SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/gpu/drm/radeon/rv770.c

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/firmware.h>
  29 #include <linux/platform_device.h>
  30 #include "drmP.h"
  31 #include "radeon.h"
  32 #include "radeon_drm.h"
  33 #include "rv770d.h"
  34 #include "atom.h"
  35 #include "avivod.h"
  36
  37 #define R700_PFP_UCODE_SIZE 848
  38 #define R700_PM4_UCODE_SIZE 1360
  39
  40 static void rv770_gpu_init(struct radeon_device *rdev);
  41 void rv770_fini(struct radeon_device *rdev);
  42
  43
  44 /*
  45  * GART
  46  */
  47 int rv770_pcie_gart_enable(struct radeon_device *rdev)
  48 {
  49         u32 tmp;
  50         int r, i;
  51
  52         if (rdev->gart.table.vram.robj == NULL) {
  53                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
  54                 return -EINVAL;
  55         }
  56         r = radeon_gart_table_vram_pin(rdev);
  57         if (r)
  58                 return r;
  59         /* Setup L2 cache */
  60         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
  61                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
  62                                 EFFECTIVE_L2_QUEUE_SIZE(7));
  63         WREG32(VM_L2_CNTL2, 0);
  64         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
  65         /* Setup TLB control */
  66         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
  67                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
  68                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
  69                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
  70         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
  71         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
  72         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
  73         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
  74         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
  75         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
  76         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
  77         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
  78         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
  79         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
  80         WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
  81                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
  82         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
  83                         (u32)(rdev->dummy_page.addr >> 12));
  84         for (i = 1; i < 7; i++)
  85                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
  86
  87         r600_pcie_gart_tlb_flush(rdev);
  88         rdev->gart.ready = true;
  89         return 0;
  90 }
  91
  92 void rv770_pcie_gart_disable(struct radeon_device *rdev)
  93 {
  94         u32 tmp;
  95         int i, r;
  96
  97         /* Disable all tables */
  98         for (i = 0; i < 7; i++)
  99                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 100
 101         /* Setup L2 cache */
 102         WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING |
 103                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 104         WREG32(VM_L2_CNTL2, 0);
 105         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 106         /* Setup TLB control */
 107         tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 108         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 109         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 110         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 111         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 112         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 113         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 114         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 115         if (rdev->gart.table.vram.robj) {
 116                 r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
 117                 if (likely(r == 0)) {
 118                         radeon_bo_kunmap(rdev->gart.table.vram.robj);
 119                         radeon_bo_unpin(rdev->gart.table.vram.robj);
 120                         radeon_bo_unreserve(rdev->gart.table.vram.robj);
 121                 }
 122         }
 123 }
 124
 125 void rv770_pcie_gart_fini(struct radeon_device *rdev)
 126 {
 127         rv770_pcie_gart_disable(rdev);
 128         radeon_gart_table_vram_free(rdev);
 129         radeon_gart_fini(rdev);
 130 }
 131
 132
 133 void rv770_agp_enable(struct radeon_device *rdev)
 134 {
 135         u32 tmp;
 136         int i;
 137
 138         /* Setup L2 cache */
 139         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 140                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 141                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 142         WREG32(VM_L2_CNTL2, 0);
 143         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 144         /* Setup TLB control */
