]> 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: fix a regression on r7xx AGP due to the HDP flush fix
[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{
3ce0a23d 270 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
771fe6b9
JG		 271}
272
3ce0a23d 273static int rv770_cp_load_microcode(struct radeon_device *rdev)
771fe6b9 274{
3ce0a23d
JG		 275 const __be32 *fw_data;
276 int i;
277
278 if (!rdev->me_fw || !rdev->pfp_fw)
279 return -EINVAL;
280
281 r700_cp_stop(rdev);
282 WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0));
283
284 /* Reset cp */
285 WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
286 RREG32(GRBM_SOFT_RESET);
287 mdelay(15);
288 WREG32(GRBM_SOFT_RESET, 0);
289
290 fw_data = (const __be32 *)rdev->pfp_fw->data;
291 WREG32(CP_PFP_UCODE_ADDR, 0);
292 for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
293 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
294 WREG32(CP_PFP_UCODE_ADDR, 0);
295
296 fw_data = (const __be32 *)rdev->me_fw->data;
297 WREG32(CP_ME_RAM_WADDR, 0);
298 for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
299 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
300
301 WREG32(CP_PFP_UCODE_ADDR, 0);
302 WREG32(CP_ME_RAM_WADDR, 0);
303 WREG32(CP_ME_RAM_RADDR, 0);
304 return 0;
771fe6b9
JG		 305}
306
fe251e2f
AD		 307void r700_cp_fini(struct radeon_device *rdev)
308{
309 r700_cp_stop(rdev);
310 radeon_ring_fini(rdev);
311}
771fe6b9
JG		 312
313/*
3ce0a23d 314 * Core functions
771fe6b9 315 */
d03f5d59
AD		 316static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev,
317 u32 num_tile_pipes,
318 u32 num_backends,
319 u32 backend_disable_mask)
771fe6b9 320{
3ce0a23d
JG		 321 u32 backend_map = 0;
322 u32 enabled_backends_mask;
323 u32 enabled_backends_count;
324 u32 cur_pipe;
325 u32 swizzle_pipe[R7XX_MAX_PIPES];
326 u32 cur_backend;
327 u32 i;
d03f5d59 328 bool force_no_swizzle;
3ce0a23d
JG		 329
330 if (num_tile_pipes > R7XX_MAX_PIPES)
331 num_tile_pipes = R7XX_MAX_PIPES;
332 if (num_tile_pipes < 1)
333 num_tile_pipes = 1;
334 if (num_backends > R7XX_MAX_BACKENDS)
335 num_backends = R7XX_MAX_BACKENDS;
336 if (num_backends < 1)
337 num_backends = 1;
338
339 enabled_backends_mask = 0;
340 enabled_backends_count = 0;
341 for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
342 if (((backend_disable_mask >> i) & 1) == 0) {
343 enabled_backends_mask |= (1 << i);
344 ++enabled_backends_count;
345 }
346 if (enabled_backends_count == num_backends)
347 break;
348 }
349
350 if (enabled_backends_count == 0) {
351 enabled_backends_mask = 1;
352 enabled_backends_count = 1;
353 }
354
355 if (enabled_backends_count != num_backends)
356 num_backends = enabled_backends_count;
357
d03f5d59
AD		 358 switch (rdev->family) {
359 case CHIP_RV770:
360 case CHIP_RV730:
361 force_no_swizzle = false;
362 break;
363 case CHIP_RV710:
364 case CHIP_RV740:
365 default:
366 force_no_swizzle = true;
367 break;
368 }
369
3ce0a23d
JG		 370 memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
371 switch (num_tile_pipes) {
372 case 1:
373 swizzle_pipe[0] = 0;
374 break;
375 case 2:
376 swizzle_pipe[0] = 0;
377 swizzle_pipe[1] = 1;
378 break;
379 case 3:
d03f5d59
AD		 380 if (force_no_swizzle) {
381 swizzle_pipe[0] = 0;
382 swizzle_pipe[1] = 1;
383 swizzle_pipe[2] = 2;
384 } else {
385 swizzle_pipe[0] = 0;
386 swizzle_pipe[1] = 2;
387 swizzle_pipe[2] = 1;
388 }
3ce0a23d
JG		 389 break;
390 case 4:
d03f5d59
AD		 391 if (force_no_swizzle) {
392 swizzle_pipe[0] = 0;
393 swizzle_pipe[1] = 1;
394 swizzle_pipe[2] = 2;
395 swizzle_pipe[3] = 3;
396 } else {
397 swizzle_pipe[0] = 0;
398 swizzle_pipe[1] = 2;
399 swizzle_pipe[2] = 3;
400 swizzle_pipe[3] = 1;
401 }
3ce0a23d
JG		 402 break;
403 case 5:
d03f5d59
AD		 404 if (force_no_swizzle) {
405 swizzle_pipe[0] = 0;
406 swizzle_pipe[1] = 1;
407 swizzle_pipe[2] = 2;
408 swizzle_pipe[3] = 3;
409 swizzle_pipe[4] = 4;
410 } else {
