> AMDGPU_GDS_SHIFT;
4687 gds_size = gds_size >> AMDGPU_GDS_SHIFT;
4688
4689 gws_base = gws_base >> AMDGPU_GWS_SHIFT;
4690 gws_size = gws_size >> AMDGPU_GWS_SHIFT;
4691
4692 oa_base = oa_base >> AMDGPU_OA_SHIFT;
4693 oa_size = oa_size >> AMDGPU_OA_SHIFT;
4694
4695 /* GDS Base */
4696 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4697 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4698 WRITE_DATA_DST_SEL(0)));
4699 amdgpu_ring_write(ring, amdgpu_gds_reg_offset[vmid].mem_base);
4700 amdgpu_ring_write(ring, 0);
4701 amdgpu_ring_write(ring, gds_base);
4702
4703 /* GDS Size */
4704 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4705 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4706 WRITE_DATA_DST_SEL(0)));
4707 amdgpu_ring_write(ring, amdgpu_gds_reg_offset[vmid].mem_size);
4708 amdgpu_ring_write(ring, 0);
4709 amdgpu_ring_write(ring, gds_size);
4710
4711 /* GWS */
4712 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4713 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4714 WRITE_DATA_DST_SEL(0)));
4715 amdgpu_ring_write(ring, amdgpu_gds_reg_offset[vmid].gws);
4716 amdgpu_ring_write(ring, 0);
4717 amdgpu_ring_write(ring, gws_size << GDS_GWS_VMID0__SIZE__SHIFT | gws_base);
4718
4719 /* OA */
4720 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4721 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4722 WRITE_DATA_DST_SEL(0)));
4723 amdgpu_ring_write(ring, amdgpu_gds_reg_offset[vmid].oa);
4724 amdgpu_ring_write(ring, 0);
4725 amdgpu_ring_write(ring, (1 << (oa_size + oa_base)) - (1 << oa_base));
4726 }
4727
4728 static int gfx_v7_0_early_init(void *handle)
4729 {
4730 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4731
4732 adev->gfx.num_gfx_rings = GFX7_NUM_GFX_RINGS;
4733 adev->gfx.num_compute_rings = GFX7_NUM_COMPUTE_RINGS;
4734 gfx_v7_0_set_ring_funcs(adev);
4735 gfx_v7_0_set_irq_funcs(adev);
4736 gfx_v7_0_set_gds_init(adev);
4737
4738 return 0;
4739 }
4740
4741 static int gfx_v7_0_sw_init(void *handle)
4742 {
4743 struct amdgpu_ring *ring;
4744 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4745 int i, r;
4746
4747 /* EOP Event */
4748 r = amdgpu_irq_add_id(adev, 181, &adev->gfx.eop_irq);
4749 if (r)
4750 return r;
4751
4752 /* Privileged reg */
4753 r = amdgpu_irq_add_id(adev, 184, &adev->gfx.priv_reg_irq);
4754 if (r)
4755 return r;
4756
4757 /* Privileged inst */
4758 r = amdgpu_irq_add_id(adev, 185, &adev->gfx.priv_inst_irq);
4759 if (r)
4760 return r;
4761
4762 gfx_v7_0_scratch_init(adev);
4763
4764 r = gfx_v7_0_init_microcode(adev);
4765 if (r) {
4766 DRM_ERROR("Failed to load gfx firmware!\n");
4767 return r;
4768 }
4769
4770 r = gfx_v7_0_rlc_init(adev);
4771 if (r) {
4772 DRM_ERROR("Failed to init rlc BOs!\n");
4773 return r;
4774 }
4775
4776 /* allocate mec buffers */
4777 r = gfx_v7_0_mec_init(adev);
4778 if (r) {
4779 DRM_ERROR("Failed to init MEC BOs!\n");
4780 return r;
4781 }
4782
4783 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
4784 ring = &adev->gfx.gfx_ring[i];
4785 ring->ring_obj = NULL;
4786 sprintf(ring->name, "gfx");
4787 r = amdgpu_ring_init(adev, ring, 1024 * 1024,
4788 PACKET3(PACKET3_NOP, 0x3FFF), 0xf,
4789 &adev->gfx.eop_irq, AMDGPU_CP_IRQ_GFX_EOP,
4790 AMDGPU_RING_TYPE_GFX);
4791 if (r)
4792 return r;
4793 }
4794
4795 /* set up the compute queues */
4796 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
4797 unsigned irq_type;
4798
4799 /* max 32 queues per MEC */
4800 if ((i >= 32) || (i >= AMDGPU_MAX_COMPUTE_RINGS)) {
4801 DRM_ERROR("Too many (%d) compute rings!\n", i);
4802 break;
4803 }
4804 ring = &adev->gfx.compute_ring[i];
4805 ring->ring_obj = NULL;
4806 ring->use_doorbell = true;
4807 ring->doorbell_index = AMDGPU_DOORBELL_MEC_RING0 + i;
4808 ring->me = 1; /* first MEC */
4809 ring->pipe = i / 8;
4810 ring->queue = i % 8;
4811 sprintf(ring->name, "comp %d.%d.%d", ring->me, ring->pipe, ring->queue);
4812 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP + ring->pipe;
4813 /* type-2 packets are deprecated on MEC, use type-3 instead */
4814 r = amdgpu_ring_init(adev, ring, 1024 * 1024,
4815 PACKET3(PACKET3_NOP, 0x3FFF), 0xf,
4816 &adev->gfx.eop_irq, irq_type,
4817 AMDGPU_RING_TYPE_COMPUTE);
