drivers/gpu/drm/radeon/rv770.c

   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  */
  28 #include <linux/firmware.h>
  29 #include <linux/platform_device.h>
  30 #include <linux/slab.h>
  31 #include "drmP.h"
  32 #include "radeon.h"
  33 #include "radeon_asic.h"
  34 #include "radeon_drm.h"
  35 #include "rv770d.h"
  36 #include "atom.h"
  37 #include "avivod.h"
  38
  39 #define R700_PFP_UCODE_SIZE 848
  40 #define R700_PM4_UCODE_SIZE 1360
  41
  42 static void rv770_gpu_init(struct radeon_device *rdev);
  43 void rv770_fini(struct radeon_device *rdev);
  44
  45 /* get temperature in millidegrees */
  46 u32 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
  60 void rv770_pm_misc(struct radeon_device *rdev)
  61 {
  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;
  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;
  71                         DRM_DEBUG("Setting: v: %d\n", voltage->voltage);
  72                 }
  73         }
  74 }
  75
  76 /*
  77  * GART
  78  */
  79 int rv770_pcie_gart_enable(struct radeon_device *rdev)
  80 {
  81         u32 tmp;
  82         int r, i;
  83
  84         if (rdev->gart.table.vram.robj == NULL) {
  85                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
  86                 return -EINVAL;
  87         }
  88         r = radeon_gart_table_vram_pin(rdev);
  89         if (r)
  90                 return r;
  91         radeon_gart_restore(rdev);
  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);
 111         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
 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);
 119
 120         r600_pcie_gart_tlb_flush(rdev);
 121         rdev->gart.ready = true;
 122         return 0;
 123 }
 124
 125 void rv770_pcie_gart_disable(struct radeon_device *rdev)
 126 {
 127         u32 tmp;
 128         int i, r;
 129
 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);
 148         if (rdev->gart.table.vram.robj) {
 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                 }
 155         }
 156 }
 157
 158 void rv770_pcie_gart_fini(struct radeon_device *rdev)
 159 {
 160         radeon_gart_fini(rdev);
 161         rv770_pcie_gart_disable(rdev);
 162         radeon_gart_table_vram_free(rdev);
 163 }
 164
 165
 166 void 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
 193 static void rv770_mc_program(struct radeon_device *rdev)
 194 {
 195         struct rv515_mc_save save;
 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         }
 207         /* r7xx hw bug.  Read from HDP_DEBUG1 rather
 208          * than writing to HDP_REG_COHERENCY_FLUSH_CNTL
 209          */
 210         tmp = RREG32(HDP_DEBUG1);
 211
 212         rv515_mc_stop(rdev, &save);
 213         if (r600_mc_wait_for_idle(rdev)) {
 214                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 215         }
 216         /* Lockout access through VGA aperture*/
 217         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
 218         /* Update configuration */
 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         }
 239         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 240         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
 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));
 245         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
 246         if (rdev->flags & RADEON_IS_AGP) {
 247                 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16);
 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         }
 255         if (r600_mc_wait_for_idle(rdev)) {
 256                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 257         }
 258         rv515_mc_resume(rdev, &save);
 259         /* we need to own VRAM, so turn off the VGA renderer here
 260          * to stop it overwriting our objects */
 261         rv515_vga_render_disable(rdev);
 262 }
 263
 264
 265 /*
 266  * CP.