411 swizzle_pipe[0] = 0;
412 swizzle_pipe[1] = 2;
413 swizzle_pipe[2] = 4;
414 swizzle_pipe[3] = 1;
415 swizzle_pipe[4] = 3;
416 }
3ce0a23d
JG		 417 break;
418 case 6:
d03f5d59
AD		 419 if (force_no_swizzle) {
420 swizzle_pipe[0] = 0;
421 swizzle_pipe[1] = 1;
422 swizzle_pipe[2] = 2;
423 swizzle_pipe[3] = 3;
424 swizzle_pipe[4] = 4;
425 swizzle_pipe[5] = 5;
426 } else {
427 swizzle_pipe[0] = 0;
428 swizzle_pipe[1] = 2;
429 swizzle_pipe[2] = 4;
430 swizzle_pipe[3] = 5;
431 swizzle_pipe[4] = 3;
432 swizzle_pipe[5] = 1;
433 }
3ce0a23d
JG		 434 break;
435 case 7:
d03f5d59
AD		 436 if (force_no_swizzle) {
437 swizzle_pipe[0] = 0;
438 swizzle_pipe[1] = 1;
439 swizzle_pipe[2] = 2;
440 swizzle_pipe[3] = 3;
441 swizzle_pipe[4] = 4;
442 swizzle_pipe[5] = 5;
443 swizzle_pipe[6] = 6;
444 } else {
445 swizzle_pipe[0] = 0;
446 swizzle_pipe[1] = 2;
447 swizzle_pipe[2] = 4;
448 swizzle_pipe[3] = 6;
449 swizzle_pipe[4] = 3;
450 swizzle_pipe[5] = 1;
451 swizzle_pipe[6] = 5;
452 }
3ce0a23d
JG		 453 break;
454 case 8:
d03f5d59
AD		 455 if (force_no_swizzle) {
456 swizzle_pipe[0] = 0;
457 swizzle_pipe[1] = 1;
458 swizzle_pipe[2] = 2;
459 swizzle_pipe[3] = 3;
460 swizzle_pipe[4] = 4;
461 swizzle_pipe[5] = 5;
462 swizzle_pipe[6] = 6;
463 swizzle_pipe[7] = 7;
464 } else {
465 swizzle_pipe[0] = 0;
466 swizzle_pipe[1] = 2;
467 swizzle_pipe[2] = 4;
468 swizzle_pipe[3] = 6;
469 swizzle_pipe[4] = 3;
470 swizzle_pipe[5] = 1;
471 swizzle_pipe[6] = 7;
472 swizzle_pipe[7] = 5;
473 }
3ce0a23d
JG		 474 break;
475 }
476
477 cur_backend = 0;
478 for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
479 while (((1 << cur_backend) & enabled_backends_mask) == 0)
480 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
481
482 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
483
484 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
485 }
486
487 return backend_map;
771fe6b9
JG		 488}
489
3ce0a23d 490static void rv770_gpu_init(struct radeon_device *rdev)
771fe6b9 491{
3ce0a23d 492 int i, j, num_qd_pipes;
d03f5d59 493 u32 ta_aux_cntl;
3ce0a23d
JG		 494 u32 sx_debug_1;
495 u32 smx_dc_ctl0;
d03f5d59 496 u32 db_debug3;
3ce0a23d
JG		 497 u32 num_gs_verts_per_thread;
498 u32 vgt_gs_per_es;
499 u32 gs_prim_buffer_depth = 0;
500 u32 sq_ms_fifo_sizes;
501 u32 sq_config;
502 u32 sq_thread_resource_mgmt;
503 u32 hdp_host_path_cntl;
504 u32 sq_dyn_gpr_size_simd_ab_0;
505 u32 backend_map;
506 u32 gb_tiling_config = 0;
507 u32 cc_rb_backend_disable = 0;
508 u32 cc_gc_shader_pipe_config = 0;
509 u32 mc_arb_ramcfg;
510 u32 db_debug4;
771fe6b9 511
3ce0a23d
JG		 512 /* setup chip specs */
513 switch (rdev->family) {
514 case CHIP_RV770:
515 rdev->config.rv770.max_pipes = 4;
516 rdev->config.rv770.max_tile_pipes = 8;
517 rdev->config.rv770.max_simds = 10;
518 rdev->config.rv770.max_backends = 4;
519 rdev->config.rv770.max_gprs = 256;
520 rdev->config.rv770.max_threads = 248;
521 rdev->config.rv770.max_stack_entries = 512;
522 rdev->config.rv770.max_hw_contexts = 8;
523 rdev->config.rv770.max_gs_threads = 16 * 2;
524 rdev->config.rv770.sx_max_export_size = 128;
525 rdev->config.rv770.sx_max_export_pos_size = 16;
526 rdev->config.rv770.sx_max_export_smx_size = 112;
527 rdev->config.rv770.sq_num_cf_insts = 2;
528
529 rdev->config.rv770.sx_num_of_sets = 7;
530 rdev->config.rv770.sc_prim_fifo_size = 0xF9;
531 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
532 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
533 break;
534 case CHIP_RV730:
535 rdev->config.rv770.max_pipes = 2;
536 rdev->config.rv770.max_tile_pipes = 4;
537 rdev->config.rv770.max_simds = 8;
538 rdev->config.rv770.max_backends = 2;
539 rdev->config.rv770.max_gprs = 128;
540 rdev->config.rv770.max_threads = 248;
541 rdev->config.rv770.max_stack_entries = 256;
542 rdev->config.rv770.max_hw_contexts = 8;
543 rdev->config.rv770.max_gs_threads = 16 * 2;
544 rdev->config.rv770.sx_max_export_size = 256;
