drivers/gpu/drm/radeon/r600_cp.c

   1 /*
   2  * Copyright 2008-2009 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * 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) AND/OR ITS SUPPLIERS 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 OTHER
  22  * DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *     Dave Airlie <airlied@redhat.com>
  26  *     Alex Deucher <alexander.deucher@amd.com>
  27  */
  28
  29 #include "drmP.h"
  30 #include "drm.h"
  31 #include "radeon_drm.h"
  32 #include "radeon_drv.h"
  33
  34 #include "r600_microcode.h"
  35
  36 # define ATI_PCIGART_PAGE_SIZE          4096    /**< PCI GART page size */
  37 # define ATI_PCIGART_PAGE_MASK          (~(ATI_PCIGART_PAGE_SIZE-1))
  38
  39 #define R600_PTE_VALID     (1 << 0)
  40 #define R600_PTE_SYSTEM    (1 << 1)
  41 #define R600_PTE_SNOOPED   (1 << 2)
  42 #define R600_PTE_READABLE  (1 << 5)
  43 #define R600_PTE_WRITEABLE (1 << 6)
  44
  45 /* MAX values used for gfx init */
  46 #define R6XX_MAX_SH_GPRS           256
  47 #define R6XX_MAX_TEMP_GPRS         16
  48 #define R6XX_MAX_SH_THREADS        256
  49 #define R6XX_MAX_SH_STACK_ENTRIES  4096
  50 #define R6XX_MAX_BACKENDS          8
  51 #define R6XX_MAX_BACKENDS_MASK     0xff
  52 #define R6XX_MAX_SIMDS             8
  53 #define R6XX_MAX_SIMDS_MASK        0xff
  54 #define R6XX_MAX_PIPES             8
  55 #define R6XX_MAX_PIPES_MASK        0xff
  56
  57 #define R7XX_MAX_SH_GPRS           256
  58 #define R7XX_MAX_TEMP_GPRS         16
  59 #define R7XX_MAX_SH_THREADS        256
  60 #define R7XX_MAX_SH_STACK_ENTRIES  4096
  61 #define R7XX_MAX_BACKENDS          8
  62 #define R7XX_MAX_BACKENDS_MASK     0xff
  63 #define R7XX_MAX_SIMDS             16
  64 #define R7XX_MAX_SIMDS_MASK        0xffff
  65 #define R7XX_MAX_PIPES             8
  66 #define R7XX_MAX_PIPES_MASK        0xff
  67
  68 static int r600_do_wait_for_fifo(drm_radeon_private_t *dev_priv, int entries)
  69 {
  70         int i;
  71
  72         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
  73
  74         for (i = 0; i < dev_priv->usec_timeout; i++) {
  75                 int slots;
  76                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
  77                         slots = (RADEON_READ(R600_GRBM_STATUS)
  78                                  & R700_CMDFIFO_AVAIL_MASK);
  79                 else
  80                         slots = (RADEON_READ(R600_GRBM_STATUS)
  81                                  & R600_CMDFIFO_AVAIL_MASK);
  82                 if (slots >= entries)
  83                         return 0;
  84                 DRM_UDELAY(1);
  85         }
  86         DRM_INFO("wait for fifo failed status : 0x%08X 0x%08X\n",
  87                  RADEON_READ(R600_GRBM_STATUS),
  88                  RADEON_READ(R600_GRBM_STATUS2));
  89
  90         return -EBUSY;
  91 }
  92
  93 static int r600_do_wait_for_idle(drm_radeon_private_t *dev_priv)
  94 {
  95         int i, ret;
  96
  97         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
  98
  99         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 100                 ret = r600_do_wait_for_fifo(dev_priv, 8);
 101         else
 102                 ret = r600_do_wait_for_fifo(dev_priv, 16);
 103         if (ret)
 104                 return ret;
 105         for (i = 0; i < dev_priv->usec_timeout; i++) {
 106                 if (!(RADEON_READ(R600_GRBM_STATUS) & R600_GUI_ACTIVE))
 107                         return 0;
 108                 DRM_UDELAY(1);
 109         }
 110         DRM_INFO("wait idle failed status : 0x%08X 0x%08X\n",
 111                  RADEON_READ(R600_GRBM_STATUS),
 112                  RADEON_READ(R600_GRBM_STATUS2));
 113
 114         return -EBUSY;
 115 }
 116
 117 void r600_page_table_cleanup(struct drm_device *dev, struct drm_ati_pcigart_info *gart_info)
 118 {
 119         struct drm_sg_mem *entry = dev->sg;
 120         int max_pages;
 121         int pages;
 122         int i;
 123
 124         if (!entry)
 125                 return;
 126
 127         if (gart_info->bus_addr) {
 128                 max_pages = (gart_info->table_size / sizeof(u64));
 129                 pages = (entry->pages <= max_pages)
 130                   ? entry->pages : max_pages;
 131
 132                 for (i = 0; i < pages; i++) {
 133                         if (!entry->busaddr[i])
 134                                 break;
 135                         pci_unmap_page(dev->pdev, entry->busaddr[i],
 136                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 137                 }
 138                 if (gart_info->gart_table_location == DRM_ATI_GART_MAIN)
 139                         gart_info->bus_addr = 0;
 140         }
 141 }
 142
 143 /* R600 has page table setup */
 144 int r600_page_table_init(struct drm_device *dev)
 145 {
 146         drm_radeon_private_t *dev_priv = dev->dev_private;
 147         struct drm_ati_pcigart_info *gart_info = &dev_priv->gart_info;
 148         struct drm_local_map *map = &gart_info->mapping;
 149         struct drm_sg_mem *entry = dev->sg;
 150         int ret = 0;
 151         int i, j;
 152         int pages;
 153         u64 page_base;
 154         dma_addr_t entry_addr;
 155         int max_ati_pages, max_real_pages, gart_idx;
 156
 157         /* okay page table is available - lets rock */
 158         max_ati_pages = (gart_info->table_size / sizeof(u64));
 159         max_real_pages = max_ati_pages / (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE);
 160
 161         pages = (entry->pages <= max_real_pages) ?
 162                 entry->pages : max_real_pages;
 163
 164         memset_io((void __iomem *)map->handle, 0, max_ati_pages * sizeof(u64));
 165
 166         gart_idx = 0;
 167         for (i = 0; i < pages; i++) {
 168                 entry->busaddr[i] = pci_map_page(dev->pdev,
 169                                                  entry->pagelist[i], 0,
 170                                                  PAGE_SIZE,
 171                                                  PCI_DMA_BIDIRECTIONAL);
 172                 if (entry->busaddr[i] == 0) {
 173                         DRM_ERROR("unable to map PCIGART pages!\n");
 174                         r600_page_table_cleanup(dev, gart_info);
 175                         goto done;
 176                 }
 177                 entry_addr = entry->busaddr[i];
 178                 for (j = 0; j < (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE); j++) {
 179                         page_base = (u64) entry_addr & ATI_PCIGART_PAGE_MASK;
 180                         page_base |= R600_PTE_VALID | R600_PTE_SYSTEM | R600_PTE_SNOOPED;
 181                         page_base |= R600_PTE_READABLE | R600_PTE_WRITEABLE;
 182
 183                         DRM_WRITE64(map, gart_idx * sizeof(u64), page_base);
 184
 185                         gart_idx++;
 186
 187                         if ((i % 128) == 0)
 188                                 DRM_DEBUG("page entry %d: 0x%016llx\n",
 189                                     i, (unsigned long long)page_base);
 190                         entry_addr += ATI_PCIGART_PAGE_SIZE;
 191                 }
 192         }
 193         ret = 1;
 194 done:
 195         return ret;
 196 }
 197
 198 static void r600_vm_flush_gart_range(struct drm_device *dev)
 199 {
 200         drm_radeon_private_t *dev_priv = dev->dev_private;
 201         u32 resp, countdown = 1000;
 202         RADEON_WRITE(R600_VM_CONTEXT0_INVALIDATION_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 203         RADEON_WRITE(R600_VM_CONTEXT0_INVALIDATION_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 204         RADEON_WRITE(R600_VM_CONTEXT0_REQUEST_RESPONSE, 2);
 205
 206         do {
 207                 resp = RADEON_READ(R600_VM_CONTEXT0_REQUEST_RESPONSE);
 208                 countdown--;
 209                 DRM_UDELAY(1);
 210         } while (((resp & 0xf0) == 0) && countdown);
 211 }
 212
 213 static void r600_vm_init(struct drm_device *dev)
 214 {
 215         drm_radeon_private_t *dev_priv = dev->dev_private;
 216         /* initialise the VM to use the page table we constructed up there */
 217         u32 vm_c0, i;
 218         u32 mc_rd_a;
 219         u32 vm_l2_cntl, vm_l2_cntl3;
 220         /* okay set up the PCIE aperture type thingo */
 221         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 222         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 223         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 224
 225         /* setup MC RD a */
 226         mc_rd_a = R600_MCD_L1_TLB | R600_MCD_L1_FRAG_PROC | R600_MCD_SYSTEM_ACCESS_MODE_IN_SYS |
 227                 R600_MCD_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | R600_MCD_EFFECTIVE_L1_TLB_SIZE(5) |
 228                 R600_MCD_EFFECTIVE_L1_QUEUE_SIZE(5) | R600_MCD_WAIT_L2_QUERY;
 229
 230         RADEON_WRITE(R600_MCD_RD_A_CNTL, mc_rd_a);
 231         RADEON_WRITE(R600_MCD_RD_B_CNTL, mc_rd_a);
 232
 233         RADEON_WRITE(R600_MCD_WR_A_CNTL, mc_rd_a);
 234         RADEON_WRITE(R600_MCD_WR_B_CNTL, mc_rd_a);
 235
 236         RADEON_WRITE(R600_MCD_RD_GFX_CNTL, mc_rd_a);
 237         RADEON_WRITE(R600_MCD_WR_GFX_CNTL, mc_rd_a);
 238
 239         RADEON_WRITE(R600_MCD_RD_SYS_CNTL, mc_rd_a);
 240         RADEON_WRITE(R600_MCD_WR_SYS_CNTL, mc_rd_a);
 241
 242         RADEON_WRITE(R600_MCD_RD_HDP_CNTL, mc_rd_a | R600_MCD_L1_STRICT_ORDERING);
 243         RADEON_WRITE(R600_MCD_WR_HDP_CNTL, mc_rd_a /*| R600_MCD_L1_STRICT_ORDERING*/);
 244
 245         RADEON_WRITE(R600_MCD_RD_PDMA_CNTL, mc_rd_a);
 246         RADEON_WRITE(R600_MCD_WR_PDMA_CNTL, mc_rd_a);
 247
 248         RADEON_WRITE(R600_MCD_RD_SEM_CNTL, mc_rd_a | R600_MCD_SEMAPHORE_MODE);
 249         RADEON_WRITE(R600_MCD_WR_SEM_CNTL, mc_rd_a);
 250
 251         vm_l2_cntl = R600_VM_L2_CACHE_EN | R600_VM_L2_FRAG_PROC | R600_VM_ENABLE_PTE_CACHE_LRU_W;
 252         vm_l2_cntl |= R600_VM_L2_CNTL_QUEUE_SIZE(7);
 253         RADEON_WRITE(R600_VM_L2_CNTL, vm_l2_cntl);
 254
 255         RADEON_WRITE(R600_VM_L2_CNTL2, 0);
 256         vm_l2_cntl3 = (R600_VM_L2_CNTL3_BANK_SELECT_0(0) |
 257                        R600_VM_L2_CNTL3_BANK_SELECT_1(1) |
 258                        R600_VM_L2_CNTL3_CACHE_UPDATE_MODE(2));
 259         RADEON_WRITE(R600_VM_L2_CNTL3, vm_l2_cntl3);
 260
 261         vm_c0 = R600_VM_ENABLE_CONTEXT | R600_VM_PAGE_TABLE_DEPTH_FLAT;
 262
 263         RADEON_WRITE(R600_VM_CONTEXT0_CNTL, vm_c0);
 264
 265         vm_c0 &= ~R600_VM_ENABLE_CONTEXT;
 266
 267         /* disable all other contexts */
 268         for (i = 1; i < 8; i++)
 269                 RADEON_WRITE(R600_VM_CONTEXT0_CNTL + (i * 4), vm_c0);
 270
 271         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, dev_priv->gart_info.bus_addr >> 12);
 272         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_START_ADDR, dev_priv->gart_vm_start >> 12);
 273         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_END_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 274
 275         r600_vm_flush_gart_range(dev);
 276 }
 277
 278 /* load r600 microcode */
 279 static void r600_cp_load_microcode(drm_radeon_private_t *dev_priv)
 280 {
 281         int i;
 282
 283         r600_do_cp_stop(dev_priv);
 284
 285         RADEON_WRITE(R600_CP_RB_CNTL,
 286                      R600_RB_NO_UPDATE |
 287                      R600_RB_BLKSZ(15) |
 288                      R600_RB_BUFSZ(3));
 289
 290         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
 291         RADEON_READ(R600_GRBM_SOFT_RESET);
 292         DRM_UDELAY(15000);
 293         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 294
 295         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 296
 297         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600)) {
 298                 DRM_INFO("Loading R600 CP Microcode\n");
 299                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 300                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 301                                      R600_cp_microcode[i][0]);
 302                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 303                                      R600_cp_microcode[i][1]);
 304                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 305                                      R600_cp_microcode[i][2]);
 306                 }
 307
 308                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 309                 DRM_INFO("Loading R600 PFP Microcode\n");
 310                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 311                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, R600_pfp_microcode[i]);
 312         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610)) {
 313                 DRM_INFO("Loading RV610 CP Microcode\n");
