]> bbs.cooldavid.org Git - net-next-2.6.git/blame - drivers/gpu/drm/radeon/rv770.c
git://bbs.cooldavid.org / net-next-2.6.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
drm/radeon/kms: avoid corner case issue with unmappable vram V2
[net-next-2.6.git] / drivers / gpu / drm / radeon / rv770.c
CommitLineData
771fe6b9
JG		 1/*
2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 * Authors: Dave Airlie
25 * Alex Deucher
26 * Jerome Glisse
27 */
3ce0a23d
JG		 28#include <linux/firmware.h>
29#include <linux/platform_device.h>
5a0e3ad6 30#include <linux/slab.h>
771fe6b9 31#include "drmP.h"
771fe6b9 32#include "radeon.h"
e6990375 33#include "radeon_asic.h"
4153e584 34#include "radeon_drm.h"
3ce0a23d 35#include "rv770d.h"
3ce0a23d 36#include "atom.h"
d39c3b89 37#include "avivod.h"
771fe6b9 38
3ce0a23d
JG		 39#define R700_PFP_UCODE_SIZE 848
40#define R700_PM4_UCODE_SIZE 1360
771fe6b9 41
3ce0a23d
JG		 42static void rv770_gpu_init(struct radeon_device *rdev);
43void rv770_fini(struct radeon_device *rdev);
771fe6b9 44
21a8122a
AD		 45/* get temperature in millidegrees */
46u32 rv770_get_temp(struct radeon_device *rdev)
47{
48 u32 temp = (RREG32(CG_MULT_THERMAL_STATUS) & ASIC_T_MASK) >>
49 ASIC_T_SHIFT;
50 u32 actual_temp = 0;
51
52 if ((temp >> 9) & 1)
53 actual_temp = 0;
54 else
55 actual_temp = (temp >> 1) & 0xff;
56
57 return actual_temp * 1000;
58}
59
49e02b73
AD		 60void rv770_pm_misc(struct radeon_device *rdev)
61{
a081a9d6
RM		 62 int req_ps_idx = rdev->pm.requested_power_state_index;
63 int req_cm_idx = rdev->pm.requested_clock_mode_index;
64 struct radeon_power_state *ps = &rdev->pm.power_state[req_ps_idx];
65 struct radeon_voltage *voltage = &ps->clock_info[req_cm_idx].voltage;
4d60173f
AD		 66
67 if ((voltage->type == VOLTAGE_SW) && voltage->voltage) {
68 if (voltage->voltage != rdev->pm.current_vddc) {
69 radeon_atom_set_voltage(rdev, voltage->voltage);
70 rdev->pm.current_vddc = voltage->voltage;
0fcbe947 71 DRM_DEBUG("Setting: v: %d\n", voltage->voltage);
4d60173f
AD		 72 }
73 }
49e02b73 74}
771fe6b9
JG		 75
76/*
3ce0a23d 77 * GART
771fe6b9 78 */
3ce0a23d 79int rv770_pcie_gart_enable(struct radeon_device *rdev)
771fe6b9 80{
3ce0a23d
JG		 81 u32 tmp;
82 int r, i;
771fe6b9 83
4aac0473
JG		 84 if (rdev->gart.table.vram.robj == NULL) {
85 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
86 return -EINVAL;
3ce0a23d 87 }
4aac0473
JG		 88 r = radeon_gart_table_vram_pin(rdev);
89 if (r)
3ce0a23d 90 return r;
82568565 91 radeon_gart_restore(rdev);
3ce0a23d
JG		 92 /* Setup L2 cache */
93 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
94 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
95 EFFECTIVE_L2_QUEUE_SIZE(7));
96 WREG32(VM_L2_CNTL2, 0);
97 WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
98 /* Setup TLB control */
99 tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
100 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
101 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
102 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
103 WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
104 WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
105 WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
106 WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
107 WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
108 WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
109 WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
110 WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
1a029b76 111 WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
3ce0a23d
JG		 112 WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
113 WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
114 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
115 WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
116 (u32)(rdev->dummy_page.addr >> 12));
117 for (i = 1; i < 7; i++)
118 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
771fe6b9 119
3ce0a23d
JG		 120 r600_pcie_gart_tlb_flush(rdev);
121 rdev->gart.ready = true;
771fe6b9
JG		 122 return 0;
123}
124
3ce0a23d 125void rv770_pcie_gart_disable(struct radeon_device *rdev)
771fe6b9 126{
3ce0a23d 127 u32 tmp;
4c788679 128 int i, r;
3ce0a23d 129
3ce0a23d
JG		 130 /* Disable all tables */
131 for (i = 0; i < 7; i++)
132 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
133
134 /* Setup L2 cache */
135 WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING |
136 EFFECTIVE_L2_QUEUE_SIZE(7));
137 WREG32(VM_L2_CNTL2, 0);
138 WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
139 /* Setup TLB control */
140 tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
141 WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
142 WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
143 WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
144 WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
145 WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
146 WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
147 WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
4aac0473 148 if (rdev->gart.table.vram.robj) {
4c788679
JG		 149 r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
150 if (likely(r == 0)) {
151 radeon_bo_kunmap(rdev->gart.table.vram.robj);
