]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - drivers/gpu/drm/amd/amdgpu/gfx_v6_0.c
projects / mirror_ubuntu-jammy-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
drm/amdgpu: remove unused header si_reg.h
[mirror_ubuntu-jammy-kernel.git] / drivers / gpu / drm / amd / amdgpu / gfx_v6_0.c
CommitLineData
2cd46ad2
KW		 1/*
2 * Copyright 2015 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 "amdgpu.h"
25#include "amdgpu_ih.h"
26#include "amdgpu_gfx.h"
27#include "amdgpu_ucode.h"
28#include "si/clearstate_si.h"
25069e06
TSD		 29#include "bif/bif_3_0_d.h"
30#include "bif/bif_3_0_sh_mask.h"
31#include "oss/oss_1_0_d.h"
32#include "oss/oss_1_0_sh_mask.h"
33#include "gca/gfx_6_0_d.h"
34#include "gca/gfx_6_0_sh_mask.h"
35#include "gmc/gmc_6_0_d.h"
36#include "gmc/gmc_6_0_sh_mask.h"
37#include "dce/dce_6_0_d.h"
38#include "dce/dce_6_0_sh_mask.h"
39#include "gca/gfx_7_2_enum.h"
40#include "si_enums.h"
2cd46ad2
KW		 41
42static void gfx_v6_0_set_ring_funcs(struct amdgpu_device *adev);
43static void gfx_v6_0_set_irq_funcs(struct amdgpu_device *adev);
44static void gfx_v6_0_get_cu_info(struct amdgpu_device *adev);
45
46MODULE_FIRMWARE("radeon/tahiti_pfp.bin");
47MODULE_FIRMWARE("radeon/tahiti_me.bin");
48MODULE_FIRMWARE("radeon/tahiti_ce.bin");
49MODULE_FIRMWARE("radeon/tahiti_rlc.bin");
50
51MODULE_FIRMWARE("radeon/pitcairn_pfp.bin");
52MODULE_FIRMWARE("radeon/pitcairn_me.bin");
53MODULE_FIRMWARE("radeon/pitcairn_ce.bin");
54MODULE_FIRMWARE("radeon/pitcairn_rlc.bin");
55
56MODULE_FIRMWARE("radeon/verde_pfp.bin");
57MODULE_FIRMWARE("radeon/verde_me.bin");
58MODULE_FIRMWARE("radeon/verde_ce.bin");
59MODULE_FIRMWARE("radeon/verde_rlc.bin");
60
61MODULE_FIRMWARE("radeon/oland_pfp.bin");
62MODULE_FIRMWARE("radeon/oland_me.bin");
63MODULE_FIRMWARE("radeon/oland_ce.bin");
64MODULE_FIRMWARE("radeon/oland_rlc.bin");
65
66MODULE_FIRMWARE("radeon/hainan_pfp.bin");
67MODULE_FIRMWARE("radeon/hainan_me.bin");
68MODULE_FIRMWARE("radeon/hainan_ce.bin");
69MODULE_FIRMWARE("radeon/hainan_rlc.bin");
70
71static u32 gfx_v6_0_get_csb_size(struct amdgpu_device *adev);
72static void gfx_v6_0_get_csb_buffer(struct amdgpu_device *adev, volatile u32 *buffer);
73//static void gfx_v6_0_init_cp_pg_table(struct amdgpu_device *adev);
74static void gfx_v6_0_init_pg(struct amdgpu_device *adev);
75
25069e06
TSD		 76#define ARRAY_MODE(x) ((x) << GB_TILE_MODE0__ARRAY_MODE__SHIFT)
77#define PIPE_CONFIG(x) ((x) << GB_TILE_MODE0__PIPE_CONFIG__SHIFT)
78#define TILE_SPLIT(x) ((x) << GB_TILE_MODE0__TILE_SPLIT__SHIFT)
79#define MICRO_TILE_MODE(x) ((x) << 0)
80#define SAMPLE_SPLIT(x) ((x) << GB_TILE_MODE0__SAMPLE_SPLIT__SHIFT)
81#define BANK_WIDTH(x) ((x) << 14)
82#define BANK_HEIGHT(x) ((x) << 16)
83#define MACRO_TILE_ASPECT(x) ((x) << 18)
84#define NUM_BANKS(x) ((x) << 20)
2cd46ad2
KW		 85
86static const u32 verde_rlc_save_restore_register_list[] =
87{
88 (0x8000 << 16) | (0x98f4 >> 2),
89 0x00000000,
90 (0x8040 << 16) | (0x98f4 >> 2),
91 0x00000000,
92 (0x8000 << 16) | (0xe80 >> 2),
93 0x00000000,
94 (0x8040 << 16) | (0xe80 >> 2),
95 0x00000000,
96 (0x8000 << 16) | (0x89bc >> 2),
97 0x00000000,
98 (0x8040 << 16) | (0x89bc >> 2),
99 0x00000000,
100 (0x8000 << 16) | (0x8c1c >> 2),
101 0x00000000,
102 (0x8040 << 16) | (0x8c1c >> 2),
103 0x00000000,
104 (0x9c00 << 16) | (0x98f0 >> 2),
105 0x00000000,
106 (0x9c00 << 16) | (0xe7c >> 2),
107 0x00000000,
108 (0x8000 << 16) | (0x9148 >> 2),
109 0x00000000,
110 (0x8040 << 16) | (0x9148 >> 2),
111 0x00000000,
112 (0x9c00 << 16) | (0x9150 >> 2),
113 0x00000000,
114 (0x9c00 << 16) | (0x897c >> 2),
115 0x00000000,
116 (0x9c00 << 16) | (0x8d8c >> 2),
117 0x00000000,
118 (0x9c00 << 16) | (0xac54 >> 2),
119 0X00000000,
120 0x3,
121 (0x9c00 << 16) | (0x98f8 >> 2),
122 0x00000000,
123 (0x9c00 << 16) | (0x9910 >> 2),
124 0x00000000,
125 (0x9c00 << 16) | (0x9914 >> 2),
126 0x00000000,
127 (0x9c00 << 16) | (0x9918 >> 2),
128 0x00000000,
129 (0x9c00 << 16) | (0x991c >> 2),
130 0x00000000,
131 (0x9c00 << 16) | (0x9920 >> 2),
132 0x00000000,
133 (0x9c00 << 16) | (0x9924 >> 2),
134 0x00000000,
135 (0x9c00 << 16) | (0x9928 >> 2),
136 0x00000000,
137 (0x9c00 << 16) | (0x992c >> 2),
138 0x00000000,
139 (0x9c00 << 16) | (0x9930 >> 2),
140 0x00000000,
141 (0x9c00 << 16) | (0x9934 >> 2),
142 0x00000000,
143 (0x9c00 << 16) | (0x9938 >> 2),
144 0x00000000,
145 (0x9c00 << 16) | (0x993c >> 2),
146 0x00000000,
147 (0x9c00 << 16) | (0x9940 >> 2),
148 0x00000000,
149 (0x9c00 << 16) | (0x9944 >> 2),
150 0x00000000,
151 (0x9c00 << 16) | (0x9948 >> 2),
152 0x00000000,
153 (0x9c00 << 16) | (0x994c >> 2),
154 0x00000000,
155 (0x9c00 << 16) | (0x9950 >> 2),
156 0x00000000,
157 (0x9c00 << 16) | (0x9954 >> 2),
158 0x00000000,
159 (0x9c00 << 16) | (0x9958 >> 2),
160 0x00000000,
161 (0x9c00 << 16) | (0x995c >> 2),
162 0x00000000,
163 (0x9c00 << 16) | (0x9960 >> 2),
164 0x00000000,
165 (0x9c00 << 16) | (0x9964 >> 2),
166 0x00000000,
167 (0x9c00 << 16) | (0x9968 >> 2),
168 0x00000000,
169 (0x9c00 << 16) | (0x996c >> 2),
170 0x00000000,
171 (0x9c00 << 16) | (0x9970 >> 2),
172 0x00000000,
173 (0x9c00 << 16) | (0x9974 >> 2),
174 0x00000000,
175 (0x9c00 << 16) | (0x9978 >> 2),
176 0x00000000,
177 (0x9c00 << 16) | (0x997c >> 2),
178 0x00000000,
179 (0x9c00 << 16) | (0x9980 >> 2),
180 0x00000000,
181 (0x9c00 << 16) | (0x9984 >> 2),
182 0x00000000,
183 (0x9c00 << 16) | (0x9988 >> 2),
184 0x00000000,
185 (0x9c00 << 16) | (0x998c >> 2),
186 0x00000000,
187 (0x9c00 << 16) | (0x8c00 >> 2),
188 0x00000000,
189 (0x9c00 << 16) | (0x8c14 >> 2),
190 0x00000000,
191 (0x9c00 << 16) | (0x8c04 >> 2),
192 0x00000000,
193 (0x9c00 << 16) | (0x8c08 >> 2),
194 0x00000000,
195 (0x8000 << 16) | (0x9b7c >> 2),
196 0x00000000,
197 (0x8040 << 16) | (0x9b7c >> 2),
198 0x00000000,
199 (0x8000 << 16) | (0xe84 >> 2),
200 0x00000000,
201 (0x8040 << 16) | (0xe84 >> 2),
202 0x00000000,
203 (0x8000 << 16) | (0x89c0 >> 2),
204 0x00000000,
205 (0x8040 << 16) | (0x89c0 >> 2),
206 0x00000000,
207 (0x8000 << 16) | (0x914c >> 2),
208 0x00000000,
209 (0x8040 << 16) | (0x914c >> 2),
210 0x00000000,
211 (0x8000 << 16) | (0x8c20 >> 2),
212 0x00000000,
213 (0x8040 << 16) | (0x8c20 >> 2),
214 0x00000000,
215 (0x8000 << 16) | (0x9354 >> 2),
216 0x00000000,
217 (0x8040 << 16) | (0x9354 >> 2),
218 0x00000000,
219 (0x9c00 << 16) | (0x9060 >> 2),
220 0x00000000,
221 (0x9c00 << 16) | (0x9364 >> 2),
222 0x00000000,
223 (0x9c00 << 16) | (0x9100 >> 2),
224 0x00000000,
225 (0x9c00 << 16) | (0x913c >> 2),
226 0x00000000,
227 (0x8000 << 16) | (0x90e0 >> 2),
228 0x00000000,
229 (0x8000 << 16) | (0x90e4 >> 2),
230 0x00000000,
231 (0x8000 << 16) | (0x90e8 >> 2),
232 0x00000000,
233 (0x8040 << 16) | (0x90e0 >> 2),
234 0x00000000,
235 (0x8040 << 16) | (0x90e4 >> 2),
236 0x00000000,
237 (0x8040 << 16) | (0x90e8 >> 2),
238 0x00000000,
239 (0x9c00 << 16) | (0x8bcc >> 2),
240 0x00000000,
241 (0x9c00 << 16) | (0x8b24 >> 2),
242 0x00000000,
243 (0x9c00 << 16) | (0x88c4 >> 2),
244 0x00000000,
245 (0x9c00 << 16) | (0x8e50 >> 2),
246 0x00000000,
247 (0x9c00 << 16) | (0x8c0c >> 2),
248 0x00000000,
249 (0x9c00 << 16) | (0x8e58 >> 2),
250 0x00000000,
251 (0x9c00 << 16) | (0x8e5c >> 2),
252 0x00000000,
253 (0x9c00 << 16) | (0x9508 >> 2),
254 0x00000000,
255 (0x9c00 << 16) | (0x950c >> 2),
256 0x00000000,
257 (0x9c00 << 16) | (0x9494 >> 2),
258 0x00000000,
259 (0x9c00 << 16) | (0xac0c >> 2),
260 0x00000000,
261 (0x9c00 << 16) | (0xac10 >> 2),
262 0x00000000,
263 (0x9c00 << 16) | (0xac14 >> 2),
264 0x00000000,
265 (0x9c00 << 16) | (0xae00 >> 2),
266 0x00000000,
267 (0x9c00 << 16) | (0xac08 >> 2),
268 0x00000000,
269 (0x9c00 << 16) | (0x88d4 >> 2),
270 0x00000000,
271 (0x9c00 << 16) | (0x88c8 >> 2),
272 0x00000000,
273 (0x9c00 << 16) | (0x88cc >> 2),
274 0x00000000,
275 (0x9c00 << 16) | (0x89b0 >> 2),
276 0x00000000,
277 (0x9c00 << 16) | (0x8b10 >> 2),
278 0x00000000,
279 (0x9c00 << 16) | (0x8a14 >> 2),
280 0x00000000,
281 (0x9c00 << 16) | (0x9830 >> 2),
282 0x00000000,
283 (0x9c00 << 16) | (0x9834 >> 2),
284 0x00000000,
285 (0x9c00 << 16) | (0x9838 >> 2),
286 0x00000000,
287 (0x9c00 << 16) | (0x9a10 >> 2),
288 0x00000000,
289 (0x8000 << 16) | (0x9870 >> 2),
290 0x00000000,
291 (0x8000 << 16) | (0x9874 >> 2),
292 0x00000000,
293 (0x8001 << 16) | (0x9870 >> 2),
294 0x00000000,
295 (0x8001 << 16) | (0x9874 >> 2),
296 0x00000000,
297 (0x8040 << 16) | (0x9870 >> 2),
298 0x00000000,
299 (0x8040 << 16) | (0x9874 >> 2),
300 0x00000000,
301 (0x8041 << 16) | (0x9870 >> 2),
302 0x00000000,
303 (0x8041 << 16) | (0x9874 >> 2),
304 0x00000000,
305 0x00000000
306};
307
308static int gfx_v6_0_init_microcode(struct amdgpu_device *adev)
309{
310 const char *chip_name;
311 char fw_name[30];
312 int err;
313 const struct gfx_firmware_header_v1_0 *cp_hdr;
314 const struct rlc_firmware_header_v1_0 *rlc_hdr;
315
316 DRM_DEBUG("\n");
317
318 switch (adev->asic_type) {
319 case CHIP_TAHITI:
320 chip_name = "tahiti";
321 break;
322 case CHIP_PITCAIRN:
323 chip_name = "pitcairn";
324 break;
325 case CHIP_VERDE:
326 chip_name = "verde";
327 break;
328 case CHIP_OLAND:
329 chip_name = "oland";
330 break;
331 case CHIP_HAINAN:
332 chip_name = "hainan";
333 break;
334 default: BUG();
335 }
336
337 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
338 err = request_firmware(&adev->gfx.pfp_fw, fw_name, adev->dev);
339 if (err)
340 goto out;
341 err = amdgpu_ucode_validate(adev->gfx.pfp_fw);
342 if (err)
343 goto out;
344 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
345 adev->gfx.pfp_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
346 adev->gfx.pfp_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
347
348 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
349 err = request_firmware(&adev->gfx.me_fw, fw_name, adev->dev);
350 if (err)
351 goto out;
352 err = amdgpu_ucode_validate(adev->gfx.me_fw);
353 if (err)
354 goto out;
355 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
356 adev->gfx.me_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
357 adev->gfx.me_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
358
359 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
360 err = request_firmware(&adev->gfx.ce_fw, fw_name, adev->dev);
361 if (err)
362 goto out;
363 err = amdgpu_ucode_validate(adev->gfx.ce_fw);
364 if (err)
365 goto out;
366 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
367 adev->gfx.ce_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
368 adev->gfx.ce_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
369
370 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", chip_name);
371 err = request_firmware(&adev->gfx.rlc_fw, fw_name, adev->dev);
372 if (err)
373 goto out;
374 err = amdgpu_ucode_validate(adev->gfx.rlc_fw);
375 rlc_hdr = (const struct rlc_firmware_header_v1_0 *)adev->gfx.rlc_fw->data;
376 adev->gfx.rlc_fw_version = le32_to_cpu(rlc_hdr->header.ucode_version);
377 adev->gfx.rlc_feature_version = le32_to_cpu(rlc_hdr->ucode_feature_version);
378
379out:
380 if (err) {
381 printk(KERN_ERR
382 "gfx6: Failed to load firmware \"%s\"\n",
383 fw_name);
384 release_firmware(adev->gfx.pfp_fw);
385 adev->gfx.pfp_fw = NULL;
386 release_firmware(adev->gfx.me_fw);
387 adev->gfx.me_fw = NULL;
388 release_firmware(adev->gfx.ce_fw);
389 adev->gfx.ce_fw = NULL;
390 release_firmware(adev->gfx.rlc_fw);
391 adev->gfx.rlc_fw = NULL;
392 }
393 return err;
394}
395
396static void gfx_v6_0_tiling_mode_table_init(struct amdgpu_device *adev)
397{
398 const u32 num_tile_mode_states = 32;
399 u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
400
401 switch (adev->gfx.config.mem_row_size_in_kb) {
402 case 1:
403 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
404 break;
405 case 2:
406 default:
407 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
408 break;
409 case 4:
410 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
411 break;
412 }
413
3548f9a8
FC		 414 if (adev->asic_type == CHIP_VERDE) {
415 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
416 switch (reg_offset) {
417 case 0:
418 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
419 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
420 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
421 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
422 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
423 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
424 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
425 NUM_BANKS(ADDR_SURF_16_BANK));
426 break;
427 case 1:
428 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
429 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
430 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
431 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
432 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
433 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
434 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
435 NUM_BANKS(ADDR_SURF_16_BANK));
436 break;
437 case 2:
438 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
439 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
440 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
441 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
442 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
443 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
444 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
445 NUM_BANKS(ADDR_SURF_16_BANK));
446 break;
447 case 3:
448 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
449 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
450 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
451 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
452 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
453 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
454 NUM_BANKS(ADDR_SURF_8_BANK) |
455 TILE_SPLIT(split_equal_to_row_size));
456 break;
457 case 4:
458 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
459 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
460 PIPE_CONFIG(ADDR_SURF_P4_8x16));
461 break;
462 case 5:
463 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
464 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
465 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
466 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
467 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
468 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
469 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
470 NUM_BANKS(ADDR_SURF_4_BANK));
471 break;
472 case 6:
473 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
474 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
475 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
476 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
477 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
478 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
479 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
480 NUM_BANKS(ADDR_SURF_4_BANK));
481 break;
482 case 7:
483 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
484 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
485 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
486 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
487 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
488 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
489 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
490 NUM_BANKS(ADDR_SURF_2_BANK));
491 break;
492 case 8:
493 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED));
494 break;
495 case 9:
496 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
497 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
498 PIPE_CONFIG(ADDR_SURF_P4_8x16));
499 break;
500 case 10:
501 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
502 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
503 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
504 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
505 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
506 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
507 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
508 NUM_BANKS(ADDR_SURF_16_BANK));
509 break;
510 case 11:
511 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
512 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
513 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
514 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
515 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
516 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
517 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
518 NUM_BANKS(ADDR_SURF_16_BANK));
519 break;
520 case 12:
521 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
522 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
523 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
524 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
525 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
526 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
527 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
528 NUM_BANKS(ADDR_SURF_16_BANK));
529 break;
530 case 13:
531 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
532 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
533 PIPE_CONFIG(ADDR_SURF_P4_8x16));
534 break;
535 case 14:
536 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
537 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
538 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
539 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
540 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