545 rdev->config.rv770.sx_max_export_pos_size = 32;
546 rdev->config.rv770.sx_max_export_smx_size = 224;
547 rdev->config.rv770.sq_num_cf_insts = 2;
548
549 rdev->config.rv770.sx_num_of_sets = 7;
550 rdev->config.rv770.sc_prim_fifo_size = 0xf9;
551 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
552 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
553 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
554 rdev->config.rv770.sx_max_export_pos_size -= 16;
555 rdev->config.rv770.sx_max_export_smx_size += 16;
556 }
557 break;
558 case CHIP_RV710:
559 rdev->config.rv770.max_pipes = 2;
560 rdev->config.rv770.max_tile_pipes = 2;
561 rdev->config.rv770.max_simds = 2;
562 rdev->config.rv770.max_backends = 1;
563 rdev->config.rv770.max_gprs = 256;
564 rdev->config.rv770.max_threads = 192;
565 rdev->config.rv770.max_stack_entries = 256;
566 rdev->config.rv770.max_hw_contexts = 4;
567 rdev->config.rv770.max_gs_threads = 8 * 2;
568 rdev->config.rv770.sx_max_export_size = 128;
569 rdev->config.rv770.sx_max_export_pos_size = 16;
570 rdev->config.rv770.sx_max_export_smx_size = 112;
571 rdev->config.rv770.sq_num_cf_insts = 1;
572
573 rdev->config.rv770.sx_num_of_sets = 7;
574 rdev->config.rv770.sc_prim_fifo_size = 0x40;
575 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
576 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
577 break;
578 case CHIP_RV740:
579 rdev->config.rv770.max_pipes = 4;
580 rdev->config.rv770.max_tile_pipes = 4;
581 rdev->config.rv770.max_simds = 8;
582 rdev->config.rv770.max_backends = 4;
583 rdev->config.rv770.max_gprs = 256;
584 rdev->config.rv770.max_threads = 248;
585 rdev->config.rv770.max_stack_entries = 512;
586 rdev->config.rv770.max_hw_contexts = 8;
587 rdev->config.rv770.max_gs_threads = 16 * 2;
588 rdev->config.rv770.sx_max_export_size = 256;
589 rdev->config.rv770.sx_max_export_pos_size = 32;
590 rdev->config.rv770.sx_max_export_smx_size = 224;
591 rdev->config.rv770.sq_num_cf_insts = 2;
592
593 rdev->config.rv770.sx_num_of_sets = 7;
594 rdev->config.rv770.sc_prim_fifo_size = 0x100;
595 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
596 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
597
598 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
599 rdev->config.rv770.sx_max_export_pos_size -= 16;
600 rdev->config.rv770.sx_max_export_smx_size += 16;
601 }
602 break;
603 default:
604 break;
605 }
606
607 /* Initialize HDP */
608 j = 0;
609 for (i = 0; i < 32; i++) {
610 WREG32((0x2c14 + j), 0x00000000);
611 WREG32((0x2c18 + j), 0x00000000);
612 WREG32((0x2c1c + j), 0x00000000);
613 WREG32((0x2c20 + j), 0x00000000);
614 WREG32((0x2c24 + j), 0x00000000);
615 j += 0x18;
616 }
617
618 WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
619
620 /* setup tiling, simd, pipe config */
621 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
622
623 switch (rdev->config.rv770.max_tile_pipes) {
624 case 1:
d03f5d59 625 default:
3ce0a23d
JG		 626 gb_tiling_config |= PIPE_TILING(0);
627 break;
628 case 2:
629 gb_tiling_config |= PIPE_TILING(1);
630 break;
631 case 4:
632 gb_tiling_config |= PIPE_TILING(2);
633 break;
634 case 8:
635 gb_tiling_config |= PIPE_TILING(3);
3ce0a23d
JG		 636 break;
637 }
d03f5d59 638 rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes;
3ce0a23d
JG		 639
640 if (rdev->family == CHIP_RV770)
641 gb_tiling_config |= BANK_TILING(1);
642 else
e29649db 643 gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
961fb597 644 rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3);
3ce0a23d
JG		 645
646 gb_tiling_config |= GROUP_SIZE(0);
961fb597 647 rdev->config.rv770.tiling_group_size = 256;
3ce0a23d 648
e29649db 649 if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) {
3ce0a23d
JG		 650 gb_tiling_config |= ROW_TILING(3);
651 gb_tiling_config |= SAMPLE_SPLIT(3);
652 } else {
653 gb_tiling_config |=
654 ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