152 radeon_bo_unpin(rdev->gart.table.vram.robj);
153 radeon_bo_unreserve(rdev->gart.table.vram.robj);
154 }
4aac0473
JG		 155 }
156}
157
158void rv770_pcie_gart_fini(struct radeon_device *rdev)
159{
f9274562 160 radeon_gart_fini(rdev);
4aac0473
JG		 161 rv770_pcie_gart_disable(rdev);
162 radeon_gart_table_vram_free(rdev);
771fe6b9
JG		 163}
164
165
1a029b76
JG		 166void rv770_agp_enable(struct radeon_device *rdev)
167{
168 u32 tmp;
169 int i;
170
171 /* Setup L2 cache */
172 WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
173 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
174 EFFECTIVE_L2_QUEUE_SIZE(7));
175 WREG32(VM_L2_CNTL2, 0);
176 WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
177 /* Setup TLB control */
178 tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
179 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
180 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
181 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
182 WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
183 WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
184 WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
185 WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
186 WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
187 WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
188 WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
189 for (i = 0; i < 7; i++)
190 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
191}
192
a3c1945a 193static void rv770_mc_program(struct radeon_device *rdev)
771fe6b9 194{
a3c1945a 195 struct rv515_mc_save save;
3ce0a23d
JG		 196 u32 tmp;
197 int i, j;
198
199 /* Initialize HDP */
200 for (i = 0, j = 0; i < 32; i++, j += 0x18) {
201 WREG32((0x2c14 + j), 0x00000000);
202 WREG32((0x2c18 + j), 0x00000000);
203 WREG32((0x2c1c + j), 0x00000000);
204 WREG32((0x2c20 + j), 0x00000000);
205 WREG32((0x2c24 + j), 0x00000000);
206 }
812d0469
AD		 207 /* r7xx hw bug. Read from HDP_DEBUG1 rather
208 * than writing to HDP_REG_COHERENCY_FLUSH_CNTL
209 */
210 tmp = RREG32(HDP_DEBUG1);
3ce0a23d 211
a3c1945a 212 rv515_mc_stop(rdev, &save);
3ce0a23d 213 if (r600_mc_wait_for_idle(rdev)) {
a3c1945a 214 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3ce0a23d 215 }
3ce0a23d
JG		 216 /* Lockout access through VGA aperture*/
217 WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
3ce0a23d 218 /* Update configuration */
1a029b76
JG		 219 if (rdev->flags & RADEON_IS_AGP) {
220 if (rdev->mc.vram_start < rdev->mc.gtt_start) {
221 /* VRAM before AGP */
222 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
223 rdev->mc.vram_start >> 12);
224 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
225 rdev->mc.gtt_end >> 12);
226 } else {
227 /* VRAM after AGP */
228 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
229 rdev->mc.gtt_start >> 12);
230 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
231 rdev->mc.vram_end >> 12);
232 }
233 } else {
234 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
235 rdev->mc.vram_start >> 12);
236 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
237 rdev->mc.vram_end >> 12);
238 }
3ce0a23d 239 WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
1a029b76 240 tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
3ce0a23d
JG		 241 tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
242 WREG32(MC_VM_FB_LOCATION, tmp);
243 WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
244 WREG32(HDP_NONSURFACE_INFO, (2 << 7));
46fcd2b3 245 WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
3ce0a23d 246 if (rdev->flags & RADEON_IS_AGP) {
1a029b76 247 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16);
3ce0a23d
JG		 248 WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16);
249 WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22);
250 } else {
251 WREG32(MC_VM_AGP_BASE, 0);
252 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
253 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
254 }
3ce0a23d 255 if (r600_mc_wait_for_idle(rdev)) {
a3c1945a 256 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
3ce0a23d 257 }
a3c1945a 258 rv515_mc_resume(rdev, &save);
698443d9
DA		 259 /* we need to own VRAM, so turn off the VGA renderer here
260 * to stop it overwriting our objects */
d39c3b89 261 rv515_vga_render_disable(rdev);
771fe6b9
JG		 262}
263
3ce0a23d
JG		 264
265/*
266 * CP.
267 */
268void r700_cp_stop(struct radeon_device *rdev)
771fe6b9 269{
c919b371 270 rdev->mc.active_vram_size = rdev->mc.visible_vram_size;
3ce0a23d 271 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
771fe6b9
JG		 272}
273
3ce0a23d 274static int rv770_cp_load_microcode(struct radeon_device *rdev)
771fe6b9 275{
3ce0a23d
JG		 276 const __be32 *fw_data;
277 int i;
278
279 if (!rdev->me_fw || !rdev->pfp_fw)
280 return -EINVAL;
281
282 r700_cp_stop(rdev);
283 WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0));
284
285 /* Reset cp */
286 WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
287 RREG32(GRBM_SOFT_RESET);
288 mdelay(15);
289 WREG32(GRBM_SOFT_RESET, 0);
290
291 fw_data = (const __be32 *)rdev->pfp_fw->data;
292 WREG32(CP_PFP_UCODE_ADDR, 0);
293 for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
294 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