541 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
542 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
543 NUM_BANKS(ADDR_SURF_16_BANK));
544 break;
545 case 15:
546 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
547 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
548 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
549 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
550 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
551 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
552 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
553 NUM_BANKS(ADDR_SURF_16_BANK));
554 break;
555 case 16:
556 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
557 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
558 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
559 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
560 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
561 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
562 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
563 NUM_BANKS(ADDR_SURF_16_BANK));
564 break;
565 case 17:
566 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
567 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
568 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
569 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
570 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
571 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
572 NUM_BANKS(ADDR_SURF_16_BANK) |
573 TILE_SPLIT(split_equal_to_row_size));
574 break;
575 case 18:
576 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
577 ARRAY_MODE(ARRAY_1D_TILED_THICK) |
578 PIPE_CONFIG(ADDR_SURF_P4_8x16));
579 break;
580 case 19:
581 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
582 ARRAY_MODE(ARRAY_2D_TILED_XTHICK) |
583 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
584 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
585 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
586 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
587 NUM_BANKS(ADDR_SURF_16_BANK) |
588 TILE_SPLIT(split_equal_to_row_size));
589 break;
590 case 20:
591 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
592 ARRAY_MODE(ARRAY_2D_TILED_THICK) |
593 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
594 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
595 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
596 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
597 NUM_BANKS(ADDR_SURF_16_BANK) |
598 TILE_SPLIT(split_equal_to_row_size));
599 break;
600 case 21:
601 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
602 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
603 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
604 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
605 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
606 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
607 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
608 NUM_BANKS(ADDR_SURF_8_BANK));
609 break;
610 case 22:
611 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
612 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
613 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
614 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
615 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
616 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
617 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
618 NUM_BANKS(ADDR_SURF_8_BANK));
619 break;
620 case 23:
621 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
622 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
623 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
624 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
625 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
626 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
627 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
628 NUM_BANKS(ADDR_SURF_4_BANK));
629 break;
630 case 24:
631 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
632 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
633 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
634 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
635 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
636 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
637 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
638 NUM_BANKS(ADDR_SURF_4_BANK));
639 break;
640 case 25:
641 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
642 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
643 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
644 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
645 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
646 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
647 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
648 NUM_BANKS(ADDR_SURF_2_BANK));
649 break;
650 case 26:
651 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
652 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
653 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
654 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
655 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
656 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
657 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
658 NUM_BANKS(ADDR_SURF_2_BANK));
659 break;
660 case 27:
661 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
662 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
663 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
664 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
665 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
666 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
667 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
668 NUM_BANKS(ADDR_SURF_2_BANK));
669 break;
670 case 28:
671 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
672 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
673 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
674 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
675 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
676 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
677 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
678 NUM_BANKS(ADDR_SURF_2_BANK));
679 break;
680 case 29:
681 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
682 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
683 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
684 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
685 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
686 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
687 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
688 NUM_BANKS(ADDR_SURF_2_BANK));
689 break;
690 case 30:
691 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
692 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
693 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
694 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_2KB) |
695 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
696 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
697 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
698 NUM_BANKS(ADDR_SURF_2_BANK));
699 break;
700 default:
701 continue;
702 }
703 adev->gfx.config.tile_mode_array[reg_offset] = gb_tile_moden;
704 WREG32(mmGB_TILE_MODE0 + reg_offset, gb_tile_moden);
705 }
706 } else if (adev->asic_type == CHIP_OLAND ||
25069e06 707 adev->asic_type == CHIP_HAINAN) {
2cd46ad2
KW		 708 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
709 switch (reg_offset) {
710 case 0:
f8d9422e
FC		 711 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
712 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
713 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 714 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2cd46ad2
KW		 715 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
716 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
f8d9422e
FC		 717 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
718 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 719 break;
45682886 720 case 1:
f8d9422e
FC		 721 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
722 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
723 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 724 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2cd46ad2
KW		 725 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
726 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
f8d9422e
FC		 727 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
728 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2
KW		 729 break;
730 case 2:
f8d9422e
FC		 731 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
732 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
733 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 734 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 735 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
736 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
f8d9422e
FC		 737 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
738 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 739 break;
45682886 740 case 3:
f8d9422e
FC		 741 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
742 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
743 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 744 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
f8d9422e
FC		 745 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
746 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
747 NUM_BANKS(ADDR_SURF_8_BANK) |
748 TILE_SPLIT(split_equal_to_row_size));
2cd46ad2 749 break;
45682886 750 case 4:
f8d9422e
FC		 751 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
752 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
753 PIPE_CONFIG(ADDR_SURF_P2));
2cd46ad2 754 break;
45682886 755 case 5:
f8d9422e
FC		 756 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
757 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
758 PIPE_CONFIG(ADDR_SURF_P2) |
759 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2 760 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
f8d9422e
FC		 761 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
762 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
763 NUM_BANKS(ADDR_SURF_8_BANK));
2cd46ad2 764 break;
45682886 765 case 6:
f8d9422e
FC		 766 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
767 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
768 PIPE_CONFIG(ADDR_SURF_P2) |
769 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2 770 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
f8d9422e
FC		 771 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
772 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
773 NUM_BANKS(ADDR_SURF_8_BANK));
2cd46ad2 774 break;
45682886 775 case 7:
f8d9422e
FC		 776 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
777 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
778 PIPE_CONFIG(ADDR_SURF_P2) |
779 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2cd46ad2 780 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
f8d9422e
FC		 781 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
782 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
783 NUM_BANKS(ADDR_SURF_4_BANK));
2cd46ad2 784 break;
45682886 785 case 8:
f8d9422e 786 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED));
2cd46ad2 787 break;
45682886 788 case 9:
f8d9422e
FC		 789 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
790 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
791 PIPE_CONFIG(ADDR_SURF_P2));
2cd46ad2 792 break;
45682886 793 case 10:
f8d9422e
FC		 794 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
795 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
796 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 797 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 798 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
799 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
f8d9422e
FC		 800 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
801 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 802 break;
45682886 803 case 11:
f8d9422e
FC		 804 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
805 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
806 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 807 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 808 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
809 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
f8d9422e
FC		 810 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
811 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 812 break;
45682886 813 case 12:
f8d9422e
FC		 814 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
815 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
816 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 817 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2
KW		 818 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
819 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
f8d9422e
FC		 820 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
821 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 822 break;
45682886 823 case 13:
f8d9422e
FC		 824 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
825 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
826 PIPE_CONFIG(ADDR_SURF_P2));
2cd46ad2 827 break;
45682886 828 case 14:
f8d9422e
FC		 829 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
830 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
831 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 832 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 833 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
834 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
f8d9422e
FC		 835 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
836 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 837 break;
45682886 838 case 15:
f8d9422e
FC		 839 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
840 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
841 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 842 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 843 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
844 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
f8d9422e
FC		 845 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
846 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 847 break;
45682886 848 case 16:
f8d9422e
FC		 849 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
850 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
851 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 852 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2
KW		 853 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
854 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
f8d9422e
FC		 855 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
856 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 857 break;
45682886 858 case 17:
f8d9422e
FC		 859 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
860 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
861 PIPE_CONFIG(ADDR_SURF_P2) |
862 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
863 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
864 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2cd46ad2 865 NUM_BANKS(ADDR_SURF_16_BANK) |
f8d9422e
FC		 866 TILE_SPLIT(split_equal_to_row_size));
867 break;
868 case 18:
869 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
870 ARRAY_MODE(ARRAY_1D_TILED_THICK) |
871 PIPE_CONFIG(ADDR_SURF_P2));
872 break;
873 case 19:
874 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
875 ARRAY_MODE(ARRAY_2D_TILED_XTHICK) |
876 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2
KW		 877 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
878 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
f8d9422e
FC		 879 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
880 NUM_BANKS(ADDR_SURF_16_BANK) |
881 TILE_SPLIT(split_equal_to_row_size));
882 break;
883 case 20:
884 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
885 ARRAY_MODE(ARRAY_2D_TILED_THICK) |
886 PIPE_CONFIG(ADDR_SURF_P2) |
887 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
888 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
889 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
890 NUM_BANKS(ADDR_SURF_16_BANK) |
891 TILE_SPLIT(split_equal_to_row_size));
2cd46ad2 892 break;
45682886 893 case 21:
f8d9422e
FC		 894 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
895 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
896 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 897 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 898 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
899 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
f8d9422e
FC		 900 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
901 NUM_BANKS(ADDR_SURF_8_BANK));
2cd46ad2 902 break;
45682886 903 case 22:
f8d9422e
FC		 904 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
905 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
906 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 907 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
f8d9422e
FC		 908 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
909 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
910 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
911 NUM_BANKS(ADDR_SURF_8_BANK));