 314                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 315                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 316                                      RV610_cp_microcode[i][0]);
 317                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 318                                      RV610_cp_microcode[i][1]);
 319                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 320                                      RV610_cp_microcode[i][2]);
 321                 }
 322
 323                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 324                 DRM_INFO("Loading RV610 PFP Microcode\n");
 325                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 326                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV610_pfp_microcode[i]);
 327         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630)) {
 328                 DRM_INFO("Loading RV630 CP Microcode\n");
 329                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 330                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 331                                      RV630_cp_microcode[i][0]);
 332                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 333                                      RV630_cp_microcode[i][1]);
 334                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 335                                      RV630_cp_microcode[i][2]);
 336                 }
 337
 338                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 339                 DRM_INFO("Loading RV630 PFP Microcode\n");
 340                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 341                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV630_pfp_microcode[i]);
 342         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620)) {
 343                 DRM_INFO("Loading RV620 CP Microcode\n");
 344                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 345                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 346                                      RV620_cp_microcode[i][0]);
 347                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 348                                      RV620_cp_microcode[i][1]);
 349                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 350                                      RV620_cp_microcode[i][2]);
 351                 }
 352
 353                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 354                 DRM_INFO("Loading RV620 PFP Microcode\n");
 355                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 356                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV620_pfp_microcode[i]);
 357         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV635)) {
 358                 DRM_INFO("Loading RV635 CP Microcode\n");
 359                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 360                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 361                                      RV635_cp_microcode[i][0]);
 362                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 363                                      RV635_cp_microcode[i][1]);
 364                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 365                                      RV635_cp_microcode[i][2]);
 366                 }
 367
 368                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 369                 DRM_INFO("Loading RV635 PFP Microcode\n");
 370                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 371                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV635_pfp_microcode[i]);
 372         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670)) {
 373                 DRM_INFO("Loading RV670 CP Microcode\n");
 374                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 375                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 376                                      RV670_cp_microcode[i][0]);
 377                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 378                                      RV670_cp_microcode[i][1]);
 379                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 380                                      RV670_cp_microcode[i][2]);
 381                 }
 382
 383                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 384                 DRM_INFO("Loading RV670 PFP Microcode\n");
 385                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 386                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV670_pfp_microcode[i]);
 387         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 388                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
 389                 DRM_INFO("Loading RS780/RS880 CP Microcode\n");
 390                 for (i = 0; i < PM4_UCODE_SIZE; i++) {
 391                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 392                                      RS780_cp_microcode[i][0]);
 393                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 394                                      RS780_cp_microcode[i][1]);
 395                         RADEON_WRITE(R600_CP_ME_RAM_DATA,
 396                                      RS780_cp_microcode[i][2]);
 397                 }
 398
 399                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 400                 DRM_INFO("Loading RS780/RS880 PFP Microcode\n");
 401                 for (i = 0; i < PFP_UCODE_SIZE; i++)
 402                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RS780_pfp_microcode[i]);
 403         }
 404         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 405         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 406         RADEON_WRITE(R600_CP_ME_RAM_RADDR, 0);
 407
 408 }
 409
 410 static void r700_vm_init(struct drm_device *dev)
 411 {
 412         drm_radeon_private_t *dev_priv = dev->dev_private;
 413         /* initialise the VM to use the page table we constructed up there */
 414         u32 vm_c0, i;
 415         u32 mc_vm_md_l1;
 416         u32 vm_l2_cntl, vm_l2_cntl3;
 417         /* okay set up the PCIE aperture type thingo */
 418         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 419         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 420         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 421
 422         mc_vm_md_l1 = R700_ENABLE_L1_TLB |
 423             R700_ENABLE_L1_FRAGMENT_PROCESSING |
 424             R700_SYSTEM_ACCESS_MODE_IN_SYS |
 425             R700_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 426             R700_EFFECTIVE_L1_TLB_SIZE(5) |
 427             R700_EFFECTIVE_L1_QUEUE_SIZE(5);
 428
 429         RADEON_WRITE(R700_MC_VM_MD_L1_TLB0_CNTL, mc_vm_md_l1);
 430         RADEON_WRITE(R700_MC_VM_MD_L1_TLB1_CNTL, mc_vm_md_l1);
 431         RADEON_WRITE(R700_MC_VM_MD_L1_TLB2_CNTL, mc_vm_md_l1);
 432         RADEON_WRITE(R700_MC_VM_MB_L1_TLB0_CNTL, mc_vm_md_l1);
 433         RADEON_WRITE(R700_MC_VM_MB_L1_TLB1_CNTL, mc_vm_md_l1);
 434         RADEON_WRITE(R700_MC_VM_MB_L1_TLB2_CNTL, mc_vm_md_l1);
 435         RADEON_WRITE(R700_MC_VM_MB_L1_TLB3_CNTL, mc_vm_md_l1);
 436
 437         vm_l2_cntl = R600_VM_L2_CACHE_EN | R600_VM_L2_FRAG_PROC | R600_VM_ENABLE_PTE_CACHE_LRU_W;
 438         vm_l2_cntl |= R700_VM_L2_CNTL_QUEUE_SIZE(7);
 439         RADEON_WRITE(R600_VM_L2_CNTL, vm_l2_cntl);
 440
 441         RADEON_WRITE(R600_VM_L2_CNTL2, 0);
 442         vm_l2_cntl3 = R700_VM_L2_CNTL3_BANK_SELECT(0) | R700_VM_L2_CNTL3_CACHE_UPDATE_MODE(2);
 443         RADEON_WRITE(R600_VM_L2_CNTL3, vm_l2_cntl3);
 444
 445         vm_c0 = R600_VM_ENABLE_CONTEXT | R600_VM_PAGE_TABLE_DEPTH_FLAT;
 446
 447         RADEON_WRITE(R600_VM_CONTEXT0_CNTL, vm_c0);
 448
 449         vm_c0 &= ~R600_VM_ENABLE_CONTEXT;
 450
 451         /* disable all other contexts */
 452         for (i = 1; i < 8; i++)
 453                 RADEON_WRITE(R600_VM_CONTEXT0_CNTL + (i * 4), vm_c0);
 454
 455         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, dev_priv->gart_info.bus_addr >> 12);
 456         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_START_ADDR, dev_priv->gart_vm_start >> 12);
 457         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_END_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 458
 459         r600_vm_flush_gart_range(dev);
 460 }
 461
 462 /* load r600 microcode */
 463 static void r700_cp_load_microcode(drm_radeon_private_t *dev_priv)
 464 {
 465         int i;
 466
 467         r600_do_cp_stop(dev_priv);
 468
 469         RADEON_WRITE(R600_CP_RB_CNTL,
 470                      R600_RB_NO_UPDATE |
 471                      (15 << 8) |
 472                      (3 << 0));
 473
 474         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
 475         RADEON_READ(R600_GRBM_SOFT_RESET);
 476         DRM_UDELAY(15000);
 477         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 478
 479
 480         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV770)) {
 481                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 482                 DRM_INFO("Loading RV770/RV790 PFP Microcode\n");
 483                 for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 484                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV770_pfp_microcode[i]);
 485                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 486
 487                 RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 488                 DRM_INFO("Loading RV770/RV790 CP Microcode\n");
 489                 for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 490                         RADEON_WRITE(R600_CP_ME_RAM_DATA, RV770_cp_microcode[i]);
 491                 RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 492
 493         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV730) ||
 494                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV740)) {
 495                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 496                 DRM_INFO("Loading RV730/RV740 PFP Microcode\n");
 497                 for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 498                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV730_pfp_microcode[i]);
 499                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 500
 501                 RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 502                 DRM_INFO("Loading RV730/RV740 CP Microcode\n");
 503                 for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 504                         RADEON_WRITE(R600_CP_ME_RAM_DATA, RV730_cp_microcode[i]);
 505                 RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 506
 507         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)) {
 508                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 509                 DRM_INFO("Loading RV710 PFP Microcode\n");
 510                 for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 511                         RADEON_WRITE(R600_CP_PFP_UCODE_DATA, RV710_pfp_microcode[i]);
 512                 RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 513
 514                 RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 515                 DRM_INFO("Loading RV710 CP Microcode\n");
 516                 for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 517                         RADEON_WRITE(R600_CP_ME_RAM_DATA, RV710_cp_microcode[i]);
 518                 RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 519
 520         }
 521         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 522         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 523         RADEON_WRITE(R600_CP_ME_RAM_RADDR, 0);
 524
 525 }
 526
 527 static void r600_test_writeback(drm_radeon_private_t *dev_priv)
 528 {
 529         u32 tmp;
 530
 531         /* Start with assuming that writeback doesn't work */
 532         dev_priv->writeback_works = 0;
 533
 534         /* Writeback doesn't seem to work everywhere, test it here and possibly
 535          * enable it if it appears to work
 536          */
 537         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(1), 0);
 538
 539         RADEON_WRITE(R600_SCRATCH_REG1, 0xdeadbeef);
 540
 541         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 542                 u32 val;
 543
 544                 val = radeon_read_ring_rptr(dev_priv, R600_SCRATCHOFF(1));
 545                 if (val == 0xdeadbeef)
 546                         break;
 547                 DRM_UDELAY(1);
 548         }
 549
 550         if (tmp < dev_priv->usec_timeout) {
 551                 dev_priv->writeback_works = 1;
 552                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 553         } else {