295 WREG32(CP_PFP_UCODE_ADDR, 0);
296
297 fw_data = (const __be32 *)rdev->me_fw->data;
298 WREG32(CP_ME_RAM_WADDR, 0);
299 for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
300 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
301
302 WREG32(CP_PFP_UCODE_ADDR, 0);
303 WREG32(CP_ME_RAM_WADDR, 0);
304 WREG32(CP_ME_RAM_RADDR, 0);
305 return 0;
771fe6b9
JG		 306}
307
fe251e2f
AD		 308void r700_cp_fini(struct radeon_device *rdev)
309{
310 r700_cp_stop(rdev);
311 radeon_ring_fini(rdev);
312}
771fe6b9
JG		 313
314/*
3ce0a23d 315 * Core functions
771fe6b9 316 */
d03f5d59
AD		 317static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev,
318 u32 num_tile_pipes,
319 u32 num_backends,
320 u32 backend_disable_mask)
771fe6b9 321{
3ce0a23d
JG		 322 u32 backend_map = 0;
323 u32 enabled_backends_mask;
324 u32 enabled_backends_count;
325 u32 cur_pipe;
326 u32 swizzle_pipe[R7XX_MAX_PIPES];
327 u32 cur_backend;
328 u32 i;
d03f5d59 329 bool force_no_swizzle;
3ce0a23d
JG		 330
331 if (num_tile_pipes > R7XX_MAX_PIPES)
332 num_tile_pipes = R7XX_MAX_PIPES;
333 if (num_tile_pipes < 1)
334 num_tile_pipes = 1;
335 if (num_backends > R7XX_MAX_BACKENDS)
336 num_backends = R7XX_MAX_BACKENDS;
337 if (num_backends < 1)
338 num_backends = 1;
339
340 enabled_backends_mask = 0;
341 enabled_backends_count = 0;
342 for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
343 if (((backend_disable_mask >> i) & 1) == 0) {
344 enabled_backends_mask |= (1 << i);
345 ++enabled_backends_count;
346 }
347 if (enabled_backends_count == num_backends)
348 break;
349 }
350
351 if (enabled_backends_count == 0) {
352 enabled_backends_mask = 1;
353 enabled_backends_count = 1;
354 }
355
356 if (enabled_backends_count != num_backends)
357 num_backends = enabled_backends_count;
358
d03f5d59
AD		 359 switch (rdev->family) {
360 case CHIP_RV770:
361 case CHIP_RV730:
362 force_no_swizzle = false;
363 break;
364 case CHIP_RV710:
365 case CHIP_RV740:
366 default:
367 force_no_swizzle = true;
368 break;
369 }
370
3ce0a23d
JG		 371 memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
372 switch (num_tile_pipes) {
373 case 1:
374 swizzle_pipe[0] = 0;
375 break;
376 case 2:
377 swizzle_pipe[0] = 0;
378 swizzle_pipe[1] = 1;
379 break;
380 case 3:
d03f5d59
AD		 381 if (force_no_swizzle) {
382 swizzle_pipe[0] = 0;
383 swizzle_pipe[1] = 1;
384 swizzle_pipe[2] = 2;
385 } else {
386 swizzle_pipe[0] = 0;
387 swizzle_pipe[1] = 2;
388 swizzle_pipe[2] = 1;
389 }
3ce0a23d
JG		 390 break;
391 case 4:
d03f5d59
AD		 392 if (force_no_swizzle) {
393 swizzle_pipe[0] = 0;
394 swizzle_pipe[1] = 1;
395 swizzle_pipe[2] = 2;
396 swizzle_pipe[3] = 3;
397 } else {
398 swizzle_pipe[0] = 0;
399 swizzle_pipe[1] = 2;
400 swizzle_pipe[2] = 3;
401 swizzle_pipe[3] = 1;
402 }
3ce0a23d
JG		 403 break;
404 case 5:
d03f5d59
AD		 405 if (force_no_swizzle) {
406 swizzle_pipe[0] = 0;
407 swizzle_pipe[1] = 1;
408 swizzle_pipe[2] = 2;
409 swizzle_pipe[3] = 3;
410 swizzle_pipe[4] = 4;
411 } else {
412 swizzle_pipe[0] = 0;
413 swizzle_pipe[1] = 2;
414 swizzle_pipe[2] = 4;
415 swizzle_pipe[3] = 1;
416 swizzle_pipe[4] = 3;
417 }
3ce0a23d
JG		 418 break;
419 case 6:
d03f5d59
AD		 420 if (force_no_swizzle) {
421 swizzle_pipe[0] = 0;
422 swizzle_pipe[1] = 1;
423 swizzle_pipe[2] = 2;
424 swizzle_pipe[3] = 3;
425 swizzle_pipe[4] = 4;
426 swizzle_pipe[5] = 5;
427 } else {
428 swizzle_pipe[0] = 0;
429 swizzle_pipe[1] = 2;
430 swizzle_pipe[2] = 4;
431 swizzle_pipe[3] = 5;
432 swizzle_pipe[4] = 3;
433 swizzle_pipe[5] = 1;
434 }
3ce0a23d
JG		 435 break;
436 case 7:
d03f5d59
AD		 437 if (force_no_swizzle) {
438 swizzle_pipe[0] = 0;
439 swizzle_pipe[1] = 1;
440 swizzle_pipe[2] = 2;
441 swizzle_pipe[3] = 3;
442 swizzle_pipe[4] = 4;
443 swizzle_pipe[5] = 5;
444 swizzle_pipe[6] = 6;
445 } else {
446 swizzle_pipe[0] = 0;
447 swizzle_pipe[1] = 2;
448 swizzle_pipe[2] = 4;
449 swizzle_pipe[3] = 6;
450 swizzle_pipe[4] = 3;
451 swizzle_pipe[5] = 1;
452 swizzle_pipe[6] = 5;
453 }
3ce0a23d
JG		 454 break;
455 case 8:
d03f5d59
AD		 456 if (force_no_swizzle) {
457 swizzle_pipe[0] = 0;
458 swizzle_pipe[1] = 1;
459 swizzle_pipe[2] = 2;
460 swizzle_pipe[3] = 3;
461 swizzle_pipe[4] = 4;
462 swizzle_pipe[5] = 5;
463 swizzle_pipe[6] = 6;
464 swizzle_pipe[7] = 7;
465 } else {
466 swizzle_pipe[0] = 0;
467 swizzle_pipe[1] = 2;
468 swizzle_pipe[2] = 4;
469 swizzle_pipe[3] = 6;
470 swizzle_pipe[4] = 3;
471 swizzle_pipe[5] = 1;
472 swizzle_pipe[6] = 7;
473 swizzle_pipe[7] = 5;
474 }
3ce0a23d
JG		 475 break;
476 }
477
478 cur_backend = 0;
479 for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
480 while (((1 << cur_backend) & enabled_backends_mask) == 0)
481 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
482
483 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
484
485 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
486 }
487
488 return backend_map;
771fe6b9
JG		 489}
490
3ce0a23d 491static void rv770_gpu_init(struct radeon_device *rdev)
771fe6b9 492{
3ce0a23d 493 int i, j, num_qd_pipes;
d03f5d59 494 u32 ta_aux_cntl;
3ce0a23d
JG		 495 u32 sx_debug_1;
496 u32 smx_dc_ctl0;
d03f5d59 497 u32 db_debug3;
3ce0a23d
JG		 498 u32 num_gs_verts_per_thread;
499 u32 vgt_gs_per_es;
500 u32 gs_prim_buffer_depth = 0;
501 u32 sq_ms_fifo_sizes;
502 u32 sq_config;