4818 if (r)
4819 return r;
4820 }
4821
4822 /* reserve GDS, GWS and OA resource for gfx */
4823 r = amdgpu_bo_create(adev, adev->gds.mem.gfx_partition_size,
4824 PAGE_SIZE, true,
4825 AMDGPU_GEM_DOMAIN_GDS, 0,
4826 NULL, NULL, &adev->gds.gds_gfx_bo);
4827 if (r)
4828 return r;
4829
4830 r = amdgpu_bo_create(adev, adev->gds.gws.gfx_partition_size,
4831 PAGE_SIZE, true,
4832 AMDGPU_GEM_DOMAIN_GWS, 0,
4833 NULL, NULL, &adev->gds.gws_gfx_bo);
4834 if (r)
4835 return r;
4836
4837 r = amdgpu_bo_create(adev, adev->gds.oa.gfx_partition_size,
4838 PAGE_SIZE, true,
4839 AMDGPU_GEM_DOMAIN_OA, 0,
4840 NULL, NULL, &adev->gds.oa_gfx_bo);
4841 if (r)
4842 return r;
4843
4844 return r;
4845 }
4846
4847 static int gfx_v7_0_sw_fini(void *handle)
4848 {
4849 int i;
4850 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4851
4852 amdgpu_bo_unref(&adev->gds.oa_gfx_bo);
4853 amdgpu_bo_unref(&adev->gds.gws_gfx_bo);
4854 amdgpu_bo_unref(&adev->gds.gds_gfx_bo);
4855
4856 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
4857 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
4858 for (i = 0; i < adev->gfx.num_compute_rings; i++)
4859 amdgpu_ring_fini(&adev->gfx.compute_ring[i]);
4860
4861 gfx_v7_0_cp_compute_fini(adev);
4862 gfx_v7_0_rlc_fini(adev);
4863 gfx_v7_0_mec_fini(adev);
4864
4865 return 0;
4866 }
4867
4868 static int gfx_v7_0_hw_init(void *handle)
4869 {
4870 int r;
4871 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4872
4873 gfx_v7_0_gpu_init(adev);
4874
4875 /* init rlc */
4876 r = gfx_v7_0_rlc_resume(adev);
4877 if (r)
4878 return r;
4879
4880 r = gfx_v7_0_cp_resume(adev);
4881 if (r)
4882 return r;
4883
4884 adev->gfx.ce_ram_size = 0x8000;
4885
4886 return r;
4887 }
4888
4889 static int gfx_v7_0_hw_fini(void *handle)
4890 {
4891 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4892
4893 gfx_v7_0_cp_enable(adev, false);
4894 gfx_v7_0_rlc_stop(adev);
4895 gfx_v7_0_fini_pg(adev);
4896
4897 return 0;
4898 }
4899
4900 static int gfx_v7_0_suspend(void *handle)
4901 {
4902 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4903
4904 return gfx_v7_0_hw_fini(adev);
4905 }
4906
4907 static int gfx_v7_0_resume(void *handle)
4908 {
4909 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4910
4911 return gfx_v7_0_hw_init(adev);
4912 }
4913
4914 static bool gfx_v7_0_is_idle(void *handle)
4915 {
4916 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4917
4918 if (RREG32(mmGRBM_STATUS) & GRBM_STATUS__GUI_ACTIVE_MASK)
4919 return false;
4920 else
4921 return true;
4922 }
4923
4924 static int gfx_v7_0_wait_for_idle(void *handle)
4925 {
4926 unsigned i;
4927 u32 tmp;
4928 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4929
4930 for (i = 0; i < adev->usec_timeout; i++) {
4931 /* read MC_STATUS */
4932 tmp = RREG32(mmGRBM_STATUS) & GRBM_STATUS__GUI_ACTIVE_MASK;
4933
4934 if (!tmp)
4935 return 0;
4936 udelay(1);
4937 }
4938 return -ETIMEDOUT;
4939 }
4940
4941 static void gfx_v7_0_print_status(void *handle)
4942 {
4943 int i;
4944 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4945
4946 dev_info(adev->dev, "GFX 7.x registers\n");
4947 dev_info(adev->dev, " GRBM_STATUS=0x%08X\n",
4948 RREG32(mmGRBM_STATUS));
4949 dev_info(adev->dev, " GRBM_STATUS2=0x%08X\n",
4950 RREG32(mmGRBM_STATUS2));
4951 dev_info(adev->dev, " GRBM_STATUS_SE0=0x%08X\n",
4952 RREG32(mmGRBM_STATUS_SE0));
4953 dev_info(adev->dev, " GRBM_STATUS_SE1=0x%08X\n",
4954 RREG32(mmGRBM_STATUS_SE1));
4955 dev_info(adev->dev, " GRBM_STATUS_SE2=0x%08X\n",
4956 RREG32(mmGRBM_STATUS_SE2));
4957 dev_info(adev->dev, " GRBM_STATUS_SE3=0x%08X\n",
4958 RREG32(mmGRBM_STATUS_SE3));
4959 dev_info(adev->dev, " CP_STAT = 0x%08x\n", RREG32(mmCP_STAT));
4960 dev_info(adev->dev, " CP_STALLED_STAT1 = 0x%08x\n",
4961 RREG32(mmCP_STALLED_STAT1));
4962 dev_info(adev->dev, " CP_STALLED_STAT2 = 0x%08x\n",
4963 RREG32(mmCP_STALLED_STAT2));
4964 dev_info(adev->dev, " CP_STALLED_STAT3 = 0x%08x\n",
4965 RREG32(mmCP_STALLED_STAT3));
4966 dev_info(adev->dev, " CP_CPF_BUSY_STAT = 0x%08x\n",
4967 RREG32(mmCP_CPF_BUSY_STAT));
4968 dev_info(adev->dev, " CP_CPF_STALLED_STAT1 = 0x%08x\n",
4969 RREG32(mmCP_CPF_STALLED_STAT1));
4970 dev_info(adev->dev, " CP_CPF_STATUS = 0x%08x\n", RREG32(mmCP_CPF_STATUS));