2cd46ad2 912 break;
45682886 913 case 23:
f8d9422e
FC		 914 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
915 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
916 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 917 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 918 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
919 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
f8d9422e
FC		 920 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
921 NUM_BANKS(ADDR_SURF_8_BANK));
2cd46ad2 922 break;
45682886 923 case 24:
f8d9422e
FC		 924 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
925 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
926 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 927 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2
KW		 928 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
929 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
f8d9422e
FC		 930 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
931 NUM_BANKS(ADDR_SURF_8_BANK));
2cd46ad2 932 break;
45682886 933 case 25:
f8d9422e
FC		 934 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
935 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
936 PIPE_CONFIG(ADDR_SURF_P2) |
937 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
938 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
939 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
940 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
941 NUM_BANKS(ADDR_SURF_4_BANK));
942 break;
943 case 26:
944 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
945 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
946 PIPE_CONFIG(ADDR_SURF_P2) |
947 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
948 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
949 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
950 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
951 NUM_BANKS(ADDR_SURF_4_BANK));
952 break;
953 case 27:
954 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
955 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
956 PIPE_CONFIG(ADDR_SURF_P2) |
2cd46ad2 957 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
f8d9422e
FC		 958 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
959 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
960 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
961 NUM_BANKS(ADDR_SURF_4_BANK));
962 break;
963 case 28:
964 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
965 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
966 PIPE_CONFIG(ADDR_SURF_P2) |
967 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
968 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
969 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
970 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
971 NUM_BANKS(ADDR_SURF_4_BANK));
972 break;
973 case 29:
974 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
975 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
976 PIPE_CONFIG(ADDR_SURF_P2) |
977 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
978 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
979 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
980 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
981 NUM_BANKS(ADDR_SURF_4_BANK));
982 break;
983 case 30:
984 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
985 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
986 PIPE_CONFIG(ADDR_SURF_P2) |
987 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_2KB) |
2cd46ad2
KW		 988 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
989 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
f8d9422e
FC		 990 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
991 NUM_BANKS(ADDR_SURF_4_BANK));
2cd46ad2
KW		 992 break;
993 default:
f8d9422e 994 continue;
2cd46ad2
KW		 995 }
996 adev->gfx.config.tile_mode_array[reg_offset] = gb_tile_moden;
25069e06 997 WREG32(mmGB_TILE_MODE0 + reg_offset, gb_tile_moden);
2cd46ad2
KW		 998 }
999 } else if ((adev->asic_type == CHIP_TAHITI) || (adev->asic_type == CHIP_PITCAIRN)) {
1000 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
1001 switch (reg_offset) {
7c0a705e
FC		 1002 case 0:
1003 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1004 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1005 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1006 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
2cd46ad2
KW		 1007 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1008 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1009 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1010 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1011 break;
7c0a705e
FC		 1012 case 1:
1013 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1014 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1015 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1016 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
2cd46ad2
KW		 1017 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1018 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1019 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1020 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1021 break;
7c0a705e
FC		 1022 case 2:
1023 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1024 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1025 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1026 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 1027 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1028 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1029 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1030 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1031 break;
7c0a705e
FC		 1032 case 3:
1033 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1034 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2 1035 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2cd46ad2
KW		 1036 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1037 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1038 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1039 NUM_BANKS(ADDR_SURF_4_BANK) |
1040 TILE_SPLIT(split_equal_to_row_size));
2cd46ad2 1041 break;
7c0a705e
FC		 1042 case 4:
1043 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1044 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1045 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16));
2cd46ad2 1046 break;
7c0a705e
FC		 1047 case 5:
1048 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1049 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2 1050 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
7c0a705e 1051 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2 1052 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
7c0a705e
FC		 1053 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1054 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1055 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1056 break;
7c0a705e
FC		 1057 case 6:
1058 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1059 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2 1060 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
7c0a705e 1061 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2 1062 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
7c0a705e
FC		 1063 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
1064 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1065 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1066 break;
7c0a705e
FC		 1067 case 7:
1068 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1069 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1070 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1071 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2cd46ad2
KW		 1072 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1073 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1074 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1075 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1076 break;
7c0a705e
FC		 1077 case 8:
1078 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED));
2cd46ad2 1079 break;
7c0a705e
FC		 1080 case 9:
1081 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1082 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1083 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16));
2cd46ad2 1084 break;
7c0a705e
FC		 1085 case 10:
1086 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1087 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1088 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1089 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 1090 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1091 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1092 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1093 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1094 break;
7c0a705e
FC		 1095 case 11:
1096 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1097 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1098 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1099 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 1100 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1101 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
7c0a705e
FC		 1102 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
1103 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1104 break;
7c0a705e
FC		 1105 case 12:
1106 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1107 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1108 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1109 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2
KW		 1110 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1111 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
7c0a705e
FC		 1112 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1113 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1114 break;
7c0a705e
FC		 1115 case 13:
1116 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1117 ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1118 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16));
2cd46ad2 1119 break;
7c0a705e
FC		 1120 case 14:
1121 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1122 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1123 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1124 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 1125 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1126 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1127 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1128 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1129 break;
7c0a705e
FC		 1130 case 15:
1131 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1132 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1133 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1134 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 1135 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1136 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
7c0a705e
FC		 1137 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1138 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1139 break;
7c0a705e
FC		 1140 case 16:
1141 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1142 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1143 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1144 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2
KW		 1145 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1146 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
7c0a705e
FC		 1147 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1148 NUM_BANKS(ADDR_SURF_16_BANK));
2cd46ad2 1149 break;
7c0a705e
FC		 1150 case 17:
1151 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1152 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2 1153 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
7c0a705e
FC		 1154 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1155 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1156 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2cd46ad2 1157 NUM_BANKS(ADDR_SURF_16_BANK) |
7c0a705e
FC		 1158 TILE_SPLIT(split_equal_to_row_size));
1159 break;
1160 case 18:
1161 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1162 ARRAY_MODE(ARRAY_1D_TILED_THICK) |
1163 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16));
1164 break;
1165 case 19:
1166 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1167 ARRAY_MODE(ARRAY_2D_TILED_XTHICK) |
1168 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2cd46ad2
KW		 1169 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1170 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
7c0a705e
FC		 1171 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1172 NUM_BANKS(ADDR_SURF_16_BANK) |
1173 TILE_SPLIT(split_equal_to_row_size));
2cd46ad2 1174 break;
7c0a705e
FC		 1175 case 20:
1176 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1177 ARRAY_MODE(ARRAY_2D_TILED_THICK) |
2cd46ad2 1178 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
7c0a705e
FC		 1179 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1180 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1181 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2cd46ad2 1182 NUM_BANKS(ADDR_SURF_16_BANK) |
7c0a705e 1183 TILE_SPLIT(split_equal_to_row_size));
2cd46ad2 1184 break;
7c0a705e
FC		 1185 case 21:
1186 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1187 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1188 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1189 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1190 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1191 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
1192 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1193 NUM_BANKS(ADDR_SURF_4_BANK));
1194 break;
1195 case 22:
1196 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1197 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1198 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1199 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2
KW		 1200 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1201 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
7c0a705e
FC		 1202 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1203 NUM_BANKS(ADDR_SURF_4_BANK));
2cd46ad2 1204 break;
7c0a705e
FC		 1205 case 23:
1206 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1207 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1208 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1209 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
2cd46ad2 1210 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
7c0a705e
FC		 1211 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
1212 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1213 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1214 break;
7c0a705e
FC		 1215 case 24:
1216 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1217 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2cd46ad2
KW		 1218 PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1219 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
2cd46ad2 1220 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
7c0a705e
FC		 1221 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1222 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1223 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1224 break;
7c0a705e
FC		 1225 case 25:
1226 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1227 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1228 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2cd46ad2 1229 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
2cd46ad2 1230 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
7c0a705e
FC		 1231 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1232 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1233 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1234 break;
7c0a705e
FC		 1235 case 26:
1236 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1237 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1238 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1239 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
1240 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1241 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1242 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1243 NUM_BANKS(ADDR_SURF_2_BANK));
1244 break;
1245 case 27:
1246 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1247 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1248 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1249 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
1250 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1251 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1252 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1253 NUM_BANKS(ADDR_SURF_2_BANK));
2cd46ad2 1254 break;
7c0a705e
FC		 1255 case 28:
1256 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1257 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1258 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1259 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
1260 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1261 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1262 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1263 NUM_BANKS(ADDR_SURF_2_BANK));
1264 break;
1265 case 29:
1266 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1267 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1268 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1269 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
1270 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1271 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1272 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1273 NUM_BANKS(ADDR_SURF_2_BANK));
1274 break;
1275 case 30:
1276 gb_tile_moden = (MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1277 ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1278 PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1279 TILE_SPLIT(ADDR_SURF_TILE_SPLIT_2KB) |
1280 BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1281 BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1282 MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
1283 NUM_BANKS(ADDR_SURF_2_BANK));
1284 break;
1285 default:
1286 continue;
2cd46ad2
KW		 1287 }
1288 adev->gfx.config.tile_mode_array[reg_offset] = gb_tile_moden;
25069e06 1289 WREG32(mmGB_TILE_MODE0 + reg_offset, gb_tile_moden);
2cd46ad2
KW		 1290 }
1291 } else{
1292
1293 DRM_ERROR("unknown asic: 0x%x\n", adev->asic_type);
1294 }
1295
1296}
1297
1298static void gfx_v6_0_select_se_sh(struct amdgpu_device *adev, u32 se_num,
1299 u32 sh_num, u32 instance)
1300{
1301 u32 data;
1302
1303 if (instance == 0xffffffff)
25069e06 1304 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_BROADCAST_WRITES, 1);
2cd46ad2 1305 else
25069e06 1306 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_INDEX, instance);
2cd46ad2
KW		 1307
1308 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
25069e06
TSD		 1309 data |= GRBM_GFX_INDEX__SH_BROADCAST_WRITES_MASK |
1310 GRBM_GFX_INDEX__SE_BROADCAST_WRITES_MASK;
2cd46ad2 1311 else if (se_num == 0xffffffff)
25069e06
TSD		 1312 data |= GRBM_GFX_INDEX__SE_BROADCAST_WRITES_MASK |
1313 (sh_num << GRBM_GFX_INDEX__SH_INDEX__SHIFT);
2cd46ad2 1314 else if (sh_num == 0xffffffff)
25069e06
TSD		 1315 data |= GRBM_GFX_INDEX__SH_BROADCAST_WRITES_MASK |
1316 (se_num << GRBM_GFX_INDEX__SE_INDEX__SHIFT);
2cd46ad2 1317 else
25069e06
TSD		 1318 data |= (sh_num << GRBM_GFX_INDEX__SH_INDEX__SHIFT) |
1319 (se_num << GRBM_GFX_INDEX__SE_INDEX__SHIFT);
1320 WREG32(mmGRBM_GFX_INDEX, data);
2cd46ad2
KW		 1321}
1322
1323static u32 gfx_v6_0_create_bitmask(u32 bit_width)
1324{
142333db 1325 return (u32)(((u64)1 << bit_width) - 1);
2cd46ad2
KW		 1326}
1327
1328static u32 gfx_v6_0_get_rb_disabled(struct amdgpu_device *adev,
1329 u32 max_rb_num_per_se,
1330 u32 sh_per_se)
1331{
1332 u32 data, mask;
1333
25069e06
TSD		 1334 data = RREG32(mmCC_RB_BACKEND_DISABLE);
1335 data &= CC_RB_BACKEND_DISABLE__BACKEND_DISABLE_MASK;
1336 data |= RREG32(mmGC_USER_RB_BACKEND_DISABLE);
2cd46ad2 1337
25069e06 1338 data >>= CC_RB_BACKEND_DISABLE__BACKEND_DISABLE__SHIFT;
2cd46ad2
KW		 1339
1340 mask = gfx_v6_0_create_bitmask(max_rb_num_per_se / sh_per_se);
1341
1342 return data & mask;
1343}
1344
865ab832
HR		 1345static void gfx_v6_0_raster_config(struct amdgpu_device *adev, u32 *rconf)
1346{
1347 switch (adev->asic_type) {
1348 case CHIP_TAHITI:
1349 case CHIP_PITCAIRN:
25069e06
TSD		 1350 *rconf |=
1351 (2 << PA_SC_RASTER_CONFIG__RB_XSEL2__SHIFT) |
1352 (1 << PA_SC_RASTER_CONFIG__RB_XSEL__SHIFT) |
1353 (2 << PA_SC_RASTER_CONFIG__PKR_MAP__SHIFT) |
1354 (1 << PA_SC_RASTER_CONFIG__PKR_YSEL__SHIFT) |
1355 (2 << PA_SC_RASTER_CONFIG__SE_MAP__SHIFT) |
1356 (2 << PA_SC_RASTER_CONFIG__SE_XSEL__SHIFT) |
1357 (2 << PA_SC_RASTER_CONFIG__SE_YSEL__SHIFT);
865ab832
HR		 1358 break;
1359 case CHIP_VERDE:
25069e06
TSD		 1360 *rconf |=
1361 (1 << PA_SC_RASTER_CONFIG__RB_XSEL__SHIFT) |
1362 (2 << PA_SC_RASTER_CONFIG__PKR_MAP__SHIFT) |
1363 (1 << PA_SC_RASTER_CONFIG__PKR_YSEL__SHIFT);
865ab832
HR		 1364 break;
1365 case CHIP_OLAND:
25069e06 1366 *rconf |= (1 << PA_SC_RASTER_CONFIG__RB_YSEL__SHIFT);
865ab832
HR		 1367 break;
1368 case CHIP_HAINAN:
1369 *rconf |= 0x0;
1370 break;
1371 default:
1372 DRM_ERROR("unknown asic: 0x%x\n", adev->asic_type);
1373 break;
1374 }
1375}
1376
1377static void gfx_v6_0_write_harvested_raster_configs(struct amdgpu_device *adev,
1378 u32 raster_config, unsigned rb_mask,
1379 unsigned num_rb)
1380{
1381 unsigned sh_per_se = max_t(unsigned, adev->gfx.config.max_sh_per_se, 1);
1382 unsigned num_se = max_t(unsigned, adev->gfx.config.max_shader_engines, 1);
1383 unsigned rb_per_pkr = min_t(unsigned, num_rb / num_se / sh_per_se, 2);
1384 unsigned rb_per_se = num_rb / num_se;
1385 unsigned se_mask[4];
1386 unsigned se;
1387
1388 se_mask[0] = ((1 << rb_per_se) - 1) & rb_mask;
1389 se_mask[1] = (se_mask[0] << rb_per_se) & rb_mask;
1390 se_mask[2] = (se_mask[1] << rb_per_se) & rb_mask;
1391 se_mask[3] = (se_mask[2] << rb_per_se) & rb_mask;
1392
1393 WARN_ON(!(num_se == 1 || num_se == 2 || num_se == 4));
1394 WARN_ON(!(sh_per_se == 1 || sh_per_se == 2));
1395 WARN_ON(!(rb_per_pkr == 1 || rb_per_pkr == 2));
1396
1397 for (se = 0; se < num_se; se++) {
1398 unsigned raster_config_se = raster_config;
1399 unsigned pkr0_mask = ((1 << rb_per_pkr) - 1) << (se * rb_per_se);
1400 unsigned pkr1_mask = pkr0_mask << rb_per_pkr;
1401 int idx = (se / 2) * 2;
1402
1403 if ((num_se > 1) && (!se_mask[idx] || !se_mask[idx + 1])) {
25069e06 1404 raster_config_se &= ~PA_SC_RASTER_CONFIG__SE_MAP_MASK;
865ab832
HR		 1405
1406 if (!se_mask[idx]) {
25069e06 1407 raster_config_se |= RASTER_CONFIG_SE_MAP_3 << PA_SC_RASTER_CONFIG__SE_MAP__SHIFT;
865ab832 1408 } else {
25069e06 1409 raster_config_se |= RASTER_CONFIG_SE_MAP_0 << PA_SC_RASTER_CONFIG__SE_MAP__SHIFT;
865ab832
HR		 1410 }
1411 }
1412
1413 pkr0_mask &= rb_mask;
1414 pkr1_mask &= rb_mask;
1415 if (rb_per_se > 2 && (!pkr0_mask || !pkr1_mask)) {
25069e06 1416 raster_config_se &= ~PA_SC_RASTER_CONFIG__PKR_MAP_MASK;
865ab832
HR		 1417
1418 if (!pkr0_mask) {
25069e06 1419 raster_config_se |= RASTER_CONFIG_PKR_MAP_3 << PA_SC_RASTER_CONFIG__PKR_MAP__SHIFT;
865ab832 1420 } else {
25069e06 1421 raster_config_se |= RASTER_CONFIG_PKR_MAP_0 << PA_SC_RASTER_CONFIG__PKR_MAP__SHIFT;
865ab832
HR		 1422 }
1423 }
1424
1425 if (rb_per_se >= 2) {
1426 unsigned rb0_mask = 1 << (se * rb_per_se);
1427 unsigned rb1_mask = rb0_mask << 1;
1428
1429 rb0_mask &= rb_mask;
1430 rb1_mask &= rb_mask;
1431 if (!rb0_mask || !rb1_mask) {
25069e06 1432 raster_config_se &= ~PA_SC_RASTER_CONFIG__RB_MAP_PKR0_MASK;
865ab832
HR		 1433
1434 if (!rb0_mask) {
1435 raster_config_se |=
25069e06 1436 RASTER_CONFIG_RB_MAP_3 << PA_SC_RASTER_CONFIG__RB_MAP_PKR0__SHIFT;
865ab832
HR		 1437 } else {
1438 raster_config_se |=
25069e06 1439 RASTER_CONFIG_RB_MAP_0 << PA_SC_RASTER_CONFIG__RB_MAP_PKR0__SHIFT;
865ab832
HR		 1440 }
1441 }
1442
1443 if (rb_per_se > 2) {
1444 rb0_mask = 1 << (se * rb_per_se + rb_per_pkr);
1445 rb1_mask = rb0_mask << 1;
1446 rb0_mask &= rb_mask;
1447 rb1_mask &= rb_mask;
1448 if (!rb0_mask || !rb1_mask) {
25069e06 1449 raster_config_se &= ~PA_SC_RASTER_CONFIG__RB_MAP_PKR1_MASK;
865ab832
HR		 1450
1451 if (!rb0_mask) {
1452 raster_config_se |=
25069e06 1453 RASTER_CONFIG_RB_MAP_3 << PA_SC_RASTER_CONFIG__RB_MAP_PKR1__SHIFT;
865ab832
HR		 1454 } else {
1455 raster_config_se |=
25069e06 1456 RASTER_CONFIG_RB_MAP_0 << PA_SC_RASTER_CONFIG__RB_MAP_PKR1__SHIFT;
865ab832
HR		 1457 }
1458 }
1459 }
1460 }
1461
1462 /* GRBM_GFX_INDEX has a different offset on SI */
1463 gfx_v6_0_select_se_sh(adev, se, 0xffffffff, 0xffffffff);
25069e06 1464 WREG32(mmPA_SC_RASTER_CONFIG, raster_config_se);
865ab832
HR		 1465 }
1466
1467 /* GRBM_GFX_INDEX has a different offset on SI */
1468 gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1469}
1470
2cd46ad2
KW		 1471static void gfx_v6_0_setup_rb(struct amdgpu_device *adev,
1472 u32 se_num, u32 sh_per_se,
1473 u32 max_rb_num_per_se)
1474{
1475 int i, j;
1476 u32 data, mask;
1477 u32 disabled_rbs = 0;
1478 u32 enabled_rbs = 0;
865ab832 1479 unsigned num_rb_pipes;
2cd46ad2 1480
deca1d1f 1481 mutex_lock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 1482 for (i = 0; i < se_num; i++) {
1483 for (j = 0; j < sh_per_se; j++) {
1484 gfx_v6_0_select_se_sh(adev, i, j, 0xffffffff);
1485 data = gfx_v6_0_get_rb_disabled(adev, max_rb_num_per_se, sh_per_se);
25069e06 1486 disabled_rbs |= data << ((i * sh_per_se + j) * 2);
2cd46ad2
KW		 1487 }
1488 }
1489 gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
deca1d1f 1490 mutex_unlock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 1491
1492 mask = 1;
1493 for (i = 0; i < max_rb_num_per_se * se_num; i++) {
1494 if (!(disabled_rbs & mask))
1495 enabled_rbs |= mask;
1496 mask <<= 1;
1497 }
1498
1499 adev->gfx.config.backend_enable_mask = enabled_rbs;
1500 adev->gfx.config.num_rbs = hweight32(enabled_rbs);
1501
865ab832
HR		 1502 num_rb_pipes = min_t(unsigned, adev->gfx.config.max_backends_per_se *
1503 adev->gfx.config.max_shader_engines, 16);
1504
deca1d1f 1505 mutex_lock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 1506 for (i = 0; i < se_num; i++) {
1507 gfx_v6_0_select_se_sh(adev, i, 0xffffffff, 0xffffffff);
1508 data = 0;
1509 for (j = 0; j < sh_per_se; j++) {
1510 switch (enabled_rbs & 3) {
1511 case 1:
1512 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
1513 break;
1514 case 2:
1515 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
1516 break;
1517 case 3:
1518 default:
1519 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
1520 break;
1521 }
1522 enabled_rbs >>= 2;
1523 }
865ab832
HR		 1524 gfx_v6_0_raster_config(adev, &data);
1525
1526 if (!adev->gfx.config.backend_enable_mask ||
1527 adev->gfx.config.num_rbs >= num_rb_pipes)
25069e06 1528 WREG32(mmPA_SC_RASTER_CONFIG, data);
865ab832
HR		 1529 else
1530 gfx_v6_0_write_harvested_raster_configs(adev, data,
1531 adev->gfx.config.backend_enable_mask,
1532 num_rb_pipes);
2cd46ad2
KW		 1533 }
1534 gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
deca1d1f 1535 mutex_unlock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 1536}
1537/*
1538static void gmc_v6_0_init_compute_vmid(struct amdgpu_device *adev)
1539{
1540}
1541*/
1542
1543static u32 gfx_v6_0_get_cu_enabled(struct amdgpu_device *adev, u32 cu_per_sh)
1544{
1545 u32 data, mask;
1546
25069e06
TSD		 1547 data = RREG32(mmCC_GC_SHADER_ARRAY_CONFIG);
1548 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_CUS_MASK;
1549 data |= RREG32(mmGC_USER_SHADER_ARRAY_CONFIG);
2cd46ad2 1550
25069e06 1551 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_CUS__SHIFT;
2cd46ad2
KW		 1552
1553 mask = gfx_v6_0_create_bitmask(cu_per_sh);
1554
1555 return ~data & mask;
1556}
1557
1558
1559static void gfx_v6_0_setup_spi(struct amdgpu_device *adev,
1560 u32 se_num, u32 sh_per_se,
1561 u32 cu_per_sh)
1562{
1563 int i, j, k;
1564 u32 data, mask;
1565 u32 active_cu = 0;
1566
deca1d1f 1567 mutex_lock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 1568 for (i = 0; i < se_num; i++) {
1569 for (j = 0; j < sh_per_se; j++) {
1570 gfx_v6_0_select_se_sh(adev, i, j, 0xffffffff);
25069e06 1571 data = RREG32(mmSPI_STATIC_THREAD_MGMT_3);
2cd46ad2
KW		 1572 active_cu = gfx_v6_0_get_cu_enabled(adev, cu_per_sh);
1573
1574 mask = 1;
1575 for (k = 0; k < 16; k++) {
1576 mask <<= k;
1577 if (active_cu & mask) {
1578 data &= ~mask;
25069e06 1579 WREG32(mmSPI_STATIC_THREAD_MGMT_3, data);
2cd46ad2
KW		 1580 break;
1581 }
1582 }
1583 }
1584 }
1585 gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
deca1d1f 1586 mutex_unlock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 1587}
1588
1589static void gfx_v6_0_gpu_init(struct amdgpu_device *adev)
1590{
1591 u32 gb_addr_config = 0;
1592 u32 mc_shared_chmap, mc_arb_ramcfg;
1593 u32 sx_debug_1;
1594 u32 hdp_host_path_cntl;
1595 u32 tmp;
1596
1597 switch (adev->asic_type) {
1598 case CHIP_TAHITI:
1599 adev->gfx.config.max_shader_engines = 2;
1600 adev->gfx.config.max_tile_pipes = 12;
1601 adev->gfx.config.max_cu_per_sh = 8;
1602 adev->gfx.config.max_sh_per_se = 2;
1603 adev->gfx.config.max_backends_per_se = 4;
1604 adev->gfx.config.max_texture_channel_caches = 12;
1605 adev->gfx.config.max_gprs = 256;
1606 adev->gfx.config.max_gs_threads = 32;
1607 adev->gfx.config.max_hw_contexts = 8;
1608
1609 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1610 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
1611 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1612 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
1613 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
1614 break;
1615 case CHIP_PITCAIRN:
1616 adev->gfx.config.max_shader_engines = 2;
1617 adev->gfx.config.max_tile_pipes = 8;
1618 adev->gfx.config.max_cu_per_sh = 5;
1619 adev->gfx.config.max_sh_per_se = 2;
1620 adev->gfx.config.max_backends_per_se = 4;
1621 adev->gfx.config.max_texture_channel_caches = 8;
1622 adev->gfx.config.max_gprs = 256;
1623 adev->gfx.config.max_gs_threads = 32;
1624 adev->gfx.config.max_hw_contexts = 8;
1625
1626 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1627 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
1628 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1629 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
1630 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
1631 break;
2cd46ad2
KW		 1632 case CHIP_VERDE:
1633 adev->gfx.config.max_shader_engines = 1;
1634 adev->gfx.config.max_tile_pipes = 4;
1635 adev->gfx.config.max_cu_per_sh = 5;
1636 adev->gfx.config.max_sh_per_se = 2;
1637 adev->gfx.config.max_backends_per_se = 4;
1638 adev->gfx.config.max_texture_channel_caches = 4;
1639 adev->gfx.config.max_gprs = 256;
1640 adev->gfx.config.max_gs_threads = 32;
1641 adev->gfx.config.max_hw_contexts = 8;
1642
1643 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1644 adev->gfx.config.sc_prim_fifo_size_backend = 0x40;
1645 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1646 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
1647 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
1648 break;
1649 case CHIP_OLAND:
1650 adev->gfx.config.max_shader_engines = 1;
1651 adev->gfx.config.max_tile_pipes = 4;
1652 adev->gfx.config.max_cu_per_sh = 6;
1653 adev->gfx.config.max_sh_per_se = 1;
1654 adev->gfx.config.max_backends_per_se = 2;
1655 adev->gfx.config.max_texture_channel_caches = 4;
1656 adev->gfx.config.max_gprs = 256;
1657 adev->gfx.config.max_gs_threads = 16;
1658 adev->gfx.config.max_hw_contexts = 8;
1659
1660 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1661 adev->gfx.config.sc_prim_fifo_size_backend = 0x40;
1662 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1663 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
1664 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
1665 break;
1666 case CHIP_HAINAN:
1667 adev->gfx.config.max_shader_engines = 1;
1668 adev->gfx.config.max_tile_pipes = 4;
1669 adev->gfx.config.max_cu_per_sh = 5;
1670 adev->gfx.config.max_sh_per_se = 1;
1671 adev->gfx.config.max_backends_per_se = 1;
1672 adev->gfx.config.max_texture_channel_caches = 2;
1673 adev->gfx.config.max_gprs = 256;
1674 adev->gfx.config.max_gs_threads = 16;
1675 adev->gfx.config.max_hw_contexts = 8;
1676
1677 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1678 adev->gfx.config.sc_prim_fifo_size_backend = 0x40;
1679 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1680 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x130;
1681 gb_addr_config = HAINAN_GB_ADDR_CONFIG_GOLDEN;
1682 break;
1683 default:
1684 BUG();
1685 break;
1686 }
1687
25069e06
TSD		 1688 WREG32(mmGRBM_CNTL, (0xff << GRBM_CNTL__READ_TIMEOUT__SHIFT));
1689 WREG32(mmSRBM_INT_CNTL, 1);
1690 WREG32(mmSRBM_INT_ACK, 1);
2cd46ad2 1691
25069e06 1692 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
2cd46ad2 1693
25069e06
TSD		 1694 mc_shared_chmap = RREG32(mmMC_SHARED_CHMAP);
1695 mc_arb_ramcfg = RREG32(mmMC_ARB_RAMCFG);
2cd46ad2
KW		 1696
1697 adev->gfx.config.num_tile_pipes = adev->gfx.config.max_tile_pipes;
1698 adev->gfx.config.mem_max_burst_length_bytes = 256;
25069e06 1699 tmp = (mc_arb_ramcfg & MC_ARB_RAMCFG__NOOFCOLS_MASK) >> MC_ARB_RAMCFG__NOOFCOLS__SHIFT;
2cd46ad2
KW		 1700 adev->gfx.config.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
1701 if (adev->gfx.config.mem_row_size_in_kb > 4)
1702 adev->gfx.config.mem_row_size_in_kb = 4;
1703 adev->gfx.config.shader_engine_tile_size = 32;
1704 adev->gfx.config.num_gpus = 1;
1705 adev->gfx.config.multi_gpu_tile_size = 64;
1706
25069e06 1707 gb_addr_config &= ~GB_ADDR_CONFIG__ROW_SIZE_MASK;
2cd46ad2
KW		 1708 switch (adev->gfx.config.mem_row_size_in_kb) {
1709 case 1:
1710 default:
25069e06 1711 gb_addr_config |= 0 << GB_ADDR_CONFIG__ROW_SIZE__SHIFT;
2cd46ad2
KW		 1712 break;
1713 case 2:
25069e06 1714 gb_addr_config |= 1 << GB_ADDR_CONFIG__ROW_SIZE__SHIFT;
2cd46ad2
KW		 1715 break;
1716 case 4:
25069e06 1717 gb_addr_config |= 2 << GB_ADDR_CONFIG__ROW_SIZE__SHIFT;
2cd46ad2
KW		 1718 break;
1719 }
1720 adev->gfx.config.gb_addr_config = gb_addr_config;
1721
25069e06
TSD		 1722 WREG32(mmGB_ADDR_CONFIG, gb_addr_config);
1723 WREG32(mmDMIF_ADDR_CONFIG, gb_addr_config);
1724 WREG32(mmDMIF_ADDR_CALC, gb_addr_config);
1725 WREG32(mmHDP_ADDR_CONFIG, gb_addr_config);
1726 WREG32(mmDMA_TILING_CONFIG + DMA0_REGISTER_OFFSET, gb_addr_config);
1727 WREG32(mmDMA_TILING_CONFIG + DMA1_REGISTER_OFFSET, gb_addr_config);
1728
2cd46ad2
KW		 1729#if 0
1730 if (adev->has_uvd) {
25069e06
TSD		 1731 WREG32(mmUVD_UDEC_ADDR_CONFIG, gb_addr_config);
1732 WREG32(mmUVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
1733 WREG32(mmUVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
2cd46ad2
KW		 1734 }
1735#endif
1736 gfx_v6_0_tiling_mode_table_init(adev);
1737
1738 gfx_v6_0_setup_rb(adev, adev->gfx.config.max_shader_engines,