 267  */
 268 void r700_cp_stop(struct radeon_device *rdev)
 269 {
 270         rdev->mc.active_vram_size = rdev->mc.visible_vram_size;
 271         WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
 272 }
 273
 274 static int rv770_cp_load_microcode(struct radeon_device *rdev)
 275 {
 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;
 306 }
 307
 308 void r700_cp_fini(struct radeon_device *rdev)
 309 {
 310         r700_cp_stop(rdev);
 311         radeon_ring_fini(rdev);
 312 }
 313
 314 /*
 315  * Core functions
 316  */
 317 static 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)
 321 {
 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;
 329         bool force_no_swizzle;
 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
 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
 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:
 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                 }
 390                 break;
 391         case 4:
 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                 }
 403                 break;
 404         case 5:
 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                 }
 418                 break;
 419         case 6:
 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                 }
 435                 break;
 436         case 7:
 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                 }
 454                 break;
 455         case 8:
 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                 }
 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;
 489 }
 490
 491 static void rv770_gpu_init(struct radeon_device *rdev)
 492 {
 493         int i, j, num_qd_pipes;
 494         u32 ta_aux_cntl;
 495         u32 sx_debug_1;
 496         u32 smx_dc_ctl0;
 497         u32 db_debug3;
 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;
 512
 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:
 626         default:
 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);
 637                 break;
 638         }
 639         rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes;
 640
 641         if (rdev->family == CHIP_RV770)
 642                 gb_tiling_config |= BANK_TILING(1);
 643         else
 644                 gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
 645         rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3);
 646         gb_tiling_config |= GROUP_SIZE((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT);
 647         if ((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT)
 648                 rdev->config.rv770.tiling_group_size = 512;
 649         else
 650                 rdev->config.rv770.tiling_group_size = 256;
 651         if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) {
 652                 gb_tiling_config |= ROW_TILING(3);
 653                 gb_tiling_config |= SAMPLE_SPLIT(3);
 654         } else {
 655                 gb_tiling_config |=
 656                         ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 657                 gb_tiling_config |=
 658                         SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 659         }
 660
 661         gb_tiling_config |= BANK_SWAPS(1);
 662
 663         cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000;
 664         cc_rb_backend_disable |=
 665                 BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK);
 666
 667         cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
 668         cc_gc_shader_pipe_config |=
 669                 INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK);
 670         cc_gc_shader_pipe_config |=
 671                 INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK);
 672
 673         if (rdev->family == CHIP_RV740)
 674                 backend_map = 0x28;
 675         else
 676                 backend_map = r700_get_tile_pipe_to_backend_map(rdev,
 677                                                                 rdev->config.rv770.max_tile_pipes,
 678                                                                 (R7XX_MAX_BACKENDS -
 679                                                                  r600_count_pipe_bits((cc_rb_backend_disable &
 680                                                                                        R7XX_MAX_BACKENDS_MASK) >> 16)),
 681                                                                 (cc_rb_backend_disable >> 16));
 682
 683         rdev->config.rv770.tile_config = gb_tiling_config;
 684         gb_tiling_config |= BACKEND_MAP(backend_map);
 685
 686         WREG32(GB_TILING_CONFIG, gb_tiling_config);
 687         WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 688         WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 689
 690         WREG32(CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
 691         WREG32(CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
 692         WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
 693         WREG32(CC_SYS_RB_BACKEND_DISABLE,  cc_rb_backend_disable);
 694
 695         WREG32(CGTS_SYS_TCC_DISABLE, 0);
 696         WREG32(CGTS_TCC_DISABLE, 0);
 697         WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
 698         WREG32(CGTS_USER_TCC_DISABLE, 0);
 699
 700         num_qd_pipes =
 701                 R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
 702         WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK);
 703         WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK);
 704
 705         /* set HW defaults for 3D engine */
 706         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