503 u32 sq_thread_resource_mgmt;
504 u32 hdp_host_path_cntl;
505 u32 sq_dyn_gpr_size_simd_ab_0;
506 u32 backend_map;
507 u32 gb_tiling_config = 0;
508 u32 cc_rb_backend_disable = 0;
509 u32 cc_gc_shader_pipe_config = 0;
510 u32 mc_arb_ramcfg;
511 u32 db_debug4;
771fe6b9 512
3ce0a23d
JG		 513 /* setup chip specs */
514 switch (rdev->family) {
515 case CHIP_RV770:
516 rdev->config.rv770.max_pipes = 4;
517 rdev->config.rv770.max_tile_pipes = 8;
518 rdev->config.rv770.max_simds = 10;
519 rdev->config.rv770.max_backends = 4;
520 rdev->config.rv770.max_gprs = 256;
521 rdev->config.rv770.max_threads = 248;
522 rdev->config.rv770.max_stack_entries = 512;
523 rdev->config.rv770.max_hw_contexts = 8;
524 rdev->config.rv770.max_gs_threads = 16 * 2;
525 rdev->config.rv770.sx_max_export_size = 128;
526 rdev->config.rv770.sx_max_export_pos_size = 16;
527 rdev->config.rv770.sx_max_export_smx_size = 112;
528 rdev->config.rv770.sq_num_cf_insts = 2;
529
530 rdev->config.rv770.sx_num_of_sets = 7;
531 rdev->config.rv770.sc_prim_fifo_size = 0xF9;
532 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
533 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
534 break;
535 case CHIP_RV730:
536 rdev->config.rv770.max_pipes = 2;
537 rdev->config.rv770.max_tile_pipes = 4;
538 rdev->config.rv770.max_simds = 8;
539 rdev->config.rv770.max_backends = 2;
540 rdev->config.rv770.max_gprs = 128;
541 rdev->config.rv770.max_threads = 248;
542 rdev->config.rv770.max_stack_entries = 256;
543 rdev->config.rv770.max_hw_contexts = 8;
544 rdev->config.rv770.max_gs_threads = 16 * 2;
545 rdev->config.rv770.sx_max_export_size = 256;
546 rdev->config.rv770.sx_max_export_pos_size = 32;
547 rdev->config.rv770.sx_max_export_smx_size = 224;
548 rdev->config.rv770.sq_num_cf_insts = 2;
549
550 rdev->config.rv770.sx_num_of_sets = 7;
551 rdev->config.rv770.sc_prim_fifo_size = 0xf9;
552 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
553 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
554 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
555 rdev->config.rv770.sx_max_export_pos_size -= 16;
556 rdev->config.rv770.sx_max_export_smx_size += 16;
557 }
558 break;
559 case CHIP_RV710:
560 rdev->config.rv770.max_pipes = 2;
561 rdev->config.rv770.max_tile_pipes = 2;
562 rdev->config.rv770.max_simds = 2;
563 rdev->config.rv770.max_backends = 1;
564 rdev->config.rv770.max_gprs = 256;
565 rdev->config.rv770.max_threads = 192;
566 rdev->config.rv770.max_stack_entries = 256;
567 rdev->config.rv770.max_hw_contexts = 4;
568 rdev->config.rv770.max_gs_threads = 8 * 2;
569 rdev->config.rv770.sx_max_export_size = 128;
570 rdev->config.rv770.sx_max_export_pos_size = 16;
571 rdev->config.rv770.sx_max_export_smx_size = 112;
572 rdev->config.rv770.sq_num_cf_insts = 1;
573
574 rdev->config.rv770.sx_num_of_sets = 7;
575 rdev->config.rv770.sc_prim_fifo_size = 0x40;
576 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
577 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
578 break;
579 case CHIP_RV740:
580 rdev->config.rv770.max_pipes = 4;
581 rdev->config.rv770.max_tile_pipes = 4;
582 rdev->config.rv770.max_simds = 8;
583 rdev->config.rv770.max_backends = 4;
584 rdev->config.rv770.max_gprs = 256;
585 rdev->config.rv770.max_threads = 248;
586 rdev->config.rv770.max_stack_entries = 512;
587 rdev->config.rv770.max_hw_contexts = 8;
588 rdev->config.rv770.max_gs_threads = 16 * 2;
589 rdev->config.rv770.sx_max_export_size = 256;
590 rdev->config.rv770.sx_max_export_pos_size = 32;
591 rdev->config.rv770.sx_max_export_smx_size = 224;
592 rdev->config.rv770.sq_num_cf_insts = 2;
593
594 rdev->config.rv770.sx_num_of_sets = 7;
595 rdev->config.rv770.sc_prim_fifo_size = 0x100;
596 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
597 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
598
599 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
600 rdev->config.rv770.sx_max_export_pos_size -= 16;
601 rdev->config.rv770.sx_max_export_smx_size += 16;
602 }
603 break;
604 default:
605 break;
606 }
607
608 /* Initialize HDP */
609 j = 0;
610 for (i = 0; i < 32; i++) {
611 WREG32((0x2c14 + j), 0x00000000);
612 WREG32((0x2c18 + j), 0x00000000);
613 WREG32((0x2c1c + j), 0x00000000);
614 WREG32((0x2c20 + j), 0x00000000);
615 WREG32((0x2c24 + j), 0x00000000);
616 j += 0x18;
617 }
618
619 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
620
621 /* setup tiling, simd, pipe config */
622 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
623
624 switch (rdev->config.rv770.max_tile_pipes) {
625 case 1:
d03f5d59 626 default:
3ce0a23d
JG		 627 gb_tiling_config |= PIPE_TILING(0);
628 break;
629 case 2:
630 gb_tiling_config |= PIPE_TILING(1);
631 break;
632 case 4:
633 gb_tiling_config |= PIPE_TILING(2);
634 break;
635 case 8:
636 gb_tiling_config |= PIPE_TILING(3);
3ce0a23d
JG		 637 break;
638 }
d03f5d59 639 rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes;
3ce0a23d
JG		 640
641 if (rdev->family == CHIP_RV770)
642 gb_tiling_config |= BANK_TILING(1);
643 else
e29649db 644 gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
961fb597 645 rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3);
3ce0a23d
JG		 646
647 gb_tiling_config |= GROUP_SIZE(0);
961fb597 648 rdev->config.rv770.tiling_group_size = 256;