4971 dev_info(adev->dev, " CP_CPC_BUSY_STAT = 0x%08x\n", RREG32(mmCP_CPC_BUSY_STAT));
4972 dev_info(adev->dev, " CP_CPC_STALLED_STAT1 = 0x%08x\n",
4973 RREG32(mmCP_CPC_STALLED_STAT1));
4974 dev_info(adev->dev, " CP_CPC_STATUS = 0x%08x\n", RREG32(mmCP_CPC_STATUS));
4975
4976 for (i = 0; i < 32; i++) {
4977 dev_info(adev->dev, " GB_TILE_MODE%d=0x%08X\n",
4978 i, RREG32(mmGB_TILE_MODE0 + (i * 4)));
4979 }
4980 for (i = 0; i < 16; i++) {
4981 dev_info(adev->dev, " GB_MACROTILE_MODE%d=0x%08X\n",
4982 i, RREG32(mmGB_MACROTILE_MODE0 + (i * 4)));
4983 }
4984 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
4985 dev_info(adev->dev, " se: %d\n", i);
4986 gfx_v7_0_select_se_sh(adev, i, 0xffffffff);
4987 dev_info(adev->dev, " PA_SC_RASTER_CONFIG=0x%08X\n",
4988 RREG32(mmPA_SC_RASTER_CONFIG));
4989 dev_info(adev->dev, " PA_SC_RASTER_CONFIG_1=0x%08X\n",
4990 RREG32(mmPA_SC_RASTER_CONFIG_1));
4991 }
4992 gfx_v7_0_select_se_sh(adev, 0xffffffff, 0xffffffff);
4993
4994 dev_info(adev->dev, " GB_ADDR_CONFIG=0x%08X\n",
4995 RREG32(mmGB_ADDR_CONFIG));
4996 dev_info(adev->dev, " HDP_ADDR_CONFIG=0x%08X\n",
4997 RREG32(mmHDP_ADDR_CONFIG));
4998 dev_info(adev->dev, " DMIF_ADDR_CALC=0x%08X\n",
4999 RREG32(mmDMIF_ADDR_CALC));
5000 dev_info(adev->dev, " SDMA0_TILING_CONFIG=0x%08X\n",
5001 RREG32(mmSDMA0_TILING_CONFIG + SDMA0_REGISTER_OFFSET));
5002 dev_info(adev->dev, " SDMA1_TILING_CONFIG=0x%08X\n",
5003 RREG32(mmSDMA0_TILING_CONFIG + SDMA1_REGISTER_OFFSET));
5004 dev_info(adev->dev, " UVD_UDEC_ADDR_CONFIG=0x%08X\n",
5005 RREG32(mmUVD_UDEC_ADDR_CONFIG));
5006 dev_info(adev->dev, " UVD_UDEC_DB_ADDR_CONFIG=0x%08X\n",
5007 RREG32(mmUVD_UDEC_DB_ADDR_CONFIG));
5008 dev_info(adev->dev, " UVD_UDEC_DBW_ADDR_CONFIG=0x%08X\n",
5009 RREG32(mmUVD_UDEC_DBW_ADDR_CONFIG));
5010
5011 dev_info(adev->dev, " CP_MEQ_THRESHOLDS=0x%08X\n",
5012 RREG32(mmCP_MEQ_THRESHOLDS));
5013 dev_info(adev->dev, " SX_DEBUG_1=0x%08X\n",
5014 RREG32(mmSX_DEBUG_1));
5015 dev_info(adev->dev, " TA_CNTL_AUX=0x%08X\n",
5016 RREG32(mmTA_CNTL_AUX));
5017 dev_info(adev->dev, " SPI_CONFIG_CNTL=0x%08X\n",
5018 RREG32(mmSPI_CONFIG_CNTL));
5019 dev_info(adev->dev, " SQ_CONFIG=0x%08X\n",
5020 RREG32(mmSQ_CONFIG));
5021 dev_info(adev->dev, " DB_DEBUG=0x%08X\n",
5022 RREG32(mmDB_DEBUG));
5023 dev_info(adev->dev, " DB_DEBUG2=0x%08X\n",
5024 RREG32(mmDB_DEBUG2));
5025 dev_info(adev->dev, " DB_DEBUG3=0x%08X\n",
5026 RREG32(mmDB_DEBUG3));
5027 dev_info(adev->dev, " CB_HW_CONTROL=0x%08X\n",
5028 RREG32(mmCB_HW_CONTROL));
5029 dev_info(adev->dev, " SPI_CONFIG_CNTL_1=0x%08X\n",
5030 RREG32(mmSPI_CONFIG_CNTL_1));
5031 dev_info(adev->dev, " PA_SC_FIFO_SIZE=0x%08X\n",
5032 RREG32(mmPA_SC_FIFO_SIZE));
5033 dev_info(adev->dev, " VGT_NUM_INSTANCES=0x%08X\n",
5034 RREG32(mmVGT_NUM_INSTANCES));
5035 dev_info(adev->dev, " CP_PERFMON_CNTL=0x%08X\n",
5036 RREG32(mmCP_PERFMON_CNTL));
5037 dev_info(adev->dev, " PA_SC_FORCE_EOV_MAX_CNTS=0x%08X\n",
5038 RREG32(mmPA_SC_FORCE_EOV_MAX_CNTS));
5039 dev_info(adev->dev, " VGT_CACHE_INVALIDATION=0x%08X\n",
5040 RREG32(mmVGT_CACHE_INVALIDATION));
5041 dev_info(adev->dev, " VGT_GS_VERTEX_REUSE=0x%08X\n",
5042 RREG32(mmVGT_GS_VERTEX_REUSE));
5043 dev_info(adev->dev, " PA_SC_LINE_STIPPLE_STATE=0x%08X\n",
5044 RREG32(mmPA_SC_LINE_STIPPLE_STATE));
5045 dev_info(adev->dev, " PA_CL_ENHANCE=0x%08X\n",
5046 RREG32(mmPA_CL_ENHANCE));
5047 dev_info(adev->dev, " PA_SC_ENHANCE=0x%08X\n",
5048 RREG32(mmPA_SC_ENHANCE));
5049
5050 dev_info(adev->dev, " CP_ME_CNTL=0x%08X\n",
5051 RREG32(mmCP_ME_CNTL));
5052 dev_info(adev->dev, " CP_MAX_CONTEXT=0x%08X\n",
5053 RREG32(mmCP_MAX_CONTEXT));
5054 dev_info(adev->dev, " CP_ENDIAN_SWAP=0x%08X\n",
5055 RREG32(mmCP_ENDIAN_SWAP));
5056 dev_info(adev->dev, " CP_DEVICE_ID=0x%08X\n",
5057 RREG32(mmCP_DEVICE_ID));
5058
5059 dev_info(adev->dev, " CP_SEM_WAIT_TIMER=0x%08X\n",
5060 RREG32(mmCP_SEM_WAIT_TIMER));
5061 if (adev->asic_type != CHIP_HAWAII)
5062 dev_info(adev->dev, " CP_SEM_INCOMPLETE_TIMER_CNTL=0x%08X\n",
5063 RREG32(mmCP_SEM_INCOMPLETE_TIMER_CNTL));
5064
5065 dev_info(adev->dev, " CP_RB_WPTR_DELAY=0x%08X\n",
5066 RREG32(mmCP_RB_WPTR_DELAY));
5067 dev_info(adev->dev, " CP_RB_VMID=0x%08X\n",