 554                 dev_priv->writeback_works = 0;
 555                 DRM_INFO("writeback test failed\n");
 556         }
 557         if (radeon_no_wb == 1) {
 558                 dev_priv->writeback_works = 0;
 559                 DRM_INFO("writeback forced off\n");
 560         }
 561
 562         if (!dev_priv->writeback_works) {
 563                 /* Disable writeback to avoid unnecessary bus master transfer */
 564                 RADEON_WRITE(R600_CP_RB_CNTL, RADEON_READ(R600_CP_RB_CNTL) |
 565                              RADEON_RB_NO_UPDATE);
 566                 RADEON_WRITE(R600_SCRATCH_UMSK, 0);
 567         }
 568 }
 569
 570 int r600_do_engine_reset(struct drm_device *dev)
 571 {
 572         drm_radeon_private_t *dev_priv = dev->dev_private;
 573         u32 cp_ptr, cp_me_cntl, cp_rb_cntl;
 574
 575         DRM_INFO("Resetting GPU\n");
 576
 577         cp_ptr = RADEON_READ(R600_CP_RB_WPTR);
 578         cp_me_cntl = RADEON_READ(R600_CP_ME_CNTL);
 579         RADEON_WRITE(R600_CP_ME_CNTL, R600_CP_ME_HALT);
 580
 581         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0x7fff);
 582         RADEON_READ(R600_GRBM_SOFT_RESET);
 583         DRM_UDELAY(50);
 584         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 585         RADEON_READ(R600_GRBM_SOFT_RESET);
 586
 587         RADEON_WRITE(R600_CP_RB_WPTR_DELAY, 0);
 588         cp_rb_cntl = RADEON_READ(R600_CP_RB_CNTL);
 589         RADEON_WRITE(R600_CP_RB_CNTL, R600_RB_RPTR_WR_ENA);
 590
 591         RADEON_WRITE(R600_CP_RB_RPTR_WR, cp_ptr);
 592         RADEON_WRITE(R600_CP_RB_WPTR, cp_ptr);
 593         RADEON_WRITE(R600_CP_RB_CNTL, cp_rb_cntl);
 594         RADEON_WRITE(R600_CP_ME_CNTL, cp_me_cntl);
 595
 596         /* Reset the CP ring */
 597         r600_do_cp_reset(dev_priv);
 598
 599         /* The CP is no longer running after an engine reset */
 600         dev_priv->cp_running = 0;
 601
 602         /* Reset any pending vertex, indirect buffers */
 603         radeon_freelist_reset(dev);
 604
 605         return 0;
 606
 607 }
 608
 609 static u32 r600_get_tile_pipe_to_backend_map(u32 num_tile_pipes,
 610                                              u32 num_backends,
 611                                              u32 backend_disable_mask)
 612 {
 613         u32 backend_map = 0;
 614         u32 enabled_backends_mask;
 615         u32 enabled_backends_count;
 616         u32 cur_pipe;
 617         u32 swizzle_pipe[R6XX_MAX_PIPES];
 618         u32 cur_backend;
 619         u32 i;
 620
 621         if (num_tile_pipes > R6XX_MAX_PIPES)
 622                 num_tile_pipes = R6XX_MAX_PIPES;
 623         if (num_tile_pipes < 1)
 624                 num_tile_pipes = 1;
 625         if (num_backends > R6XX_MAX_BACKENDS)
 626                 num_backends = R6XX_MAX_BACKENDS;
 627         if (num_backends < 1)
 628                 num_backends = 1;
 629
 630         enabled_backends_mask = 0;
 631         enabled_backends_count = 0;
 632         for (i = 0; i < R6XX_MAX_BACKENDS; ++i) {
 633                 if (((backend_disable_mask >> i) & 1) == 0) {
 634                         enabled_backends_mask |= (1 << i);
 635                         ++enabled_backends_count;
 636                 }
 637                 if (enabled_backends_count == num_backends)
 638                         break;
 639         }
 640
 641         if (enabled_backends_count == 0) {
 642                 enabled_backends_mask = 1;
 643                 enabled_backends_count = 1;
 644         }
 645
 646         if (enabled_backends_count != num_backends)
 647                 num_backends = enabled_backends_count;
 648
 649         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R6XX_MAX_PIPES);
 650         switch (num_tile_pipes) {
 651         case 1:
 652                 swizzle_pipe[0] = 0;
 653                 break;
 654         case 2:
 655                 swizzle_pipe[0] = 0;
 656                 swizzle_pipe[1] = 1;
 657                 break;
 658         case 3:
 659                 swizzle_pipe[0] = 0;
 660                 swizzle_pipe[1] = 1;
 661                 swizzle_pipe[2] = 2;
 662                 break;
 663         case 4:
 664                 swizzle_pipe[0] = 0;
 665                 swizzle_pipe[1] = 1;
 666                 swizzle_pipe[2] = 2;
 667                 swizzle_pipe[3] = 3;
 668                 break;
 669         case 5:
 670                 swizzle_pipe[0] = 0;
 671                 swizzle_pipe[1] = 1;
 672                 swizzle_pipe[2] = 2;
 673                 swizzle_pipe[3] = 3;
 674                 swizzle_pipe[4] = 4;
 675                 break;
 676         case 6:
 677                 swizzle_pipe[0] = 0;
 678                 swizzle_pipe[1] = 2;
 679                 swizzle_pipe[2] = 4;
 680                 swizzle_pipe[3] = 5;
 681                 swizzle_pipe[4] = 1;
 682                 swizzle_pipe[5] = 3;
 683                 break;
 684         case 7:
 685                 swizzle_pipe[0] = 0;
 686                 swizzle_pipe[1] = 2;
 687                 swizzle_pipe[2] = 4;
 688                 swizzle_pipe[3] = 6;
 689                 swizzle_pipe[4] = 1;
 690                 swizzle_pipe[5] = 3;
 691                 swizzle_pipe[6] = 5;
 692                 break;
 693         case 8:
 694                 swizzle_pipe[0] = 0;
 695                 swizzle_pipe[1] = 2;
 696                 swizzle_pipe[2] = 4;
 697                 swizzle_pipe[3] = 6;
 698                 swizzle_pipe[4] = 1;
 699                 swizzle_pipe[5] = 3;
 700                 swizzle_pipe[6] = 5;
 701                 swizzle_pipe[7] = 7;
 702                 break;
 703         }
 704
 705         cur_backend = 0;
 706         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
 707                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
 708                         cur_backend = (cur_backend + 1) % R6XX_MAX_BACKENDS;
 709
 710                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
 711
 712                 cur_backend = (cur_backend + 1) % R6XX_MAX_BACKENDS;
 713         }
 714
 715         return backend_map;
 716 }
 717
 718 static int r600_count_pipe_bits(uint32_t val)
 719 {
 720         int i, ret = 0;
 721         for (i = 0; i < 32; i++) {
 722                 ret += val & 1;
 723                 val >>= 1;
 724         }
 725         return ret;
 726 }
 727
 728 static void r600_gfx_init(struct drm_device *dev,
 729                           drm_radeon_private_t *dev_priv)
 730 {
 731         int i, j, num_qd_pipes;
 732         u32 sx_debug_1;
 733         u32 tc_cntl;
 734         u32 arb_pop;
 735         u32 num_gs_verts_per_thread;
 736         u32 vgt_gs_per_es;
 737         u32 gs_prim_buffer_depth = 0;
 738         u32 sq_ms_fifo_sizes;
 739         u32 sq_config;
 740         u32 sq_gpr_resource_mgmt_1 = 0;
 741         u32 sq_gpr_resource_mgmt_2 = 0;
 742         u32 sq_thread_resource_mgmt = 0;
 743         u32 sq_stack_resource_mgmt_1 = 0;
 744         u32 sq_stack_resource_mgmt_2 = 0;
 745         u32 hdp_host_path_cntl;
 746         u32 backend_map;
 747         u32 gb_tiling_config = 0;
 748         u32 cc_rb_backend_disable = 0;
 749         u32 cc_gc_shader_pipe_config = 0;
 750         u32 ramcfg;
 751
 752         /* setup chip specs */
 753         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
 754         case CHIP_R600:
 755                 dev_priv->r600_max_pipes = 4;
 756                 dev_priv->r600_max_tile_pipes = 8;
 757                 dev_priv->r600_max_simds = 4;
 758                 dev_priv->r600_max_backends = 4;
 759                 dev_priv->r600_max_gprs = 256;
 760                 dev_priv->r600_max_threads = 192;
 761                 dev_priv->r600_max_stack_entries = 256;
 762                 dev_priv->r600_max_hw_contexts = 8;
 763                 dev_priv->r600_max_gs_threads = 16;
 764                 dev_priv->r600_sx_max_export_size = 128;
 765                 dev_priv->r600_sx_max_export_pos_size = 16;
 766                 dev_priv->r600_sx_max_export_smx_size = 128;
 767                 dev_priv->r600_sq_num_cf_insts = 2;
 768                 break;
 769         case CHIP_RV630:
 770         case CHIP_RV635:
 771                 dev_priv->r600_max_pipes = 2;
 772                 dev_priv->r600_max_tile_pipes = 2;
 773                 dev_priv->r600_max_simds = 3;
 774                 dev_priv->r600_max_backends = 1;
 775                 dev_priv->r600_max_gprs = 128;
 776                 dev_priv->r600_max_threads = 192;
 777                 dev_priv->r600_max_stack_entries = 128;
 778                 dev_priv->r600_max_hw_contexts = 8;
 779                 dev_priv->r600_max_gs_threads = 4;
 780                 dev_priv->r600_sx_max_export_size = 128;
 781                 dev_priv->r600_sx_max_export_pos_size = 16;
 782                 dev_priv->r600_sx_max_export_smx_size = 128;
 783                 dev_priv->r600_sq_num_cf_insts = 2;
 784                 break;
 785         case CHIP_RV610:
 786         case CHIP_RS780:
 787         case CHIP_RS880:
 788         case CHIP_RV620:
 789                 dev_priv->r600_max_pipes = 1;
 790                 dev_priv->r600_max_tile_pipes = 1;
 791                 dev_priv->r600_max_simds = 2;
 792                 dev_priv->r600_max_backends = 1;
 793                 dev_priv->r600_max_gprs = 128;
 794                 dev_priv->r600_max_threads = 192;
 795                 dev_priv->r600_max_stack_entries = 128;
 796                 dev_priv->r600_max_hw_contexts = 4;
 797                 dev_priv->r600_max_gs_threads = 4;
 798                 dev_priv->r600_sx_max_export_size = 128;
 799                 dev_priv->r600_sx_max_export_pos_size = 16;
 800                 dev_priv->r600_sx_max_export_smx_size = 128;
 801                 dev_priv->r600_sq_num_cf_insts = 1;
 802                 break;
 803         case CHIP_RV670:
 804                 dev_priv->r600_max_pipes = 4;
 805                 dev_priv->r600_max_tile_pipes = 4;
 806                 dev_priv->r600_max_simds = 4;
 807                 dev_priv->r600_max_backends = 4;
 808                 dev_priv->r600_max_gprs = 192;
 809                 dev_priv->r600_max_threads = 192;
 810                 dev_priv->r600_max_stack_entries = 256;
 811                 dev_priv->r600_max_hw_contexts = 8;
 812                 dev_priv->r600_max_gs_threads = 16;
 813                 dev_priv->r600_sx_max_export_size = 128;
 814                 dev_priv->r600_sx_max_export_pos_size = 16;
 815                 dev_priv->r600_sx_max_export_smx_size = 128;
 816                 dev_priv->r600_sq_num_cf_insts = 2;
 817                 break;
 818         default:
 819                 break;
 820         }
 821
 822         /* Initialize HDP */
 823         j = 0;
 824         for (i = 0; i < 32; i++) {
 825                 RADEON_WRITE((0x2c14 + j), 0x00000000);
 826                 RADEON_WRITE((0x2c18 + j), 0x00000000);
 827                 RADEON_WRITE((0x2c1c + j), 0x00000000);
 828                 RADEON_WRITE((0x2c20 + j), 0x00000000);
 829                 RADEON_WRITE((0x2c24 + j), 0x00000000);
 830                 j += 0x18;
 831         }
 832
 833         RADEON_WRITE(R600_GRBM_CNTL, R600_GRBM_READ_TIMEOUT(0xff));
 834
 835         /* setup tiling, simd, pipe config */
 836         ramcfg = RADEON_READ(R600_RAMCFG);
 837
 838         switch (dev_priv->r600_max_tile_pipes) {
 839         case 1:
 840                 gb_tiling_config |= R600_PIPE_TILING(0);
 841                 break;
 842         case 2:
 843                 gb_tiling_config |= R600_PIPE_TILING(1);
 844                 break;
 845         case 4:
 846                 gb_tiling_config |= R600_PIPE_TILING(2);
 847                 break;
 848         case 8:
 849                 gb_tiling_config |= R600_PIPE_TILING(3);
 850                 break;
 851         default:
 852                 break;
 853         }
 854
 855         gb_tiling_config |= R600_BANK_TILING((ramcfg >> R600_NOOFBANK_SHIFT) & R600_NOOFBANK_MASK);
 856
 857         gb_tiling_config |= R600_GROUP_SIZE(0);
 858
 859         if (((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK) > 3) {
 860                 gb_tiling_config |= R600_ROW_TILING(3);
 861                 gb_tiling_config |= R600_SAMPLE_SPLIT(3);
 862         } else {
 863                 gb_tiling_config |=
 864                         R600_ROW_TILING(((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK));
 865                 gb_tiling_config |=
 866                         R600_SAMPLE_SPLIT(((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK));
 867         }
 868
 869         gb_tiling_config |= R600_BANK_SWAPS(1);
 870
 871         backend_map = r600_get_tile_pipe_to_backend_map(dev_priv->r600_max_tile_pipes,
 872                                                         dev_priv->r600_max_backends,