 145         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 146                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 147                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 148                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 149         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 150         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 151         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 152         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 153         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 154         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 155         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 156         for (i = 0; i < 7; i++)
 157                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 158 }
 159
 160 static void rv770_mc_program(struct radeon_device *rdev)
 161 {
 162         struct rv515_mc_save save;
 163         u32 tmp;
 164         int i, j;
 165
 166         /* Initialize HDP */
 167         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
 168                 WREG32((0x2c14 + j), 0x00000000);
 169                 WREG32((0x2c18 + j), 0x00000000);
 170                 WREG32((0x2c1c + j), 0x00000000);
 171                 WREG32((0x2c20 + j), 0x00000000);
 172                 WREG32((0x2c24 + j), 0x00000000);
 173         }
 174         WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
 175
 176         rv515_mc_stop(rdev, &save);
 177         if (r600_mc_wait_for_idle(rdev)) {
 178                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 179         }
 180         /* Lockout access through VGA aperture*/
 181         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
 182         /* Update configuration */
 183         if (rdev->flags & RADEON_IS_AGP) {
 184                 if (rdev->mc.vram_start < rdev->mc.gtt_start) {
 185                         /* VRAM before AGP */
 186                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 187                                 rdev->mc.vram_start >> 12);
 188                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 189                                 rdev->mc.gtt_end >> 12);
 190                 } else {
 191                         /* VRAM after AGP */
 192                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 193                                 rdev->mc.gtt_start >> 12);
 194                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 195                                 rdev->mc.vram_end >> 12);
 196                 }
 197         } else {
 198                 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 199                         rdev->mc.vram_start >> 12);
 200                 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 201                         rdev->mc.vram_end >> 12);
 202         }
 203         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 204         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
 205         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
 206         WREG32(MC_VM_FB_LOCATION, tmp);
 207         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
 208         WREG32(HDP_NONSURFACE_INFO, (2 << 7));
 209         WREG32(HDP_NONSURFACE_SIZE, (rdev->mc.mc_vram_size - 1) | 0x3FF);
 210         if (rdev->flags & RADEON_IS_AGP) {
 211                 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16);
 212                 WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16);
 213                 WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22);
 214         } else {
 215                 WREG32(MC_VM_AGP_BASE, 0);
 216                 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
 217                 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
 218         }
 219         if (r600_mc_wait_for_idle(rdev)) {
 220                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 221         }
 222         rv515_mc_resume(rdev, &save);
 223         /* we need to own VRAM, so turn off the VGA renderer here
 224          * to stop it overwriting our objects */
 225         rv515_vga_render_disable(rdev);
 226 }
 227
 228
 229 /*
 230  * CP.
 231  */
 232 void r700_cp_stop(struct radeon_device *rdev)
 233 {
 234         WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
 235 }
 236
 237
 238 static int rv770_cp_load_microcode(struct radeon_device *rdev)
 239 {
 240         const __be32 *fw_data;
 241         int i;
 242
 243         if (!rdev->me_fw || !rdev->pfp_fw)
 244                 return -EINVAL;
 245
 246         r700_cp_stop(rdev);
 247         WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0));
 248
 249         /* Reset cp */
 250         WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
 251         RREG32(GRBM_SOFT_RESET);
 252         mdelay(15);
 253         WREG32(GRBM_SOFT_RESET, 0);
 254
 255         fw_data = (const __be32 *)rdev->pfp_fw->data;
 256         WREG32(CP_PFP_UCODE_ADDR, 0);
 257         for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 258                 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
 259         WREG32(CP_PFP_UCODE_ADDR, 0);
 260
 261         fw_data = (const __be32 *)rdev->me_fw->data;
 262         WREG32(CP_ME_RAM_WADDR, 0);