 707                                      ROQ_IB2_START(0x2b)));
 708
 709         WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
 710
 711         ta_aux_cntl = RREG32(TA_CNTL_AUX);
 712         WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO);
 713
 714         sx_debug_1 = RREG32(SX_DEBUG_1);
 715         sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
 716         WREG32(SX_DEBUG_1, sx_debug_1);
 717
 718         smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
 719         smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff);
 720         smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1);
 721         WREG32(SMX_DC_CTL0, smx_dc_ctl0);
 722
 723         if (rdev->family != CHIP_RV740)
 724                 WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) |
 725                                        GS_FLUSH_CTL(4) |
 726                                        ACK_FLUSH_CTL(3) |
 727                                        SYNC_FLUSH_CTL));
 728
 729         db_debug3 = RREG32(DB_DEBUG3);
 730         db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f);
 731         switch (rdev->family) {
 732         case CHIP_RV770:
 733         case CHIP_RV740:
 734                 db_debug3 |= DB_CLK_OFF_DELAY(0x1f);
 735                 break;
 736         case CHIP_RV710:
 737         case CHIP_RV730:
 738         default:
 739                 db_debug3 |= DB_CLK_OFF_DELAY(2);
 740                 break;
 741         }
 742         WREG32(DB_DEBUG3, db_debug3);
 743
 744         if (rdev->family != CHIP_RV770) {
 745                 db_debug4 = RREG32(DB_DEBUG4);
 746                 db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER;
 747                 WREG32(DB_DEBUG4, db_debug4);
 748         }
 749
 750         WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) |
 751                                         POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) |
 752                                         SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1)));
 753
 754         WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) |
 755                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) |
 756                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize)));
 757
 758         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 759
 760         WREG32(VGT_NUM_INSTANCES, 1);
 761
 762         WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0));
 763
 764         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
 765
 766         WREG32(CP_PERFMON_CNTL, 0);
 767
 768         sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) |
 769                             DONE_FIFO_HIWATER(0xe0) |
 770                             ALU_UPDATE_FIFO_HIWATER(0x8));
 771         switch (rdev->family) {
 772         case CHIP_RV770:
 773         case CHIP_RV730:
 774         case CHIP_RV710:
 775                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1);
 776                 break;
 777         case CHIP_RV740:
 778         default:
 779                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4);
 780                 break;
 781         }
 782         WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 783
 784         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 785          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 786          */
 787         sq_config = RREG32(SQ_CONFIG);
 788         sq_config &= ~(PS_PRIO(3) |
 789                        VS_PRIO(3) |
 790                        GS_PRIO(3) |
 791                        ES_PRIO(3));
 792         sq_config |= (DX9_CONSTS |
 793                       VC_ENABLE |
 794                       EXPORT_SRC_C |
 795                       PS_PRIO(0) |
 796                       VS_PRIO(1) |
 797                       GS_PRIO(2) |
 798                       ES_PRIO(3));
 799         if (rdev->family == CHIP_RV710)
 800                 /* no vertex cache */
 801                 sq_config &= ~VC_ENABLE;
 802
 803         WREG32(SQ_CONFIG, sq_config);
 804
 805         WREG32(SQ_GPR_RESOURCE_MGMT_1,  (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 806                                          NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 807                                          NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2)));
 808
 809         WREG32(SQ_GPR_RESOURCE_MGMT_2,  (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) |
 810                                          NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64)));
 811
 812         sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) |
 813                                    NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) |
 814                                    NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8));
 815         if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads)
 816                 sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads);
 817         else
 818                 sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8);
 819         WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
 820
 821         WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 822                                                      NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 823
 824         WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 825                                                      NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 826
 827         sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 828                                      SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) |