 873                                                         (0xff << dev_priv->r600_max_backends) & 0xff);
 874         gb_tiling_config |= R600_BACKEND_MAP(backend_map);
 875
 876         cc_gc_shader_pipe_config =
 877                 R600_INACTIVE_QD_PIPES((R6XX_MAX_PIPES_MASK << dev_priv->r600_max_pipes) & R6XX_MAX_PIPES_MASK);
 878         cc_gc_shader_pipe_config |=
 879                 R600_INACTIVE_SIMDS((R6XX_MAX_SIMDS_MASK << dev_priv->r600_max_simds) & R6XX_MAX_SIMDS_MASK);
 880
 881         cc_rb_backend_disable =
 882                 R600_BACKEND_DISABLE((R6XX_MAX_BACKENDS_MASK << dev_priv->r600_max_backends) & R6XX_MAX_BACKENDS_MASK);
 883
 884         RADEON_WRITE(R600_GB_TILING_CONFIG,      gb_tiling_config);
 885         RADEON_WRITE(R600_DCP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
 886         RADEON_WRITE(R600_HDP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
 887
 888         RADEON_WRITE(R600_CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
 889         RADEON_WRITE(R600_CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
 890         RADEON_WRITE(R600_GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
 891
 892         num_qd_pipes =
 893                 R6XX_MAX_BACKENDS - r600_count_pipe_bits(cc_gc_shader_pipe_config & R600_INACTIVE_QD_PIPES_MASK);
 894         RADEON_WRITE(R600_VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & R600_DEALLOC_DIST_MASK);
 895         RADEON_WRITE(R600_VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & R600_VTX_REUSE_DEPTH_MASK);
 896
 897         /* set HW defaults for 3D engine */
 898         RADEON_WRITE(R600_CP_QUEUE_THRESHOLDS, (R600_ROQ_IB1_START(0x16) |
 899                                                 R600_ROQ_IB2_START(0x2b)));
 900
 901         RADEON_WRITE(R600_CP_MEQ_THRESHOLDS, (R600_MEQ_END(0x40) |
 902                                               R600_ROQ_END(0x40)));
 903
 904         RADEON_WRITE(R600_TA_CNTL_AUX, (R600_DISABLE_CUBE_ANISO |
 905                                         R600_SYNC_GRADIENT |
 906                                         R600_SYNC_WALKER |
 907                                         R600_SYNC_ALIGNER));
 908
 909         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670)
 910                 RADEON_WRITE(R600_ARB_GDEC_RD_CNTL, 0x00000021);
 911
 912         sx_debug_1 = RADEON_READ(R600_SX_DEBUG_1);
 913         sx_debug_1 |= R600_SMX_EVENT_RELEASE;
 914         if (((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_R600))
 915                 sx_debug_1 |= R600_ENABLE_NEW_SMX_ADDRESS;
 916         RADEON_WRITE(R600_SX_DEBUG_1, sx_debug_1);
 917
 918         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) ||
 919             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630) ||
 920             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 921             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 922             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 923             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880))
 924                 RADEON_WRITE(R600_DB_DEBUG, R600_PREZ_MUST_WAIT_FOR_POSTZ_DONE);
 925         else
 926                 RADEON_WRITE(R600_DB_DEBUG, 0);
 927
 928         RADEON_WRITE(R600_DB_WATERMARKS, (R600_DEPTH_FREE(4) |
 929                                           R600_DEPTH_FLUSH(16) |
 930                                           R600_DEPTH_PENDING_FREE(4) |
 931                                           R600_DEPTH_CACHELINE_FREE(16)));
 932         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
 933         RADEON_WRITE(R600_VGT_NUM_INSTANCES, 0);
 934
 935         RADEON_WRITE(R600_SPI_CONFIG_CNTL, R600_GPR_WRITE_PRIORITY(0));
 936         RADEON_WRITE(R600_SPI_CONFIG_CNTL_1, R600_VTX_DONE_DELAY(0));
 937
 938         sq_ms_fifo_sizes = RADEON_READ(R600_SQ_MS_FIFO_SIZES);
 939         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 940             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 941             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 942             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
 943                 sq_ms_fifo_sizes = (R600_CACHE_FIFO_SIZE(0xa) |
 944                                     R600_FETCH_FIFO_HIWATER(0xa) |
 945                                     R600_DONE_FIFO_HIWATER(0xe0) |
 946                                     R600_ALU_UPDATE_FIFO_HIWATER(0x8));
 947         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) ||
 948                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630)) {
 949                 sq_ms_fifo_sizes &= ~R600_DONE_FIFO_HIWATER(0xff);
 950                 sq_ms_fifo_sizes |= R600_DONE_FIFO_HIWATER(0x4);
 951         }
 952         RADEON_WRITE(R600_SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 953
 954         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 955          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 956          */
 957         sq_config = RADEON_READ(R600_SQ_CONFIG);
 958         sq_config &= ~(R600_PS_PRIO(3) |
 959                        R600_VS_PRIO(3) |
 960                        R600_GS_PRIO(3) |
 961                        R600_ES_PRIO(3));
 962         sq_config |= (R600_DX9_CONSTS |
 963                       R600_VC_ENABLE |
 964                       R600_PS_PRIO(0) |
 965                       R600_VS_PRIO(1) |
 966                       R600_GS_PRIO(2) |
 967                       R600_ES_PRIO(3));
 968
 969         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) {
 970                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(124) |
 971                                           R600_NUM_VS_GPRS(124) |
 972                                           R600_NUM_CLAUSE_TEMP_GPRS(4));
 973                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(0) |
 974                                           R600_NUM_ES_GPRS(0));
 975                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(136) |
 976                                            R600_NUM_VS_THREADS(48) |
 977                                            R600_NUM_GS_THREADS(4) |
 978                                            R600_NUM_ES_THREADS(4));
 979                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(128) |
 980                                             R600_NUM_VS_STACK_ENTRIES(128));
 981                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(0) |
 982                                             R600_NUM_ES_STACK_ENTRIES(0));
 983         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 984                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 985                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 986                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
 987                 /* no vertex cache */
 988                 sq_config &= ~R600_VC_ENABLE;
 989
 990                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
 991                                           R600_NUM_VS_GPRS(44) |
 992                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
 993                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(17) |
 994                                           R600_NUM_ES_GPRS(17));
 995                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
 996                                            R600_NUM_VS_THREADS(78) |
 997                                            R600_NUM_GS_THREADS(4) |
 998                                            R600_NUM_ES_THREADS(31));
 999                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(40) |
1000                                             R600_NUM_VS_STACK_ENTRIES(40));
1001                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(32) |
1002                                             R600_NUM_ES_STACK_ENTRIES(16));
1003         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630) ||
1004                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV635)) {
1005                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1006                                           R600_NUM_VS_GPRS(44) |
1007                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1008                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(18) |
1009                                           R600_NUM_ES_GPRS(18));
1010                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1011                                            R600_NUM_VS_THREADS(78) |
1012                                            R600_NUM_GS_THREADS(4) |
1013                                            R600_NUM_ES_THREADS(31));
1014                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(40) |
1015                                             R600_NUM_VS_STACK_ENTRIES(40));
1016                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(32) |
1017                                             R600_NUM_ES_STACK_ENTRIES(16));
1018         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670) {
1019                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1020                                           R600_NUM_VS_GPRS(44) |
1021                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1022                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(17) |
1023                                           R600_NUM_ES_GPRS(17));
1024                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1025                                            R600_NUM_VS_THREADS(78) |
1026                                            R600_NUM_GS_THREADS(4) |
1027                                            R600_NUM_ES_THREADS(31));
1028                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(64) |
1029                                             R600_NUM_VS_STACK_ENTRIES(64));
1030                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(64) |
1031                                             R600_NUM_ES_STACK_ENTRIES(64));
1032         }
1033
1034         RADEON_WRITE(R600_SQ_CONFIG, sq_config);
1035         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_1,  sq_gpr_resource_mgmt_1);
1036         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_2,  sq_gpr_resource_mgmt_2);
1037         RADEON_WRITE(R600_SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1038         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_1, sq_stack_resource_mgmt_1);
1039         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_2, sq_stack_resource_mgmt_2);
1040
1041         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
1042             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
1043             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
1044             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880))
1045                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, R600_CACHE_INVALIDATION(R600_TC_ONLY));
1046         else
1047                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, R600_CACHE_INVALIDATION(R600_VC_AND_TC));
1048
1049         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_2S, (R600_S0_X(0xc) |
1050                                                     R600_S0_Y(0x4) |
1051                                                     R600_S1_X(0x4) |
1052                                                     R600_S1_Y(0xc)));
1053         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_4S, (R600_S0_X(0xe) |
1054                                                     R600_S0_Y(0xe) |
1055                                                     R600_S1_X(0x2) |
1056                                                     R600_S1_Y(0x2) |
1057                                                     R600_S2_X(0xa) |
1058                                                     R600_S2_Y(0x6) |
1059                                                     R600_S3_X(0x6) |
1060                                                     R600_S3_Y(0xa)));
1061         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_8S_WD0, (R600_S0_X(0xe) |
1062                                                         R600_S0_Y(0xb) |
1063                                                         R600_S1_X(0x4) |
1064                                                         R600_S1_Y(0xc) |
1065                                                         R600_S2_X(0x1) |
1066                                                         R600_S2_Y(0x6) |
1067                                                         R600_S3_X(0xa) |
1068                                                         R600_S3_Y(0xe)));
1069         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_8S_WD1, (R600_S4_X(0x6) |