655 gb_tiling_config |=
656 SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
657 }
658
659 gb_tiling_config |= BANK_SWAPS(1);
660
d03f5d59
AD		 661 cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000;
662 cc_rb_backend_disable |=
663 BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK);
3ce0a23d 664
d03f5d59
AD		 665 cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
666 cc_gc_shader_pipe_config |=
3ce0a23d
JG		 667 INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK);
668 cc_gc_shader_pipe_config |=
669 INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK);
670
d03f5d59
AD		 671 if (rdev->family == CHIP_RV740)
672 backend_map = 0x28;
673 else
674 backend_map = r700_get_tile_pipe_to_backend_map(rdev,
675 rdev->config.rv770.max_tile_pipes,
676 (R7XX_MAX_BACKENDS -
677 r600_count_pipe_bits((cc_rb_backend_disable &
678 R7XX_MAX_BACKENDS_MASK) >> 16)),
679 (cc_rb_backend_disable >> 16));
d03f5d59 680
e7aeeba6
AD		 681 rdev->config.rv770.tile_config = gb_tiling_config;
682 gb_tiling_config |= BACKEND_MAP(backend_map);
3ce0a23d
JG		 683
684 WREG32(GB_TILING_CONFIG, gb_tiling_config);
685 WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
686 WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
687
688 WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable);
689 WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
f867c60d 690 WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
d03f5d59 691 WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable);
3ce0a23d 692
3ce0a23d
JG		 693 WREG32(CGTS_SYS_TCC_DISABLE, 0);
694 WREG32(CGTS_TCC_DISABLE, 0);
f867c60d
AD		 695 WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
696 WREG32(CGTS_USER_TCC_DISABLE, 0);
3ce0a23d
JG		 697
698 num_qd_pipes =
d03f5d59 699 R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
3ce0a23d
JG		 700 WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK);
701 WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK);
702
703 /* set HW defaults for 3D engine */
704 WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
e29649db 705 ROQ_IB2_START(0x2b)));
3ce0a23d
JG		 706
707 WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
708
d03f5d59
AD		 709 ta_aux_cntl = RREG32(TA_CNTL_AUX);
710 WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO);
3ce0a23d
JG		 711
712 sx_debug_1 = RREG32(SX_DEBUG_1);
713 sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
714 WREG32(SX_DEBUG_1, sx_debug_1);
715
716 smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
717 smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff);
718 smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1);
719 WREG32(SMX_DC_CTL0, smx_dc_ctl0);
720
d03f5d59
AD		 721 if (rdev->family != CHIP_RV740)
722 WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) |
723 GS_FLUSH_CTL(4) |
724 ACK_FLUSH_CTL(3) |
725 SYNC_FLUSH_CTL));
3ce0a23d 726
d03f5d59
AD		 727 db_debug3 = RREG32(DB_DEBUG3);
728 db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f);
729 switch (rdev->family) {
730 case CHIP_RV770:
731 case CHIP_RV740:
732 db_debug3 |= DB_CLK_OFF_DELAY(0x1f);
733 break;
734 case CHIP_RV710:
735 case CHIP_RV730:
736 default:
737 db_debug3 |= DB_CLK_OFF_DELAY(2);
738 break;
739 }
740 WREG32(DB_DEBUG3, db_debug3);
741
742 if (rdev->family != CHIP_RV770) {
3ce0a23d
JG		 743 db_debug4 = RREG32(DB_DEBUG4);
744 db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER;
745 WREG32(DB_DEBUG4, db_debug4);
746 }
747
748 WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) |
e29649db
AD		 749 POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) |
750 SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1)));
3ce0a23d
JG		 751
752 WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) |
e29649db
AD		 753 SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) |
754 SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize)));