 263         for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 264                 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
 265
 266         WREG32(CP_PFP_UCODE_ADDR, 0);
 267         WREG32(CP_ME_RAM_WADDR, 0);
 268         WREG32(CP_ME_RAM_RADDR, 0);
 269         return 0;
 270 }
 271
 272
 273 /*
 274  * Core functions
 275  */
 276 static u32 r700_get_tile_pipe_to_backend_map(u32 num_tile_pipes,
 277                                                 u32 num_backends,
 278                                                 u32 backend_disable_mask)
 279 {
 280         u32 backend_map = 0;
 281         u32 enabled_backends_mask;
 282         u32 enabled_backends_count;
 283         u32 cur_pipe;
 284         u32 swizzle_pipe[R7XX_MAX_PIPES];
 285         u32 cur_backend;
 286         u32 i;
 287
 288         if (num_tile_pipes > R7XX_MAX_PIPES)
 289                 num_tile_pipes = R7XX_MAX_PIPES;
 290         if (num_tile_pipes < 1)
 291                 num_tile_pipes = 1;
 292         if (num_backends > R7XX_MAX_BACKENDS)
 293                 num_backends = R7XX_MAX_BACKENDS;
 294         if (num_backends < 1)
 295                 num_backends = 1;
 296
 297         enabled_backends_mask = 0;
 298         enabled_backends_count = 0;
 299         for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
 300                 if (((backend_disable_mask >> i) & 1) == 0) {
 301                         enabled_backends_mask |= (1 << i);
 302                         ++enabled_backends_count;
 303                 }
 304                 if (enabled_backends_count == num_backends)
 305                         break;
 306         }
 307
 308         if (enabled_backends_count == 0) {
 309                 enabled_backends_mask = 1;
 310                 enabled_backends_count = 1;
 311         }
 312
 313         if (enabled_backends_count != num_backends)
 314                 num_backends = enabled_backends_count;
 315
 316         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
 317         switch (num_tile_pipes) {
 318         case 1:
 319                 swizzle_pipe[0] = 0;
 320                 break;
 321         case 2:
 322                 swizzle_pipe[0] = 0;
 323                 swizzle_pipe[1] = 1;
 324                 break;
 325         case 3:
 326                 swizzle_pipe[0] = 0;
 327                 swizzle_pipe[1] = 2;
 328                 swizzle_pipe[2] = 1;
 329                 break;
 330         case 4:
 331                 swizzle_pipe[0] = 0;
 332                 swizzle_pipe[1] = 2;
 333                 swizzle_pipe[2] = 3;
 334                 swizzle_pipe[3] = 1;
 335                 break;
 336         case 5:
 337                 swizzle_pipe[0] = 0;
 338                 swizzle_pipe[1] = 2;
 339                 swizzle_pipe[2] = 4;
 340                 swizzle_pipe[3] = 1;
 341                 swizzle_pipe[4] = 3;
 342                 break;
 343         case 6:
 344                 swizzle_pipe[0] = 0;
 345                 swizzle_pipe[1] = 2;
 346                 swizzle_pipe[2] = 4;
 347                 swizzle_pipe[3] = 5;
 348                 swizzle_pipe[4] = 3;
 349                 swizzle_pipe[5] = 1;
 350                 break;
 351         case 7:
 352                 swizzle_pipe[0] = 0;
 353                 swizzle_pipe[1] = 2;
 354                 swizzle_pipe[2] = 4;
 355                 swizzle_pipe[3] = 6;
 356                 swizzle_pipe[4] = 3;
 357                 swizzle_pipe[5] = 1;
 358                 swizzle_pipe[6] = 5;
 359                 break;
 360         case 8:
 361                 swizzle_pipe[0] = 0;
 362                 swizzle_pipe[1] = 2;
 363                 swizzle_pipe[2] = 4;
 364                 swizzle_pipe[3] = 6;
 365                 swizzle_pipe[4] = 3;
 366                 swizzle_pipe[5] = 1;
 367                 swizzle_pipe[6] = 7;
 368                 swizzle_pipe[7] = 5;
 369                 break;
 370         }
 371
 372         cur_backend = 0;
 373         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
 374                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
 375                         cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
 376
 377                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
 378
 379                 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
 380         }
 381
 382         return backend_map;
 383 }
 384
 385 static void rv770_gpu_init(struct radeon_device *rdev)
 386 {
 387         int i, j, num_qd_pipes;
 388         u32 sx_debug_1;
 389         u32 smx_dc_ctl0;
 390         u32 num_gs_verts_per_thread;
 391         u32 vgt_gs_per_es;
 392         u32 gs_prim_buffer_depth = 0;
 393         u32 sq_ms_fifo_sizes;
 394         u32 sq_config;
 395         u32 sq_thread_resource_mgmt;
 396         u32 hdp_host_path_cntl;
 397         u32 sq_dyn_gpr_size_simd_ab_0;
 398         u32 backend_map;
 399         u32 gb_tiling_config = 0;
 400         u32 cc_rb_backend_disable = 0;
 401         u32 cc_gc_shader_pipe_config = 0;
 402         u32 mc_arb_ramcfg;
 403         u32 db_debug4;
 404
 405         /* setup chip specs */
 406         switch (rdev->family) {
 407         case CHIP_RV770:
 408                 rdev->config.rv770.max_pipes = 4;