3ce0a23d 649
e29649db 650 if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) {
3ce0a23d
JG		 651 gb_tiling_config |= ROW_TILING(3);
652 gb_tiling_config |= SAMPLE_SPLIT(3);
653 } else {
654 gb_tiling_config |=
655 ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
656 gb_tiling_config |=
657 SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
658 }
659
660 gb_tiling_config |= BANK_SWAPS(1);
661
d03f5d59
AD		 662 cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000;
663 cc_rb_backend_disable |=
664 BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK);
3ce0a23d 665
d03f5d59
AD		 666 cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
667 cc_gc_shader_pipe_config |=
3ce0a23d
JG		 668 INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK);
669 cc_gc_shader_pipe_config |=
670 INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK);
671
d03f5d59
AD		 672 if (rdev->family == CHIP_RV740)
673 backend_map = 0x28;
674 else
675 backend_map = r700_get_tile_pipe_to_backend_map(rdev,
676 rdev->config.rv770.max_tile_pipes,
677 (R7XX_MAX_BACKENDS -
678 r600_count_pipe_bits((cc_rb_backend_disable &
679 R7XX_MAX_BACKENDS_MASK) >> 16)),
680 (cc_rb_backend_disable >> 16));
d03f5d59 681
e7aeeba6
AD		 682 rdev->config.rv770.tile_config = gb_tiling_config;
683 gb_tiling_config |= BACKEND_MAP(backend_map);
3ce0a23d
JG		 684
685 WREG32(GB_TILING_CONFIG, gb_tiling_config);
686 WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
687 WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
688
689 WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable);
690 WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
f867c60d 691 WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
d03f5d59 692 WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable);
3ce0a23d 693
3ce0a23d
JG		 694 WREG32(CGTS_SYS_TCC_DISABLE, 0);
695 WREG32(CGTS_TCC_DISABLE, 0);
f867c60d
AD		 696 WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
697 WREG32(CGTS_USER_TCC_DISABLE, 0);
3ce0a23d
JG		 698
699 num_qd_pipes =
d03f5d59 700 R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
3ce0a23d
JG		 701 WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK);
702 WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK);
703
704 /* set HW defaults for 3D engine */
705 WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
e29649db 706 ROQ_IB2_START(0x2b)));
3ce0a23d
JG		 707
708 WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
709
d03f5d59
AD		 710 ta_aux_cntl = RREG32(TA_CNTL_AUX);
711 WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO);
3ce0a23d
JG		 712
713 sx_debug_1 = RREG32(SX_DEBUG_1);
714 sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
715 WREG32(SX_DEBUG_1, sx_debug_1);
716
717 smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
718 smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff);
719 smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1);
720 WREG32(SMX_DC_CTL0, smx_dc_ctl0);
721
d03f5d59
AD		 722 if (rdev->family != CHIP_RV740)
723 WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) |
724 GS_FLUSH_CTL(4) |
725 ACK_FLUSH_CTL(3) |
726 SYNC_FLUSH_CTL));
3ce0a23d 727
d03f5d59
AD		 728 db_debug3 = RREG32(DB_DEBUG3);
729 db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f);
730 switch (rdev->family) {
731 case CHIP_RV770:
732 case CHIP_RV740:
733 db_debug3 |= DB_CLK_OFF_DELAY(0x1f);
734 break;
735 case CHIP_RV710:
736 case CHIP_RV730:
737 default:
738 db_debug3 |= DB_CLK_OFF_DELAY(2);
739 break;
740 }
741 WREG32(DB_DEBUG3, db_debug3);
742
743 if (rdev->family != CHIP_RV770) {
3ce0a23d
JG		 744 db_debug4 = RREG32(DB_DEBUG4);
745 db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER;
746 WREG32(DB_DEBUG4, db_debug4);
747 }
748
749 WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) |
e29649db
AD		 750 POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) |
751 SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1)));
3ce0a23d
JG		 752
753 WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) |
e29649db
AD		 754 SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) |
755 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize)));
3ce0a23d
JG		 756
757 WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
758
759 WREG32(VGT_NUM_INSTANCES, 1);
760
761 WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0));
762
763 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
764
765 WREG32(CP_PERFMON_CNTL, 0);
766
767 sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) |
768 DONE_FIFO_HIWATER(0xe0) |
769 ALU_UPDATE_FIFO_HIWATER(0x8));
770 switch (rdev->family) {
771 case CHIP_RV770:
3ce0a23d
JG		 772 case CHIP_RV730:
773 case CHIP_RV710:
d03f5d59
AD		 774 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1);
775 break;
3ce0a23d
JG		 776 case CHIP_RV740:
777 default:
778 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4);
779 break;
780 }
781 WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
782
783 /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
784 * should be adjusted as needed by the 2D/3D drivers. This just sets default values
785 */
786 sq_config = RREG32(SQ_CONFIG);
787 sq_config &= ~(PS_PRIO(3) |
788 VS_PRIO(3) |
789 GS_PRIO(3) |
790 ES_PRIO(3));
791 sq_config |= (DX9_CONSTS |
792 VC_ENABLE |