3ce0a23d
JG		 755
756 WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
757
758 WREG32(VGT_NUM_INSTANCES, 1);
759
760 WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0));
761
762 WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
763
764 WREG32(CP_PERFMON_CNTL, 0);
765
766 sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) |
767 DONE_FIFO_HIWATER(0xe0) |
768 ALU_UPDATE_FIFO_HIWATER(0x8));
769 switch (rdev->family) {
770 case CHIP_RV770:
3ce0a23d
JG		 771 case CHIP_RV730:
772 case CHIP_RV710:
d03f5d59
AD		 773 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1);
774 break;
3ce0a23d
JG		 775 case CHIP_RV740:
776 default:
777 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4);
778 break;
779 }
780 WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
781
782 /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
783 * should be adjusted as needed by the 2D/3D drivers. This just sets default values
784 */
785 sq_config = RREG32(SQ_CONFIG);
786 sq_config &= ~(PS_PRIO(3) |
787 VS_PRIO(3) |
788 GS_PRIO(3) |
789 ES_PRIO(3));
790 sq_config |= (DX9_CONSTS |
791 VC_ENABLE |
792 EXPORT_SRC_C |
793 PS_PRIO(0) |
794 VS_PRIO(1) |
795 GS_PRIO(2) |
796 ES_PRIO(3));
797 if (rdev->family == CHIP_RV710)
798 /* no vertex cache */
799 sq_config &= ~VC_ENABLE;
800
801 WREG32(SQ_CONFIG, sq_config);
802
803 WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
fe62e1a4
DA		 804 NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
805 NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2)));
3ce0a23d
JG		 806
807 WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) |
fe62e1a4 808 NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64)));
3ce0a23d
JG		 809
810 sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) |
811 NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) |
812 NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8));
813 if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads)
814 sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads);
815 else
816 sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8);
817 WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
818
819 WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
820 NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
821
822 WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
823 NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
824
825 sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) |
826 SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) |
827 SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) |
828 SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64));
829
830 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
831 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
832 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
833 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
834 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
835 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
836 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
837 WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
838
839 WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
fe62e1a4 840 FORCE_EOV_MAX_REZ_CNT(255)));
3ce0a23d
JG		 841
842 if (rdev->family == CHIP_RV710)
843 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) |
fe62e1a4 844 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
3ce0a23d
JG		 845 else
846 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) |
fe62e1a4 847 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
3ce0a23d
JG		 848
849 switch (rdev->family) {
850 case CHIP_RV770:
851 case CHIP_RV730:
852 case CHIP_RV740:
853 gs_prim_buffer_depth = 384;
854 break;
855 case CHIP_RV710:
856 gs_prim_buffer_depth = 128;
857 break;
858 default:
859 break;
860 }
861
862 num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16;