 409                 rdev->config.rv770.max_tile_pipes = 8;
 410                 rdev->config.rv770.max_simds = 10;
 411                 rdev->config.rv770.max_backends = 4;
 412                 rdev->config.rv770.max_gprs = 256;
 413                 rdev->config.rv770.max_threads = 248;
 414                 rdev->config.rv770.max_stack_entries = 512;
 415                 rdev->config.rv770.max_hw_contexts = 8;
 416                 rdev->config.rv770.max_gs_threads = 16 * 2;
 417                 rdev->config.rv770.sx_max_export_size = 128;
 418                 rdev->config.rv770.sx_max_export_pos_size = 16;
 419                 rdev->config.rv770.sx_max_export_smx_size = 112;
 420                 rdev->config.rv770.sq_num_cf_insts = 2;
 421
 422                 rdev->config.rv770.sx_num_of_sets = 7;
 423                 rdev->config.rv770.sc_prim_fifo_size = 0xF9;
 424                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 425                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 426                 break;
 427         case CHIP_RV730:
 428                 rdev->config.rv770.max_pipes = 2;
 429                 rdev->config.rv770.max_tile_pipes = 4;
 430                 rdev->config.rv770.max_simds = 8;
 431                 rdev->config.rv770.max_backends = 2;
 432                 rdev->config.rv770.max_gprs = 128;
 433                 rdev->config.rv770.max_threads = 248;
 434                 rdev->config.rv770.max_stack_entries = 256;
 435                 rdev->config.rv770.max_hw_contexts = 8;
 436                 rdev->config.rv770.max_gs_threads = 16 * 2;
 437                 rdev->config.rv770.sx_max_export_size = 256;
 438                 rdev->config.rv770.sx_max_export_pos_size = 32;
 439                 rdev->config.rv770.sx_max_export_smx_size = 224;
 440                 rdev->config.rv770.sq_num_cf_insts = 2;
 441
 442                 rdev->config.rv770.sx_num_of_sets = 7;
 443                 rdev->config.rv770.sc_prim_fifo_size = 0xf9;
 444                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 445                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 446                 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
 447                         rdev->config.rv770.sx_max_export_pos_size -= 16;
 448                         rdev->config.rv770.sx_max_export_smx_size += 16;
 449                 }
 450                 break;
 451         case CHIP_RV710:
 452                 rdev->config.rv770.max_pipes = 2;
 453                 rdev->config.rv770.max_tile_pipes = 2;
 454                 rdev->config.rv770.max_simds = 2;
 455                 rdev->config.rv770.max_backends = 1;
 456                 rdev->config.rv770.max_gprs = 256;
 457                 rdev->config.rv770.max_threads = 192;
 458                 rdev->config.rv770.max_stack_entries = 256;
 459                 rdev->config.rv770.max_hw_contexts = 4;
 460                 rdev->config.rv770.max_gs_threads = 8 * 2;
 461                 rdev->config.rv770.sx_max_export_size = 128;
 462                 rdev->config.rv770.sx_max_export_pos_size = 16;
 463                 rdev->config.rv770.sx_max_export_smx_size = 112;
 464                 rdev->config.rv770.sq_num_cf_insts = 1;
 465
 466                 rdev->config.rv770.sx_num_of_sets = 7;
 467                 rdev->config.rv770.sc_prim_fifo_size = 0x40;
 468                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 469                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 470                 break;
 471         case CHIP_RV740:
 472                 rdev->config.rv770.max_pipes = 4;
 473                 rdev->config.rv770.max_tile_pipes = 4;
 474                 rdev->config.rv770.max_simds = 8;
 475                 rdev->config.rv770.max_backends = 4;
 476                 rdev->config.rv770.max_gprs = 256;
 477                 rdev->config.rv770.max_threads = 248;
 478                 rdev->config.rv770.max_stack_entries = 512;
 479                 rdev->config.rv770.max_hw_contexts = 8;
 480                 rdev->config.rv770.max_gs_threads = 16 * 2;
 481                 rdev->config.rv770.sx_max_export_size = 256;
 482                 rdev->config.rv770.sx_max_export_pos_size = 32;
 483                 rdev->config.rv770.sx_max_export_smx_size = 224;
 484                 rdev->config.rv770.sq_num_cf_insts = 2;
 485
 486                 rdev->config.rv770.sx_num_of_sets = 7;
 487                 rdev->config.rv770.sc_prim_fifo_size = 0x100;
 488                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 489                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 490
 491                 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
 492                         rdev->config.rv770.sx_max_export_pos_size -= 16;
 493                         rdev->config.rv770.sx_max_export_smx_size += 16;
 494                 }
 495                 break;
 496         default:
 497                 break;
 498         }
 499
 500         /* Initialize HDP */
 501         j = 0;
 502         for (i = 0; i < 32; i++) {
 503                 WREG32((0x2c14 + j), 0x00000000);
 504                 WREG32((0x2c18 + j), 0x00000000);
 505                 WREG32((0x2c1c + j), 0x00000000);