1739 adev->gfx.config.max_sh_per_se,
1740 adev->gfx.config.max_backends_per_se);
1741
1742 gfx_v6_0_setup_spi(adev, adev->gfx.config.max_shader_engines,
1743 adev->gfx.config.max_sh_per_se,
1744 adev->gfx.config.max_cu_per_sh);
1745
1746 gfx_v6_0_get_cu_info(adev);
1747
25069e06
TSD		 1748 WREG32(mmCP_QUEUE_THRESHOLDS, ((0x16 << CP_QUEUE_THRESHOLDS__ROQ_IB1_START__SHIFT) |
1749 (0x2b << CP_QUEUE_THRESHOLDS__ROQ_IB2_START__SHIFT)));
1750 WREG32(mmCP_MEQ_THRESHOLDS, (0x30 << CP_MEQ_THRESHOLDS__MEQ1_START__SHIFT) |
1751 (0x60 << CP_MEQ_THRESHOLDS__MEQ2_START__SHIFT));
2cd46ad2 1752
25069e06
TSD		 1753 sx_debug_1 = RREG32(mmSX_DEBUG_1);
1754 WREG32(mmSX_DEBUG_1, sx_debug_1);
2cd46ad2 1755
25069e06 1756 WREG32(mmSPI_CONFIG_CNTL_1, (4 << SPI_CONFIG_CNTL_1__VTX_DONE_DELAY__SHIFT));
2cd46ad2 1757
25069e06
TSD		 1758 WREG32(mmPA_SC_FIFO_SIZE, ((adev->gfx.config.sc_prim_fifo_size_frontend << PA_SC_FIFO_SIZE__SC_FRONTEND_PRIM_FIFO_SIZE__SHIFT) |
1759 (adev->gfx.config.sc_prim_fifo_size_backend << PA_SC_FIFO_SIZE__SC_BACKEND_PRIM_FIFO_SIZE__SHIFT) |
1760 (adev->gfx.config.sc_hiz_tile_fifo_size << PA_SC_FIFO_SIZE__SC_HIZ_TILE_FIFO_SIZE__SHIFT) |
1761 (adev->gfx.config.sc_earlyz_tile_fifo_size << PA_SC_FIFO_SIZE__SC_EARLYZ_TILE_FIFO_SIZE__SHIFT)));
2cd46ad2 1762
25069e06
TSD		 1763 WREG32(mmVGT_NUM_INSTANCES, 1);
1764 WREG32(mmCP_PERFMON_CNTL, 0);
1765 WREG32(mmSQ_CONFIG, 0);
1766 WREG32(mmPA_SC_FORCE_EOV_MAX_CNTS, ((4095 << PA_SC_FORCE_EOV_MAX_CNTS__FORCE_EOV_MAX_CLK_CNT__SHIFT) |
1767 (255 << PA_SC_FORCE_EOV_MAX_CNTS__FORCE_EOV_MAX_REZ_CNT__SHIFT)));
2cd46ad2 1768
25069e06
TSD		 1769 WREG32(mmVGT_CACHE_INVALIDATION,
1770 (VC_AND_TC << VGT_CACHE_INVALIDATION__CACHE_INVALIDATION__SHIFT) |
1771 (ES_AND_GS_AUTO << VGT_CACHE_INVALIDATION__AUTO_INVLD_EN__SHIFT));
2cd46ad2 1772
25069e06
TSD		 1773 WREG32(mmVGT_GS_VERTEX_REUSE, 16);
1774 WREG32(mmPA_SC_LINE_STIPPLE_STATE, 0);
2cd46ad2 1775
25069e06
TSD		 1776 WREG32(mmCB_PERFCOUNTER0_SELECT0, 0);
1777 WREG32(mmCB_PERFCOUNTER0_SELECT1, 0);
1778 WREG32(mmCB_PERFCOUNTER1_SELECT0, 0);
1779 WREG32(mmCB_PERFCOUNTER1_SELECT1, 0);
1780 WREG32(mmCB_PERFCOUNTER2_SELECT0, 0);
1781 WREG32(mmCB_PERFCOUNTER2_SELECT1, 0);
1782 WREG32(mmCB_PERFCOUNTER3_SELECT0, 0);
1783 WREG32(mmCB_PERFCOUNTER3_SELECT1, 0);
2cd46ad2 1784
25069e06
TSD		 1785 hdp_host_path_cntl = RREG32(mmHDP_HOST_PATH_CNTL);
1786 WREG32(mmHDP_HOST_PATH_CNTL, hdp_host_path_cntl);
2cd46ad2 1787
25069e06
TSD		 1788 WREG32(mmPA_CL_ENHANCE, PA_CL_ENHANCE__CLIP_VTX_REORDER_ENA_MASK |
1789 (3 << PA_CL_ENHANCE__NUM_CLIP_SEQ__SHIFT));
2cd46ad2
KW		 1790
1791 udelay(50);
2cd46ad2
KW		 1792}
1793
1794
1795static void gfx_v6_0_scratch_init(struct amdgpu_device *adev)
1796{
2cd46ad2 1797 adev->gfx.scratch.num_reg = 7;
25069e06 1798 adev->gfx.scratch.reg_base = mmSCRATCH_REG0;
50261151 1799 adev->gfx.scratch.free_mask = (1u << adev->gfx.scratch.num_reg) - 1;
2cd46ad2
KW		 1800}
1801
1802static int gfx_v6_0_ring_test_ring(struct amdgpu_ring *ring)
1803{
1804 struct amdgpu_device *adev = ring->adev;
1805 uint32_t scratch;
1806 uint32_t tmp = 0;
1807 unsigned i;
1808 int r;
1809
1810 r = amdgpu_gfx_scratch_get(adev, &scratch);
1811 if (r) {
1812 DRM_ERROR("amdgpu: cp failed to get scratch reg (%d).\n", r);
1813 return r;
1814 }
1815 WREG32(scratch, 0xCAFEDEAD);
1816
1817 r = amdgpu_ring_alloc(ring, 3);
1818 if (r) {
1819 DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r);
1820 amdgpu_gfx_scratch_free(adev, scratch);
1821 return r;
1822 }
1823 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
1824 amdgpu_ring_write(ring, (scratch - PACKET3_SET_CONFIG_REG_START));
1825 amdgpu_ring_write(ring, 0xDEADBEEF);
1826 amdgpu_ring_commit(ring);
1827
1828 for (i = 0; i < adev->usec_timeout; i++) {
1829 tmp = RREG32(scratch);
1830 if (tmp == 0xDEADBEEF)
1831 break;
1832 DRM_UDELAY(1);
1833 }
1834 if (i < adev->usec_timeout) {
1835 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
1836 } else {
1837 DRM_ERROR("amdgpu: ring %d test failed (scratch(0x%04X)=0x%08X)\n",
1838 ring->idx, scratch, tmp);
1839 r = -EINVAL;
1840 }
1841 amdgpu_gfx_scratch_free(adev, scratch);
1842 return r;
1843}
1844
1845static void gfx_v6_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
1846{
1847 /* flush hdp cache */
1848 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
1849 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
1850 WRITE_DATA_DST_SEL(0)));
25069e06 1851 amdgpu_ring_write(ring, mmHDP_MEM_COHERENCY_FLUSH_CNTL);
2cd46ad2
KW		 1852 amdgpu_ring_write(ring, 0);
1853 amdgpu_ring_write(ring, 0x1);
1854}
1855
45682886
ML		 1856static void gfx_v6_0_ring_emit_vgt_flush(struct amdgpu_ring *ring)
1857{
1858 amdgpu_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE, 0));
1859 amdgpu_ring_write(ring, EVENT_TYPE(VGT_FLUSH) |
1860 EVENT_INDEX(0));
1861}
1862
2cd46ad2
KW		 1863/**
1864 * gfx_v6_0_ring_emit_hdp_invalidate - emit an hdp invalidate on the cp
1865 *
1866 * @adev: amdgpu_device pointer
1867 * @ridx: amdgpu ring index
1868 *
1869 * Emits an hdp invalidate on the cp.
1870 */
1871static void gfx_v6_0_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
1872{
1873 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
1874 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
1875 WRITE_DATA_DST_SEL(0)));
25069e06 1876 amdgpu_ring_write(ring, mmHDP_DEBUG0);
2cd46ad2
KW		 1877 amdgpu_ring_write(ring, 0);
1878 amdgpu_ring_write(ring, 0x1);
1879}
1880
668f52c3
AD		 1881static void gfx_v6_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
1882 u64 seq, unsigned flags)
2cd46ad2
KW		 1883{
1884 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
1885 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
1886 /* flush read cache over gart */
1887 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
25069e06 1888 amdgpu_ring_write(ring, (mmCP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START));
2cd46ad2
KW		 1889 amdgpu_ring_write(ring, 0);
1890 amdgpu_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
1891 amdgpu_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
1892 PACKET3_TC_ACTION_ENA |
1893 PACKET3_SH_KCACHE_ACTION_ENA |
1894 PACKET3_SH_ICACHE_ACTION_ENA);
1895 amdgpu_ring_write(ring, 0xFFFFFFFF);
1896 amdgpu_ring_write(ring, 0);
1897 amdgpu_ring_write(ring, 10); /* poll interval */
1898 /* EVENT_WRITE_EOP - flush caches, send int */
1899 amdgpu_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
1900 amdgpu_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
1901 amdgpu_ring_write(ring, addr & 0xfffffffc);
1902 amdgpu_ring_write(ring, (upper_32_bits(addr) & 0xffff) |
25069e06
TSD		 1903 ((write64bit ? 2 : 1) << CP_EOP_DONE_DATA_CNTL__DATA_SEL__SHIFT) |
1904 ((int_sel ? 2 : 0) << CP_EOP_DONE_DATA_CNTL__INT_SEL__SHIFT));
2cd46ad2
KW		 1905 amdgpu_ring_write(ring, lower_32_bits(seq));
1906 amdgpu_ring_write(ring, upper_32_bits(seq));
1907}
1908
668f52c3
AD		 1909static void gfx_v6_0_ring_emit_ib(struct amdgpu_ring *ring,
1910 struct amdgpu_ib *ib,
1911 unsigned vm_id, bool ctx_switch)
2cd46ad2
KW		 1912{
1913 u32 header, control = 0;
1914
1915 /* insert SWITCH_BUFFER packet before first IB in the ring frame */
1916 if (ctx_switch) {
1917 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
1918 amdgpu_ring_write(ring, 0);
1919 }
1920
1921 if (ib->flags & AMDGPU_IB_FLAG_CE)
1922 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
1923 else
1924 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
1925
1926 control |= ib->length_dw | (vm_id << 24);
1927
1928 amdgpu_ring_write(ring, header);
1929 amdgpu_ring_write(ring,
1930#ifdef __BIG_ENDIAN
1931 (2 << 0) |
1932#endif
1933 (ib->gpu_addr & 0xFFFFFFFC));
1934 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
1935 amdgpu_ring_write(ring, control);
1936}
1937
2cd46ad2
KW		 1938/**
1939 * gfx_v6_0_ring_test_ib - basic ring IB test
1940 *
1941 * @ring: amdgpu_ring structure holding ring information
1942 *
1943 * Allocate an IB and execute it on the gfx ring (SI).
1944 * Provides a basic gfx ring test to verify that IBs are working.
1945 * Returns 0 on success, error on failure.
1946 */
1947static int gfx_v6_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1948{
1949 struct amdgpu_device *adev = ring->adev;
1950 struct amdgpu_ib ib;
f54d1867 1951 struct dma_fence *f = NULL;
2cd46ad2
KW		 1952 uint32_t scratch;
1953 uint32_t tmp = 0;
1954 long r;
1955
1956 r = amdgpu_gfx_scratch_get(adev, &scratch);
1957 if (r) {
1958 DRM_ERROR("amdgpu: failed to get scratch reg (%ld).\n", r);
1959 return r;
1960 }
1961 WREG32(scratch, 0xCAFEDEAD);
1962 memset(&ib, 0, sizeof(ib));
1963 r = amdgpu_ib_get(adev, NULL, 256, &ib);
1964 if (r) {
1965 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1966 goto err1;
1967 }
1968 ib.ptr[0] = PACKET3(PACKET3_SET_CONFIG_REG, 1);
1969 ib.ptr[1] = ((scratch - PACKET3_SET_CONFIG_REG_START));
1970 ib.ptr[2] = 0xDEADBEEF;
1971 ib.length_dw = 3;
1972
50ddc75e 1973 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
2cd46ad2
KW		 1974 if (r)
1975 goto err2;
1976
f54d1867 1977 r = dma_fence_wait_timeout(f, false, timeout);
2cd46ad2
KW		 1978 if (r == 0) {
1979 DRM_ERROR("amdgpu: IB test timed out\n");
1980 r = -ETIMEDOUT;
1981 goto err2;
1982 } else if (r < 0) {
1983 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1984 goto err2;
1985 }
1986 tmp = RREG32(scratch);
1987 if (tmp == 0xDEADBEEF) {
1988 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
1989 r = 0;
1990 } else {
1991 DRM_ERROR("amdgpu: ib test failed (scratch(0x%04X)=0x%08X)\n",
1992 scratch, tmp);
1993 r = -EINVAL;
1994 }
1995
1996err2:
1997 amdgpu_ib_free(adev, &ib, NULL);
f54d1867 1998 dma_fence_put(f);
2cd46ad2
KW		 1999err1:
2000 amdgpu_gfx_scratch_free(adev, scratch);
2001 return r;
2002}
2003
2004static void gfx_v6_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable)
2005{
2006 int i;
25069e06
TSD		 2007 if (enable) {
2008 WREG32(mmCP_ME_CNTL, 0);
2009 } else {
2010 WREG32(mmCP_ME_CNTL, (CP_ME_CNTL__ME_HALT_MASK |
2011 CP_ME_CNTL__PFP_HALT_MASK |
2012 CP_ME_CNTL__CE_HALT_MASK));
2013 WREG32(mmSCRATCH_UMSK, 0);
2cd46ad2
KW		 2014 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
2015 adev->gfx.gfx_ring[i].ready = false;
2016 for (i = 0; i < adev->gfx.num_compute_rings; i++)
2017 adev->gfx.compute_ring[i].ready = false;
2018 }
2019 udelay(50);
2020}
2021
2022static int gfx_v6_0_cp_gfx_load_microcode(struct amdgpu_device *adev)
2023{
2024 unsigned i;
2025 const struct gfx_firmware_header_v1_0 *pfp_hdr;
2026 const struct gfx_firmware_header_v1_0 *ce_hdr;
2027 const struct gfx_firmware_header_v1_0 *me_hdr;
2028 const __le32 *fw_data;
2029 u32 fw_size;
2030
2031 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw || !adev->gfx.ce_fw)
2032 return -EINVAL;
2033
2034 gfx_v6_0_cp_gfx_enable(adev, false);
2035 pfp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
2036 ce_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
2037 me_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
2038
2039 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header);
2040 amdgpu_ucode_print_gfx_hdr(&ce_hdr->header);
2041 amdgpu_ucode_print_gfx_hdr(&me_hdr->header);
2042
2043 /* PFP */
2044 fw_data = (const __le32 *)
2045 (adev->gfx.pfp_fw->data + le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
2046 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
25069e06 2047 WREG32(mmCP_PFP_UCODE_ADDR, 0);
2cd46ad2 2048 for (i = 0; i < fw_size; i++)
25069e06
TSD		 2049 WREG32(mmCP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
2050 WREG32(mmCP_PFP_UCODE_ADDR, 0);
2cd46ad2
KW		 2051
2052 /* CE */
2053 fw_data = (const __le32 *)
2054 (adev->gfx.ce_fw->data + le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
2055 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
25069e06 2056 WREG32(mmCP_CE_UCODE_ADDR, 0);
2cd46ad2 2057 for (i = 0; i < fw_size; i++)
25069e06
TSD		 2058 WREG32(mmCP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
2059 WREG32(mmCP_CE_UCODE_ADDR, 0);
2cd46ad2
KW		 2060
2061 /* ME */
2062 fw_data = (const __be32 *)
2063 (adev->gfx.me_fw->data + le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
2064 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
25069e06 2065 WREG32(mmCP_ME_RAM_WADDR, 0);
2cd46ad2 2066 for (i = 0; i < fw_size; i++)
25069e06
TSD		 2067 WREG32(mmCP_ME_RAM_DATA, le32_to_cpup(fw_data++));
2068 WREG32(mmCP_ME_RAM_WADDR, 0);
2cd46ad2 2069
25069e06
TSD		 2070 WREG32(mmCP_PFP_UCODE_ADDR, 0);
2071 WREG32(mmCP_CE_UCODE_ADDR, 0);
2072 WREG32(mmCP_ME_RAM_WADDR, 0);
2073 WREG32(mmCP_ME_RAM_RADDR, 0);
2cd46ad2
KW		 2074 return 0;
2075}
2076
2077static int gfx_v6_0_cp_gfx_start(struct amdgpu_device *adev)
2078{
2079 const struct cs_section_def *sect = NULL;
2080 const struct cs_extent_def *ext = NULL;
2081 struct amdgpu_ring *ring = &adev->gfx.gfx_ring[0];
2082 int r, i;
2083
2084 r = amdgpu_ring_alloc(ring, 7 + 4);
2085 if (r) {
2086 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r);
2087 return r;
2088 }
2089 amdgpu_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
2090 amdgpu_ring_write(ring, 0x1);
2091 amdgpu_ring_write(ring, 0x0);
2092 amdgpu_ring_write(ring, adev->gfx.config.max_hw_contexts - 1);
2093 amdgpu_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
2094 amdgpu_ring_write(ring, 0);
2095 amdgpu_ring_write(ring, 0);
2096
2097 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
2098 amdgpu_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
2099 amdgpu_ring_write(ring, 0xc000);
2100 amdgpu_ring_write(ring, 0xe000);
2101 amdgpu_ring_commit(ring);
2102
2103 gfx_v6_0_cp_gfx_enable(adev, true);
2104
2105 r = amdgpu_ring_alloc(ring, gfx_v6_0_get_csb_size(adev) + 10);
2106 if (r) {
2107 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r);
2108 return r;
2109 }
2110
2111 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
2112 amdgpu_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
2113
2114 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
2115 for (ext = sect->section; ext->extent != NULL; ++ext) {
2116 if (sect->id == SECT_CONTEXT) {
2117 amdgpu_ring_write(ring,
2118 PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
2119 amdgpu_ring_write(ring, ext->reg_index - PACKET3_SET_CONTEXT_REG_START);
2120 for (i = 0; i < ext->reg_count; i++)
2121 amdgpu_ring_write(ring, ext->extent[i]);
2122 }
2123 }
2124 }
2125
2126 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
2127 amdgpu_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
2128
2129 amdgpu_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
2130 amdgpu_ring_write(ring, 0);
2131
2132 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
2133 amdgpu_ring_write(ring, 0x00000316);
2134 amdgpu_ring_write(ring, 0x0000000e);
2135 amdgpu_ring_write(ring, 0x00000010);
2136
2137 amdgpu_ring_commit(ring);
2138
2139 return 0;
2140}
2141
2142static int gfx_v6_0_cp_gfx_resume(struct amdgpu_device *adev)
2143{
2144 struct amdgpu_ring *ring;
2145 u32 tmp;
2146 u32 rb_bufsz;
2147 int r;
2148 u64 rptr_addr;
2149
25069e06
TSD		 2150 WREG32(mmCP_SEM_WAIT_TIMER, 0x0);
2151 WREG32(mmCP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
2cd46ad2
KW		 2152
2153 /* Set the write pointer delay */
25069e06 2154 WREG32(mmCP_RB_WPTR_DELAY, 0);
2cd46ad2 2155
25069e06
TSD		 2156 WREG32(mmCP_DEBUG, 0);
2157 WREG32(mmSCRATCH_ADDR, 0);
2cd46ad2
KW		 2158
2159 /* ring 0 - compute and gfx */
2160 /* Set ring buffer size */
2161 ring = &adev->gfx.gfx_ring[0];
2162 rb_bufsz = order_base_2(ring->ring_size / 8);
2163 tmp = (order_base_2(AMDGPU_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2164
2165#ifdef __BIG_ENDIAN
2166 tmp |= BUF_SWAP_32BIT;
2167#endif
25069e06 2168 WREG32(mmCP_RB0_CNTL, tmp);
2cd46ad2
KW		 2169
2170 /* Initialize the ring buffer's read and write pointers */
25069e06 2171 WREG32(mmCP_RB0_CNTL, tmp | CP_RB0_CNTL__RB_RPTR_WR_ENA_MASK);
2cd46ad2 2172 ring->wptr = 0;
25069e06 2173 WREG32(mmCP_RB0_WPTR, ring->wptr);
2cd46ad2
KW		 2174
2175 /* set the wb address whether it's enabled or not */
2176 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
25069e06
TSD		 2177 WREG32(mmCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr));
2178 WREG32(mmCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 0xFF);
2cd46ad2 2179
25069e06 2180 WREG32(mmSCRATCH_UMSK, 0);
2cd46ad2
KW		 2181
2182 mdelay(1);
25069e06 2183 WREG32(mmCP_RB0_CNTL, tmp);
2cd46ad2 2184
25069e06 2185 WREG32(mmCP_RB0_BASE, ring->gpu_addr >> 8);
2cd46ad2
KW		 2186
2187 /* start the rings */
2188 gfx_v6_0_cp_gfx_start(adev);
2189 ring->ready = true;
2190 r = amdgpu_ring_test_ring(ring);
2191 if (r) {
2192 ring->ready = false;
2193 return r;
2194 }
2195
2196 return 0;
2197}
2198
6f924e20 2199static u32 gfx_v6_0_ring_get_rptr(struct amdgpu_ring *ring)
2cd46ad2 2200{
4aeacf0f 2201 return ring->adev->wb.wb[ring->rptr_offs];
2cd46ad2
KW		 2202}
2203
832c6ef7 2204static u32 gfx_v6_0_ring_get_wptr(struct amdgpu_ring *ring)
2cd46ad2
KW		 2205{
2206 struct amdgpu_device *adev = ring->adev;
2cd46ad2 2207
832c6ef7 2208 if (ring == &adev->gfx.gfx_ring[0])
25069e06 2209 return RREG32(mmCP_RB0_WPTR);
832c6ef7 2210 else if (ring == &adev->gfx.compute_ring[0])
25069e06 2211 return RREG32(mmCP_RB1_WPTR);
832c6ef7 2212 else if (ring == &adev->gfx.compute_ring[1])
25069e06 2213 return RREG32(mmCP_RB2_WPTR);
832c6ef7
TSD		 2214 else
2215 BUG();
2cd46ad2
KW		 2216}
2217
2218static void gfx_v6_0_ring_set_wptr_gfx(struct amdgpu_ring *ring)
2219{
2220 struct amdgpu_device *adev = ring->adev;
2221
25069e06