 829                                      SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 830                                      SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64));
 831
 832         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
 833         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
 834         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
 835         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
 836         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
 837         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
 838         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
 839         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
 840
 841         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
 842                                           FORCE_EOV_MAX_REZ_CNT(255)));
 843
 844         if (rdev->family == CHIP_RV710)
 845                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) |
 846                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 847         else
 848                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) |
 849                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 850
 851         switch (rdev->family) {
 852         case CHIP_RV770:
 853         case CHIP_RV730:
 854         case CHIP_RV740:
 855                 gs_prim_buffer_depth = 384;
 856                 break;
 857         case CHIP_RV710:
 858                 gs_prim_buffer_depth = 128;
 859                 break;
 860         default:
 861                 break;
 862         }
 863
 864         num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16;
 865         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
 866         /* Max value for this is 256 */
 867         if (vgt_gs_per_es > 256)
 868                 vgt_gs_per_es = 256;
 869
 870         WREG32(VGT_ES_PER_GS, 128);
 871         WREG32(VGT_GS_PER_ES, vgt_gs_per_es);
 872         WREG32(VGT_GS_PER_VS, 2);
 873
 874         /* more default values. 2D/3D driver should adjust as needed */
 875         WREG32(VGT_GS_VERTEX_REUSE, 16);
 876         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
 877         WREG32(VGT_STRMOUT_EN, 0);
 878         WREG32(SX_MISC, 0);
 879         WREG32(PA_SC_MODE_CNTL, 0);
 880         WREG32(PA_SC_EDGERULE, 0xaaaaaaaa);
 881         WREG32(PA_SC_AA_CONFIG, 0);
 882         WREG32(PA_SC_CLIPRECT_RULE, 0xffff);
 883         WREG32(PA_SC_LINE_STIPPLE, 0);
 884         WREG32(SPI_INPUT_Z, 0);
 885         WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
 886         WREG32(CB_COLOR7_FRAG, 0);
 887
 888         /* clear render buffer base addresses */
 889         WREG32(CB_COLOR0_BASE, 0);
 890         WREG32(CB_COLOR1_BASE, 0);
 891         WREG32(CB_COLOR2_BASE, 0);
 892         WREG32(CB_COLOR3_BASE, 0);
 893         WREG32(CB_COLOR4_BASE, 0);
 894         WREG32(CB_COLOR5_BASE, 0);
 895         WREG32(CB_COLOR6_BASE, 0);
 896         WREG32(CB_COLOR7_BASE, 0);
 897
 898         WREG32(TCP_CNTL, 0);
 899
 900         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
 901         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
 902
 903         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 904
 905         WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
 906                                           NUM_CLIP_SEQ(3)));
 907
 908 }
 909
 910 static int rv770_vram_scratch_init(struct radeon_device *rdev)
 911 {
 912         int r;
 913         u64 gpu_addr;
 914
 915         if (rdev->vram_scratch.robj == NULL) {
 916                 r = radeon_bo_create(rdev, NULL, RADEON_GPU_PAGE_SIZE,
 917                                         true, RADEON_GEM_DOMAIN_VRAM,
 918                                         &rdev->vram_scratch.robj);
 919                 if (r) {
 920                         return r;
 921                 }
 922         }
 923
 924         r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
 925         if (unlikely(r != 0))
 926                 return r;
 927         r = radeon_bo_pin(rdev->vram_scratch.robj,
 928                           RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
 929         if (r) {
 930                 radeon_bo_unreserve(rdev->vram_scratch.robj);
 931                 return r;
 932         }
 933         r = radeon_bo_kmap(rdev->vram_scratch.robj,
 934                                 (void **)&rdev->vram_scratch.ptr);
 935         if (r)
 936                 radeon_bo_unpin(rdev->vram_scratch.robj);
 937         radeon_bo_unreserve(rdev->vram_scratch.robj);
 938
 939         return r;
 940 }
 941
 942 static void rv770_vram_scratch_fini(struct radeon_device *rdev)
 943 {
 944         int r;
 945
 946         if (rdev->vram_scratch.robj == NULL) {
 947                 return;
 948         }
 949         r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
 950         if (likely(r == 0)) {
 951                 radeon_bo_kunmap(rdev->vram_scratch.robj);
 952                 radeon_bo_unpin(rdev->vram_scratch.robj);
 953                 radeon_bo_unreserve(rdev->vram_scratch.robj);
 954         }
 955         radeon_bo_unref(&rdev->vram_scratch.robj);
 956 }
 957
 958 int rv770_mc_init(struct radeon_device *rdev)
 959 {
 960         u32 tmp;
 961         int chansize, numchan;
 962
 963         /* Get VRAM informations */
 964         rdev->mc.vram_is_ddr = true;
 965         tmp = RREG32(MC_ARB_RAMCFG);
 966         if (tmp & CHANSIZE_OVERRIDE) {
 967                 chansize = 16;
 968         } else if (tmp & CHANSIZE_MASK) {
 969                 chansize = 64;