863 vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
864 /* Max value for this is 256 */
865 if (vgt_gs_per_es > 256)
866 vgt_gs_per_es = 256;
867
868 WREG32(VGT_ES_PER_GS, 128);
869 WREG32(VGT_GS_PER_ES, vgt_gs_per_es);
870 WREG32(VGT_GS_PER_VS, 2);
871
872 /* more default values. 2D/3D driver should adjust as needed */
873 WREG32(VGT_GS_VERTEX_REUSE, 16);
874 WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
875 WREG32(VGT_STRMOUT_EN, 0);
876 WREG32(SX_MISC, 0);
877 WREG32(PA_SC_MODE_CNTL, 0);
878 WREG32(PA_SC_EDGERULE, 0xaaaaaaaa);
879 WREG32(PA_SC_AA_CONFIG, 0);
880 WREG32(PA_SC_CLIPRECT_RULE, 0xffff);
881 WREG32(PA_SC_LINE_STIPPLE, 0);
882 WREG32(SPI_INPUT_Z, 0);
883 WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
884 WREG32(CB_COLOR7_FRAG, 0);
885
886 /* clear render buffer base addresses */
887 WREG32(CB_COLOR0_BASE, 0);
888 WREG32(CB_COLOR1_BASE, 0);
889 WREG32(CB_COLOR2_BASE, 0);
890 WREG32(CB_COLOR3_BASE, 0);
891 WREG32(CB_COLOR4_BASE, 0);
892 WREG32(CB_COLOR5_BASE, 0);
893 WREG32(CB_COLOR6_BASE, 0);
894 WREG32(CB_COLOR7_BASE, 0);
895
896 WREG32(TCP_CNTL, 0);
897
898 hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
899 WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
900
901 WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
902
903 WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
904 NUM_CLIP_SEQ(3)));
905
906}
907
87cbf8f2
AD		 908static int rv770_vram_scratch_init(struct radeon_device *rdev)
909{
910 int r;
911 u64 gpu_addr;
912
913 if (rdev->vram_scratch.robj == NULL) {
914 r = radeon_bo_create(rdev, NULL, RADEON_GPU_PAGE_SIZE,
915 true, RADEON_GEM_DOMAIN_VRAM,
916 &rdev->vram_scratch.robj);
917 if (r) {
918 return r;
919 }
920 }
921
922 r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
923 if (unlikely(r != 0))
924 return r;
925 r = radeon_bo_pin(rdev->vram_scratch.robj,
926 RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
927 if (r) {
928 radeon_bo_unreserve(rdev->vram_scratch.robj);
929 return r;
930 }
931 r = radeon_bo_kmap(rdev->vram_scratch.robj,
932 (void **)&rdev->vram_scratch.ptr);
933 if (r)
934 radeon_bo_unpin(rdev->vram_scratch.robj);
935 radeon_bo_unreserve(rdev->vram_scratch.robj);
936
937 return r;
938}
939
940static void rv770_vram_scratch_fini(struct radeon_device *rdev)
941{
942 int r;
943
944 if (rdev->vram_scratch.robj == NULL) {
945 return;
946 }
947 r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
948 if (likely(r == 0)) {
949 radeon_bo_kunmap(rdev->vram_scratch.robj);
950 radeon_bo_unpin(rdev->vram_scratch.robj);
951 radeon_bo_unreserve(rdev->vram_scratch.robj);
952 }
953 radeon_bo_unref(&rdev->vram_scratch.robj);
954}
955
3ce0a23d
JG		 956int rv770_mc_init(struct radeon_device *rdev)
957{
3ce0a23d 958 u32 tmp;
5885b7a9 959 int chansize, numchan;
3ce0a23d
JG		 960
961 /* Get VRAM informations */
3ce0a23d 962 rdev->mc.vram_is_ddr = true;
5885b7a9
AD		 963 tmp = RREG32(MC_ARB_RAMCFG);
964 if (tmp & CHANSIZE_OVERRIDE) {
965 chansize = 16;
966 } else if (tmp & CHANSIZE_MASK) {
967 chansize = 64;
968 } else {
969 chansize = 32;
970 }
971 tmp = RREG32(MC_SHARED_CHMAP);
972 switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
973 case 0:
974 default:
975 numchan = 1;
976 break;
977 case 1:
978 numchan = 2;
979 break;
980 case 2:
981 numchan = 4;
982 break;
983 case 3:
984 numchan = 8;
985 break;
986 }
987 rdev->mc.vram_width = numchan * chansize;
771fe6b9 988 /* Could aper size report 0 ? */
01d73a69
JC		 989 rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
990 rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
3ce0a23d
JG		 991 /* Setup GPU memory space */
992 rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
993 rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
51e5fcd3 994 rdev->mc.visible_vram_size = rdev->mc.aper_size;
d594e46a 995 r600_vram_gtt_location(rdev, &rdev->mc);
f47299c5
AD		 996 radeon_update_bandwidth_info(rdev);