TSD		 2222 WREG32(mmCP_RB0_WPTR, ring->wptr);
2223 (void)RREG32(mmCP_RB0_WPTR);
2cd46ad2
KW		 2224}
2225
2cd46ad2
KW		 2226static void gfx_v6_0_ring_set_wptr_compute(struct amdgpu_ring *ring)
2227{
2228 struct amdgpu_device *adev = ring->adev;
2229
2230 if (ring == &adev->gfx.compute_ring[0]) {
25069e06
TSD		 2231 WREG32(mmCP_RB1_WPTR, ring->wptr);
2232 (void)RREG32(mmCP_RB1_WPTR);
2cd46ad2 2233 } else if (ring == &adev->gfx.compute_ring[1]) {
25069e06
TSD		 2234 WREG32(mmCP_RB2_WPTR, ring->wptr);
2235 (void)RREG32(mmCP_RB2_WPTR);
2cd46ad2
KW		 2236 } else {
2237 BUG();
2238 }
2239
2240}
2241
2cd46ad2
KW		 2242static int gfx_v6_0_cp_compute_resume(struct amdgpu_device *adev)
2243{
2244 struct amdgpu_ring *ring;
2245 u32 tmp;
2246 u32 rb_bufsz;
25069e06 2247 int i, r;
2cd46ad2
KW		 2248 u64 rptr_addr;
2249
2250 /* ring1 - compute only */
2251 /* Set ring buffer size */
2252
2253 ring = &adev->gfx.compute_ring[0];
2254 rb_bufsz = order_base_2(ring->ring_size / 8);
2255 tmp = (order_base_2(AMDGPU_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2256#ifdef __BIG_ENDIAN
2257 tmp |= BUF_SWAP_32BIT;
2258#endif
25069e06 2259 WREG32(mmCP_RB1_CNTL, tmp);
2cd46ad2 2260
25069e06 2261 WREG32(mmCP_RB1_CNTL, tmp | CP_RB1_CNTL__RB_RPTR_WR_ENA_MASK);
2cd46ad2 2262 ring->wptr = 0;
25069e06 2263 WREG32(mmCP_RB1_WPTR, ring->wptr);
2cd46ad2 2264
2cd46ad2 2265 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
25069e06
TSD		 2266 WREG32(mmCP_RB1_RPTR_ADDR, lower_32_bits(rptr_addr));
2267 WREG32(mmCP_RB1_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 0xFF);
2cd46ad2
KW		 2268
2269 mdelay(1);
25069e06
TSD		 2270 WREG32(mmCP_RB1_CNTL, tmp);
2271 WREG32(mmCP_RB1_BASE, ring->gpu_addr >> 8);
2cd46ad2
KW		 2272
2273 ring = &adev->gfx.compute_ring[1];
2274 rb_bufsz = order_base_2(ring->ring_size / 8);
2275 tmp = (order_base_2(AMDGPU_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2276#ifdef __BIG_ENDIAN
2277 tmp |= BUF_SWAP_32BIT;
2278#endif
25069e06 2279 WREG32(mmCP_RB2_CNTL, tmp);
2cd46ad2 2280
25069e06 2281 WREG32(mmCP_RB2_CNTL, tmp | CP_RB2_CNTL__RB_RPTR_WR_ENA_MASK);
2cd46ad2 2282 ring->wptr = 0;
25069e06 2283 WREG32(mmCP_RB2_WPTR, ring->wptr);
2cd46ad2 2284 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
25069e06
TSD		 2285 WREG32(mmCP_RB2_RPTR_ADDR, lower_32_bits(rptr_addr));
2286 WREG32(mmCP_RB2_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 0xFF);
2cd46ad2
KW		 2287
2288 mdelay(1);
25069e06
TSD		 2289 WREG32(mmCP_RB2_CNTL, tmp);
2290 WREG32(mmCP_RB2_BASE, ring->gpu_addr >> 8);
2cd46ad2 2291
25069e06
TSD		 2292 adev->gfx.compute_ring[0].ready = false;
2293 adev->gfx.compute_ring[1].ready = false;
2cd46ad2 2294
25069e06
TSD		 2295 for (i = 0; i < 2; i++) {
2296 r = amdgpu_ring_test_ring(&adev->gfx.compute_ring[i]);
2297 if (r)
2298 return r;
2299 adev->gfx.compute_ring[i].ready = true;
2cd46ad2
KW		 2300 }
2301
2302 return 0;
2303}
2304
2305static void gfx_v6_0_cp_enable(struct amdgpu_device *adev, bool enable)
2306{
2307 gfx_v6_0_cp_gfx_enable(adev, enable);
2308}
2309
2310static int gfx_v6_0_cp_load_microcode(struct amdgpu_device *adev)
2311{
4aeacf0f 2312 return gfx_v6_0_cp_gfx_load_microcode(adev);
2cd46ad2
KW		 2313}
2314
2315static void gfx_v6_0_enable_gui_idle_interrupt(struct amdgpu_device *adev,
2316 bool enable)
45682886 2317{
25069e06 2318 u32 tmp = RREG32(mmCP_INT_CNTL_RING0);
2cd46ad2
KW		 2319 u32 mask;
2320 int i;
2321
2322 if (enable)
25069e06
TSD		 2323 tmp |= (CP_INT_CNTL__CNTX_BUSY_INT_ENABLE_MASK |
2324 CP_INT_CNTL__CNTX_EMPTY_INT_ENABLE_MASK);
2cd46ad2 2325 else
25069e06
TSD		 2326 tmp &= ~(CP_INT_CNTL__CNTX_BUSY_INT_ENABLE_MASK |
2327 CP_INT_CNTL__CNTX_EMPTY_INT_ENABLE_MASK);
2328 WREG32(mmCP_INT_CNTL_RING0, tmp);
2cd46ad2
KW		 2329
2330 if (!enable) {
2331 /* read a gfx register */
25069e06 2332 tmp = RREG32(mmDB_DEPTH_INFO);
2cd46ad2
KW		 2333
2334 mask = RLC_BUSY_STATUS | GFX_POWER_STATUS | GFX_CLOCK_STATUS | GFX_LS_STATUS;
2335 for (i = 0; i < adev->usec_timeout; i++) {
25069e06 2336 if ((RREG32(mmRLC_STAT) & mask) == (GFX_CLOCK_STATUS | GFX_POWER_STATUS))
2cd46ad2
KW		 2337 break;
2338 udelay(1);
2339 }
2340 }
2341}
2342
2343static int gfx_v6_0_cp_resume(struct amdgpu_device *adev)
2344{
2345 int r;
2346
2347 gfx_v6_0_enable_gui_idle_interrupt(adev, false);
2348
2349 r = gfx_v6_0_cp_load_microcode(adev);
2350 if (r)
2351 return r;
2352
2353 r = gfx_v6_0_cp_gfx_resume(adev);
2354 if (r)
2355 return r;
2356 r = gfx_v6_0_cp_compute_resume(adev);
2357 if (r)
2358 return r;
2359
2360 gfx_v6_0_enable_gui_idle_interrupt(adev, true);
2361
2362 return 0;
2363}
2364
2365static void gfx_v6_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
2366{
21cd942e 2367 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
2cd46ad2
KW		 2368 uint32_t seq = ring->fence_drv.sync_seq;
2369 uint64_t addr = ring->fence_drv.gpu_addr;
2370
2371 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
2372 amdgpu_ring_write(ring, (WAIT_REG_MEM_MEM_SPACE(1) | /* memory */
2373 WAIT_REG_MEM_FUNCTION(3) | /* equal */
2374 WAIT_REG_MEM_ENGINE(usepfp))); /* pfp or me */
2375 amdgpu_ring_write(ring, addr & 0xfffffffc);
2376 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
2377 amdgpu_ring_write(ring, seq);
2378 amdgpu_ring_write(ring, 0xffffffff);
2379 amdgpu_ring_write(ring, 4); /* poll interval */
2380
2381 if (usepfp) {
2382 /* synce CE with ME to prevent CE fetch CEIB before context switch done */
2383 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
2384 amdgpu_ring_write(ring, 0);
2385 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
2386 amdgpu_ring_write(ring, 0);
2387 }
2388}
2389
2390static void gfx_v6_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
2391 unsigned vm_id, uint64_t pd_addr)
2392{
21cd942e 2393 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
2cd46ad2
KW		 2394
2395 /* write new base address */
2396 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
2397 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
2398 WRITE_DATA_DST_SEL(0)));
2399 if (vm_id < 8) {
25069e06 2400 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id ));
2cd46ad2 2401 } else {
25069e06 2402 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + (vm_id - 8)));
2cd46ad2
KW		 2403 }
2404 amdgpu_ring_write(ring, 0);
2405 amdgpu_ring_write(ring, pd_addr >> 12);
2406
2407 /* bits 0-15 are the VM contexts0-15 */
2408 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
2409 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
2410 WRITE_DATA_DST_SEL(0)));
25069e06 2411 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
2cd46ad2
KW		 2412 amdgpu_ring_write(ring, 0);
2413 amdgpu_ring_write(ring, 1 << vm_id);
2414
2415 /* wait for the invalidate to complete */
2416 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
2417 amdgpu_ring_write(ring, (WAIT_REG_MEM_FUNCTION(0) | /* always */
2418 WAIT_REG_MEM_ENGINE(0))); /* me */
25069e06 2419 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
2cd46ad2
KW		 2420 amdgpu_ring_write(ring, 0);
2421 amdgpu_ring_write(ring, 0); /* ref */
2422 amdgpu_ring_write(ring, 0); /* mask */
2423 amdgpu_ring_write(ring, 0x20); /* poll interval */
2424
2425 if (usepfp) {
2426 /* sync PFP to ME, otherwise we might get invalid PFP reads */
2427 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
2428 amdgpu_ring_write(ring, 0x0);
2429
2430 /* synce CE with ME to prevent CE fetch CEIB before context switch done */
2431 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
2432 amdgpu_ring_write(ring, 0);
2433 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
2434 amdgpu_ring_write(ring, 0);
2435 }
2436}
2437
2438
2439static void gfx_v6_0_rlc_fini(struct amdgpu_device *adev)
2440{
2441 int r;
2442
2443 if (adev->gfx.rlc.save_restore_obj) {
2444 r = amdgpu_bo_reserve(adev->gfx.rlc.save_restore_obj, false);
2445 if (unlikely(r != 0))
2446 dev_warn(adev->dev, "(%d) reserve RLC sr bo failed\n", r);
2447 amdgpu_bo_unpin(adev->gfx.rlc.save_restore_obj);
2448 amdgpu_bo_unreserve(adev->gfx.rlc.save_restore_obj);
2449
2450 amdgpu_bo_unref(&adev->gfx.rlc.save_restore_obj);
2451 adev->gfx.rlc.save_restore_obj = NULL;
2452 }
2453
2454 if (adev->gfx.rlc.clear_state_obj) {
2455 r = amdgpu_bo_reserve(adev->gfx.rlc.clear_state_obj, false);
2456 if (unlikely(r != 0))
2457 dev_warn(adev->dev, "(%d) reserve RLC c bo failed\n", r);
2458 amdgpu_bo_unpin(adev->gfx.rlc.clear_state_obj);
2459 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
2460
2461 amdgpu_bo_unref(&adev->gfx.rlc.clear_state_obj);
2462 adev->gfx.rlc.clear_state_obj = NULL;
2463 }
2464
2465 if (adev->gfx.rlc.cp_table_obj) {
2466 r = amdgpu_bo_reserve(adev->gfx.rlc.cp_table_obj, false);
2467 if (unlikely(r != 0))
2468 dev_warn(adev->dev, "(%d) reserve RLC cp table bo failed\n", r);
2469 amdgpu_bo_unpin(adev->gfx.rlc.cp_table_obj);
2470 amdgpu_bo_unreserve(adev->gfx.rlc.cp_table_obj);
2471
2472 amdgpu_bo_unref(&adev->gfx.rlc.cp_table_obj);
2473 adev->gfx.rlc.cp_table_obj = NULL;
2474 }
2475}
2476
2477static int gfx_v6_0_rlc_init(struct amdgpu_device *adev)
2478{
2479 const u32 *src_ptr;
2480 volatile u32 *dst_ptr;
2481 u32 dws, i;
2482 u64 reg_list_mc_addr;
2483 const struct cs_section_def *cs_data;
2484 int r;
2485
2486 adev->gfx.rlc.reg_list = verde_rlc_save_restore_register_list;
2487 adev->gfx.rlc.reg_list_size =
2488 (u32)ARRAY_SIZE(verde_rlc_save_restore_register_list);
2489
2490 adev->gfx.rlc.cs_data = si_cs_data;
2491 src_ptr = adev->gfx.rlc.reg_list;
2492 dws = adev->gfx.rlc.reg_list_size;
2493 cs_data = adev->gfx.rlc.cs_data;
2494
2495 if (src_ptr) {
2496 /* save restore block */
2497 if (adev->gfx.rlc.save_restore_obj == NULL) {
2cd46ad2
KW		 2498 r = amdgpu_bo_create(adev, dws * 4, PAGE_SIZE, true,
2499 AMDGPU_GEM_DOMAIN_VRAM,
2500 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
2501 NULL, NULL,
2502 &adev->gfx.rlc.save_restore_obj);
2503
2504 if (r) {
2505 dev_warn(adev->dev, "(%d) create RLC sr bo failed\n", r);
2506 return r;
2507 }
2508 }
2509
2510 r = amdgpu_bo_reserve(adev->gfx.rlc.save_restore_obj, false);
2511 if (unlikely(r != 0)) {
2512 gfx_v6_0_rlc_fini(adev);
2513 return r;
2514 }
2515 r = amdgpu_bo_pin(adev->gfx.rlc.save_restore_obj, AMDGPU_GEM_DOMAIN_VRAM,
2516 &adev->gfx.rlc.save_restore_gpu_addr);
2517 if (r) {
2518 amdgpu_bo_unreserve(adev->gfx.rlc.save_restore_obj);
2519 dev_warn(adev->dev, "(%d) pin RLC sr bo failed\n", r);
2520 gfx_v6_0_rlc_fini(adev);
2521 return r;
2522 }
2523
2524 r = amdgpu_bo_kmap(adev->gfx.rlc.save_restore_obj, (void **)&adev->gfx.rlc.sr_ptr);
2525 if (r) {
2526 dev_warn(adev->dev, "(%d) map RLC sr bo failed\n", r);
2527 gfx_v6_0_rlc_fini(adev);
2528 return r;
2529 }
2530 /* write the sr buffer */
2531 dst_ptr = adev->gfx.rlc.sr_ptr;
2532 for (i = 0; i < adev->gfx.rlc.reg_list_size; i++)
2533 dst_ptr[i] = cpu_to_le32(src_ptr[i]);
2534 amdgpu_bo_kunmap(adev->gfx.rlc.save_restore_obj);
2535 amdgpu_bo_unreserve(adev->gfx.rlc.save_restore_obj);
2536 }
2537
2538 if (cs_data) {
2539 /* clear state block */
2540 adev->gfx.rlc.clear_state_size = gfx_v6_0_get_csb_size(adev);
2541 dws = adev->gfx.rlc.clear_state_size + (256 / 4);
2542
2543 if (adev->gfx.rlc.clear_state_obj == NULL) {
2544 r = amdgpu_bo_create(adev, dws * 4, PAGE_SIZE, true,
2545 AMDGPU_GEM_DOMAIN_VRAM,
2546 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
2547 NULL, NULL,
2548 &adev->gfx.rlc.clear_state_obj);
2549
2550 if (r) {
2551 dev_warn(adev->dev, "(%d) create RLC c bo failed\n", r);
2552 gfx_v6_0_rlc_fini(adev);
2553 return r;
2554 }
2555 }
2556 r = amdgpu_bo_reserve(adev->gfx.rlc.clear_state_obj, false);
2557 if (unlikely(r != 0)) {
2558 gfx_v6_0_rlc_fini(adev);
2559 return r;
2560 }
2561 r = amdgpu_bo_pin(adev->gfx.rlc.clear_state_obj, AMDGPU_GEM_DOMAIN_VRAM,
2562 &adev->gfx.rlc.clear_state_gpu_addr);
2563 if (r) {
2564 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
2565 dev_warn(adev->dev, "(%d) pin RLC c bo failed\n", r);
2566 gfx_v6_0_rlc_fini(adev);
2567 return r;
2568 }
2569
2570 r = amdgpu_bo_kmap(adev->gfx.rlc.clear_state_obj, (void **)&adev->gfx.rlc.cs_ptr);
2571 if (r) {
2572 dev_warn(adev->dev, "(%d) map RLC c bo failed\n", r);
2573 gfx_v6_0_rlc_fini(adev);
2574 return r;
2575 }
2576 /* set up the cs buffer */
2577 dst_ptr = adev->gfx.rlc.cs_ptr;
2578 reg_list_mc_addr = adev->gfx.rlc.clear_state_gpu_addr + 256;
2579 dst_ptr[0] = cpu_to_le32(upper_32_bits(reg_list_mc_addr));
2580 dst_ptr[1] = cpu_to_le32(lower_32_bits(reg_list_mc_addr));
2581 dst_ptr[2] = cpu_to_le32(adev->gfx.rlc.clear_state_size);
2582 gfx_v6_0_get_csb_buffer(adev, &dst_ptr[(256/4)]);
2583 amdgpu_bo_kunmap(adev->gfx.rlc.clear_state_obj);
2584 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
2585 }
2586
2587 return 0;
2588}
2589
2590static void gfx_v6_0_enable_lbpw(struct amdgpu_device *adev, bool enable)
2591{
25069e06 2592 WREG32_FIELD(RLC_LB_CNTL, LOAD_BALANCE_ENABLE, enable ? 1 : 0);
2cd46ad2
KW		 2593
2594 if (!enable) {
2595 gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
25069e06 2596 WREG32(mmSPI_LB_CU_MASK, 0x00ff);
2cd46ad2 2597 }
2cd46ad2
KW		 2598}
2599
2600static void gfx_v6_0_wait_for_rlc_serdes(struct amdgpu_device *adev)
2601{
2602 int i;
2603
2604 for (i = 0; i < adev->usec_timeout; i++) {
25069e06 2605 if (RREG32(mmRLC_SERDES_MASTER_BUSY_0) == 0)
2cd46ad2
KW		 2606 break;
2607 udelay(1);
2608 }
2609
2610 for (i = 0; i < adev->usec_timeout; i++) {
25069e06 2611 if (RREG32(mmRLC_SERDES_MASTER_BUSY_1) == 0)
2cd46ad2
KW		 2612 break;
2613 udelay(1);
2614 }
2615}
2616
2617static void gfx_v6_0_update_rlc(struct amdgpu_device *adev, u32 rlc)
2618{
2619 u32 tmp;
2620
25069e06 2621 tmp = RREG32(mmRLC_CNTL);
2cd46ad2 2622 if (tmp != rlc)
25069e06 2623 WREG32(mmRLC_CNTL, rlc);
2cd46ad2
KW		 2624}
2625
2626static u32 gfx_v6_0_halt_rlc(struct amdgpu_device *adev)
2627{
2628 u32 data, orig;
2629
25069e06 2630 orig = data = RREG32(mmRLC_CNTL);
2cd46ad2 2631
25069e06
TSD		 2632 if (data & RLC_CNTL__RLC_ENABLE_F32_MASK) {
2633 data &= ~RLC_CNTL__RLC_ENABLE_F32_MASK;
2634 WREG32(mmRLC_CNTL, data);
2cd46ad2
KW		 2635
2636 gfx_v6_0_wait_for_rlc_serdes(adev);
2637 }
2638
2639 return orig;
2640}
2641
2642static void gfx_v6_0_rlc_stop(struct amdgpu_device *adev)
2643{
25069e06 2644 WREG32(mmRLC_CNTL, 0);
2cd46ad2
KW		 2645
2646 gfx_v6_0_enable_gui_idle_interrupt(adev, false);
2cd46ad2
KW		 2647 gfx_v6_0_wait_for_rlc_serdes(adev);
2648}
2649
2650static void gfx_v6_0_rlc_start(struct amdgpu_device *adev)
2651{
25069e06 2652 WREG32(mmRLC_CNTL, RLC_CNTL__RLC_ENABLE_F32_MASK);
2cd46ad2
KW		 2653
2654 gfx_v6_0_enable_gui_idle_interrupt(adev, true);
2655
2656 udelay(50);
2657}
2658
2659static void gfx_v6_0_rlc_reset(struct amdgpu_device *adev)
2660{
25069e06 2661 WREG32_FIELD(GRBM_SOFT_RESET, SOFT_RESET_RLC, 1);
2cd46ad2 2662 udelay(50);
25069e06 2663 WREG32_FIELD(GRBM_SOFT_RESET, SOFT_RESET_RLC, 0);
2cd46ad2
KW		 2664 udelay(50);
2665}
2666
2667static bool gfx_v6_0_lbpw_supported(struct amdgpu_device *adev)
2668{
2669 u32 tmp;
2670
2671 /* Enable LBPW only for DDR3 */
25069e06 2672 tmp = RREG32(mmMC_SEQ_MISC0);
2cd46ad2
KW		 2673 if ((tmp & 0xF0000000) == 0xB0000000)
2674 return true;
2675 return false;
2676}
25069e06 2677
2cd46ad2
KW		 2678static void gfx_v6_0_init_cg(struct amdgpu_device *adev)
2679{
2680}
2681
2682static int gfx_v6_0_rlc_resume(struct amdgpu_device *adev)
2683{
2684 u32 i;
2685 const struct rlc_firmware_header_v1_0 *hdr;
2686 const __le32 *fw_data;
2687 u32 fw_size;
2688
2689
2690 if (!adev->gfx.rlc_fw)
2691 return -EINVAL;
2692
2693 gfx_v6_0_rlc_stop(adev);
2cd46ad2 2694 gfx_v6_0_rlc_reset(adev);
2cd46ad2 2695 gfx_v6_0_init_pg(adev);
2cd46ad2
KW		 2696 gfx_v6_0_init_cg(adev);
2697
25069e06
TSD		 2698 WREG32(mmRLC_RL_BASE, 0);
2699 WREG32(mmRLC_RL_SIZE, 0);
2700 WREG32(mmRLC_LB_CNTL, 0);
2701 WREG32(mmRLC_LB_CNTR_MAX, 0xffffffff);
2702 WREG32(mmRLC_LB_CNTR_INIT, 0);
2703 WREG32(mmRLC_LB_INIT_CU_MASK, 0xffffffff);
2cd46ad2 2704
25069e06
TSD		 2705 WREG32(mmRLC_MC_CNTL, 0);
2706 WREG32(mmRLC_UCODE_CNTL, 0);
2cd46ad2
KW		 2707
2708 hdr = (const struct rlc_firmware_header_v1_0 *)adev->gfx.rlc_fw->data;
2709 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
2710 fw_data = (const __le32 *)
2711 (adev->gfx.rlc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2712
2713 amdgpu_ucode_print_rlc_hdr(&hdr->header);
2714
2715 for (i = 0; i < fw_size; i++) {
25069e06
TSD		 2716 WREG32(mmRLC_UCODE_ADDR, i);
2717 WREG32(mmRLC_UCODE_DATA, le32_to_cpup(fw_data++));
2cd46ad2 2718 }
25069e06 2719 WREG32(mmRLC_UCODE_ADDR, 0);
2cd46ad2
KW		 2720
2721 gfx_v6_0_enable_lbpw(adev, gfx_v6_0_lbpw_supported(adev));
2cd46ad2
KW		 2722 gfx_v6_0_rlc_start(adev);
2723
2724 return 0;
2725}
2726
2727static void gfx_v6_0_enable_cgcg(struct amdgpu_device *adev, bool enable)
2728{
2729 u32 data, orig, tmp;
2730
25069e06 2731 orig = data = RREG32(mmRLC_CGCG_CGLS_CTRL);
2cd46ad2
KW		 2732
2733 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG)) {
2734 gfx_v6_0_enable_gui_idle_interrupt(adev, true);
2735
25069e06 2736 WREG32(mmRLC_GCPM_GENERAL_3, 0x00000080);
2cd46ad2
KW		 2737
2738 tmp = gfx_v6_0_halt_rlc(adev);
2739
25069e06
TSD		 2740 WREG32(mmRLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
2741 WREG32(mmRLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
2742 WREG32(mmRLC_SERDES_WR_CTRL, 0x00b000ff);
2cd46ad2
KW		 2743
2744 gfx_v6_0_wait_for_rlc_serdes(adev);
2cd46ad2
KW		 2745 gfx_v6_0_update_rlc(adev, tmp);
2746
25069e06 2747 WREG32(mmRLC_SERDES_WR_CTRL, 0x007000ff);
2cd46ad2 2748