1070                                                         R600_S4_Y(0x1) |
1071                                                         R600_S5_X(0x0) |
1072                                                         R600_S5_Y(0x0) |
1073                                                         R600_S6_X(0xb) |
1074                                                         R600_S6_Y(0x4) |
1075                                                         R600_S7_X(0x7) |
1076                                                         R600_S7_Y(0x8)));
1077
1078
1079         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1080         case CHIP_R600:
1081         case CHIP_RV630:
1082         case CHIP_RV635:
1083                 gs_prim_buffer_depth = 0;
1084                 break;
1085         case CHIP_RV610:
1086         case CHIP_RS780:
1087         case CHIP_RS880:
1088         case CHIP_RV620:
1089                 gs_prim_buffer_depth = 32;
1090                 break;
1091         case CHIP_RV670:
1092                 gs_prim_buffer_depth = 128;
1093                 break;
1094         default:
1095                 break;
1096         }
1097
1098         num_gs_verts_per_thread = dev_priv->r600_max_pipes * 16;
1099         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
1100         /* Max value for this is 256 */
1101         if (vgt_gs_per_es > 256)
1102                 vgt_gs_per_es = 256;
1103
1104         RADEON_WRITE(R600_VGT_ES_PER_GS, 128);
1105         RADEON_WRITE(R600_VGT_GS_PER_ES, vgt_gs_per_es);
1106         RADEON_WRITE(R600_VGT_GS_PER_VS, 2);
1107         RADEON_WRITE(R600_VGT_GS_VERTEX_REUSE, 16);
1108
1109         /* more default values. 2D/3D driver should adjust as needed */
1110         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE_STATE, 0);
1111         RADEON_WRITE(R600_VGT_STRMOUT_EN, 0);
1112         RADEON_WRITE(R600_SX_MISC, 0);
1113         RADEON_WRITE(R600_PA_SC_MODE_CNTL, 0);
1114         RADEON_WRITE(R600_PA_SC_AA_CONFIG, 0);
1115         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE, 0);
1116         RADEON_WRITE(R600_SPI_INPUT_Z, 0);
1117         RADEON_WRITE(R600_SPI_PS_IN_CONTROL_0, R600_NUM_INTERP(2));
1118         RADEON_WRITE(R600_CB_COLOR7_FRAG, 0);
1119
1120         /* clear render buffer base addresses */
1121         RADEON_WRITE(R600_CB_COLOR0_BASE, 0);
1122         RADEON_WRITE(R600_CB_COLOR1_BASE, 0);
1123         RADEON_WRITE(R600_CB_COLOR2_BASE, 0);
1124         RADEON_WRITE(R600_CB_COLOR3_BASE, 0);
1125         RADEON_WRITE(R600_CB_COLOR4_BASE, 0);
1126         RADEON_WRITE(R600_CB_COLOR5_BASE, 0);
1127         RADEON_WRITE(R600_CB_COLOR6_BASE, 0);
1128         RADEON_WRITE(R600_CB_COLOR7_BASE, 0);
1129
1130         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1131         case CHIP_RV610:
1132         case CHIP_RS780:
1133         case CHIP_RS880:
1134         case CHIP_RV620:
1135                 tc_cntl = R600_TC_L2_SIZE(8);
1136                 break;
1137         case CHIP_RV630:
1138         case CHIP_RV635:
1139                 tc_cntl = R600_TC_L2_SIZE(4);
1140                 break;
1141         case CHIP_R600:
1142                 tc_cntl = R600_TC_L2_SIZE(0) | R600_L2_DISABLE_LATE_HIT;
1143                 break;
1144         default:
1145                 tc_cntl = R600_TC_L2_SIZE(0);
1146                 break;
1147         }
1148
1149         RADEON_WRITE(R600_TC_CNTL, tc_cntl);
1150
1151         hdp_host_path_cntl = RADEON_READ(R600_HDP_HOST_PATH_CNTL);
1152         RADEON_WRITE(R600_HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1153
1154         arb_pop = RADEON_READ(R600_ARB_POP);
1155         arb_pop |= R600_ENABLE_TC128;
1156         RADEON_WRITE(R600_ARB_POP, arb_pop);
1157
1158         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1159         RADEON_WRITE(R600_PA_CL_ENHANCE, (R600_CLIP_VTX_REORDER_ENA |
1160                                           R600_NUM_CLIP_SEQ(3)));
1161         RADEON_WRITE(R600_PA_SC_ENHANCE, R600_FORCE_EOV_MAX_CLK_CNT(4095));
1162
1163 }
1164
1165 static u32 r700_get_tile_pipe_to_backend_map(u32 num_tile_pipes,
1166                                              u32 num_backends,
1167                                              u32 backend_disable_mask)
1168 {
1169         u32 backend_map = 0;
1170         u32 enabled_backends_mask;
1171         u32 enabled_backends_count;
1172         u32 cur_pipe;
1173         u32 swizzle_pipe[R7XX_MAX_PIPES];
1174         u32 cur_backend;
1175         u32 i;
1176
1177         if (num_tile_pipes > R7XX_MAX_PIPES)
1178                 num_tile_pipes = R7XX_MAX_PIPES;
1179         if (num_tile_pipes < 1)
1180                 num_tile_pipes = 1;
1181         if (num_backends > R7XX_MAX_BACKENDS)
1182                 num_backends = R7XX_MAX_BACKENDS;
1183         if (num_backends < 1)
1184                 num_backends = 1;
1185
1186         enabled_backends_mask = 0;
1187         enabled_backends_count = 0;
1188         for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
1189                 if (((backend_disable_mask >> i) & 1) == 0) {
1190                         enabled_backends_mask |= (1 << i);
1191                         ++enabled_backends_count;
1192                 }
1193                 if (enabled_backends_count == num_backends)
1194                         break;
1195         }
1196
1197         if (enabled_backends_count == 0) {
1198                 enabled_backends_mask = 1;
1199                 enabled_backends_count = 1;
1200         }
1201
1202         if (enabled_backends_count != num_backends)
1203                 num_backends = enabled_backends_count;
1204
1205         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
1206         switch (num_tile_pipes) {
1207         case 1:
1208                 swizzle_pipe[0] = 0;
1209                 break;
1210         case 2:
1211                 swizzle_pipe[0] = 0;
1212                 swizzle_pipe[1] = 1;
1213                 break;
1214         case 3:
1215                 swizzle_pipe[0] = 0;
1216                 swizzle_pipe[1] = 2;
1217                 swizzle_pipe[2] = 1;
1218                 break;
1219         case 4:
1220                 swizzle_pipe[0] = 0;
1221                 swizzle_pipe[1] = 2;
1222                 swizzle_pipe[2] = 3;
1223                 swizzle_pipe[3] = 1;
1224                 break;
1225         case 5:
1226                 swizzle_pipe[0] = 0;
1227                 swizzle_pipe[1] = 2;
1228                 swizzle_pipe[2] = 4;
1229                 swizzle_pipe[3] = 1;
1230                 swizzle_pipe[4] = 3;
1231                 break;
1232         case 6:
1233                 swizzle_pipe[0] = 0;
1234                 swizzle_pipe[1] = 2;
1235                 swizzle_pipe[2] = 4;
1236                 swizzle_pipe[3] = 5;
1237                 swizzle_pipe[4] = 3;
1238                 swizzle_pipe[5] = 1;
1239                 break;
1240         case 7:
1241                 swizzle_pipe[0] = 0;
1242                 swizzle_pipe[1] = 2;
1243                 swizzle_pipe[2] = 4;
1244                 swizzle_pipe[3] = 6;
1245                 swizzle_pipe[4] = 3;
1246                 swizzle_pipe[5] = 1;
1247                 swizzle_pipe[6] = 5;
1248                 break;
1249         case 8:
1250                 swizzle_pipe[0] = 0;
1251                 swizzle_pipe[1] = 2;
1252                 swizzle_pipe[2] = 4;
1253                 swizzle_pipe[3] = 6;
1254                 swizzle_pipe[4] = 3;
1255                 swizzle_pipe[5] = 1;
1256                 swizzle_pipe[6] = 7;
1257                 swizzle_pipe[7] = 5;
1258                 break;
1259         }
1260
1261         cur_backend = 0;
1262         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
1263                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
1264                         cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
1265
1266                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
1267
1268                 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
1269         }
1270
1271         return backend_map;
1272 }
1273
1274 static void r700_gfx_init(struct drm_device *dev,
1275                           drm_radeon_private_t *dev_priv)
1276 {
1277         int i, j, num_qd_pipes;
1278         u32 sx_debug_1;
1279         u32 smx_dc_ctl0;
1280         u32 num_gs_verts_per_thread;
1281         u32 vgt_gs_per_es;
1282         u32 gs_prim_buffer_depth = 0;
1283         u32 sq_ms_fifo_sizes;
1284         u32 sq_config;
1285         u32 sq_thread_resource_mgmt;
1286         u32 hdp_host_path_cntl;
1287         u32 sq_dyn_gpr_size_simd_ab_0;
1288         u32 backend_map;
1289         u32 gb_tiling_config = 0;
1290         u32 cc_rb_backend_disable = 0;
1291         u32 cc_gc_shader_pipe_config = 0;
1292         u32 mc_arb_ramcfg;
1293         u32 db_debug4;
1294
1295         /* setup chip specs */
1296         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1297         case CHIP_RV770:
1298                 dev_priv->r600_max_pipes = 4;
1299                 dev_priv->r600_max_tile_pipes = 8;
1300                 dev_priv->r600_max_simds = 10;
1301                 dev_priv->r600_max_backends = 4;
1302                 dev_priv->r600_max_gprs = 256;
1303                 dev_priv->r600_max_threads = 248;
1304                 dev_priv->r600_max_stack_entries = 512;
1305                 dev_priv->r600_max_hw_contexts = 8;
1306                 dev_priv->r600_max_gs_threads = 16 * 2;
1307                 dev_priv->r600_sx_max_export_size = 128;
1308                 dev_priv->r600_sx_max_export_pos_size = 16;
1309                 dev_priv->r600_sx_max_export_smx_size = 112;
1310                 dev_priv->r600_sq_num_cf_insts = 2;
1311
1312                 dev_priv->r700_sx_num_of_sets = 7;
1313                 dev_priv->r700_sc_prim_fifo_size = 0xF9;
1314                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1315                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1316                 break;
1317         case CHIP_RV730:
1318                 dev_priv->r600_max_pipes = 2;
1319                 dev_priv->r600_max_tile_pipes = 4;
1320                 dev_priv->r600_max_simds = 8;
1321                 dev_priv->r600_max_backends = 2;
1322                 dev_priv->r600_max_gprs = 128;
1323                 dev_priv->r600_max_threads = 248;
1324                 dev_priv->r600_max_stack_entries = 256;
1325                 dev_priv->r600_max_hw_contexts = 8;
1326                 dev_priv->r600_max_gs_threads = 16 * 2;
1327                 dev_priv->r600_sx_max_export_size = 256;
1328                 dev_priv->r600_sx_max_export_pos_size = 32;
1329                 dev_priv->r600_sx_max_export_smx_size = 224;
1330                 dev_priv->r600_sq_num_cf_insts = 2;
1331
1332                 dev_priv->r700_sx_num_of_sets = 7;
1333                 dev_priv->r700_sc_prim_fifo_size = 0xf9;
1334                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1335                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1336                 if (dev_priv->r600_sx_max_export_pos_size > 16) {
1337                         dev_priv->r600_sx_max_export_pos_size -= 16;
1338                         dev_priv->r600_sx_max_export_smx_size += 16;
1339                 }
1340                 break;
1341         case CHIP_RV710:
1342                 dev_priv->r600_max_pipes = 2;
1343                 dev_priv->r600_max_tile_pipes = 2;
1344                 dev_priv->r600_max_simds = 2;
1345                 dev_priv->r600_max_backends = 1;
1346                 dev_priv->r600_max_gprs = 256;
1347                 dev_priv->r600_max_threads = 192;
1348                 dev_priv->r600_max_stack_entries = 256;
1349                 dev_priv->r600_max_hw_contexts = 4;
1350                 dev_priv->r600_max_gs_threads = 8 * 2;
1351                 dev_priv->r600_sx_max_export_size = 128;
1352                 dev_priv->r600_sx_max_export_pos_size = 16;
1353                 dev_priv->r600_sx_max_export_smx_size = 112;
1354                 dev_priv->r600_sq_num_cf_insts = 1;
1355
1356                 dev_priv->r700_sx_num_of_sets = 7;
1357                 dev_priv->r700_sc_prim_fifo_size = 0x40;
1358                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1359                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1360                 break;
1361         case CHIP_RV740:
1362                 dev_priv->r600_max_pipes = 4;
1363                 dev_priv->r600_max_tile_pipes = 4;
1364                 dev_priv->r600_max_simds = 8;
1365                 dev_priv->r600_max_backends = 4;
1366                 dev_priv->r600_max_gprs = 256;
1367                 dev_priv->r600_max_threads = 248;
1368                 dev_priv->r600_max_stack_entries = 512;
1369                 dev_priv->r600_max_hw_contexts = 8;