997
3ce0a23d
JG		 998 return 0;
999}
d594e46a 1000
fc30b8ef 1001static int rv770_startup(struct radeon_device *rdev)
3ce0a23d
JG		 1002{
1003 int r;
1004
779720a3
AD		 1005 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
1006 r = r600_init_microcode(rdev);
1007 if (r) {
1008 DRM_ERROR("Failed to load firmware!\n");
1009 return r;
1010 }
1011 }
1012
a3c1945a 1013 rv770_mc_program(rdev);
1a029b76
JG		 1014 if (rdev->flags & RADEON_IS_AGP) {
1015 rv770_agp_enable(rdev);
1016 } else {
1017 r = rv770_pcie_gart_enable(rdev);
1018 if (r)
1019 return r;
1020 }
87cbf8f2
AD		 1021 r = rv770_vram_scratch_init(rdev);
1022 if (r)
1023 return r;
3ce0a23d 1024 rv770_gpu_init(rdev);
c38c7b64
JG		 1025 r = r600_blit_init(rdev);
1026 if (r) {
1027 r600_blit_fini(rdev);
1028 rdev->asic->copy = NULL;
1029 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
1030 }
ff82f052
JG		 1031 /* pin copy shader into vram */
1032 if (rdev->r600_blit.shader_obj) {
1033 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1034 if (unlikely(r != 0))
1035 return r;
1036 r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM,
1037 &rdev->r600_blit.shader_gpu_addr);
1038 radeon_bo_unreserve(rdev->r600_blit.shader_obj);
7923c615 1039 if (r) {
ff82f052 1040 DRM_ERROR("failed to pin blit object %d\n", r);
7923c615
AD		 1041 return r;
1042 }
1043 }
d8f60cfc 1044 /* Enable IRQ */
d8f60cfc
AD		 1045 r = r600_irq_init(rdev);
1046 if (r) {
1047 DRM_ERROR("radeon: IH init failed (%d).\n", r);
1048 radeon_irq_kms_fini(rdev);
1049 return r;
1050 }
1051 r600_irq_set(rdev);
1052
3ce0a23d
JG		 1053 r = radeon_ring_init(rdev, rdev->cp.ring_size);
1054 if (r)
1055 return r;
1056 r = rv770_cp_load_microcode(rdev);
1057 if (r)
1058 return r;
1059 r = r600_cp_resume(rdev);
1060 if (r)
1061 return r;
81cc35bf
JG		 1062 /* write back buffer are not vital so don't worry about failure */
1063 r600_wb_enable(rdev);
3ce0a23d
JG		 1064 return 0;
1065}
1066
fc30b8ef
DA		 1067int rv770_resume(struct radeon_device *rdev)
1068{
1069 int r;
1070
1a029b76
JG		 1071 /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
1072 * posting will perform necessary task to bring back GPU into good
1073 * shape.
1074 */
fc30b8ef 1075 /* post card */
e7d40b9a 1076 atom_asic_init(rdev->mode_info.atom_context);
fc30b8ef
DA		 1077 /* Initialize clocks */
1078 r = radeon_clocks_init(rdev);
1079 if (r) {
1080 return r;
1081 }
1082
1083 r = rv770_startup(rdev);
1084 if (r) {
1085 DRM_ERROR("r600 startup failed on resume\n");
1086 return r;
1087 }
1088
62a8ea3f 1089 r = r600_ib_test(rdev);
fc30b8ef
DA		 1090 if (r) {
1091 DRM_ERROR("radeon: failled testing IB (%d).\n", r);
1092 return r;
1093 }
8a8c6e7c
RM		 1094
1095 r = r600_audio_init(rdev);
1096 if (r) {
1097 dev_err(rdev->dev, "radeon: audio init failed\n");
1098 return r;
1099 }
1100
fc30b8ef
DA		 1101 return r;
1102
1103}
1104
3ce0a23d
JG		 1105int rv770_suspend(struct radeon_device *rdev)
1106{
4c788679
JG		 1107 int r;
1108
8a8c6e7c 1109 r600_audio_fini(rdev);
3ce0a23d
JG		 1110 /* FIXME: we should wait for ring to be empty */
1111 r700_cp_stop(rdev);
4153e584 1112 rdev->cp.ready = false;
0c45249f 1113 r600_irq_suspend(rdev);
81cc35bf 1114 r600_wb_disable(rdev);
4aac0473 1115 rv770_pcie_gart_disable(rdev);
4153e584 1116 /* unpin shaders bo */
30d2d9a5
JG		 1117 if (rdev->r600_blit.shader_obj) {
1118 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1119 if (likely(r == 0)) {
1120 radeon_bo_unpin(rdev->r600_blit.shader_obj);
1121 radeon_bo_unreserve(rdev->r600_blit.shader_obj);
1122 }
4c788679 1123 }
3ce0a23d
JG		 1124 return 0;
1125}
1126
1127/* Plan is to move initialization in that function and use
1128 * helper function so that radeon_device_init pretty much
1129 * do nothing more than calling asic specific function. This
1130 * should also allow to remove a bunch of callback function
1131 * like vram_info.