25069e06 2749 data |= RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK | RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK;
2cd46ad2
KW		 2750 } else {
2751 gfx_v6_0_enable_gui_idle_interrupt(adev, false);
2752
25069e06
TSD		 2753 RREG32(mmCB_CGTT_SCLK_CTRL);
2754 RREG32(mmCB_CGTT_SCLK_CTRL);
2755 RREG32(mmCB_CGTT_SCLK_CTRL);
2756 RREG32(mmCB_CGTT_SCLK_CTRL);
2cd46ad2 2757
25069e06 2758 data &= ~(RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK | RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK);
2cd46ad2
KW		 2759 }
2760
2761 if (orig != data)
25069e06 2762 WREG32(mmRLC_CGCG_CGLS_CTRL, data);
2cd46ad2
KW		 2763
2764}
2765
2766static void gfx_v6_0_enable_mgcg(struct amdgpu_device *adev, bool enable)
2767{
2768
2769 u32 data, orig, tmp = 0;
2770
2771 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG)) {
25069e06 2772 orig = data = RREG32(mmCGTS_SM_CTRL_REG);
2cd46ad2
KW		 2773 data = 0x96940200;
2774 if (orig != data)
25069e06 2775 WREG32(mmCGTS_SM_CTRL_REG, data);
2cd46ad2
KW		 2776
2777 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CP_LS) {
25069e06
TSD		 2778 orig = data = RREG32(mmCP_MEM_SLP_CNTL);
2779 data |= CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
2cd46ad2 2780 if (orig != data)
25069e06 2781 WREG32(mmCP_MEM_SLP_CNTL, data);
2cd46ad2
KW		 2782 }
2783
25069e06 2784 orig = data = RREG32(mmRLC_CGTT_MGCG_OVERRIDE);
2cd46ad2
KW		 2785 data &= 0xffffffc0;
2786 if (orig != data)
25069e06 2787 WREG32(mmRLC_CGTT_MGCG_OVERRIDE, data);
2cd46ad2
KW		 2788
2789 tmp = gfx_v6_0_halt_rlc(adev);
2790
25069e06
TSD		 2791 WREG32(mmRLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
2792 WREG32(mmRLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
2793 WREG32(mmRLC_SERDES_WR_CTRL, 0x00d000ff);
2cd46ad2
KW		 2794
2795 gfx_v6_0_update_rlc(adev, tmp);
2796 } else {
25069e06 2797 orig = data = RREG32(mmRLC_CGTT_MGCG_OVERRIDE);
2cd46ad2
KW		 2798 data |= 0x00000003;
2799 if (orig != data)
25069e06 2800 WREG32(mmRLC_CGTT_MGCG_OVERRIDE, data);
2cd46ad2 2801
25069e06
TSD		 2802 data = RREG32(mmCP_MEM_SLP_CNTL);
2803 if (data & CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK) {
2804 data &= ~CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
2805 WREG32(mmCP_MEM_SLP_CNTL, data);
2cd46ad2 2806 }
25069e06
TSD		 2807 orig = data = RREG32(mmCGTS_SM_CTRL_REG);
2808 data |= CGTS_SM_CTRL_REG__LS_OVERRIDE_MASK | CGTS_SM_CTRL_REG__OVERRIDE_MASK;
2cd46ad2 2809 if (orig != data)
25069e06 2810 WREG32(mmCGTS_SM_CTRL_REG, data);
2cd46ad2
KW		 2811
2812 tmp = gfx_v6_0_halt_rlc(adev);
2813
25069e06
TSD		 2814 WREG32(mmRLC_SERDES_WR_MASTER_MASK_0, 0xffffffff);
2815 WREG32(mmRLC_SERDES_WR_MASTER_MASK_1, 0xffffffff);
2816 WREG32(mmRLC_SERDES_WR_CTRL, 0x00e000ff);
2cd46ad2
KW		 2817
2818 gfx_v6_0_update_rlc(adev, tmp);
2819 }
2820}
2821/*
2822static void gfx_v6_0_update_cg(struct amdgpu_device *adev,
2823 bool enable)
2824{
2825 gfx_v6_0_enable_gui_idle_interrupt(adev, false);
2826 if (enable) {
2827 gfx_v6_0_enable_mgcg(adev, true);
2828 gfx_v6_0_enable_cgcg(adev, true);
2829 } else {
2830 gfx_v6_0_enable_cgcg(adev, false);
2831 gfx_v6_0_enable_mgcg(adev, false);
2832 }
2833 gfx_v6_0_enable_gui_idle_interrupt(adev, true);
2834}
2835*/
25069e06 2836
2cd46ad2
KW		 2837static void gfx_v6_0_enable_sclk_slowdown_on_pu(struct amdgpu_device *adev,
2838 bool enable)
2839{
2840}
2841
2842static void gfx_v6_0_enable_sclk_slowdown_on_pd(struct amdgpu_device *adev,
2843 bool enable)
2844{
2845}
2846
2847static void gfx_v6_0_enable_cp_pg(struct amdgpu_device *adev, bool enable)
2848{
2849 u32 data, orig;
2850
25069e06 2851 orig = data = RREG32(mmRLC_PG_CNTL);
2cd46ad2
KW		 2852 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_CP))
2853 data &= ~0x8000;
2854 else
2855 data |= 0x8000;
2856 if (orig != data)
25069e06 2857 WREG32(mmRLC_PG_CNTL, data);
2cd46ad2
KW		 2858}
2859
2860static void gfx_v6_0_enable_gds_pg(struct amdgpu_device *adev, bool enable)
2861{
2862}
2863/*
2864static void gfx_v6_0_init_cp_pg_table(struct amdgpu_device *adev)
2865{
2866 const __le32 *fw_data;
2867 volatile u32 *dst_ptr;
2868 int me, i, max_me = 4;
2869 u32 bo_offset = 0;
2870 u32 table_offset, table_size;
2871
2872 if (adev->asic_type == CHIP_KAVERI)
2873 max_me = 5;
2874
2875 if (adev->gfx.rlc.cp_table_ptr == NULL)
2876 return;
2877
2878 dst_ptr = adev->gfx.rlc.cp_table_ptr;
2879 for (me = 0; me < max_me; me++) {
2880 if (me == 0) {
2881 const struct gfx_firmware_header_v1_0 *hdr =
2882 (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
2883 fw_data = (const __le32 *)
2884 (adev->gfx.ce_fw->data +
2885 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2886 table_offset = le32_to_cpu(hdr->jt_offset);
2887 table_size = le32_to_cpu(hdr->jt_size);
2888 } else if (me == 1) {
2889 const struct gfx_firmware_header_v1_0 *hdr =
2890 (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
2891 fw_data = (const __le32 *)
2892 (adev->gfx.pfp_fw->data +
2893 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2894 table_offset = le32_to_cpu(hdr->jt_offset);
2895 table_size = le32_to_cpu(hdr->jt_size);
2896 } else if (me == 2) {
2897 const struct gfx_firmware_header_v1_0 *hdr =
2898 (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
2899 fw_data = (const __le32 *)
2900 (adev->gfx.me_fw->data +
2901 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2902 table_offset = le32_to_cpu(hdr->jt_offset);
2903 table_size = le32_to_cpu(hdr->jt_size);
2904 } else if (me == 3) {
2905 const struct gfx_firmware_header_v1_0 *hdr =
2906 (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
2907 fw_data = (const __le32 *)
2908 (adev->gfx.mec_fw->data +
2909 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2910 table_offset = le32_to_cpu(hdr->jt_offset);
2911 table_size = le32_to_cpu(hdr->jt_size);
2912 } else {
2913 const struct gfx_firmware_header_v1_0 *hdr =
2914 (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec2_fw->data;
2915 fw_data = (const __le32 *)
2916 (adev->gfx.mec2_fw->data +
2917 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2918 table_offset = le32_to_cpu(hdr->jt_offset);
2919 table_size = le32_to_cpu(hdr->jt_size);
2920 }
2921
2922 for (i = 0; i < table_size; i ++) {
2923 dst_ptr[bo_offset + i] =
2924 cpu_to_le32(le32_to_cpu(fw_data[table_offset + i]));
2925 }
2926
2927 bo_offset += table_size;
2928 }
2929}
2930*/
2931static void gfx_v6_0_enable_gfx_cgpg(struct amdgpu_device *adev,
2932 bool enable)
2933{
2cd46ad2 2934 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_GFX_PG)) {
25069e06
TSD		 2935 WREG32(mmRLC_TTOP_D, RLC_PUD(0x10) | RLC_PDD(0x10) | RLC_TTPD(0x10) | RLC_MSD(0x10));
2936 WREG32_FIELD(RLC_PG_CNTL, GFX_POWER_GATING_ENABLE, 1);
2937 WREG32_FIELD(RLC_AUTO_PG_CTRL, AUTO_PG_EN, 1);
2cd46ad2 2938 } else {
25069e06
TSD		 2939 WREG32_FIELD(RLC_AUTO_PG_CTRL, AUTO_PG_EN, 0);
2940 (void)RREG32(mmDB_RENDER_CONTROL);
2cd46ad2
KW		 2941 }
2942}
2943
2944static u32 gfx_v6_0_get_cu_active_bitmap(struct amdgpu_device *adev,
2945 u32 se, u32 sh)
2946{
2947
2948 u32 mask = 0, tmp, tmp1;
2949 int i;
2950
deca1d1f 2951 mutex_lock(&adev->grbm_idx_mutex);
2cd46ad2 2952 gfx_v6_0_select_se_sh(adev, se, sh, 0xffffffff);
25069e06
TSD		 2953 tmp = RREG32(mmCC_GC_SHADER_ARRAY_CONFIG);
2954 tmp1 = RREG32(mmGC_USER_SHADER_ARRAY_CONFIG);
2cd46ad2 2955 gfx_v6_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
deca1d1f 2956 mutex_unlock(&adev->grbm_idx_mutex);
2cd46ad2
KW		 2957
2958 tmp &= 0xffff0000;
2959
2960 tmp |= tmp1;
2961 tmp >>= 16;
2962
2963 for (i = 0; i < adev->gfx.config.max_cu_per_sh; i ++) {
2964 mask <<= 1;
2965 mask |= 1;
2966 }
2967
2968 return (~tmp) & mask;
2969}
2970
2971static void gfx_v6_0_init_ao_cu_mask(struct amdgpu_device *adev)
2972{
2973 u32 i, j, k, active_cu_number = 0;
2974
2975 u32 mask, counter, cu_bitmap;
2976 u32 tmp = 0;
2977
2978 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
2979 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
2980 mask = 1;
2981 cu_bitmap = 0;
2982 counter = 0;
2983 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k++) {
2984 if (gfx_v6_0_get_cu_active_bitmap(adev, i, j) & mask) {
2985 if (counter < 2)
2986 cu_bitmap |= mask;
2987 counter++;
2988 }
2989 mask <<= 1;
2990 }
2991
2992 active_cu_number += counter;
2993 tmp |= (cu_bitmap << (i * 16 + j * 8));
2994 }
2995 }
2996
25069e06
TSD		 2997 WREG32(mmRLC_PG_AO_CU_MASK, tmp);
2998 WREG32_FIELD(RLC_MAX_PG_CU, MAX_POWERED_UP_CU, active_cu_number);
2cd46ad2
KW		 2999}
3000
3001static void gfx_v6_0_enable_gfx_static_mgpg(struct amdgpu_device *adev,
3002 bool enable)
3003{
3004 u32 data, orig;
3005
25069e06 3006 orig = data = RREG32(mmRLC_PG_CNTL);
2cd46ad2 3007 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_GFX_SMG))
25069e06 3008 data |= RLC_PG_CNTL__STATIC_PER_CU_PG_ENABLE_MASK;
2cd46ad2 3009 else
25069e06 3010 data &= ~RLC_PG_CNTL__STATIC_PER_CU_PG_ENABLE_MASK;
2cd46ad2 3011 if (orig != data)
25069e06 3012 WREG32(mmRLC_PG_CNTL, data);
2cd46ad2
KW		 3013}
3014
3015static void gfx_v6_0_enable_gfx_dynamic_mgpg(struct amdgpu_device *adev,
3016 bool enable)
3017{
3018 u32 data, orig;
3019
25069e06 3020 orig = data = RREG32(mmRLC_PG_CNTL);
2cd46ad2 3021 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_GFX_DMG))
25069e06 3022 data |= RLC_PG_CNTL__DYN_PER_CU_PG_ENABLE_MASK;
2cd46ad2 3023 else
25069e06 3024 data &= ~RLC_PG_CNTL__DYN_PER_CU_PG_ENABLE_MASK;
2cd46ad2 3025 if (orig != data)
25069e06 3026 WREG32(mmRLC_PG_CNTL, data);
2cd46ad2
KW		 3027}
3028
3029static void gfx_v6_0_init_gfx_cgpg(struct amdgpu_device *adev)
3030{
3031 u32 tmp;
3032
25069e06
TSD		 3033 WREG32(mmRLC_SAVE_AND_RESTORE_BASE, adev->gfx.rlc.save_restore_gpu_addr >> 8);
3034 WREG32_FIELD(RLC_PG_CNTL, GFX_POWER_GATING_SRC, 1);
3035 WREG32(mmRLC_CLEAR_STATE_RESTORE_BASE, adev->gfx.rlc.clear_state_gpu_addr >> 8);
2cd46ad2 3036
25069e06
TSD		 3037 tmp = RREG32(mmRLC_AUTO_PG_CTRL);
3038 tmp &= ~RLC_AUTO_PG_CTRL__GRBM_REG_SAVE_GFX_IDLE_THRESHOLD_MASK;
3039 tmp |= (0x700 << RLC_AUTO_PG_CTRL__GRBM_REG_SAVE_GFX_IDLE_THRESHOLD__SHIFT);
3040 tmp &= ~RLC_AUTO_PG_CTRL__PG_AFTER_GRBM_REG_SAVE_THRESHOLD_MASK;
3041 WREG32(mmRLC_AUTO_PG_CTRL, tmp);
2cd46ad2
KW		 3042}
3043
3044static void gfx_v6_0_update_gfx_pg(struct amdgpu_device *adev, bool enable)
3045{
3046 gfx_v6_0_enable_gfx_cgpg(adev, enable);
3047 gfx_v6_0_enable_gfx_static_mgpg(adev, enable);
3048 gfx_v6_0_enable_gfx_dynamic_mgpg(adev, enable);
3049}
3050
3051static u32 gfx_v6_0_get_csb_size(struct amdgpu_device *adev)
3052{
3053 u32 count = 0;
3054 const struct cs_section_def *sect = NULL;
3055 const struct cs_extent_def *ext = NULL;
3056
3057 if (adev->gfx.rlc.cs_data == NULL)
3058 return 0;
3059
3060 /* begin clear state */
3061 count += 2;
3062 /* context control state */
3063 count += 3;
3064
3065 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
3066 for (ext = sect->section; ext->extent != NULL; ++ext) {
3067 if (sect->id == SECT_CONTEXT)
3068 count += 2 + ext->reg_count;
3069 else
3070 return 0;
3071 }
3072 }
3073 /* pa_sc_raster_config */
3074 count += 3;
3075 /* end clear state */
3076 count += 2;
3077 /* clear state */
3078 count += 2;
3079
3080 return count;
3081}
3082
3083static void gfx_v6_0_get_csb_buffer(struct amdgpu_device *adev,
3084 volatile u32 *buffer)
3085{
3086 u32 count = 0, i;
3087 const struct cs_section_def *sect = NULL;
3088 const struct cs_extent_def *ext = NULL;
3089
3090 if (adev->gfx.rlc.cs_data == NULL)
3091 return;
3092 if (buffer == NULL)
3093 return;
3094
3095 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3096 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
2cd46ad2
KW		 3097 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
3098 buffer[count++] = cpu_to_le32(0x80000000);
3099 buffer[count++] = cpu_to_le32(0x80000000);
3100
3101 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
3102 for (ext = sect->section; ext->extent != NULL; ++ext) {
3103 if (sect->id == SECT_CONTEXT) {
3104 buffer[count++] =
3105 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
3106 buffer[count++] = cpu_to_le32(ext->reg_index - 0xa000);
3107 for (i = 0; i < ext->reg_count; i++)
3108 buffer[count++] = cpu_to_le32(ext->extent[i]);
3109 } else {
3110 return;
3111 }
3112 }
3113 }
3114
3115 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 1));
25069e06 3116 buffer[count++] = cpu_to_le32(mmPA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
2cd46ad2
KW		 3117
3118 switch (adev->asic_type) {
3119 case CHIP_TAHITI:
3120 case CHIP_PITCAIRN:
3121 buffer[count++] = cpu_to_le32(0x2a00126a);
3122 break;
3123 case CHIP_VERDE:
3124 buffer[count++] = cpu_to_le32(0x0000124a);
3125 break;
3126 case CHIP_OLAND:
3127 buffer[count++] = cpu_to_le32(0x00000082);
3128 break;
3129 case CHIP_HAINAN:
3130 buffer[count++] = cpu_to_le32(0x00000000);
3131 break;
3132 default:
3133 buffer[count++] = cpu_to_le32(0x00000000);
3134 break;
3135 }
3136
3137 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
3138 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
3139
3140 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
3141 buffer[count++] = cpu_to_le32(0);
3142}
3143
3144static void gfx_v6_0_init_pg(struct amdgpu_device *adev)
3145{
3146 if (adev->pg_flags & (AMD_PG_SUPPORT_GFX_PG |
3147 AMD_PG_SUPPORT_GFX_SMG |
3148 AMD_PG_SUPPORT_GFX_DMG |
3149 AMD_PG_SUPPORT_CP |
3150 AMD_PG_SUPPORT_GDS |
3151 AMD_PG_SUPPORT_RLC_SMU_HS)) {
3152 gfx_v6_0_enable_sclk_slowdown_on_pu(adev, true);
3153 gfx_v6_0_enable_sclk_slowdown_on_pd(adev, true);
3154 if (adev->pg_flags & AMD_PG_SUPPORT_GFX_PG) {
3155 gfx_v6_0_init_gfx_cgpg(adev);
3156 gfx_v6_0_enable_cp_pg(adev, true);
3157 gfx_v6_0_enable_gds_pg(adev, true);
3158 } else {
25069e06
TSD		 3159 WREG32(mmRLC_SAVE_AND_RESTORE_BASE, adev->gfx.rlc.save_restore_gpu_addr >> 8);
3160 WREG32(mmRLC_CLEAR_STATE_RESTORE_BASE, adev->gfx.rlc.clear_state_gpu_addr >> 8);
2cd46ad2
KW		 3161
3162 }
3163 gfx_v6_0_init_ao_cu_mask(adev);
3164 gfx_v6_0_update_gfx_pg(adev, true);
3165 } else {
3166
25069e06
TSD		 3167 WREG32(mmRLC_SAVE_AND_RESTORE_BASE, adev->gfx.rlc.save_restore_gpu_addr >> 8);
3168 WREG32(mmRLC_CLEAR_STATE_RESTORE_BASE, adev->gfx.rlc.clear_state_gpu_addr >> 8);
2cd46ad2
KW		 3169 }
3170}
3171
3172static void gfx_v6_0_fini_pg(struct amdgpu_device *adev)
3173{
3174 if (adev->pg_flags & (AMD_PG_SUPPORT_GFX_PG |
3175 AMD_PG_SUPPORT_GFX_SMG |
3176 AMD_PG_SUPPORT_GFX_DMG |
3177 AMD_PG_SUPPORT_CP |
3178 AMD_PG_SUPPORT_GDS |
3179 AMD_PG_SUPPORT_RLC_SMU_HS)) {
3180 gfx_v6_0_update_gfx_pg(adev, false);
3181 if (adev->pg_flags & AMD_PG_SUPPORT_GFX_PG) {
3182 gfx_v6_0_enable_cp_pg(adev, false);
3183 gfx_v6_0_enable_gds_pg(adev, false);
3184 }
3185 }
3186}
3187
3188static uint64_t gfx_v6_0_get_gpu_clock_counter(struct amdgpu_device *adev)
3189{
3190 uint64_t clock;
3191
3192 mutex_lock(&adev->gfx.gpu_clock_mutex);
25069e06
TSD		 3193 WREG32(mmRLC_CAPTURE_GPU_CLOCK_COUNT, 1);
3194 clock = (uint64_t)RREG32(mmRLC_GPU_CLOCK_COUNT_LSB) |
3195 ((uint64_t)RREG32(mmRLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
2cd46ad2
KW		 3196 mutex_unlock(&adev->gfx.gpu_clock_mutex);
3197 return clock;
3198}
3199
0f444c24
AD		 3200static void gfx_v6_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)
3201{
45682886
ML		 3202 if (flags & AMDGPU_HAVE_CTX_SWITCH)
3203 gfx_v6_0_ring_emit_vgt_flush(ring);
0f444c24
AD		 3204 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
3205 amdgpu_ring_write(ring, 0x80000000);
3206 amdgpu_ring_write(ring, 0);
3207}
3208
3ee73ed8
TSD		 3209
3210static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t address)
3211{
3212 WREG32(mmSQ_IND_INDEX,
3213 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
3214 (simd << SQ_IND_INDEX__SIMD_ID__SHIFT) |
3215 (address << SQ_IND_INDEX__INDEX__SHIFT) |
3216 (SQ_IND_INDEX__FORCE_READ_MASK));
3217 return RREG32(mmSQ_IND_DATA);
3218}
3219
34e646f4
TSD		 3220static void wave_read_regs(struct amdgpu_device *adev, uint32_t simd,
3221 uint32_t wave, uint32_t thread,
3222 uint32_t regno, uint32_t num, uint32_t *out)
3223{
3224 WREG32(mmSQ_IND_INDEX,
3225 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
3226 (simd << SQ_IND_INDEX__SIMD_ID__SHIFT) |
3227 (regno << SQ_IND_INDEX__INDEX__SHIFT) |
3228 (thread << SQ_IND_INDEX__THREAD_ID__SHIFT) |
3229 (SQ_IND_INDEX__FORCE_READ_MASK) |
3230 (SQ_IND_INDEX__AUTO_INCR_MASK));
3231 while (num--)
3232 *(out++) = RREG32(mmSQ_IND_DATA);
3233}
3234
3ee73ed8
TSD		 3235static void gfx_v6_0_read_wave_data(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t *dst, int *no_fields)
3236{
3237 /* type 0 wave data */
3238 dst[(*no_fields)++] = 0;
3239 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_STATUS);
3240 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_PC_LO);
3241 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_PC_HI);
3242 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_EXEC_LO);
3243 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_EXEC_HI);
3244 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_HW_ID);
3245 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_INST_DW0);
3246 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_INST_DW1);
3247 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_GPR_ALLOC);
3248 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_LDS_ALLOC);
3249 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TRAPSTS);
3250 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_IB_STS);
3251 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TBA_LO);
3252 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TBA_HI);