1370                 dev_priv->r600_max_gs_threads = 16 * 2;
1371                 dev_priv->r600_sx_max_export_size = 256;
1372                 dev_priv->r600_sx_max_export_pos_size = 32;
1373                 dev_priv->r600_sx_max_export_smx_size = 224;
1374                 dev_priv->r600_sq_num_cf_insts = 2;
1375
1376                 dev_priv->r700_sx_num_of_sets = 7;
1377                 dev_priv->r700_sc_prim_fifo_size = 0x100;
1378                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1379                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1380
1381                 if (dev_priv->r600_sx_max_export_pos_size > 16) {
1382                         dev_priv->r600_sx_max_export_pos_size -= 16;
1383                         dev_priv->r600_sx_max_export_smx_size += 16;
1384                 }
1385                 break;
1386         default:
1387                 break;
1388         }
1389
1390         /* Initialize HDP */
1391         j = 0;
1392         for (i = 0; i < 32; i++) {
1393                 RADEON_WRITE((0x2c14 + j), 0x00000000);
1394                 RADEON_WRITE((0x2c18 + j), 0x00000000);
1395                 RADEON_WRITE((0x2c1c + j), 0x00000000);
1396                 RADEON_WRITE((0x2c20 + j), 0x00000000);
1397                 RADEON_WRITE((0x2c24 + j), 0x00000000);
1398                 j += 0x18;
1399         }
1400
1401         RADEON_WRITE(R600_GRBM_CNTL, R600_GRBM_READ_TIMEOUT(0xff));
1402
1403         /* setup tiling, simd, pipe config */
1404         mc_arb_ramcfg = RADEON_READ(R700_MC_ARB_RAMCFG);
1405
1406         switch (dev_priv->r600_max_tile_pipes) {
1407         case 1:
1408                 gb_tiling_config |= R600_PIPE_TILING(0);
1409                 break;
1410         case 2:
1411                 gb_tiling_config |= R600_PIPE_TILING(1);
1412                 break;
1413         case 4:
1414                 gb_tiling_config |= R600_PIPE_TILING(2);
1415                 break;
1416         case 8:
1417                 gb_tiling_config |= R600_PIPE_TILING(3);
1418                 break;
1419         default:
1420                 break;
1421         }
1422
1423         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV770)
1424                 gb_tiling_config |= R600_BANK_TILING(1);
1425         else
1426                 gb_tiling_config |= R600_BANK_TILING((mc_arb_ramcfg >> R700_NOOFBANK_SHIFT) & R700_NOOFBANK_MASK);
1427
1428         gb_tiling_config |= R600_GROUP_SIZE(0);
1429
1430         if (((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK) > 3) {
1431                 gb_tiling_config |= R600_ROW_TILING(3);
1432                 gb_tiling_config |= R600_SAMPLE_SPLIT(3);
1433         } else {
1434                 gb_tiling_config |=
1435                         R600_ROW_TILING(((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK));
1436                 gb_tiling_config |=
1437                         R600_SAMPLE_SPLIT(((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK));
1438         }
1439
1440         gb_tiling_config |= R600_BANK_SWAPS(1);
1441
1442         backend_map = r700_get_tile_pipe_to_backend_map(dev_priv->r600_max_tile_pipes,
1443                                                         dev_priv->r600_max_backends,
1444                                                         (0xff << dev_priv->r600_max_backends) & 0xff);
1445         gb_tiling_config |= R600_BACKEND_MAP(backend_map);
1446
1447         cc_gc_shader_pipe_config =
1448                 R600_INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << dev_priv->r600_max_pipes) & R7XX_MAX_PIPES_MASK);
1449         cc_gc_shader_pipe_config |=
1450                 R600_INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << dev_priv->r600_max_simds) & R7XX_MAX_SIMDS_MASK);
1451
1452         cc_rb_backend_disable =
1453                 R600_BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << dev_priv->r600_max_backends) & R7XX_MAX_BACKENDS_MASK);
1454
1455         RADEON_WRITE(R600_GB_TILING_CONFIG,      gb_tiling_config);
1456         RADEON_WRITE(R600_DCP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
1457         RADEON_WRITE(R600_HDP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
1458
1459         RADEON_WRITE(R600_CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
1460         RADEON_WRITE(R600_CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
1461         RADEON_WRITE(R600_GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
1462
1463         RADEON_WRITE(R700_CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable);
1464         RADEON_WRITE(R700_CGTS_SYS_TCC_DISABLE, 0);
1465         RADEON_WRITE(R700_CGTS_TCC_DISABLE, 0);
1466         RADEON_WRITE(R700_CGTS_USER_SYS_TCC_DISABLE, 0);
1467         RADEON_WRITE(R700_CGTS_USER_TCC_DISABLE, 0);
1468
1469         num_qd_pipes =
1470                 R7XX_MAX_BACKENDS - r600_count_pipe_bits(cc_gc_shader_pipe_config & R600_INACTIVE_QD_PIPES_MASK);
1471         RADEON_WRITE(R600_VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & R600_DEALLOC_DIST_MASK);
1472         RADEON_WRITE(R600_VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & R600_VTX_REUSE_DEPTH_MASK);
1473
1474         /* set HW defaults for 3D engine */
1475         RADEON_WRITE(R600_CP_QUEUE_THRESHOLDS, (R600_ROQ_IB1_START(0x16) |
1476                                                 R600_ROQ_IB2_START(0x2b)));
1477
1478         RADEON_WRITE(R600_CP_MEQ_THRESHOLDS, R700_STQ_SPLIT(0x30));
1479
1480         RADEON_WRITE(R600_TA_CNTL_AUX, (R600_DISABLE_CUBE_ANISO |
1481                                         R600_SYNC_GRADIENT |
1482                                         R600_SYNC_WALKER |
1483                                         R600_SYNC_ALIGNER));
1484
1485         sx_debug_1 = RADEON_READ(R700_SX_DEBUG_1);
1486         sx_debug_1 |= R700_ENABLE_NEW_SMX_ADDRESS;
1487         RADEON_WRITE(R700_SX_DEBUG_1, sx_debug_1);
1488
1489         smx_dc_ctl0 = RADEON_READ(R600_SMX_DC_CTL0);
1490         smx_dc_ctl0 &= ~R700_CACHE_DEPTH(0x1ff);
1491         smx_dc_ctl0 |= R700_CACHE_DEPTH((dev_priv->r700_sx_num_of_sets * 64) - 1);
1492         RADEON_WRITE(R600_SMX_DC_CTL0, smx_dc_ctl0);
1493
1494         RADEON_WRITE(R700_SMX_EVENT_CTL, (R700_ES_FLUSH_CTL(4) |
1495                                           R700_GS_FLUSH_CTL(4) |
1496                                           R700_ACK_FLUSH_CTL(3) |
1497                                           R700_SYNC_FLUSH_CTL));
1498
1499         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV770)
1500                 RADEON_WRITE(R700_DB_DEBUG3, R700_DB_CLK_OFF_DELAY(0x1f));
1501         else {
1502                 db_debug4 = RADEON_READ(RV700_DB_DEBUG4);
1503                 db_debug4 |= RV700_DISABLE_TILE_COVERED_FOR_PS_ITER;
1504                 RADEON_WRITE(RV700_DB_DEBUG4, db_debug4);
1505         }
1506
1507         RADEON_WRITE(R600_SX_EXPORT_BUFFER_SIZES, (R600_COLOR_BUFFER_SIZE((dev_priv->r600_sx_max_export_size / 4) - 1) |
1508                                                    R600_POSITION_BUFFER_SIZE((dev_priv->r600_sx_max_export_pos_size / 4) - 1) |
1509                                                    R600_SMX_BUFFER_SIZE((dev_priv->r600_sx_max_export_smx_size / 4) - 1)));
1510
1511         RADEON_WRITE(R700_PA_SC_FIFO_SIZE_R7XX, (R700_SC_PRIM_FIFO_SIZE(dev_priv->r700_sc_prim_fifo_size) |
1512                                                  R700_SC_HIZ_TILE_FIFO_SIZE(dev_priv->r700_sc_hiz_tile_fifo_size) |
1513                                                  R700_SC_EARLYZ_TILE_FIFO_SIZE(dev_priv->r700_sc_earlyz_tile_fifo_fize)));
1514
1515         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1516
1517         RADEON_WRITE(R600_VGT_NUM_INSTANCES, 1);
1518
1519         RADEON_WRITE(R600_SPI_CONFIG_CNTL, R600_GPR_WRITE_PRIORITY(0));
1520
1521         RADEON_WRITE(R600_SPI_CONFIG_CNTL_1, R600_VTX_DONE_DELAY(4));
1522
1523         RADEON_WRITE(R600_CP_PERFMON_CNTL, 0);
1524
1525         sq_ms_fifo_sizes = (R600_CACHE_FIFO_SIZE(16 * dev_priv->r600_sq_num_cf_insts) |
1526                             R600_DONE_FIFO_HIWATER(0xe0) |
1527                             R600_ALU_UPDATE_FIFO_HIWATER(0x8));
1528         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1529         case CHIP_RV770:
1530                 sq_ms_fifo_sizes |= R600_FETCH_FIFO_HIWATER(0x1);
1531                 break;
1532         case CHIP_RV730:
1533         case CHIP_RV710:
1534         case CHIP_RV740:
1535         default:
1536                 sq_ms_fifo_sizes |= R600_FETCH_FIFO_HIWATER(0x4);
1537                 break;
1538         }
1539         RADEON_WRITE(R600_SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
1540
1541         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
1542          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
1543          */
1544         sq_config = RADEON_READ(R600_SQ_CONFIG);
1545         sq_config &= ~(R600_PS_PRIO(3) |
1546                        R600_VS_PRIO(3) |
1547                        R600_GS_PRIO(3) |
1548                        R600_ES_PRIO(3));
1549         sq_config |= (R600_DX9_CONSTS |
1550                       R600_VC_ENABLE |
1551                       R600_EXPORT_SRC_C |
1552                       R600_PS_PRIO(0) |
1553                       R600_VS_PRIO(1) |
1554                       R600_GS_PRIO(2) |
1555                       R600_ES_PRIO(3));
1556         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)
1557                 /* no vertex cache */
1558                 sq_config &= ~R600_VC_ENABLE;
1559
1560         RADEON_WRITE(R600_SQ_CONFIG, sq_config);
1561
1562         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_1,  (R600_NUM_PS_GPRS((dev_priv->r600_max_gprs * 24)/64) |
1563                                                     R600_NUM_VS_GPRS((dev_priv->r600_max_gprs * 24)/64) |
1564                                                     R600_NUM_CLAUSE_TEMP_GPRS(((dev_priv->r600_max_gprs * 24)/64)/2)));
1565
1566         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_2,  (R600_NUM_GS_GPRS((dev_priv->r600_max_gprs * 7)/64) |
1567                                                     R600_NUM_ES_GPRS((dev_priv->r600_max_gprs * 7)/64)));
1568
1569         sq_thread_resource_mgmt = (R600_NUM_PS_THREADS((dev_priv->r600_max_threads * 4)/8) |
1570                                    R600_NUM_VS_THREADS((dev_priv->r600_max_threads * 2)/8) |
1571                                    R600_NUM_ES_THREADS((dev_priv->r600_max_threads * 1)/8));
1572         if (((dev_priv->r600_max_threads * 1) / 8) > dev_priv->r600_max_gs_threads)
1573                 sq_thread_resource_mgmt |= R600_NUM_GS_THREADS(dev_priv->r600_max_gs_threads);
1574         else
1575                 sq_thread_resource_mgmt |= R600_NUM_GS_THREADS((dev_priv->r600_max_gs_threads * 1)/8);
1576         RADEON_WRITE(R600_SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1577
1578         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_1, (R600_NUM_PS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4) |
1579                                                      R600_NUM_VS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4)));
1580
1581         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_2, (R600_NUM_GS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4) |
1582                                                      R600_NUM_ES_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4)));
1583
1584         sq_dyn_gpr_size_simd_ab_0 = (R700_SIMDA_RING0((dev_priv->r600_max_gprs * 38)/64) |
1585                                      R700_SIMDA_RING1((dev_priv->r600_max_gprs * 38)/64) |
1586                                      R700_SIMDB_RING0((dev_priv->r600_max_gprs * 38)/64) |
1587                                      R700_SIMDB_RING1((dev_priv->r600_max_gprs * 38)/64));
1588
1589         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
1590         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
1591         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
1592         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
1593         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
1594         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
1595         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
1596         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
1597
1598         RADEON_WRITE(R700_PA_SC_FORCE_EOV_MAX_CNTS, (R700_FORCE_EOV_MAX_CLK_CNT(4095) |
1599                                                      R700_FORCE_EOV_MAX_REZ_CNT(255)));
1600
1601         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)
1602                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, (R600_CACHE_INVALIDATION(R600_TC_ONLY) |
1603                                                            R700_AUTO_INVLD_EN(R700_ES_AND_GS_AUTO)));