 506                 WREG32((0x2c20 + j), 0x00000000);
 507                 WREG32((0x2c24 + j), 0x00000000);
 508                 j += 0x18;
 509         }
 510
 511         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
 512
 513         /* setup tiling, simd, pipe config */
 514         mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
 515
 516         switch (rdev->config.rv770.max_tile_pipes) {
 517         case 1:
 518                 gb_tiling_config |= PIPE_TILING(0);
 519                 break;
 520         case 2:
 521                 gb_tiling_config |= PIPE_TILING(1);
 522                 break;
 523         case 4:
 524                 gb_tiling_config |= PIPE_TILING(2);
 525                 break;
 526         case 8:
 527                 gb_tiling_config |= PIPE_TILING(3);
 528                 break;
 529         default:
 530                 break;
 531         }
 532
 533         if (rdev->family == CHIP_RV770)
 534                 gb_tiling_config |= BANK_TILING(1);
 535         else
 536                 gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
 537
 538         gb_tiling_config |= GROUP_SIZE(0);
 539
 540         if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) {
 541                 gb_tiling_config |= ROW_TILING(3);
 542                 gb_tiling_config |= SAMPLE_SPLIT(3);
 543         } else {
 544                 gb_tiling_config |=
 545                         ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 546                 gb_tiling_config |=
 547                         SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 548         }
 549
 550         gb_tiling_config |= BANK_SWAPS(1);
 551
 552         if (rdev->family == CHIP_RV740)
 553                 backend_map = 0x28;
 554         else
 555                 backend_map = r700_get_tile_pipe_to_backend_map(rdev->config.rv770.max_tile_pipes,
 556                                                                 rdev->config.rv770.max_backends,
 557                                                                 (0xff << rdev->config.rv770.max_backends) & 0xff);
 558         gb_tiling_config |= BACKEND_MAP(backend_map);
 559
 560         cc_gc_shader_pipe_config =
 561                 INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK);
 562         cc_gc_shader_pipe_config |=
 563                 INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK);
 564
 565         cc_rb_backend_disable =
 566                 BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK);
 567
 568         WREG32(GB_TILING_CONFIG, gb_tiling_config);
 569         WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 570         WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 571
 572         WREG32(CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
 573         WREG32(CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
 574         WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
 575
 576         WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable);
 577         WREG32(CGTS_SYS_TCC_DISABLE, 0);
 578         WREG32(CGTS_TCC_DISABLE, 0);
 579         WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
 580         WREG32(CGTS_USER_TCC_DISABLE, 0);
 581
 582         num_qd_pipes =
 583                 R7XX_MAX_BACKENDS - r600_count_pipe_bits(cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK);
 584         WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK);
 585         WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK);
 586
 587         /* set HW defaults for 3D engine */
 588         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
 589                                      ROQ_IB2_START(0x2b)));
 590
 591         WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
 592
 593         WREG32(TA_CNTL_AUX, (DISABLE_CUBE_ANISO |
 594                              SYNC_GRADIENT |
 595                              SYNC_WALKER |
 596                              SYNC_ALIGNER));
 597
 598         sx_debug_1 = RREG32(SX_DEBUG_1);
 599         sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
 600         WREG32(SX_DEBUG_1, sx_debug_1);
 601
 602         smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
 603         smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff);
 604         smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1);
 605         WREG32(SMX_DC_CTL0, smx_dc_ctl0);
 606
 607         WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) |
 608                                GS_FLUSH_CTL(4) |
 609                                ACK_FLUSH_CTL(3) |
 610                                SYNC_FLUSH_CTL));
 611
 612         if (rdev->family == CHIP_RV770)
 613                 WREG32(DB_DEBUG3, DB_CLK_OFF_DELAY(0x1f));
 614         else {
 615                 db_debug4 = RREG32(DB_DEBUG4);
 616                 db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER;
 617                 WREG32(DB_DEBUG4, db_debug4);
 618         }
 619
 620         WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) |
 621                                         POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) |
 622                                         SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1)));