5068 RREG32(mmCP_RB_VMID));
5069 dev_info(adev->dev, " CP_RB0_CNTL=0x%08X\n",
5070 RREG32(mmCP_RB0_CNTL));
5071 dev_info(adev->dev, " CP_RB0_WPTR=0x%08X\n",
5072 RREG32(mmCP_RB0_WPTR));
5073 dev_info(adev->dev, " CP_RB0_RPTR_ADDR=0x%08X\n",
5074 RREG32(mmCP_RB0_RPTR_ADDR));
5075 dev_info(adev->dev, " CP_RB0_RPTR_ADDR_HI=0x%08X\n",
5076 RREG32(mmCP_RB0_RPTR_ADDR_HI));
5077 dev_info(adev->dev, " CP_RB0_CNTL=0x%08X\n",
5078 RREG32(mmCP_RB0_CNTL));
5079 dev_info(adev->dev, " CP_RB0_BASE=0x%08X\n",
5080 RREG32(mmCP_RB0_BASE));
5081 dev_info(adev->dev, " CP_RB0_BASE_HI=0x%08X\n",
5082 RREG32(mmCP_RB0_BASE_HI));
5083 dev_info(adev->dev, " CP_MEC_CNTL=0x%08X\n",
5084 RREG32(mmCP_MEC_CNTL));
5085 dev_info(adev->dev, " CP_CPF_DEBUG=0x%08X\n",
5086 RREG32(mmCP_CPF_DEBUG));
5087
5088 dev_info(adev->dev, " SCRATCH_ADDR=0x%08X\n",
5089 RREG32(mmSCRATCH_ADDR));
5090 dev_info(adev->dev, " SCRATCH_UMSK=0x%08X\n",
5091 RREG32(mmSCRATCH_UMSK));
5092
5093 /* init the pipes */
5094 mutex_lock(&adev->srbm_mutex);
5095 for (i = 0; i < (adev->gfx.mec.num_pipe * adev->gfx.mec.num_mec); i++) {
5096 int me = (i < 4) ? 1 : 2;
5097 int pipe = (i < 4) ? i : (i - 4);
5098 int queue;
5099
5100 dev_info(adev->dev, " me: %d, pipe: %d\n", me, pipe);
5101 cik_srbm_select(adev, me, pipe, 0, 0);
5102 dev_info(adev->dev, " CP_HPD_EOP_BASE_ADDR=0x%08X\n",
5103 RREG32(mmCP_HPD_EOP_BASE_ADDR));
5104 dev_info(adev->dev, " CP_HPD_EOP_BASE_ADDR_HI=0x%08X\n",
5105 RREG32(mmCP_HPD_EOP_BASE_ADDR_HI));
5106 dev_info(adev->dev, " CP_HPD_EOP_VMID=0x%08X\n",
5107 RREG32(mmCP_HPD_EOP_VMID));
5108 dev_info(adev->dev, " CP_HPD_EOP_CONTROL=0x%08X\n",
5109 RREG32(mmCP_HPD_EOP_CONTROL));
5110
5111 for (queue = 0; queue < 8; queue++) {
5112 cik_srbm_select(adev, me, pipe, queue, 0);
5113 dev_info(adev->dev, " queue: %d\n", queue);
5114 dev_info(adev->dev, " CP_PQ_WPTR_POLL_CNTL=0x%08X\n",
5115 RREG32(mmCP_PQ_WPTR_POLL_CNTL));
5116 dev_info(adev->dev, " CP_HQD_PQ_DOORBELL_CONTROL=0x%08X\n",
5117 RREG32(mmCP_HQD_PQ_DOORBELL_CONTROL));
5118 dev_info(adev->dev, " CP_HQD_ACTIVE=0x%08X\n",
5119 RREG32(mmCP_HQD_ACTIVE));
5120 dev_info(adev->dev, " CP_HQD_DEQUEUE_REQUEST=0x%08X\n",
5121 RREG32(mmCP_HQD_DEQUEUE_REQUEST));
5122 dev_info(adev->dev, " CP_HQD_PQ_RPTR=0x%08X\n",
5123 RREG32(mmCP_HQD_PQ_RPTR));
5124 dev_info(adev->dev, " CP_HQD_PQ_WPTR=0x%08X\n",
5125 RREG32(mmCP_HQD_PQ_WPTR));
5126 dev_info(adev->dev, " CP_HQD_PQ_BASE=0x%08X\n",
5127 RREG32(mmCP_HQD_PQ_BASE));
5128 dev_info(adev->dev, " CP_HQD_PQ_BASE_HI=0x%08X\n",
5129 RREG32(mmCP_HQD_PQ_BASE_HI));
5130 dev_info(adev->dev, " CP_HQD_PQ_CONTROL=0x%08X\n",
5131 RREG32(mmCP_HQD_PQ_CONTROL));
5132 dev_info(adev->dev, " CP_HQD_PQ_WPTR_POLL_ADDR=0x%08X\n",
5133 RREG32(mmCP_HQD_PQ_WPTR_POLL_ADDR));
5134 dev_info(adev->dev, " CP_HQD_PQ_WPTR_POLL_ADDR_HI=0x%08X\n",
5135 RREG32(mmCP_HQD_PQ_WPTR_POLL_ADDR_HI));
5136 dev_info(adev->dev, " CP_HQD_PQ_RPTR_REPORT_ADDR=0x%08X\n",
5137 RREG32(mmCP_HQD_PQ_RPTR_REPORT_ADDR));
5138 dev_info(adev->dev, " CP_HQD_PQ_RPTR_REPORT_ADDR_HI=0x%08X\n",
5139 RREG32(mmCP_HQD_PQ_RPTR_REPORT_ADDR_HI));
5140 dev_info(adev->dev, " CP_HQD_PQ_DOORBELL_CONTROL=0x%08X\n",
5141 RREG32(mmCP_HQD_PQ_DOORBELL_CONTROL));
5142 dev_info(adev->dev, " CP_HQD_PQ_WPTR=0x%08X\n",
5143 RREG32(mmCP_HQD_PQ_WPTR));
5144 dev_info(adev->dev, " CP_HQD_VMID=0x%08X\n",
5145 RREG32(mmCP_HQD_VMID));
5146 dev_info(adev->dev, " CP_MQD_BASE_ADDR=0x%08X\n",
5147 RREG32(mmCP_MQD_BASE_ADDR));
5148 dev_info(adev->dev, " CP_MQD_BASE_ADDR_HI=0x%08X\n",
5149 RREG32(mmCP_MQD_BASE_ADDR_HI));
5150 dev_info(adev->dev, " CP_MQD_CONTROL=0x%08X\n",
5151 RREG32(mmCP_MQD_CONTROL));
5152 }
5153 }
5154 cik_srbm_select(adev, 0, 0, 0, 0);
5155 mutex_unlock(&adev->srbm_mutex);
5156
5157 dev_info(adev->dev, " CP_INT_CNTL_RING0=0x%08X\n",
5158 RREG32(mmCP_INT_CNTL_RING0));
5159 dev_info(adev->dev, " RLC_LB_CNTL=0x%08X\n",
5160 RREG32(mmRLC_LB_CNTL));
5161 dev_info(adev->dev, " RLC_CNTL=0x%08X\n",
5162 RREG32(mmRLC_CNTL));
5163 dev_info(adev->dev, " RLC_CGCG_CGLS_CTRL=0x%08X\n",
5164 RREG32(mmRLC_CGCG_CGLS_CTRL));
5165 dev_info(adev->dev, " RLC_LB_CNTR_INIT=0x%08X\n",
5166 RREG32(mmRLC_LB_CNTR_INIT));
5167 dev_info(adev->dev, " RLC_LB_CNTR_MAX=0x%08X\n",