1132 */
1133int rv770_init(struct radeon_device *rdev)
1134{
1135 int r;
1136
3ce0a23d
JG		 1137 r = radeon_dummy_page_init(rdev);
1138 if (r)
1139 return r;
1140 /* This don't do much */
1141 r = radeon_gem_init(rdev);
1142 if (r)
1143 return r;
1144 /* Read BIOS */
1145 if (!radeon_get_bios(rdev)) {
1146 if (ASIC_IS_AVIVO(rdev))
1147 return -EINVAL;
1148 }
1149 /* Must be an ATOMBIOS */
e7d40b9a
JG		 1150 if (!rdev->is_atom_bios) {
1151 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
3ce0a23d 1152 return -EINVAL;
e7d40b9a 1153 }
3ce0a23d
JG		 1154 r = radeon_atombios_init(rdev);
1155 if (r)
1156 return r;
1157 /* Post card if necessary */
72542d77
DA		 1158 if (!r600_card_posted(rdev)) {
1159 if (!rdev->bios) {
1160 dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1161 return -EINVAL;
1162 }
3ce0a23d
JG		 1163 DRM_INFO("GPU not posted. posting now...\n");
1164 atom_asic_init(rdev->mode_info.atom_context);
1165 }
1166 /* Initialize scratch registers */
1167 r600_scratch_init(rdev);
1168 /* Initialize surface registers */
1169 radeon_surface_init(rdev);
7433874e 1170 /* Initialize clocks */
5e6dde7e 1171 radeon_get_clock_info(rdev->ddev);
3ce0a23d
JG		 1172 r = radeon_clocks_init(rdev);
1173 if (r)
1174 return r;
1175 /* Fence driver */
1176 r = radeon_fence_driver_init(rdev);
1177 if (r)
1178 return r;
d594e46a 1179 /* initialize AGP */
700a0cc0
JG		 1180 if (rdev->flags & RADEON_IS_AGP) {
1181 r = radeon_agp_init(rdev);
1182 if (r)
1183 radeon_agp_disable(rdev);
1184 }
3ce0a23d 1185 r = rv770_mc_init(rdev);
b574f251 1186 if (r)
3ce0a23d 1187 return r;
3ce0a23d 1188 /* Memory manager */
4c788679 1189 r = radeon_bo_init(rdev);
3ce0a23d
JG		 1190 if (r)
1191 return r;
d8f60cfc
AD		 1192
1193 r = radeon_irq_kms_init(rdev);
1194 if (r)
1195 return r;
1196
3ce0a23d
JG		 1197 rdev->cp.ring_obj = NULL;
1198 r600_ring_init(rdev, 1024 * 1024);
1199
d8f60cfc
AD		 1200 rdev->ih.ring_obj = NULL;
1201 r600_ih_ring_init(rdev, 64 * 1024);
1202
4aac0473
JG		 1203 r = r600_pcie_gart_init(rdev);
1204 if (r)
1205 return r;
1206
779720a3 1207 rdev->accel_working = true;
fc30b8ef 1208 r = rv770_startup(rdev);
3ce0a23d 1209 if (r) {
655efd3d 1210 dev_err(rdev->dev, "disabling GPU acceleration\n");
fe251e2f 1211 r700_cp_fini(rdev);
75c81298 1212 r600_wb_fini(rdev);
655efd3d
JG		 1213 r600_irq_fini(rdev);
1214 radeon_irq_kms_fini(rdev);
75c81298 1215 rv770_pcie_gart_fini(rdev);
733289c2 1216 rdev->accel_working = false;
3ce0a23d 1217 }
733289c2 1218 if (rdev->accel_working) {
733289c2
JG		 1219 r = radeon_ib_pool_init(rdev);
1220 if (r) {
db96380e 1221 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
733289c2 1222 rdev->accel_working = false;
db96380e
JG		 1223 } else {
1224 r = r600_ib_test(rdev);
1225 if (r) {
1226 dev_err(rdev->dev, "IB test failed (%d).\n", r);
1227 rdev->accel_working = false;
1228 }
733289c2 1229 }
3ce0a23d 1230 }
8a8c6e7c
RM		 1231
1232 r = r600_audio_init(rdev);
1233 if (r) {
1234 dev_err(rdev->dev, "radeon: audio init failed\n");
1235 return r;
1236 }
1237
3ce0a23d
JG		 1238 return 0;
1239}
1240
1241void rv770_fini(struct radeon_device *rdev)
1242{
1243 r600_blit_fini(rdev);
fe251e2f 1244 r700_cp_fini(rdev);
655efd3d 1245 r600_wb_fini(rdev);
d8f60cfc
AD		 1246 r600_irq_fini(rdev);
1247 radeon_irq_kms_fini(rdev);
4aac0473 1248 rv770_pcie_gart_fini(rdev);
87cbf8f2 1249 rv770_vram_scratch_fini(rdev);
3ce0a23d
JG		 1250 radeon_gem_fini(rdev);
1251 radeon_fence_driver_fini(rdev);
1252 radeon_clocks_fini(rdev);
d0269ed8 1253 radeon_agp_fini(rdev);
4c788679 1254 radeon_bo_fini(rdev);
e7d40b9a 1255 radeon_atombios_fini(rdev);
3ce0a23d
JG		 1256 kfree(rdev->bios);
1257 rdev->bios = NULL;
1258 radeon_dummy_page_fini(rdev);
771fe6b9 1259}
Clone of davem's net-next-2.6 tree