 623
 624         WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) |
 625                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) |
 626                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize)));
 627
 628         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 629
 630         WREG32(VGT_NUM_INSTANCES, 1);
 631
 632         WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0));
 633
 634         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
 635
 636         WREG32(CP_PERFMON_CNTL, 0);
 637
 638         sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) |
 639                             DONE_FIFO_HIWATER(0xe0) |
 640                             ALU_UPDATE_FIFO_HIWATER(0x8));
 641         switch (rdev->family) {
 642         case CHIP_RV770:
 643                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1);
 644                 break;
 645         case CHIP_RV730:
 646         case CHIP_RV710:
 647         case CHIP_RV740:
 648         default:
 649                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4);
 650                 break;
 651         }
 652         WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 653
 654         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 655          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 656          */
 657         sq_config = RREG32(SQ_CONFIG);
 658         sq_config &= ~(PS_PRIO(3) |
 659                        VS_PRIO(3) |
 660                        GS_PRIO(3) |
 661                        ES_PRIO(3));
 662         sq_config |= (DX9_CONSTS |
 663                       VC_ENABLE |
 664                       EXPORT_SRC_C |
 665                       PS_PRIO(0) |
 666                       VS_PRIO(1) |
 667                       GS_PRIO(2) |
 668                       ES_PRIO(3));
 669         if (rdev->family == CHIP_RV710)
 670                 /* no vertex cache */
 671                 sq_config &= ~VC_ENABLE;
 672
 673         WREG32(SQ_CONFIG, sq_config);
 674
 675         WREG32(SQ_GPR_RESOURCE_MGMT_1,  (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 676                                          NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 677                                          NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2)));
 678
 679         WREG32(SQ_GPR_RESOURCE_MGMT_2,  (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) |
 680                                          NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64)));
 681
 682         sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) |
 683                                    NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) |
 684                                    NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8));
 685         if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads)
 686                 sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads);
 687         else
 688                 sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8);
 689         WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
 690
 691         WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 692                                                      NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 693
 694         WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 695                                                      NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 696
 697         sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 698                                      SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) |
 699                                      SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 700                                      SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64));
 701
 702         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
 703         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
 704         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
 705         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
 706         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
 707         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
 708         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
 709         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
 710
 711         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
 712                                           FORCE_EOV_MAX_REZ_CNT(255)));
 713
 714         if (rdev->family == CHIP_RV710)
 715                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) |
 716                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 717         else
 718                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) |
 719                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 720
 721         switch (rdev->family) {
 722         case CHIP_RV770:
 723         case CHIP_RV730:
 724         case CHIP_RV740:
 725                 gs_prim_buffer_depth = 384;
 726                 break;
 727         case CHIP_RV710:
 728                 gs_prim_buffer_depth = 128;