 970         } else {
 971                 chansize = 32;
 972         }
 973         tmp = RREG32(MC_SHARED_CHMAP);
 974         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
 975         case 0:
 976         default:
 977                 numchan = 1;
 978                 break;
 979         case 1:
 980                 numchan = 2;
 981                 break;
 982         case 2:
 983                 numchan = 4;
 984                 break;
 985         case 3:
 986                 numchan = 8;
 987                 break;
 988         }
 989         rdev->mc.vram_width = numchan * chansize;
 990         /* Could aper size report 0 ? */
 991         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
 992         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
 993         /* Setup GPU memory space */
 994         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
 995         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
 996         rdev->mc.visible_vram_size = rdev->mc.aper_size;
 997         rdev->mc.active_vram_size = rdev->mc.visible_vram_size;
 998         r600_vram_gtt_location(rdev, &rdev->mc);
 999         radeon_update_bandwidth_info(rdev);
1000
1001         return 0;
1002 }
1003
1004 static int rv770_startup(struct radeon_device *rdev)
1005 {
1006         int r;
1007
1008         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
1009                 r = r600_init_microcode(rdev);
1010                 if (r) {
1011                         DRM_ERROR("Failed to load firmware!\n");
1012                         return r;
1013                 }
1014         }
1015
1016         rv770_mc_program(rdev);
1017         if (rdev->flags & RADEON_IS_AGP) {
1018                 rv770_agp_enable(rdev);
1019         } else {
1020                 r = rv770_pcie_gart_enable(rdev);
1021                 if (r)
1022                         return r;
1023         }
1024         r = rv770_vram_scratch_init(rdev);
1025         if (r)
1026                 return r;
1027         rv770_gpu_init(rdev);
1028         r = r600_blit_init(rdev);
1029         if (r) {
1030                 r600_blit_fini(rdev);
1031                 rdev->asic->copy = NULL;
1032                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
1033         }
1034         /* pin copy shader into vram */
1035         if (rdev->r600_blit.shader_obj) {
1036                 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1037                 if (unlikely(r != 0))
1038                         return r;
1039                 r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM,
1040                                 &rdev->r600_blit.shader_gpu_addr);
1041                 radeon_bo_unreserve(rdev->r600_blit.shader_obj);
1042                 if (r) {
1043                         DRM_ERROR("failed to pin blit object %d\n", r);
1044                         return r;
1045                 }
1046         }
1047         /* Enable IRQ */
1048         r = r600_irq_init(rdev);
1049         if (r) {
1050                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
1051                 radeon_irq_kms_fini(rdev);
1052                 return r;
1053         }
1054         r600_irq_set(rdev);
1055
1056         r = radeon_ring_init(rdev, rdev->cp.ring_size);
1057         if (r)
1058                 return r;
1059         r = rv770_cp_load_microcode(rdev);
1060         if (r)
1061                 return r;
1062         r = r600_cp_resume(rdev);
1063         if (r)
1064                 return r;
1065         /* write back buffer are not vital so don't worry about failure */
1066         r600_wb_enable(rdev);
1067         return 0;
1068 }
1069
1070 int rv770_resume(struct radeon_device *rdev)
1071 {
1072         int r;
1073
1074         /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
1075          * posting will perform necessary task to bring back GPU into good
1076          * shape.
1077          */
1078         /* post card */
1079         atom_asic_init(rdev->mode_info.atom_context);
1080
1081         r = rv770_startup(rdev);
1082         if (r) {
1083                 DRM_ERROR("r600 startup failed on resume\n");
1084                 return r;
1085         }
1086
1087         r = r600_ib_test(rdev);
1088         if (r) {
1089                 DRM_ERROR("radeon: failled testing IB (%d).\n", r);
1090                 return r;
1091         }
1092
1093         r = r600_audio_init(rdev);
1094         if (r) {
1095                 dev_err(rdev->dev, "radeon: audio init failed\n");
1096                 return r;
1097         }
1098
1099         return r;
1100
1101 }
1102
1103 int rv770_suspend(struct radeon_device *rdev)
1104 {
1105         int r;
1106
1107         r600_audio_fini(rdev);
1108         /* FIXME: we should wait for ring to be empty */
1109         r700_cp_stop(rdev);
1110         rdev->cp.ready = false;
1111         r600_irq_suspend(rdev);
1112         r600_wb_disable(rdev);
1113         rv770_pcie_gart_disable(rdev);
1114         /* unpin shaders bo */
1115         if (rdev->r600_blit.shader_obj) {
1116                 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1117                 if (likely(r == 0)) {
1118                         radeon_bo_unpin(rdev->r600_blit.shader_obj);
1119                         radeon_bo_unreserve(rdev->r600_blit.shader_obj);
1120                 }
1121         }
1122         return 0;
1123 }
1124
1125 /* Plan is to move initialization in that function and use
1126  * helper function so that radeon_device_init pretty much
1127  * do nothing more than calling asic specific function. This
1128  * should also allow to remove a bunch of callback function
1129  * like vram_info.