5168 RREG32(mmRLC_LB_CNTR_MAX));
5169 dev_info(adev->dev, " RLC_LB_INIT_CU_MASK=0x%08X\n",
5170 RREG32(mmRLC_LB_INIT_CU_MASK));
5171 dev_info(adev->dev, " RLC_LB_PARAMS=0x%08X\n",
5172 RREG32(mmRLC_LB_PARAMS));
5173 dev_info(adev->dev, " RLC_LB_CNTL=0x%08X\n",
5174 RREG32(mmRLC_LB_CNTL));
5175 dev_info(adev->dev, " RLC_MC_CNTL=0x%08X\n",
5176 RREG32(mmRLC_MC_CNTL));
5177 dev_info(adev->dev, " RLC_UCODE_CNTL=0x%08X\n",
5178 RREG32(mmRLC_UCODE_CNTL));
5179
5180 if (adev->asic_type == CHIP_BONAIRE)
5181 dev_info(adev->dev, " RLC_DRIVER_CPDMA_STATUS=0x%08X\n",
5182 RREG32(mmRLC_DRIVER_CPDMA_STATUS));
5183
5184 mutex_lock(&adev->srbm_mutex);
5185 for (i = 0; i < 16; i++) {
5186 cik_srbm_select(adev, 0, 0, 0, i);
5187 dev_info(adev->dev, " VM %d:\n", i);
5188 dev_info(adev->dev, " SH_MEM_CONFIG=0x%08X\n",
5189 RREG32(mmSH_MEM_CONFIG));
5190 dev_info(adev->dev, " SH_MEM_APE1_BASE=0x%08X\n",
5191 RREG32(mmSH_MEM_APE1_BASE));
5192 dev_info(adev->dev, " SH_MEM_APE1_LIMIT=0x%08X\n",
5193 RREG32(mmSH_MEM_APE1_LIMIT));
5194 dev_info(adev->dev, " SH_MEM_BASES=0x%08X\n",
5195 RREG32(mmSH_MEM_BASES));
5196 }
5197 cik_srbm_select(adev, 0, 0, 0, 0);
5198 mutex_unlock(&adev->srbm_mutex);
5199 }
5200
5201 static int gfx_v7_0_soft_reset(void *handle)
5202 {
5203 u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
5204 u32 tmp;
5205 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
5206
5207 /* GRBM_STATUS */
5208 tmp = RREG32(mmGRBM_STATUS);
5209 if (tmp & (GRBM_STATUS__PA_BUSY_MASK | GRBM_STATUS__SC_BUSY_MASK |
5210 GRBM_STATUS__BCI_BUSY_MASK | GRBM_STATUS__SX_BUSY_MASK |
5211 GRBM_STATUS__TA_BUSY_MASK | GRBM_STATUS__VGT_BUSY_MASK |
5212 GRBM_STATUS__DB_BUSY_MASK | GRBM_STATUS__CB_BUSY_MASK |
5213 GRBM_STATUS__GDS_BUSY_MASK | GRBM_STATUS__SPI_BUSY_MASK |
5214 GRBM_STATUS__IA_BUSY_MASK | GRBM_STATUS__IA_BUSY_NO_DMA_MASK))
5215 grbm_soft_reset |= GRBM_SOFT_RESET__SOFT_RESET_CP_MASK |
5216 GRBM_SOFT_RESET__SOFT_RESET_GFX_MASK;
5217
5218 if (tmp & (GRBM_STATUS__CP_BUSY_MASK | GRBM_STATUS__CP_COHERENCY_BUSY_MASK)) {
5219 grbm_soft_reset |= GRBM_SOFT_RESET__SOFT_RESET_CP_MASK;
5220 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_GRBM_MASK;
5221 }
5222
5223 /* GRBM_STATUS2 */
5224 tmp = RREG32(mmGRBM_STATUS2);
5225 if (tmp & GRBM_STATUS2__RLC_BUSY_MASK)
5226 grbm_soft_reset |= GRBM_SOFT_RESET__SOFT_RESET_RLC_MASK;
5227
5228 /* SRBM_STATUS */
5229 tmp = RREG32(mmSRBM_STATUS);
5230 if (tmp & SRBM_STATUS__GRBM_RQ_PENDING_MASK)
5231 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_GRBM_MASK;
5232
5233 if (grbm_soft_reset || srbm_soft_reset) {
5234 gfx_v7_0_print_status((void *)adev);
5235 /* disable CG/PG */
5236 gfx_v7_0_fini_pg(adev);
5237 gfx_v7_0_update_cg(adev, false);
5238
5239 /* stop the rlc */
5240 gfx_v7_0_rlc_stop(adev);
5241
5242 /* Disable GFX parsing/prefetching */
5243 WREG32(mmCP_ME_CNTL, CP_ME_CNTL__ME_HALT_MASK | CP_ME_CNTL__PFP_HALT_MASK | CP_ME_CNTL__CE_HALT_MASK);
5244
5245 /* Disable MEC parsing/prefetching */
5246 WREG32(mmCP_MEC_CNTL, CP_MEC_CNTL__MEC_ME1_HALT_MASK | CP_MEC_CNTL__MEC_ME2_HALT_MASK);
5247
5248 if (grbm_soft_reset) {
5249 tmp = RREG32(mmGRBM_SOFT_RESET);
5250 tmp |= grbm_soft_reset;
5251 dev_info(adev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
5252 WREG32(mmGRBM_SOFT_RESET, tmp);
5253 tmp = RREG32(mmGRBM_SOFT_RESET);
5254
5255 udelay(50);
5256
5257 tmp &= ~grbm_soft_reset;
5258 WREG32(mmGRBM_SOFT_RESET, tmp);
5259 tmp = RREG32(mmGRBM_SOFT_RESET);
5260 }
5261
5262 if (srbm_soft_reset) {
5263 tmp = RREG32(mmSRBM_SOFT_RESET);
5264 tmp |= srbm_soft_reset;
5265 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
5266 WREG32(mmSRBM_SOFT_RESET, tmp);
5267 tmp = RREG32(mmSRBM_SOFT_RESET);
5268
5269 udelay(50);
5270
5271 tmp &= ~srbm_soft_reset;
5272 WREG32(mmSRBM_SOFT_RESET, tmp);
5273 tmp = RREG32(mmSRBM_SOFT_RESET);
5274 }
5275 /* Wait a little for things to settle down */
5276 udelay(50);
5277 gfx_v7_0_print_status((void *)adev);
5278 }
5279 return 0;
5280 }
5281
5282 static void gfx_v7_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev,
5283 enum amdgpu_interrupt_state state)
5284 {
5285 u32 cp_int_cntl;
5286
5287 switch (state) {