3253 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TMA_LO);
3254 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TMA_HI);
3255 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_IB_DBG0);
3256 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_M0);
3257}
3258
34e646f4
TSD		 3259static void gfx_v6_0_read_wave_sgprs(struct amdgpu_device *adev, uint32_t simd,
3260 uint32_t wave, uint32_t start,
3261 uint32_t size, uint32_t *dst)
3262{
3263 wave_read_regs(
3264 adev, simd, wave, 0,
3265 start + SQIND_WAVE_SGPRS_OFFSET, size, dst);
3266}
3267
2cd46ad2
KW		 3268static const struct amdgpu_gfx_funcs gfx_v6_0_gfx_funcs = {
3269 .get_gpu_clock_counter = &gfx_v6_0_get_gpu_clock_counter,
3270 .select_se_sh = &gfx_v6_0_select_se_sh,
3ee73ed8 3271 .read_wave_data = &gfx_v6_0_read_wave_data,
34e646f4 3272 .read_wave_sgprs = &gfx_v6_0_read_wave_sgprs,
2cd46ad2
KW		 3273};
3274
3275static int gfx_v6_0_early_init(void *handle)
3276{
3277 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3278
3279 adev->gfx.num_gfx_rings = GFX6_NUM_GFX_RINGS;
3280 adev->gfx.num_compute_rings = GFX6_NUM_COMPUTE_RINGS;
3281 adev->gfx.funcs = &gfx_v6_0_gfx_funcs;
3282 gfx_v6_0_set_ring_funcs(adev);
3283 gfx_v6_0_set_irq_funcs(adev);
3284
3285 return 0;
3286}
3287
3288static int gfx_v6_0_sw_init(void *handle)
3289{
3290 struct amdgpu_ring *ring;
3291 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3292 int i, r;
3293
3294 r = amdgpu_irq_add_id(adev, 181, &adev->gfx.eop_irq);
3295 if (r)
3296 return r;
3297
3298 r = amdgpu_irq_add_id(adev, 184, &adev->gfx.priv_reg_irq);
3299 if (r)
3300 return r;
3301
3302 r = amdgpu_irq_add_id(adev, 185, &adev->gfx.priv_inst_irq);
3303 if (r)
3304 return r;
3305
3306 gfx_v6_0_scratch_init(adev);
3307
3308 r = gfx_v6_0_init_microcode(adev);
3309 if (r) {
3310 DRM_ERROR("Failed to load gfx firmware!\n");
3311 return r;
3312 }
3313
3314 r = gfx_v6_0_rlc_init(adev);
3315 if (r) {
3316 DRM_ERROR("Failed to init rlc BOs!\n");
3317 return r;
3318 }
3319
3320 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
3321 ring = &adev->gfx.gfx_ring[i];
3322 ring->ring_obj = NULL;
3323 sprintf(ring->name, "gfx");
3324 r = amdgpu_ring_init(adev, ring, 1024,
21cd942e 3325 &adev->gfx.eop_irq, AMDGPU_CP_IRQ_GFX_EOP);
2cd46ad2
KW		 3326 if (r)
3327 return r;
3328 }
3329
3330 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
3331 unsigned irq_type;
3332
3333 if ((i >= 32) || (i >= AMDGPU_MAX_COMPUTE_RINGS)) {
3334 DRM_ERROR("Too many (%d) compute rings!\n", i);
3335 break;
3336 }
3337 ring = &adev->gfx.compute_ring[i];
3338 ring->ring_obj = NULL;
3339 ring->use_doorbell = false;
3340 ring->doorbell_index = 0;
3341 ring->me = 1;
3342 ring->pipe = i;
3343 ring->queue = i;
3344 sprintf(ring->name, "comp %d.%d.%d", ring->me, ring->pipe, ring->queue);
3345 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP + ring->pipe;
2255e8c1 3346 r = amdgpu_ring_init(adev, ring, 1024,
21cd942e 3347 &adev->gfx.eop_irq, irq_type);
2cd46ad2
KW		 3348 if (r)
3349 return r;
3350 }
3351
3352 return r;
3353}
3354
3355static int gfx_v6_0_sw_fini(void *handle)
3356{
3357 int i;
3358 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3359
3360 amdgpu_bo_unref(&adev->gds.oa_gfx_bo);
3361 amdgpu_bo_unref(&adev->gds.gws_gfx_bo);
3362 amdgpu_bo_unref(&adev->gds.gds_gfx_bo);
3363
3364 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
3365 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
3366 for (i = 0; i < adev->gfx.num_compute_rings; i++)
3367 amdgpu_ring_fini(&adev->gfx.compute_ring[i]);
3368
2cd46ad2
KW		 3369 gfx_v6_0_rlc_fini(adev);
3370
3371 return 0;
3372}
3373
3374static int gfx_v6_0_hw_init(void *handle)
3375{
3376 int r;
3377 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3378
3379 gfx_v6_0_gpu_init(adev);
3380
3381 r = gfx_v6_0_rlc_resume(adev);
3382 if (r)
3383 return r;
3384
3385 r = gfx_v6_0_cp_resume(adev);
3386 if (r)
3387 return r;
3388
3389 adev->gfx.ce_ram_size = 0x8000;
3390
3391 return r;
3392}
3393
3394static int gfx_v6_0_hw_fini(void *handle)
3395{
3396 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3397
3398 gfx_v6_0_cp_enable(adev, false);
3399 gfx_v6_0_rlc_stop(adev);
3400 gfx_v6_0_fini_pg(adev);
3401
3402 return 0;
3403}
3404
3405static int gfx_v6_0_suspend(void *handle)
3406{
3407 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3408
3409 return gfx_v6_0_hw_fini(adev);
3410}
3411
3412static int gfx_v6_0_resume(void *handle)
3413{
3414 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3415
3416 return gfx_v6_0_hw_init(adev);
3417}
3418
3419static bool gfx_v6_0_is_idle(void *handle)
3420{
3421 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3422
25069e06 3423 if (RREG32(mmGRBM_STATUS) & GRBM_STATUS__GUI_ACTIVE_MASK)
2cd46ad2
KW		 3424 return false;
3425 else
3426 return true;
3427}
3428
3429static int gfx_v6_0_wait_for_idle(void *handle)
3430{
3431 unsigned i;
2cd46ad2
KW		 3432 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3433
3434 for (i = 0; i < adev->usec_timeout; i++) {
4aeacf0f 3435 if (gfx_v6_0_is_idle(handle))
2cd46ad2
KW		 3436 return 0;
3437 udelay(1);
3438 }
3439 return -ETIMEDOUT;
3440}
3441
3442static int gfx_v6_0_soft_reset(void *handle)
3443{
3444 return 0;
3445}
3446
3447static void gfx_v6_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev,
3448 enum amdgpu_interrupt_state state)
3449{
3450 u32 cp_int_cntl;
3451
3452 switch (state) {
3453 case AMDGPU_IRQ_STATE_DISABLE:
25069e06
TSD		 3454 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
3455 cp_int_cntl &= ~CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK;
3456 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
2cd46ad2
KW		 3457 break;
3458 case AMDGPU_IRQ_STATE_ENABLE:
25069e06
TSD		 3459 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
3460 cp_int_cntl |= CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK;
3461 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
2cd46ad2
KW		 3462 break;
3463 default:
3464 break;
3465 }
3466}
3467
3468static void gfx_v6_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev,
3469 int ring,
3470 enum amdgpu_interrupt_state state)
3471{
3472 u32 cp_int_cntl;
3473 switch (state){
3474 case AMDGPU_IRQ_STATE_DISABLE:
3475 if (ring == 0) {
25069e06
TSD		 3476 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING1);
3477 cp_int_cntl &= ~CP_INT_CNTL_RING1__TIME_STAMP_INT_ENABLE_MASK;
3478 WREG32(mmCP_INT_CNTL_RING1, cp_int_cntl);
2cd46ad2
KW		 3479 break;
3480 } else {
25069e06
TSD		 3481 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING2);
3482 cp_int_cntl &= ~CP_INT_CNTL_RING2__TIME_STAMP_INT_ENABLE_MASK;
3483 WREG32(mmCP_INT_CNTL_RING2, cp_int_cntl);
2cd46ad2
KW		 3484 break;
3485
3486 }
3487 case AMDGPU_IRQ_STATE_ENABLE:
3488 if (ring == 0) {
25069e06
TSD		 3489 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING1);
3490 cp_int_cntl |= CP_INT_CNTL_RING1__TIME_STAMP_INT_ENABLE_MASK;
3491 WREG32(mmCP_INT_CNTL_RING1, cp_int_cntl);
2cd46ad2
KW		 3492 break;
3493 } else {
25069e06
TSD		 3494 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING2);
3495 cp_int_cntl |= CP_INT_CNTL_RING2__TIME_STAMP_INT_ENABLE_MASK;
3496 WREG32(mmCP_INT_CNTL_RING2, cp_int_cntl);
2cd46ad2
KW		 3497 break;
3498
3499 }
3500
3501 default:
3502 BUG();
3503 break;
3504
3505 }
3506}
3507
3508static int gfx_v6_0_set_priv_reg_fault_state(struct amdgpu_device *adev,
3509 struct amdgpu_irq_src *src,
3510 unsigned type,
3511 enum amdgpu_interrupt_state state)
3512{
3513 u32 cp_int_cntl;
3514
3515 switch (state) {
3516 case AMDGPU_IRQ_STATE_DISABLE:
25069e06 3517 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
2cd46ad2 3518 cp_int_cntl &= ~CP_INT_CNTL_RING0__PRIV_REG_INT_ENABLE_MASK;
25069e06 3519 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
2cd46ad2
KW		 3520 break;
3521 case AMDGPU_IRQ_STATE_ENABLE:
25069e06 3522 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
2cd46ad2 3523 cp_int_cntl |= CP_INT_CNTL_RING0__PRIV_REG_INT_ENABLE_MASK;
25069e06 3524 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
2cd46ad2
KW		 3525 break;
3526 default:
3527 break;
3528 }
3529
3530 return 0;
3531}
3532
3533static int gfx_v6_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
3534 struct amdgpu_irq_src *src,
3535 unsigned type,
3536 enum amdgpu_interrupt_state state)
3537{
3538 u32 cp_int_cntl;
3539
3540 switch (state) {
3541 case AMDGPU_IRQ_STATE_DISABLE:
25069e06 3542 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
2cd46ad2 3543 cp_int_cntl &= ~CP_INT_CNTL_RING0__PRIV_INSTR_INT_ENABLE_MASK;
25069e06 3544 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
2cd46ad2
KW		 3545 break;
3546 case AMDGPU_IRQ_STATE_ENABLE:
25069e06 3547 cp_int_cntl = RREG32(mmCP_INT_CNTL_RING0);
2cd46ad2 3548 cp_int_cntl |= CP_INT_CNTL_RING0__PRIV_INSTR_INT_ENABLE_MASK;
25069e06 3549 WREG32(mmCP_INT_CNTL_RING0, cp_int_cntl);
2cd46ad2
KW		 3550 break;
3551 default:
3552 break;
3553 }
3554
3555 return 0;
3556}
3557
3558static int gfx_v6_0_set_eop_interrupt_state(struct amdgpu_device *adev,
3559 struct amdgpu_irq_src *src,
3560 unsigned type,
3561 enum amdgpu_interrupt_state state)
3562{
3563 switch (type) {
3564 case AMDGPU_CP_IRQ_GFX_EOP:
3565 gfx_v6_0_set_gfx_eop_interrupt_state(adev, state);
3566 break;
3567 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP:
3568 gfx_v6_0_set_compute_eop_interrupt_state(adev, 0, state);
3569 break;
3570 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP:
3571 gfx_v6_0_set_compute_eop_interrupt_state(adev, 1, state);
3572 break;
3573 default:
3574 break;
3575 }
3576 return 0;
3577}
3578
3579static int gfx_v6_0_eop_irq(struct amdgpu_device *adev,
3580 struct amdgpu_irq_src *source,
3581 struct amdgpu_iv_entry *entry)
3582{
3583 switch (entry->ring_id) {
3584 case 0:
3585 amdgpu_fence_process(&adev->gfx.gfx_ring[0]);
3586 break;
3587 case 1:
3588 case 2:
25069e06 3589 amdgpu_fence_process(&adev->gfx.compute_ring[entry->ring_id - 1]);
2cd46ad2
KW		 3590 break;
3591 default:
3592 break;
3593 }
3594 return 0;
3595}
3596
3597static int gfx_v6_0_priv_reg_irq(struct amdgpu_device *adev,
3598 struct amdgpu_irq_src *source,
3599 struct amdgpu_iv_entry *entry)
3600{
3601 DRM_ERROR("Illegal register access in command stream\n");
3602 schedule_work(&adev->reset_work);
3603 return 0;
3604}
3605
3606static int gfx_v6_0_priv_inst_irq(struct amdgpu_device *adev,
3607 struct amdgpu_irq_src *source,
3608 struct amdgpu_iv_entry *entry)
3609{
3610 DRM_ERROR("Illegal instruction in command stream\n");
3611 schedule_work(&adev->reset_work);
3612 return 0;
3613}
3614
3615static int gfx_v6_0_set_clockgating_state(void *handle,
3616 enum amd_clockgating_state state)
3617{
3618 bool gate = false;
3619 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3620
3621 if (state == AMD_CG_STATE_GATE)
3622 gate = true;
3623
3624 gfx_v6_0_enable_gui_idle_interrupt(adev, false);
3625 if (gate) {
3626 gfx_v6_0_enable_mgcg(adev, true);
3627 gfx_v6_0_enable_cgcg(adev, true);
3628 } else {
3629 gfx_v6_0_enable_cgcg(adev, false);
3630 gfx_v6_0_enable_mgcg(adev, false);
3631 }
3632 gfx_v6_0_enable_gui_idle_interrupt(adev, true);
3633
3634 return 0;
3635}
3636
3637static int gfx_v6_0_set_powergating_state(void *handle,
3638 enum amd_powergating_state state)
3639{
3640 bool gate = false;
3641 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3642
3643 if (state == AMD_PG_STATE_GATE)
3644 gate = true;
3645
3646 if (adev->pg_flags & (AMD_PG_SUPPORT_GFX_PG |
3647 AMD_PG_SUPPORT_GFX_SMG |
3648 AMD_PG_SUPPORT_GFX_DMG |
3649 AMD_PG_SUPPORT_CP |
3650 AMD_PG_SUPPORT_GDS |
3651 AMD_PG_SUPPORT_RLC_SMU_HS)) {
3652 gfx_v6_0_update_gfx_pg(adev, gate);
3653 if (adev->pg_flags & AMD_PG_SUPPORT_GFX_PG) {
3654 gfx_v6_0_enable_cp_pg(adev, gate);
3655 gfx_v6_0_enable_gds_pg(adev, gate);
3656 }
3657 }
3658
3659 return 0;
3660}
3661
a1255107 3662static const struct amd_ip_funcs gfx_v6_0_ip_funcs = {
2cd46ad2
KW		 3663 .name = "gfx_v6_0",
3664 .early_init = gfx_v6_0_early_init,
3665 .late_init = NULL,
3666 .sw_init = gfx_v6_0_sw_init,
3667 .sw_fini = gfx_v6_0_sw_fini,
3668 .hw_init = gfx_v6_0_hw_init,
3669 .hw_fini = gfx_v6_0_hw_fini,
3670 .suspend = gfx_v6_0_suspend,
3671 .resume = gfx_v6_0_resume,
3672 .is_idle = gfx_v6_0_is_idle,
3673 .wait_for_idle = gfx_v6_0_wait_for_idle,
3674 .soft_reset = gfx_v6_0_soft_reset,
3675 .set_clockgating_state = gfx_v6_0_set_clockgating_state,
3676 .set_powergating_state = gfx_v6_0_set_powergating_state,
3677};
3678
3679static const struct amdgpu_ring_funcs gfx_v6_0_ring_funcs_gfx = {
21cd942e 3680 .type = AMDGPU_RING_TYPE_GFX,
79887142
CK		 3681 .align_mask = 0xff,
3682 .nop = 0x80000000,
6f924e20 3683 .get_rptr = gfx_v6_0_ring_get_rptr,
832c6ef7 3684 .get_wptr = gfx_v6_0_ring_get_wptr,
2cd46ad2 3685 .set_wptr = gfx_v6_0_ring_set_wptr_gfx,
e12f3d7a
CK		 3686 .emit_frame_size =
3687 5 + /* gfx_v6_0_ring_emit_hdp_flush */
3688 5 + /* gfx_v6_0_ring_emit_hdp_invalidate */
3689 14 + 14 + 14 + /* gfx_v6_0_ring_emit_fence x3 for user fence, vm fence */
3690 7 + 4 + /* gfx_v6_0_ring_emit_pipeline_sync */
3691 17 + 6 + /* gfx_v6_0_ring_emit_vm_flush */
45682886 3692 3 + 2, /* gfx_v6_ring_emit_cntxcntl including vgt flush */
e12f3d7a 3693 .emit_ib_size = 6, /* gfx_v6_0_ring_emit_ib */
668f52c3
AD		 3694 .emit_ib = gfx_v6_0_ring_emit_ib,
3695 .emit_fence = gfx_v6_0_ring_emit_fence,
2cd46ad2
KW		 3696 .emit_pipeline_sync = gfx_v6_0_ring_emit_pipeline_sync,
3697 .emit_vm_flush = gfx_v6_0_ring_emit_vm_flush,
2cd46ad2
KW		 3698 .emit_hdp_flush = gfx_v6_0_ring_emit_hdp_flush,
3699 .emit_hdp_invalidate = gfx_v6_0_ring_emit_hdp_invalidate,
3700 .test_ring = gfx_v6_0_ring_test_ring,
3701 .test_ib = gfx_v6_0_ring_test_ib,
3702 .insert_nop = amdgpu_ring_insert_nop,
0f444c24 3703 .emit_cntxcntl = gfx_v6_ring_emit_cntxcntl,
2cd46ad2
KW		 3704};
3705
3706static const struct amdgpu_ring_funcs gfx_v6_0_ring_funcs_compute = {
21cd942e 3707 .type = AMDGPU_RING_TYPE_COMPUTE,
79887142
CK		 3708 .align_mask = 0xff,
3709 .nop = 0x80000000,
6f924e20 3710 .get_rptr = gfx_v6_0_ring_get_rptr,
832c6ef7 3711 .get_wptr = gfx_v6_0_ring_get_wptr,
2cd46ad2 3712 .set_wptr = gfx_v6_0_ring_set_wptr_compute,
e12f3d7a
CK		 3713 .emit_frame_size =
3714 5 + /* gfx_v6_0_ring_emit_hdp_flush */
3715 5 + /* gfx_v6_0_ring_emit_hdp_invalidate */
3716 7 + /* gfx_v6_0_ring_emit_pipeline_sync */
3717 17 + /* gfx_v6_0_ring_emit_vm_flush */
3718 14 + 14 + 14, /* gfx_v6_0_ring_emit_fence x3 for user fence, vm fence */
3719 .emit_ib_size = 6, /* gfx_v6_0_ring_emit_ib */
668f52c3
AD		 3720 .emit_ib = gfx_v6_0_ring_emit_ib,
3721 .emit_fence = gfx_v6_0_ring_emit_fence,
2cd46ad2
KW		 3722 .emit_pipeline_sync = gfx_v6_0_ring_emit_pipeline_sync,
3723 .emit_vm_flush = gfx_v6_0_ring_emit_vm_flush,
2cd46ad2
KW		 3724 .emit_hdp_flush = gfx_v6_0_ring_emit_hdp_flush,
3725 .emit_hdp_invalidate = gfx_v6_0_ring_emit_hdp_invalidate,
3726 .test_ring = gfx_v6_0_ring_test_ring,
3727 .test_ib = gfx_v6_0_ring_test_ib,
3728 .insert_nop = amdgpu_ring_insert_nop,
3729};
3730
3731static void gfx_v6_0_set_ring_funcs(struct amdgpu_device *adev)
3732{
3733 int i;
3734
3735 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
3736 adev->gfx.gfx_ring[i].funcs = &gfx_v6_0_ring_funcs_gfx;
3737 for (i = 0; i < adev->gfx.num_compute_rings; i++)
3738 adev->gfx.compute_ring[i].funcs = &gfx_v6_0_ring_funcs_compute;
3739}
3740
3741static const struct amdgpu_irq_src_funcs gfx_v6_0_eop_irq_funcs = {
3742 .set = gfx_v6_0_set_eop_interrupt_state,
3743 .process = gfx_v6_0_eop_irq,
3744};
3745
3746static const struct amdgpu_irq_src_funcs gfx_v6_0_priv_reg_irq_funcs = {
3747 .set = gfx_v6_0_set_priv_reg_fault_state,
3748 .process = gfx_v6_0_priv_reg_irq,
3749};
3750
3751static const struct amdgpu_irq_src_funcs gfx_v6_0_priv_inst_irq_funcs = {
3752 .set = gfx_v6_0_set_priv_inst_fault_state,
3753 .process = gfx_v6_0_priv_inst_irq,
3754};
3755
3756static void gfx_v6_0_set_irq_funcs(struct amdgpu_device *adev)
3757{
3758 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST;
3759 adev->gfx.eop_irq.funcs = &gfx_v6_0_eop_irq_funcs;
3760
3761 adev->gfx.priv_reg_irq.num_types = 1;
3762 adev->gfx.priv_reg_irq.funcs = &gfx_v6_0_priv_reg_irq_funcs;
3763
3764 adev->gfx.priv_inst_irq.num_types = 1;
3765 adev->gfx.priv_inst_irq.funcs = &gfx_v6_0_priv_inst_irq_funcs;
3766}
3767
3768static void gfx_v6_0_get_cu_info(struct amdgpu_device *adev)
3769{
3770 int i, j, k, counter, active_cu_number = 0;
3771 u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
3772 struct amdgpu_cu_info *cu_info = &adev->gfx.cu_info;
3773
3774 memset(cu_info, 0, sizeof(*cu_info));
3775
2cd46ad2
KW		 3776 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
3777 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
3778 mask = 1;
3779 ao_bitmap = 0;
3780 counter = 0;
2cd46ad2
KW		 3781 bitmap = gfx_v6_0_get_cu_active_bitmap(adev, i, j);
3782 cu_info->bitmap[i][j] = bitmap;
3783
3784 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k ++) {
3785 if (bitmap & mask) {
3786 if (counter < 2)
3787 ao_bitmap |= mask;
3788 counter ++;
3789 }
3790 mask <<= 1;
3791 }
3792 active_cu_number += counter;
3793 ao_cu_mask |= (ao_bitmap << (i * 16 + j * 8));
3794 }
3795 }
2cd46ad2
KW		 3796
3797 cu_info->number = active_cu_number;
3798 cu_info->ao_cu_mask = ao_cu_mask;
3799}
a1255107
AD		 3800
3801const struct amdgpu_ip_block_version gfx_v6_0_ip_block =
3802{
3803 .type = AMD_IP_BLOCK_TYPE_GFX,
3804 .major = 6,
3805 .minor = 0,
3806 .rev = 0,
3807 .funcs = &gfx_v6_0_ip_funcs,
3808};
Ubuntu Jammy Linux kernel mirror