1130  */
1131 int rv770_init(struct radeon_device *rdev)
1132 {
1133         int r;
1134
1135         r = radeon_dummy_page_init(rdev);
1136         if (r)
1137                 return r;
1138         /* This don't do much */
1139         r = radeon_gem_init(rdev);
1140         if (r)
1141                 return r;
1142         /* Read BIOS */
1143         if (!radeon_get_bios(rdev)) {
1144                 if (ASIC_IS_AVIVO(rdev))
1145                         return -EINVAL;
1146         }
1147         /* Must be an ATOMBIOS */
1148         if (!rdev->is_atom_bios) {
1149                 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
1150                 return -EINVAL;
1151         }
1152         r = radeon_atombios_init(rdev);
1153         if (r)
1154                 return r;
1155         /* Post card if necessary */
1156         if (!r600_card_posted(rdev)) {
1157                 if (!rdev->bios) {
1158                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1159                         return -EINVAL;
1160                 }
1161                 DRM_INFO("GPU not posted. posting now...\n");
1162                 atom_asic_init(rdev->mode_info.atom_context);
1163         }
1164         /* Initialize scratch registers */
1165         r600_scratch_init(rdev);
1166         /* Initialize surface registers */
1167         radeon_surface_init(rdev);
1168         /* Initialize clocks */
1169         radeon_get_clock_info(rdev->ddev);
1170         /* Fence driver */
1171         r = radeon_fence_driver_init(rdev);
1172         if (r)
1173                 return r;
1174         /* initialize AGP */
1175         if (rdev->flags & RADEON_IS_AGP) {
1176                 r = radeon_agp_init(rdev);
1177                 if (r)
1178                         radeon_agp_disable(rdev);
1179         }
1180         r = rv770_mc_init(rdev);
1181         if (r)
1182                 return r;
1183         /* Memory manager */
1184         r = radeon_bo_init(rdev);
1185         if (r)
1186                 return r;
1187
1188         r = radeon_irq_kms_init(rdev);
1189         if (r)
1190                 return r;
1191
1192         rdev->cp.ring_obj = NULL;
1193         r600_ring_init(rdev, 1024 * 1024);
1194
1195         rdev->ih.ring_obj = NULL;
1196         r600_ih_ring_init(rdev, 64 * 1024);
1197
1198         r = r600_pcie_gart_init(rdev);
1199         if (r)
1200                 return r;
1201
1202         rdev->accel_working = true;
1203         r = rv770_startup(rdev);
1204         if (r) {
1205                 dev_err(rdev->dev, "disabling GPU acceleration\n");
1206                 r700_cp_fini(rdev);
1207                 r600_wb_fini(rdev);
1208                 r600_irq_fini(rdev);
1209                 radeon_irq_kms_fini(rdev);
1210                 rv770_pcie_gart_fini(rdev);
1211                 rdev->accel_working = false;
1212         }
1213         if (rdev->accel_working) {
1214                 r = radeon_ib_pool_init(rdev);
1215                 if (r) {
1216                         dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
1217                         rdev->accel_working = false;
1218                 } else {
1219                         r = r600_ib_test(rdev);
1220                         if (r) {
1221                                 dev_err(rdev->dev, "IB test failed (%d).\n", r);
1222                                 rdev->accel_working = false;
1223                         }
1224                 }
1225         }
1226
1227         r = r600_audio_init(rdev);
1228         if (r) {
1229                 dev_err(rdev->dev, "radeon: audio init failed\n");
1230                 return r;
1231         }
1232
1233         return 0;
1234 }
1235
1236 void rv770_fini(struct radeon_device *rdev)
1237 {
1238         r600_blit_fini(rdev);
1239         r700_cp_fini(rdev);
1240         r600_wb_fini(rdev);
1241         r600_irq_fini(rdev);
1242         radeon_irq_kms_fini(rdev);
1243         rv770_pcie_gart_fini(rdev);
1244         rv770_vram_scratch_fini(rdev);
1245         radeon_gem_fini(rdev);
1246         radeon_fence_driver_fini(rdev);
1247         radeon_agp_fini(rdev);
1248         radeon_bo_fini(rdev);
1249         radeon_atombios_fini(rdev);
1250         kfree(rdev->bios);
1251         rdev->bios = NULL;
1252         radeon_dummy_page_fini(rdev);
1253 }