5288 case AMDGPU_IRQ_STATE_DISABLE:
5289 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
5290 cp_int_cntl &= ~CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK;
5291 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
5292 break;
5293 case AMDGPU_IRQ_STATE_ENABLE:
5294 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
5295 cp_int_cntl |= CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK;
5296 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
5297 break;
5298 default:
5299 break;
5300 }
5301 }
5302
5303 static void gfx_v7_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev,
5304 int me, int pipe,
5305 enum amdgpu_interrupt_state state)
5306 {
5307 u32 mec_int_cntl, mec_int_cntl_reg;
5308
5309 /*
5310 * amdgpu controls only pipe 0 of MEC1. That's why this function only
5311 * handles the setting of interrupts for this specific pipe. All other
5312 * pipes' interrupts are set by amdkfd.
5313 */
5314
5315 if (me == 1) {
5316 switch (pipe) {
5317 case 0:
5318 mec_int_cntl_reg = mmCP_ME1_PIPE0_INT_CNTL;
5319 break;
5320 default:
5321 DRM_DEBUG("invalid pipe %d\n", pipe);
5322 return;
5323 }
5324 } else {
5325 DRM_DEBUG("invalid me %d\n", me);
5326 return;
5327 }
5328
5329 switch (state) {
5330 case AMDGPU_IRQ_STATE_DISABLE:
5331 mec_int_cntl = RREG32(mec_int_cntl_reg);
5332 mec_int_cntl &= ~CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK;
5333 WREG32(mec_int_cntl_reg, mec_int_cntl);
5334 break;
5335 case AMDGPU_IRQ_STATE_ENABLE:
5336 mec_int_cntl = RREG32(mec_int_cntl_reg);
5337 mec_int_cntl |= CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK;
5338 WREG32(mec_int_cntl_reg, mec_int_cntl);
5339 break;
5340 default:
5341 break;
5342 }
5343 }
5344
5345 static int gfx_v7_0_set_priv_reg_fault_state(struct amdgpu_device *adev,
5346 struct amdgpu_irq_src *src,
5347 unsigned type,
5348 enum amdgpu_interrupt_state state)
5349 {
5350 u32 cp_int_cntl;
5351
5352 switch (state) {
5353 case AMDGPU_IRQ_STATE_DISABLE:
5354 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
5355 cp_int_cntl &= ~CP_INT_CNTL_RING0__PRIV_REG_INT_ENABLE_MASK;
5356 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
5357 break;
5358 case AMDGPU_IRQ_STATE_ENABLE:
5359 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
5360 cp_int_cntl |= CP_INT_CNTL_RING0__PRIV_REG_INT_ENABLE_MASK;
5361 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
5362 break;
5363 default:
5364 break;
5365 }
5366
5367 return 0;
5368 }
5369
5370 static int gfx_v7_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
5371 struct amdgpu_irq_src *src,
5372 unsigned type,
5373 enum amdgpu_interrupt_state state)
5374 {
5375 u32 cp_int_cntl;
5376
5377 switch (state) {
5378 case AMDGPU_IRQ_STATE_DISABLE:
5379 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
5380 cp_int_cntl &= ~CP_INT_CNTL_RING0__PRIV_INSTR_INT_ENABLE_MASK;
5381 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
5382 break;
5383 case AMDGPU_IRQ_STATE_ENABLE:
5384 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
5385 cp_int_cntl |= CP_INT_CNTL_RING0__PRIV_INSTR_INT_ENABLE_MASK;
5386 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
5387 break;
5388 default:
5389 break;
5390 }
5391
5392 return 0;
5393 }
5394
5395 static int gfx_v7_0_set_eop_interrupt_state(struct amdgpu_device *adev,
5396 struct amdgpu_irq_src *src,
5397 unsigned type,
5398 enum amdgpu_interrupt_state state)
5399 {
5400 switch (type) {
5401 case AMDGPU_CP_IRQ_GFX_EOP:
5402 gfx_v7_0_set_gfx_eop_interrupt_state(adev, state);
5403 break;
5404 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP:
5405 gfx_v7_0_set_compute_eop_interrupt_state(adev, 1, 0, state);
5406 break;
5407 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP:
5408 gfx_v7_0_set_compute_eop_interrupt_state(adev, 1, 1, state);
5409 break;
5410 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP:
5411 gfx_v7_0_set_compute_eop_interrupt_state(adev, 1, 2, state);
5412 break;
5413 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP:
5414 gfx_v7_0_set_compute_eop_interrupt_state(adev, 1, 3, state);
5415 break;
5416 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE0_EOP:
5417 gfx_v7_0_set_compute_eop_interrupt_state(adev, 2, 0, state);
5418 break;
5419 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE1_EOP:
5420 gfx_v7_0_set_compute_eop_interrupt_state(adev, 2, 1, state);
5421 break;
5422 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE2_EOP:
5423 gfx_v7_0_set_compute_eop_interrupt_state(adev, 2, 2, state);
5424 break;
5425 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE3_EOP:
5426 gfx_v7_0_set_compute_eop_interrupt_state(adev, 2, 3, state);
5427 break;
5428 default:
5429 break;
5430 }
5431 return 0;
5432 }
5433
5434 static int gfx_v7_0_eop_irq(struct amdgpu_device *adev,
5435 struct amdgpu_irq_src *source,
5436 struct amdgpu_iv_entry *entry)
5437 {
5438 u8 me_id, pipe_id;
5439 struct amdgpu_ring *ring;
5440 int i;
5441
5442 DRM_DEBUG("IH: CP EOP\n");
5443 me_id = (entry->ring_id & 0x0c) >> 2;
5444 pipe_id = (entry->ring_id & 0x03) >> 0;
5445 switch (me_id) {
5446 case 0:
5447 amdgpu_fence_process(&adev->gfx.gfx_ring[0]);
5448 break;
5449 case 1:
5450 case 2:
5451 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
5452 ring = &adev->gfx.compute_ring[i];
5453 if ((ring->me == me_id) & (ring->pipe == pipe_id))
5454 amdgpu_fence_process(ring);
5455 }
5456 break;
5457 }
5458 return 0;
5459 }
5460
5461 static int gfx_v7_0_priv_reg_irq(struct amdgpu_device *adev,
5462 struct amdgpu_irq_src *source,
5463 struct amdgpu_iv_entry *entry)
5464 {
5465 DRM_ERROR("Illegal register access in command stream\n");
5466 schedule_work(&adev->reset_work);
5467 return 0;
5468 }
5469
5470 static int gfx_v7_0_priv_inst_irq(struct amdgpu_device *adev,
5471 struct amdgpu_irq_src *source,
5472 struct amdgpu_iv_entry *entry)
5473 {
5474 DRM_ERROR("Illegal instruction in command stream\n");
5475 // XXX soft reset the gfx block only
5476 schedule_work(&adev->reset_work);
5477 return 0;
5478 }
5479
5480 static int gfx_v7_0_set_clockgating_state(void *handle,
5481 enum amd_clockgating_state state)
5482 {
5483 bool gate = false;
5484 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
5485
5486 if (state == AMD_CG_STATE_GATE)
5487 gate = true;
5488
5489 gfx_v7_0_enable_gui_idle_interrupt(adev, false);
5490 /* order matters! */
5491 if (gate) {
5492 gfx_v7_0_enable_mgcg(adev, true);
5493 gfx_v7_0_enable_cgcg(adev, true);
5494 } else {
5495 gfx_v7_0_enable_cgcg(adev, false);
5496 gfx_v7_0_enable_mgcg(adev, false);
5497 }
5498 gfx_v7_0_enable_gui_idle_interrupt(adev, true);
5499
5500 return 0;
5501 }
5502
5503 static int gfx_v7_0_set_powergating_state(void *handle,
5504 enum amd_powergating_state state)
5505 {
5506 bool gate = false;
5507 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
5508
5509 if (state == AMD_PG_STATE_GATE)
5510 gate = true;
5511
5512 if (adev->pg_flags & (AMDGPU_PG_SUPPORT_GFX_PG |
5513 AMDGPU_PG_SUPPORT_GFX_SMG |
5514 AMDGPU_PG_SUPPORT_GFX_DMG |
5515 AMDGPU_PG_SUPPORT_CP |
5516 AMDGPU_PG_SUPPORT_GDS |
5517 AMDGPU_PG_SUPPORT_RLC_SMU_HS)) {
5518 gfx_v7_0_update_gfx_pg(adev, gate);
5519 if (adev->pg_flags & AMDGPU_PG_SUPPORT_GFX_PG) {
5520 gfx_v7_0_enable_cp_pg(adev, gate);
5521 gfx_v7_0_enable_gds_pg(adev, gate);
5522 }
5523 }
5524
5525 return 0;
5526 }
5527
5528 const struct amd_ip_funcs gfx_v7_0_ip_funcs = {
5529 .early_init = gfx_v7_0_early_init,
5530 .late_init = NULL,
5531 .sw_init = gfx_v7_0_sw_init,
5532 .sw_fini = gfx_v7_0_sw_fini,
5533 .hw_init = gfx_v7_0_hw_init,
5534 .hw_fini = gfx_v7_0_hw_fini,
5535 .suspend = gfx_v7_0_suspend,
5536 .resume = gfx_v7_0_resume,
5537 .is_idle = gfx_v7_0_is_idle,
5538 .wait_for_idle = gfx_v7_0_wait_for_idle,
5539 .soft_reset = gfx_v7_0_soft_reset,
5540 .print_status = gfx_v7_0_print_status,
5541 .set_clockgating_state = gfx_v7_0_set_clockgating_state,
5542 .set_powergating_state = gfx_v7_0_set_powergating_state,
5543 };
5544
5545 static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_gfx = {
5546 .get_rptr = gfx_v7_0_ring_get_rptr_gfx,
5547 .get_wptr = gfx_v7_0_ring_get_wptr_gfx,
5548 .set_wptr = gfx_v7_0_ring_set_wptr_gfx,
5549 .parse_cs = NULL,
5550 .emit_ib = gfx_v7_0_ring_emit_ib_gfx,
5551 .emit_fence = gfx_v7_0_ring_emit_fence_gfx,
5552 .emit_semaphore = gfx_v7_0_ring_emit_semaphore,
5553 .emit_vm_flush = gfx_v7_0_ring_emit_vm_flush,