1604         else
1605                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, (R600_CACHE_INVALIDATION(R600_VC_AND_TC) |
1606                                                            R700_AUTO_INVLD_EN(R700_ES_AND_GS_AUTO)));
1607
1608         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1609         case CHIP_RV770:
1610         case CHIP_RV730:
1611         case CHIP_RV740:
1612                 gs_prim_buffer_depth = 384;
1613                 break;
1614         case CHIP_RV710:
1615                 gs_prim_buffer_depth = 128;
1616                 break;
1617         default:
1618                 break;
1619         }
1620
1621         num_gs_verts_per_thread = dev_priv->r600_max_pipes * 16;
1622         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
1623         /* Max value for this is 256 */
1624         if (vgt_gs_per_es > 256)
1625                 vgt_gs_per_es = 256;
1626
1627         RADEON_WRITE(R600_VGT_ES_PER_GS, 128);
1628         RADEON_WRITE(R600_VGT_GS_PER_ES, vgt_gs_per_es);
1629         RADEON_WRITE(R600_VGT_GS_PER_VS, 2);
1630
1631         /* more default values. 2D/3D driver should adjust as needed */
1632         RADEON_WRITE(R600_VGT_GS_VERTEX_REUSE, 16);
1633         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE_STATE, 0);
1634         RADEON_WRITE(R600_VGT_STRMOUT_EN, 0);
1635         RADEON_WRITE(R600_SX_MISC, 0);
1636         RADEON_WRITE(R600_PA_SC_MODE_CNTL, 0);
1637         RADEON_WRITE(R700_PA_SC_EDGERULE, 0xaaaaaaaa);
1638         RADEON_WRITE(R600_PA_SC_AA_CONFIG, 0);
1639         RADEON_WRITE(R600_PA_SC_CLIPRECT_RULE, 0xffff);
1640         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE, 0);
1641         RADEON_WRITE(R600_SPI_INPUT_Z, 0);
1642         RADEON_WRITE(R600_SPI_PS_IN_CONTROL_0, R600_NUM_INTERP(2));
1643         RADEON_WRITE(R600_CB_COLOR7_FRAG, 0);
1644
1645         /* clear render buffer base addresses */
1646         RADEON_WRITE(R600_CB_COLOR0_BASE, 0);
1647         RADEON_WRITE(R600_CB_COLOR1_BASE, 0);
1648         RADEON_WRITE(R600_CB_COLOR2_BASE, 0);
1649         RADEON_WRITE(R600_CB_COLOR3_BASE, 0);
1650         RADEON_WRITE(R600_CB_COLOR4_BASE, 0);
1651         RADEON_WRITE(R600_CB_COLOR5_BASE, 0);
1652         RADEON_WRITE(R600_CB_COLOR6_BASE, 0);
1653         RADEON_WRITE(R600_CB_COLOR7_BASE, 0);
1654
1655         RADEON_WRITE(R700_TCP_CNTL, 0);
1656
1657         hdp_host_path_cntl = RADEON_READ(R600_HDP_HOST_PATH_CNTL);
1658         RADEON_WRITE(R600_HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1659
1660         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1661
1662         RADEON_WRITE(R600_PA_CL_ENHANCE, (R600_CLIP_VTX_REORDER_ENA |
1663                                           R600_NUM_CLIP_SEQ(3)));
1664
1665 }
1666
1667 static void r600_cp_init_ring_buffer(struct drm_device *dev,
1668                                        drm_radeon_private_t *dev_priv,
1669                                        struct drm_file *file_priv)
1670 {
1671         struct drm_radeon_master_private *master_priv;
1672         u32 ring_start;
1673         u64 rptr_addr;
1674
1675         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
1676                 r700_gfx_init(dev, dev_priv);
1677         else
1678                 r600_gfx_init(dev, dev_priv);
1679
1680         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
1681         RADEON_READ(R600_GRBM_SOFT_RESET);
1682         DRM_UDELAY(15000);
1683         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
1684
1685
1686         /* Set ring buffer size */
1687 #ifdef __BIG_ENDIAN
1688         RADEON_WRITE(R600_CP_RB_CNTL,
1689                      RADEON_BUF_SWAP_32BIT |
1690                      RADEON_RB_NO_UPDATE |
1691                      (dev_priv->ring.rptr_update_l2qw << 8) |
1692                      dev_priv->ring.size_l2qw);
1693 #else
1694         RADEON_WRITE(R600_CP_RB_CNTL,
1695                      RADEON_RB_NO_UPDATE |
1696                      (dev_priv->ring.rptr_update_l2qw << 8) |
1697                      dev_priv->ring.size_l2qw);
1698 #endif
1699
1700         RADEON_WRITE(R600_CP_SEM_WAIT_TIMER, 0x4);
1701
1702         /* Set the write pointer delay */
1703         RADEON_WRITE(R600_CP_RB_WPTR_DELAY, 0);
1704
1705 #ifdef __BIG_ENDIAN
1706         RADEON_WRITE(R600_CP_RB_CNTL,
1707                      RADEON_BUF_SWAP_32BIT |
1708                      RADEON_RB_NO_UPDATE |
1709                      RADEON_RB_RPTR_WR_ENA |
1710                      (dev_priv->ring.rptr_update_l2qw << 8) |
1711                      dev_priv->ring.size_l2qw);
1712 #else
1713         RADEON_WRITE(R600_CP_RB_CNTL,
1714                      RADEON_RB_NO_UPDATE |
1715                      RADEON_RB_RPTR_WR_ENA |
1716                      (dev_priv->ring.rptr_update_l2qw << 8) |
1717                      dev_priv->ring.size_l2qw);
1718 #endif
1719
1720         /* Initialize the ring buffer's read and write pointers */
1721         RADEON_WRITE(R600_CP_RB_RPTR_WR, 0);
1722         RADEON_WRITE(R600_CP_RB_WPTR, 0);
1723         SET_RING_HEAD(dev_priv, 0);
1724         dev_priv->ring.tail = 0;
1725
1726 #if __OS_HAS_AGP
1727         if (dev_priv->flags & RADEON_IS_AGP) {
1728                 rptr_addr = dev_priv->ring_rptr->offset
1729                         - dev->agp->base +
1730                         dev_priv->gart_vm_start;
1731         } else
1732 #endif
1733         {
1734                 rptr_addr = dev_priv->ring_rptr->offset
1735                         - ((unsigned long) dev->sg->virtual)
1736                         + dev_priv->gart_vm_start;
1737         }
1738         RADEON_WRITE(R600_CP_RB_RPTR_ADDR,
1739                      rptr_addr & 0xffffffff);
1740         RADEON_WRITE(R600_CP_RB_RPTR_ADDR_HI,
1741                      upper_32_bits(rptr_addr));
1742
1743 #ifdef __BIG_ENDIAN
1744         RADEON_WRITE(R600_CP_RB_CNTL,
1745                      RADEON_BUF_SWAP_32BIT |
1746                      (dev_priv->ring.rptr_update_l2qw << 8) |
1747                      dev_priv->ring.size_l2qw);
1748 #else
1749         RADEON_WRITE(R600_CP_RB_CNTL,
1750                      (dev_priv->ring.rptr_update_l2qw << 8) |
1751                      dev_priv->ring.size_l2qw);
1752 #endif
1753
1754 #if __OS_HAS_AGP
1755         if (dev_priv->flags & RADEON_IS_AGP) {
1756                 /* XXX */
1757                 radeon_write_agp_base(dev_priv, dev->agp->base);
1758
1759                 /* XXX */
1760                 radeon_write_agp_location(dev_priv,
1761                              (((dev_priv->gart_vm_start - 1 +
1762                                 dev_priv->gart_size) & 0xffff0000) |
1763                               (dev_priv->gart_vm_start >> 16)));
1764
1765                 ring_start = (dev_priv->cp_ring->offset
1766                               - dev->agp->base
1767                               + dev_priv->gart_vm_start);
1768         } else
1769 #endif
1770                 ring_start = (dev_priv->cp_ring->offset
1771                               - (unsigned long)dev->sg->virtual
1772                               + dev_priv->gart_vm_start);
1773
1774         RADEON_WRITE(R600_CP_RB_BASE, ring_start >> 8);
1775
1776         RADEON_WRITE(R600_CP_ME_CNTL, 0xff);
1777
1778         RADEON_WRITE(R600_CP_DEBUG, (1 << 27) | (1 << 28));
1779
1780         /* Initialize the scratch register pointer.  This will cause
1781          * the scratch register values to be written out to memory
1782          * whenever they are updated.
1783          *
1784          * We simply put this behind the ring read pointer, this works
1785          * with PCI GART as well as (whatever kind of) AGP GART
1786          */
1787         {
1788                 u64 scratch_addr;
1789
1790                 scratch_addr = RADEON_READ(R600_CP_RB_RPTR_ADDR);
1791                 scratch_addr |= ((u64)RADEON_READ(R600_CP_RB_RPTR_ADDR_HI)) << 32;
1792                 scratch_addr += R600_SCRATCH_REG_OFFSET;
1793                 scratch_addr >>= 8;
1794                 scratch_addr &= 0xffffffff;
1795
1796                 RADEON_WRITE(R600_SCRATCH_ADDR, (uint32_t)scratch_addr);
1797         }
1798
1799         RADEON_WRITE(R600_SCRATCH_UMSK, 0x7);
1800
1801         /* Turn on bus mastering */
1802         radeon_enable_bm(dev_priv);
1803
1804         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(0), 0);
1805         RADEON_WRITE(R600_LAST_FRAME_REG, 0);
1806
1807         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(1), 0);
1808         RADEON_WRITE(R600_LAST_DISPATCH_REG, 0);
1809
1810         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(2), 0);
1811         RADEON_WRITE(R600_LAST_CLEAR_REG, 0);
1812
1813         /* reset sarea copies of these */
1814         master_priv = file_priv->master->driver_priv;
1815         if (master_priv->sarea_priv) {
1816                 master_priv->sarea_priv->last_frame = 0;
1817                 master_priv->sarea_priv->last_dispatch = 0;
1818                 master_priv->sarea_priv->last_clear = 0;
1819         }
1820
1821         r600_do_wait_for_idle(dev_priv);
1822
1823 }
1824
1825 int r600_do_cleanup_cp(struct drm_device *dev)
1826 {
1827         drm_radeon_private_t *dev_priv = dev->dev_private;
1828         DRM_DEBUG("\n");
1829
1830         /* Make sure interrupts are disabled here because the uninstall ioctl
1831          * may not have been called from userspace and after dev_private
1832          * is freed, it's too late.
1833          */
1834         if (dev->irq_enabled)
1835                 drm_irq_uninstall(dev);
1836
1837 #if __OS_HAS_AGP
1838         if (dev_priv->flags & RADEON_IS_AGP) {
1839                 if (dev_priv->cp_ring != NULL) {
1840                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1841                         dev_priv->cp_ring = NULL;
1842                 }
1843                 if (dev_priv->ring_rptr != NULL) {
1844                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1845                         dev_priv->ring_rptr = NULL;
1846                 }
1847                 if (dev->agp_buffer_map != NULL) {
1848                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1849                         dev->agp_buffer_map = NULL;
1850                 }
1851         } else
1852 #endif
1853         {
1854
1855                 if (dev_priv->gart_info.bus_addr)
1856                         r600_page_table_cleanup(dev, &dev_priv->gart_info);
1857
1858                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB) {
1859                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1860                         dev_priv->gart_info.addr = NULL;
1861                 }
1862         }
1863         /* only clear to the start of flags */
1864         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1865
1866         return 0;
1867 }
1868
1869 int r600_do_init_cp(struct drm_device *dev, drm_radeon_init_t *init,
1870                     struct drm_file *file_priv)
1871 {
1872         drm_radeon_private_t *dev_priv = dev->dev_private;
1873         struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
1874
1875         DRM_DEBUG("\n");
1876
1877         /* if we require new memory map but we don't have it fail */
1878         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
1879                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
1880                 r600_do_cleanup_cp(dev);
1881                 return -EINVAL;
1882         }
1883
1884         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
1885                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
1886                 dev_priv->flags &= ~RADEON_IS_AGP;
1887                 /* The writeback test succeeds, but when writeback is enabled,
1888                  * the ring buffer read ptr update fails after first 128 bytes.
1889                  */
1890                 radeon_no_wb = 1;
1891         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
1892                  && !init->is_pci) {
1893                 DRM_DEBUG("Restoring AGP flag\n");
1894                 dev_priv->flags |= RADEON_IS_AGP;
1895         }
1896
1897         dev_priv->usec_timeout = init->usec_timeout;
1898         if (dev_priv->usec_timeout < 1 ||
1899             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
1900                 DRM_DEBUG("TIMEOUT problem!\n");
1901                 r600_do_cleanup_cp(dev);
1902                 return -EINVAL;
1903         }
1904
1905         /* Enable vblank on CRTC1 for older X servers
1906          */
1907         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
1908
1909         dev_priv->cp_mode = init->cp_mode;
1910
1911         /* We don't support anything other than bus-mastering ring mode,
1912          * but the ring can be in either AGP or PCI space for the ring
1913          * read pointer.