793 EXPORT_SRC_C |
794 PS_PRIO(0) |
795 VS_PRIO(1) |
796 GS_PRIO(2) |
797 ES_PRIO(3));
798 if (rdev->family == CHIP_RV710)
799 /* no vertex cache */
800 sq_config &= ~VC_ENABLE;
801
802 WREG32(SQ_CONFIG, sq_config);
803
804 WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
fe62e1a4
DA		 805 NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
806 NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2)));
3ce0a23d
JG		 807
808 WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) |
fe62e1a4 809 NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64)));
3ce0a23d
JG		 810
811 sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) |
812 NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) |
813 NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8));
814 if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads)
815 sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads);
816 else
817 sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8);
818 WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
819
820 WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
821 NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
822
823 WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
824 NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
825
826 sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) |
827 SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) |
828 SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) |
829 SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64));
830
831 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
832 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
833 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
834 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
835 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
836 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
837 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
838 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
839
840 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
fe62e1a4 841 FORCE_EOV_MAX_REZ_CNT(255)));
3ce0a23d
JG		 842
843 if (rdev->family == CHIP_RV710)
844 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) |
fe62e1a4 845 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
3ce0a23d
JG		 846 else
847 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) |
fe62e1a4 848 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
3ce0a23d
JG		 849
850 switch (rdev->family) {
851 case CHIP_RV770:
852 case CHIP_RV730:
853 case CHIP_RV740:
854 gs_prim_buffer_depth = 384;
855 break;
856 case CHIP_RV710:
857 gs_prim_buffer_depth = 128;
858 break;
859 default:
860 break;
861 }
862
863 num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16;
864 vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
865 /* Max value for this is 256 */
866 if (vgt_gs_per_es > 256)
867 vgt_gs_per_es = 256;
868
869 WREG32(VGT_ES_PER_GS, 128);
870 WREG32(VGT_GS_PER_ES, vgt_gs_per_es);
871 WREG32(VGT_GS_PER_VS, 2);
872
873 /* more default values. 2D/3D driver should adjust as needed */
874 WREG32(VGT_GS_VERTEX_REUSE, 16);
875 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
876 WREG32(VGT_STRMOUT_EN, 0);
877 WREG32(SX_MISC, 0);
878 WREG32(PA_SC_MODE_CNTL, 0);
879 WREG32(PA_SC_EDGERULE, 0xaaaaaaaa);
880 WREG32(PA_SC_AA_CONFIG, 0);
881 WREG32(PA_SC_CLIPRECT_RULE, 0xffff);
882 WREG32(PA_SC_LINE_STIPPLE, 0);
883 WREG32(SPI_INPUT_Z, 0);
884 WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
885 WREG32(CB_COLOR7_FRAG, 0);
886
887 /* clear render buffer base addresses */
888 WREG32(CB_COLOR0_BASE, 0);
889 WREG32(CB_COLOR1_BASE, 0);
890 WREG32(CB_COLOR2_BASE, 0);
891 WREG32(CB_COLOR3_BASE, 0);
892 WREG32(CB_COLOR4_BASE, 0);
893 WREG32(CB_COLOR5_BASE, 0);
894 WREG32(CB_COLOR6_BASE, 0);
895 WREG32(CB_COLOR7_BASE, 0);
896
897 WREG32(TCP_CNTL, 0);
898
899 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
900 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
901
902 WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
903
904 WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
905 NUM_CLIP_SEQ(3)));
906
907}
908
87cbf8f2
AD		 909static int rv770_vram_scratch_init(struct radeon_device *rdev)
910{
911 int r;
912 u64 gpu_addr;
913
914 if (rdev->vram_scratch.robj == NULL) {
915 r = radeon_bo_create(rdev, NULL, RADEON_GPU_PAGE_SIZE,
916 true, RADEON_GEM_DOMAIN_VRAM,
917 &rdev->vram_scratch.robj);
918 if (r) {
919 return r;
920 }
921 }
922
923 r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
924 if (unlikely(r != 0))
925 return r;
926 r = radeon_bo_pin(rdev->vram_scratch.robj,
927 RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
928 if (r) {
929 radeon_bo_unreserve(rdev->vram_scratch.robj);
930 return r;
931 }
932 r = radeon_bo_kmap(rdev->vram_scratch.robj,
933 (void **)&rdev->vram_scratch.ptr);
934 if (r)
935 radeon_bo_unpin(rdev->vram_scratch.robj);
936 radeon_bo_unreserve(rdev->vram_scratch.robj);
937
938 return r;
939}
940
941static void rv770_vram_scratch_fini(struct radeon_device *rdev)
942{
943 int r;
944
945 if (rdev->vram_scratch.robj == NULL) {
946 return;
947 }
948 r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
949 if (likely(r == 0)) {
950 radeon_bo_kunmap(rdev->vram_scratch.robj);