5554 .emit_gds_switch = gfx_v7_0_ring_emit_gds_switch,
5555 .emit_hdp_flush = gfx_v7_0_ring_emit_hdp_flush,
5556 .test_ring = gfx_v7_0_ring_test_ring,
5557 .test_ib = gfx_v7_0_ring_test_ib,
5558 .insert_nop = amdgpu_ring_insert_nop,
5559 };
5560
5561 static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_compute = {
5562 .get_rptr = gfx_v7_0_ring_get_rptr_compute,
5563 .get_wptr = gfx_v7_0_ring_get_wptr_compute,
5564 .set_wptr = gfx_v7_0_ring_set_wptr_compute,
5565 .parse_cs = NULL,
5566 .emit_ib = gfx_v7_0_ring_emit_ib_compute,
5567 .emit_fence = gfx_v7_0_ring_emit_fence_compute,
5568 .emit_semaphore = gfx_v7_0_ring_emit_semaphore,
5569 .emit_vm_flush = gfx_v7_0_ring_emit_vm_flush,
5570 .emit_gds_switch = gfx_v7_0_ring_emit_gds_switch,
5571 .emit_hdp_flush = gfx_v7_0_ring_emit_hdp_flush,
5572 .test_ring = gfx_v7_0_ring_test_ring,
5573 .test_ib = gfx_v7_0_ring_test_ib,
5574 .insert_nop = amdgpu_ring_insert_nop,
5575 };
5576
5577 static void gfx_v7_0_set_ring_funcs(struct amdgpu_device *adev)
5578 {
5579 int i;
5580
5581 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
5582 adev->gfx.gfx_ring[i].funcs = &gfx_v7_0_ring_funcs_gfx;
5583 for (i = 0; i < adev->gfx.num_compute_rings; i++)
5584 adev->gfx.compute_ring[i].funcs = &gfx_v7_0_ring_funcs_compute;
5585 }
5586
5587 static const struct amdgpu_irq_src_funcs gfx_v7_0_eop_irq_funcs = {
5588 .set = gfx_v7_0_set_eop_interrupt_state,
5589 .process = gfx_v7_0_eop_irq,
5590 };
5591
5592 static const struct amdgpu_irq_src_funcs gfx_v7_0_priv_reg_irq_funcs = {
5593 .set = gfx_v7_0_set_priv_reg_fault_state,
5594 .process = gfx_v7_0_priv_reg_irq,
5595 };
5596
5597 static const struct amdgpu_irq_src_funcs gfx_v7_0_priv_inst_irq_funcs = {
5598 .set = gfx_v7_0_set_priv_inst_fault_state,
5599 .process = gfx_v7_0_priv_inst_irq,
5600 };
5601
5602 static void gfx_v7_0_set_irq_funcs(struct amdgpu_device *adev)
5603 {
5604 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST;
5605 adev->gfx.eop_irq.funcs = &gfx_v7_0_eop_irq_funcs;
5606
5607 adev->gfx.priv_reg_irq.num_types = 1;
5608 adev->gfx.priv_reg_irq.funcs = &gfx_v7_0_priv_reg_irq_funcs;
5609
5610 adev->gfx.priv_inst_irq.num_types = 1;
5611 adev->gfx.priv_inst_irq.funcs = &gfx_v7_0_priv_inst_irq_funcs;
5612 }
5613
5614 static void gfx_v7_0_set_gds_init(struct amdgpu_device *adev)
5615 {
5616 /* init asci gds info */
5617 adev->gds.mem.total_size = RREG32(mmGDS_VMID0_SIZE);
5618 adev->gds.gws.total_size = 64;
5619 adev->gds.oa.total_size = 16;
5620
5621 if (adev->gds.mem.total_size == 64 * 1024) {
5622 adev->gds.mem.gfx_partition_size = 4096;
5623 adev->gds.mem.cs_partition_size = 4096;
5624
5625 adev->gds.gws.gfx_partition_size = 4;
5626 adev->gds.gws.cs_partition_size = 4;
5627
5628 adev->gds.oa.gfx_partition_size = 4;
5629 adev->gds.oa.cs_partition_size = 1;
5630 } else {
5631 adev->gds.mem.gfx_partition_size = 1024;
5632 adev->gds.mem.cs_partition_size = 1024;
5633
5634 adev->gds.gws.gfx_partition_size = 16;
5635 adev->gds.gws.cs_partition_size = 16;
5636
5637 adev->gds.oa.gfx_partition_size = 4;
5638 adev->gds.oa.cs_partition_size = 4;
5639 }
5640 }
5641
5642
5643 int gfx_v7_0_get_cu_info(struct amdgpu_device *adev,
5644 struct amdgpu_cu_info *cu_info)
5645 {
5646 int i, j, k, counter, active_cu_number = 0;
5647 u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
5648
5649 if (!adev || !cu_info)
5650 return -EINVAL;
5651
5652 mutex_lock(&adev->grbm_idx_mutex);
5653 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
5654 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
5655 mask = 1;
5656 ao_bitmap = 0;
5657 counter = 0;
5658 bitmap = gfx_v7_0_get_cu_active_bitmap(adev, i, j);
5659 cu_info->bitmap[i][j] = bitmap;
5660
5661 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k ++) {
5662 if (bitmap & mask) {
5663 if (counter < 2)
5664 ao_bitmap |= mask;
5665 counter ++;
5666 }
5667 mask <<= 1;
5668 }
5669 active_cu_number += counter;
5670 ao_cu_mask |= (ao_bitmap << (i * 16 + j * 8));
5671 }
5672 }
5673
5674 cu_info->number = active_cu_number;
5675 cu_info->ao_cu_mask = ao_cu_mask;
5676 mutex_unlock(&adev->grbm_idx_mutex);
5677 return 0;
5678 }
Linux with added FPC-III support