1914          */
1915         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
1916             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
1917                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
1918                 r600_do_cleanup_cp(dev);
1919                 return -EINVAL;
1920         }
1921
1922         switch (init->fb_bpp) {
1923         case 16:
1924                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
1925                 break;
1926         case 32:
1927         default:
1928                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
1929                 break;
1930         }
1931         dev_priv->front_offset = init->front_offset;
1932         dev_priv->front_pitch = init->front_pitch;
1933         dev_priv->back_offset = init->back_offset;
1934         dev_priv->back_pitch = init->back_pitch;
1935
1936         dev_priv->ring_offset = init->ring_offset;
1937         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
1938         dev_priv->buffers_offset = init->buffers_offset;
1939         dev_priv->gart_textures_offset = init->gart_textures_offset;
1940
1941         master_priv->sarea = drm_getsarea(dev);
1942         if (!master_priv->sarea) {
1943                 DRM_ERROR("could not find sarea!\n");
1944                 r600_do_cleanup_cp(dev);
1945                 return -EINVAL;
1946         }
1947
1948         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
1949         if (!dev_priv->cp_ring) {
1950                 DRM_ERROR("could not find cp ring region!\n");
1951                 r600_do_cleanup_cp(dev);
1952                 return -EINVAL;
1953         }
1954         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
1955         if (!dev_priv->ring_rptr) {
1956                 DRM_ERROR("could not find ring read pointer!\n");
1957                 r600_do_cleanup_cp(dev);
1958                 return -EINVAL;
1959         }
1960         dev->agp_buffer_token = init->buffers_offset;
1961         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
1962         if (!dev->agp_buffer_map) {
1963                 DRM_ERROR("could not find dma buffer region!\n");
1964                 r600_do_cleanup_cp(dev);
1965                 return -EINVAL;
1966         }
1967
1968         if (init->gart_textures_offset) {
1969                 dev_priv->gart_textures =
1970                     drm_core_findmap(dev, init->gart_textures_offset);
1971                 if (!dev_priv->gart_textures) {
1972                         DRM_ERROR("could not find GART texture region!\n");
1973                         r600_do_cleanup_cp(dev);
1974                         return -EINVAL;
1975                 }
1976         }
1977
1978 #if __OS_HAS_AGP
1979         /* XXX */
1980         if (dev_priv->flags & RADEON_IS_AGP) {
1981                 drm_core_ioremap_wc(dev_priv->cp_ring, dev);
1982                 drm_core_ioremap_wc(dev_priv->ring_rptr, dev);
1983                 drm_core_ioremap_wc(dev->agp_buffer_map, dev);
1984                 if (!dev_priv->cp_ring->handle ||
1985                     !dev_priv->ring_rptr->handle ||
1986                     !dev->agp_buffer_map->handle) {
1987                         DRM_ERROR("could not find ioremap agp regions!\n");
1988                         r600_do_cleanup_cp(dev);
1989                         return -EINVAL;
1990                 }
1991         } else
1992 #endif
1993         {
1994                 dev_priv->cp_ring->handle = (void *)dev_priv->cp_ring->offset;
1995                 dev_priv->ring_rptr->handle =
1996                     (void *)dev_priv->ring_rptr->offset;
1997                 dev->agp_buffer_map->handle =
1998                     (void *)dev->agp_buffer_map->offset;
1999
2000                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
2001                           dev_priv->cp_ring->handle);
2002                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
2003                           dev_priv->ring_rptr->handle);
2004                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
2005                           dev->agp_buffer_map->handle);
2006         }
2007
2008         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 24;
2009         dev_priv->fb_size =
2010                 (((radeon_read_fb_location(dev_priv) & 0xffff0000u) << 8) + 0x1000000)
2011                 - dev_priv->fb_location;
2012
2013         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
2014                                         ((dev_priv->front_offset
2015                                           + dev_priv->fb_location) >> 10));
2016
2017         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
2018                                        ((dev_priv->back_offset
2019                                          + dev_priv->fb_location) >> 10));
2020
2021         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
2022                                         ((dev_priv->depth_offset
2023                                           + dev_priv->fb_location) >> 10));
2024
2025         dev_priv->gart_size = init->gart_size;
2026
2027         /* New let's set the memory map ... */
2028         if (dev_priv->new_memmap) {
2029                 u32 base = 0;
2030
2031                 DRM_INFO("Setting GART location based on new memory map\n");
2032
2033                 /* If using AGP, try to locate the AGP aperture at the same
2034                  * location in the card and on the bus, though we have to
2035                  * align it down.
2036                  */
2037 #if __OS_HAS_AGP
2038                 /* XXX */
2039                 if (dev_priv->flags & RADEON_IS_AGP) {
2040                         base = dev->agp->base;
2041                         /* Check if valid */
2042                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
2043                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
2044                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
2045                                          dev->agp->base);
2046                                 base = 0;
2047                         }
2048                 }
2049 #endif
2050                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
2051                 if (base == 0) {
2052                         base = dev_priv->fb_location + dev_priv->fb_size;
2053                         if (base < dev_priv->fb_location ||
2054                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
2055                                 base = dev_priv->fb_location
2056                                         - dev_priv->gart_size;
2057                 }
2058                 dev_priv->gart_vm_start = base & 0xffc00000u;
2059                 if (dev_priv->gart_vm_start != base)
2060                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
2061                                  base, dev_priv->gart_vm_start);
2062         }
2063
2064 #if __OS_HAS_AGP
2065         /* XXX */
2066         if (dev_priv->flags & RADEON_IS_AGP)
2067                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
2068                                                  - dev->agp->base
2069                                                  + dev_priv->gart_vm_start);
2070         else
2071 #endif
2072                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
2073                                                  - (unsigned long)dev->sg->virtual
2074                                                  + dev_priv->gart_vm_start);
2075
2076         DRM_DEBUG("fb 0x%08x size %d\n",
2077                   (unsigned int) dev_priv->fb_location,
2078                   (unsigned int) dev_priv->fb_size);
2079         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
2080         DRM_DEBUG("dev_priv->gart_vm_start 0x%08x\n",
2081                   (unsigned int) dev_priv->gart_vm_start);
2082         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%08lx\n",
2083                   dev_priv->gart_buffers_offset);
2084
2085         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
2086         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
2087                               + init->ring_size / sizeof(u32));
2088         dev_priv->ring.size = init->ring_size;
2089         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
2090
2091         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
2092         dev_priv->ring.rptr_update_l2qw = drm_order(/* init->rptr_update */ 4096 / 8);
2093
2094         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
2095         dev_priv->ring.fetch_size_l2ow = drm_order(/* init->fetch_size */ 32 / 16);
2096
2097         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
2098
2099         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
2100
2101 #if __OS_HAS_AGP
2102         if (dev_priv->flags & RADEON_IS_AGP) {
2103                 /* XXX turn off pcie gart */
2104         } else
2105 #endif
2106         {
2107                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
2108                 /* if we have an offset set from userspace */
2109                 if (!dev_priv->pcigart_offset_set) {
2110                         DRM_ERROR("Need gart offset from userspace\n");
2111                         r600_do_cleanup_cp(dev);
2112                         return -EINVAL;
2113                 }
2114
2115                 DRM_DEBUG("Using gart offset 0x%08lx\n", dev_priv->pcigart_offset);
2116
2117                 dev_priv->gart_info.bus_addr =
2118                         dev_priv->pcigart_offset + dev_priv->fb_location;
2119                 dev_priv->gart_info.mapping.offset =
2120                         dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
2121                 dev_priv->gart_info.mapping.size =
2122                         dev_priv->gart_info.table_size;
2123
2124                 drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
2125                 if (!dev_priv->gart_info.mapping.handle) {
2126                         DRM_ERROR("ioremap failed.\n");
2127                         r600_do_cleanup_cp(dev);
2128                         return -EINVAL;
2129                 }
2130
2131                 dev_priv->gart_info.addr =
2132                         dev_priv->gart_info.mapping.handle;
2133
2134                 DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
2135                           dev_priv->gart_info.addr,
2136                           dev_priv->pcigart_offset);
2137
2138                 if (!r600_page_table_init(dev)) {
2139                         DRM_ERROR("Failed to init GART table\n");
2140                         r600_do_cleanup_cp(dev);
2141                         return -EINVAL;
2142                 }
2143
2144                 if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
2145                         r700_vm_init(dev);
2146                 else
2147                         r600_vm_init(dev);
2148         }
2149
2150         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
2151                 r700_cp_load_microcode(dev_priv);
2152         else
2153                 r600_cp_load_microcode(dev_priv);
2154
2155         r600_cp_init_ring_buffer(dev, dev_priv, file_priv);
2156
2157         dev_priv->last_buf = 0;
2158
2159         r600_do_engine_reset(dev);
2160         r600_test_writeback(dev_priv);
2161
2162         return 0;
2163 }
2164
2165 int r600_do_resume_cp(struct drm_device *dev, struct drm_file *file_priv)
2166 {
2167         drm_radeon_private_t *dev_priv = dev->dev_private;
2168
2169         DRM_DEBUG("\n");
2170         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)) {
2171                 r700_vm_init(dev);
2172                 r700_cp_load_microcode(dev_priv);
2173         } else {
2174                 r600_vm_init(dev);
2175                 r600_cp_load_microcode(dev_priv);
2176         }
2177         r600_cp_init_ring_buffer(dev, dev_priv, file_priv);
2178         r600_do_engine_reset(dev);
2179
2180         return 0;
2181 }
2182
2183 /* Wait for the CP to go idle.
2184  */
2185 int r600_do_cp_idle(drm_radeon_private_t *dev_priv)
2186 {
2187         RING_LOCALS;
2188         DRM_DEBUG("\n");
2189
2190         BEGIN_RING(5);
2191         OUT_RING(CP_PACKET3(R600_IT_EVENT_WRITE, 0));
2192         OUT_RING(R600_CACHE_FLUSH_AND_INV_EVENT);
2193         /* wait for 3D idle clean */
2194         OUT_RING(CP_PACKET3(R600_IT_SET_CONFIG_REG, 1));
2195         OUT_RING((R600_WAIT_UNTIL - R600_SET_CONFIG_REG_OFFSET) >> 2);
2196         OUT_RING(RADEON_WAIT_3D_IDLE | RADEON_WAIT_3D_IDLECLEAN);
2197
2198         ADVANCE_RING();
2199         COMMIT_RING();
2200
2201         return r600_do_wait_for_idle(dev_priv);
2202 }
2203
2204 /* Start the Command Processor.
2205  */
2206 void r600_do_cp_start(drm_radeon_private_t *dev_priv)
2207 {
2208         u32 cp_me;
2209         RING_LOCALS;
2210         DRM_DEBUG("\n");
2211
2212         BEGIN_RING(7);
2213         OUT_RING(CP_PACKET3(R600_IT_ME_INITIALIZE, 5));
2214         OUT_RING(0x00000001);
2215         if (((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV770))
2216                 OUT_RING(0x00000003);
2217         else
2218                 OUT_RING(0x00000000);
2219         OUT_RING((dev_priv->r600_max_hw_contexts - 1));
2220         OUT_RING(R600_ME_INITIALIZE_DEVICE_ID(1));
2221         OUT_RING(0x00000000);
2222         OUT_RING(0x00000000);
2223         ADVANCE_RING();
2224         COMMIT_RING();
2225
2226         /* set the mux and reset the halt bit */
2227         cp_me = 0xff;
2228         RADEON_WRITE(R600_CP_ME_CNTL, cp_me);
2229
2230         dev_priv->cp_running = 1;
2231
2232 }
2233
2234 void r600_do_cp_reset(drm_radeon_private_t *dev_priv)
2235 {
2236         u32 cur_read_ptr;
2237         DRM_DEBUG("\n");
2238
2239         cur_read_ptr = RADEON_READ(R600_CP_RB_RPTR);
2240         RADEON_WRITE(R600_CP_RB_WPTR, cur_read_ptr);
2241         SET_RING_HEAD(dev_priv, cur_read_ptr);
2242         dev_priv->ring.tail = cur_read_ptr;
2243 }
2244
2245 void r600_do_cp_stop(drm_radeon_private_t *dev_priv)
2246 {
2247         uint32_t cp_me;
2248
2249         DRM_DEBUG("\n");
2250
2251         cp_me = 0xff | R600_CP_ME_HALT;
2252
2253         RADEON_WRITE(R600_CP_ME_CNTL, cp_me);
2254
2255         dev_priv->cp_running = 0;
2256 }
2257
2258 int r600_cp_dispatch_indirect(struct drm_device *dev,
2259                               struct drm_buf *buf, int start, int end)
2260 {
2261         drm_radeon_private_t *dev_priv = dev->dev_private;
2262         RING_LOCALS;
2263
2264         if (start != end) {
2265                 unsigned long offset = (dev_priv->gart_buffers_offset
2266                                         + buf->offset + start);
2267                 int dwords = (end - start + 3) / sizeof(u32);
2268
2269                 DRM_DEBUG("dwords:%d\n", dwords);
2270                 DRM_DEBUG("offset 0x%lx\n", offset);
2271
2272
2273                 /* Indirect buffer data must be a multiple of 16 dwords.
2274                  * pad the data with a Type-2 CP packet.
2275                  */
2276                 while (dwords & 0xf) {
2277                         u32 *data = (u32 *)
2278                             ((char *)dev->agp_buffer_map->handle
2279                              + buf->offset + start);
2280                         data[dwords++] = RADEON_CP_PACKET2;
2281                 }
2282
2283                 /* Fire off the indirect buffer */
2284                 BEGIN_RING(4);
2285                 OUT_RING(CP_PACKET3(R600_IT_INDIRECT_BUFFER, 2));
2286                 OUT_RING((offset & 0xfffffffc));
2287                 OUT_RING((upper_32_bits(offset) & 0xff));
2288                 OUT_RING(dwords);
2289                 ADVANCE_RING();
2290         }
2291
2292         return 0;
2293 }