951 radeon_bo_unpin(rdev->vram_scratch.robj);
952 radeon_bo_unreserve(rdev->vram_scratch.robj);
953 }
954 radeon_bo_unref(&rdev->vram_scratch.robj);
955}
956
3ce0a23d
JG		 957int rv770_mc_init(struct radeon_device *rdev)
958{
3ce0a23d 959 u32 tmp;
5885b7a9 960 int chansize, numchan;
3ce0a23d
JG		 961
962 /* Get VRAM informations */
3ce0a23d 963 rdev->mc.vram_is_ddr = true;
5885b7a9
AD		 964 tmp = RREG32(MC_ARB_RAMCFG);
965 if (tmp & CHANSIZE_OVERRIDE) {
966 chansize = 16;
967 } else if (tmp & CHANSIZE_MASK) {
968 chansize = 64;
969 } else {
970 chansize = 32;
971 }
972 tmp = RREG32(MC_SHARED_CHMAP);
973 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
974 case 0:
975 default:
976 numchan = 1;
977 break;
978 case 1:
979 numchan = 2;
980 break;
981 case 2:
982 numchan = 4;
983 break;
984 case 3:
985 numchan = 8;
986 break;
987 }
988 rdev->mc.vram_width = numchan * chansize;
771fe6b9 989 /* Could aper size report 0 ? */
01d73a69
JC		 990 rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
991 rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
3ce0a23d
JG		 992 /* Setup GPU memory space */
993 rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
994 rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
51e5fcd3 995 rdev->mc.visible_vram_size = rdev->mc.aper_size;
c919b371 996 rdev->mc.active_vram_size = rdev->mc.visible_vram_size;
d594e46a 997 r600_vram_gtt_location(rdev, &rdev->mc);
f47299c5
AD		 998 radeon_update_bandwidth_info(rdev);
999
3ce0a23d
JG		 1000 return 0;
1001}
d594e46a 1002
fc30b8ef 1003static int rv770_startup(struct radeon_device *rdev)
3ce0a23d
JG		 1004{
1005 int r;
1006
779720a3
AD		 1007 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
1008 r = r600_init_microcode(rdev);
1009 if (r) {
1010 DRM_ERROR("Failed to load firmware!\n");
1011 return r;
1012 }
1013 }
1014
a3c1945a 1015 rv770_mc_program(rdev);
1a029b76
JG		 1016 if (rdev->flags & RADEON_IS_AGP) {
1017 rv770_agp_enable(rdev);
1018 } else {
1019 r = rv770_pcie_gart_enable(rdev);
1020 if (r)
1021 return r;
1022 }
87cbf8f2
AD		 1023 r = rv770_vram_scratch_init(rdev);
1024 if (r)
1025 return r;
3ce0a23d 1026 rv770_gpu_init(rdev);
c38c7b64
JG		 1027 r = r600_blit_init(rdev);
1028 if (r) {
1029 r600_blit_fini(rdev);
1030 rdev->asic->copy = NULL;
1031 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
1032 }
ff82f052
JG		 1033 /* pin copy shader into vram */
1034 if (rdev->r600_blit.shader_obj) {
1035 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1036 if (unlikely(r != 0))
1037 return r;
1038 r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM,
1039 &rdev->r600_blit.shader_gpu_addr);
1040 radeon_bo_unreserve(rdev->r600_blit.shader_obj);
7923c615 1041 if (r) {
ff82f052 1042 DRM_ERROR("failed to pin blit object %d\n", r);
7923c615
AD		 1043 return r;
1044 }
1045 }
d8f60cfc 1046 /* Enable IRQ */
d8f60cfc
AD		 1047 r = r600_irq_init(rdev);
1048 if (r) {
1049 DRM_ERROR("radeon: IH init failed (%d).\n", r);
1050 radeon_irq_kms_fini(rdev);
1051 return r;
1052 }
1053 r600_irq_set(rdev);
1054
3ce0a23d
JG		 1055 r = radeon_ring_init(rdev, rdev->cp.ring_size);
1056 if (r)
1057 return r;
1058 r = rv770_cp_load_microcode(rdev);
1059 if (r)
1060 return r;
1061 r = r600_cp_resume(rdev);
1062 if (r)
1063 return r;
81cc35bf
JG		 1064 /* write back buffer are not vital so don't worry about failure */
1065 r600_wb_enable(rdev);
3ce0a23d
JG		 1066 return 0;
1067}
1068
fc30b8ef
DA		 1069int rv770_resume(struct radeon_device *rdev)
1070{
1071 int r;
1072
1a029b76
JG		 1073 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
1074 * posting will perform necessary task to bring back GPU into good
1075 * shape.
1076 */
fc30b8ef 1077 /* post card */
e7d40b9a 1078 atom_asic_init(rdev->mode_info.atom_context);
fc30b8ef
DA		 1079
1080 r = rv770_startup(rdev);
1081 if (r) {
1082 DRM_ERROR("r600 startup failed on resume\n");
1083 return r;
1084 }
1085
62a8ea3f 1086 r = r600_ib_test(rdev);
fc30b8ef
DA		 1087 if (r) {
1088 DRM_ERROR("radeon: failled testing IB (%d).\n", r);
1089 return r;
1090 }
8a8c6e7c
RM		 1091
1092 r = r600_audio_init(rdev);
1093 if (r) {
1094 dev_err(rdev->dev, "radeon: audio init failed\n");
1095 return r;
1096 }
1097
fc30b8ef
DA		 1098 return r;
1099
1100}
1101
3ce0a23d
JG		 1102int rv770_suspend(struct radeon_device *rdev)
1103{
4c788679
JG		 1104 int r;
1105
8a8c6e7c 1106 r600_audio_fini(rdev);
3ce0a23d
JG		 1107 /* FIXME: we should wait for ring to be empty */
1108 r700_cp_stop(rdev);
4153e584 1109 rdev->cp.ready = false;
0c45249f 1110 r600_irq_suspend(rdev);
81cc35bf 1111 r600_wb_disable(rdev);
4aac0473 1112 rv770_pcie_gart_disable(rdev);
4153e584 1113 /* unpin shaders bo */
30d2d9a5
JG		 1114 if (rdev->r600_blit.shader_obj) {
1115 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1116 if (likely(r == 0)) {
1117 radeon_bo_unpin(rdev->r600_blit.shader_obj);
1118 radeon_bo_unreserve(rdev->r600_blit.shader_obj);
1119 }
4c788679 1120 }
3ce0a23d
JG		 1121 return 0;
1122}
1123
1124/* Plan is to move initialization in that function and use
1125 * helper function so that radeon_device_init pretty much
1126 * do nothing more than calling asic specific function. This
1127 * should also allow to remove a bunch of callback function
1128 * like vram_info.
1129 */
1130int rv770_init(struct radeon_device *rdev)
1131{
1132 int r;
1133
3ce0a23d
JG		 1134 r = radeon_dummy_page_init(rdev);
1135 if (r)
1136 return r;
1137 /* This don't do much */
1138 r = radeon_gem_init(rdev);
1139 if (r)
1140 return r;
1141 /* Read BIOS */
1142 if (!radeon_get_bios(rdev)) {
1143 if (ASIC_IS_AVIVO(rdev))
1144 return -EINVAL;
1145 }
1146 /* Must be an ATOMBIOS */
e7d40b9a
JG		 1147 if (!rdev->is_atom_bios) {
1148 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
3ce0a23d 1149 return -EINVAL;
e7d40b9a 1150 }
3ce0a23d
JG		 1151 r = radeon_atombios_init(rdev);
1152 if (r)
1153 return r;
1154 /* Post card if necessary */
72542d77
DA		 1155 if (!r600_card_posted(rdev)) {
1156 if (!rdev->bios) {
1157 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1158 return -EINVAL;
1159 }
3ce0a23d
JG		 1160 DRM_INFO("GPU not posted. posting now...\n");
1161 atom_asic_init(rdev->mode_info.atom_context);
1162 }
1163 /* Initialize scratch registers */
1164 r600_scratch_init(rdev);
1165 /* Initialize surface registers */
1166 radeon_surface_init(rdev);
7433874e 1167 /* Initialize clocks */
5e6dde7e 1168 radeon_get_clock_info(rdev->ddev);
3ce0a23d
JG		 1169 /* Fence driver */
1170 r = radeon_fence_driver_init(rdev);
1171 if (r)
1172 return r;
d594e46a 1173 /* initialize AGP */
700a0cc0
JG		 1174 if (rdev->flags & RADEON_IS_AGP) {
1175 r = radeon_agp_init(rdev);
1176 if (r)
1177 radeon_agp_disable(rdev);
1178 }
3ce0a23d 1179 r = rv770_mc_init(rdev);
b574f251 1180 if (r)
3ce0a23d 1181 return r;
3ce0a23d 1182 /* Memory manager */
4c788679 1183 r = radeon_bo_init(rdev);
3ce0a23d
JG		 1184 if (r)
1185 return r;
d8f60cfc
AD		 1186
1187 r = radeon_irq_kms_init(rdev);
1188 if (r)
1189 return r;
1190
3ce0a23d
JG		 1191 rdev->cp.ring_obj = NULL;
1192 r600_ring_init(rdev, 1024 * 1024);
1193
d8f60cfc
AD		 1194 rdev->ih.ring_obj = NULL;
1195 r600_ih_ring_init(rdev, 64 * 1024);
1196
4aac0473
JG		 1197 r = r600_pcie_gart_init(rdev);
1198 if (r)
1199 return r;
1200
779720a3 1201 rdev->accel_working = true;
fc30b8ef 1202 r = rv770_startup(rdev);
3ce0a23d 1203 if (r) {
655efd3d 1204 dev_err(rdev->dev, "disabling GPU acceleration\n");
fe251e2f 1205 r700_cp_fini(rdev);
75c81298 1206 r600_wb_fini(rdev);
655efd3d
JG		 1207 r600_irq_fini(rdev);
1208 radeon_irq_kms_fini(rdev);
75c81298 1209 rv770_pcie_gart_fini(rdev);
733289c2 1210 rdev->accel_working = false;
3ce0a23d 1211 }
733289c2 1212 if (rdev->accel_working) {
733289c2
JG		 1213 r = radeon_ib_pool_init(rdev);
1214 if (r) {
db96380e 1215 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
733289c2 1216 rdev->accel_working = false;
db96380e
JG		 1217 } else {
1218 r = r600_ib_test(rdev);
1219 if (r) {
1220 dev_err(rdev->dev, "IB test failed (%d).\n", r);
1221 rdev->accel_working = false;
1222 }
733289c2 1223 }
3ce0a23d 1224 }
8a8c6e7c
RM		 1225
1226 r = r600_audio_init(rdev);
1227 if (r) {
1228 dev_err(rdev->dev, "radeon: audio init failed\n");
1229 return r;
1230 }
1231
3ce0a23d
JG		 1232 return 0;
1233}
1234
1235void rv770_fini(struct radeon_device *rdev)
1236{
1237 r600_blit_fini(rdev);
fe251e2f 1238 r700_cp_fini(rdev);
655efd3d 1239 r600_wb_fini(rdev);
d8f60cfc
AD		 1240 r600_irq_fini(rdev);
1241 radeon_irq_kms_fini(rdev);
4aac0473 1242 rv770_pcie_gart_fini(rdev);
87cbf8f2 1243 rv770_vram_scratch_fini(rdev);
3ce0a23d
JG		 1244 radeon_gem_fini(rdev);
1245 radeon_fence_driver_fini(rdev);
d0269ed8 1246 radeon_agp_fini(rdev);
4c788679 1247 radeon_bo_fini(rdev);
e7d40b9a 1248 radeon_atombios_fini(rdev);
3ce0a23d
JG		 1249 kfree(rdev->bios);
1250 rdev->bios = NULL;
1251 radeon_dummy_page_fini(rdev);
771fe6b9 1252}
Clone of davem's net-next-2.6 tree