 729                 break;
 730         default:
 731                 break;
 732         }
 733
 734         num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16;
 735         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
 736         /* Max value for this is 256 */
 737         if (vgt_gs_per_es > 256)
 738                 vgt_gs_per_es = 256;
 739
 740         WREG32(VGT_ES_PER_GS, 128);
 741         WREG32(VGT_GS_PER_ES, vgt_gs_per_es);
 742         WREG32(VGT_GS_PER_VS, 2);
 743
 744         /* more default values. 2D/3D driver should adjust as needed */
 745         WREG32(VGT_GS_VERTEX_REUSE, 16);
 746         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
 747         WREG32(VGT_STRMOUT_EN, 0);
 748         WREG32(SX_MISC, 0);
 749         WREG32(PA_SC_MODE_CNTL, 0);
 750         WREG32(PA_SC_EDGERULE, 0xaaaaaaaa);
 751         WREG32(PA_SC_AA_CONFIG, 0);
 752         WREG32(PA_SC_CLIPRECT_RULE, 0xffff);
 753         WREG32(PA_SC_LINE_STIPPLE, 0);
 754         WREG32(SPI_INPUT_Z, 0);
 755         WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
 756         WREG32(CB_COLOR7_FRAG, 0);
 757
 758         /* clear render buffer base addresses */
 759         WREG32(CB_COLOR0_BASE, 0);
 760         WREG32(CB_COLOR1_BASE, 0);
 761         WREG32(CB_COLOR2_BASE, 0);
 762         WREG32(CB_COLOR3_BASE, 0);
 763         WREG32(CB_COLOR4_BASE, 0);
 764         WREG32(CB_COLOR5_BASE, 0);
 765         WREG32(CB_COLOR6_BASE, 0);
 766         WREG32(CB_COLOR7_BASE, 0);
 767
 768         WREG32(TCP_CNTL, 0);
 769
 770         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
 771         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
 772
 773         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 774
 775         WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
 776                                           NUM_CLIP_SEQ(3)));
 777
 778 }
 779
 780 int rv770_mc_init(struct radeon_device *rdev)
 781 {
 782         fixed20_12 a;
 783         u32 tmp;
 784         int chansize, numchan;
 785
 786         /* Get VRAM informations */
 787         rdev->mc.vram_is_ddr = true;
 788         tmp = RREG32(MC_ARB_RAMCFG);
 789         if (tmp & CHANSIZE_OVERRIDE) {
 790                 chansize = 16;
 791         } else if (tmp & CHANSIZE_MASK) {
 792                 chansize = 64;
 793         } else {
 794                 chansize = 32;
 795         }
 796         tmp = RREG32(MC_SHARED_CHMAP);
 797         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
 798         case 0:
 799         default:
 800                 numchan = 1;
 801                 break;
 802         case 1:
 803                 numchan = 2;
 804                 break;
 805         case 2:
 806                 numchan = 4;
 807                 break;
 808         case 3:
 809                 numchan = 8;
 810                 break;
 811         }
 812         rdev->mc.vram_width = numchan * chansize;
 813         /* Could aper size report 0 ? */
 814         rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
 815         rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
 816         /* Setup GPU memory space */
 817         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
 818         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
 819
 820         if (rdev->mc.mc_vram_size > rdev->mc.aper_size)
 821                 rdev->mc.mc_vram_size = rdev->mc.aper_size;
 822
 823         if (rdev->mc.real_vram_size > rdev->mc.aper_size)
 824                 rdev->mc.real_vram_size = rdev->mc.aper_size;
 825
 826         if (rdev->flags & RADEON_IS_AGP) {
 827                 /* gtt_size is setup by radeon_agp_init */
 828                 rdev->mc.gtt_location = rdev->mc.agp_base;
 829                 tmp = 0xFFFFFFFFUL - rdev->mc.agp_base - rdev->mc.gtt_size;
 830                 /* Try to put vram before or after AGP because we
 831                  * we want SYSTEM_APERTURE to cover both VRAM and
 832                  * AGP so that GPU can catch out of VRAM/AGP access
 833                  */
 834                 if (rdev->mc.gtt_location > rdev->mc.mc_vram_size) {
 835                         /* Enough place before */
 836                         rdev->mc.vram_location = rdev->mc.gtt_location -
 837                                                         rdev->mc.mc_vram_size;
 838                 } else if (tmp > rdev->mc.mc_vram_size) {
 839                         /* Enough place after */
 840                         rdev->mc.vram_location = rdev->mc.gtt_location +
 841                                                         rdev->mc.gtt_size;
 842                 } else {
 843                         /* Try to setup VRAM then AGP might not
 844                          * not work on some card
 845                          */
 846                         rdev->mc.vram_location = 0x00000000UL;
 847                         rdev->mc.gtt_location = rdev->mc.mc_vram_size;
 848                 }
 849         } else {
 850                 rdev->mc.vram_location = 0x00000000UL;
 851                 rdev->mc.gtt_location = rdev->mc.mc_vram_size;
 852                 rdev->mc.gtt_size = radeon_gart_size * 1024 * 1024;
 853         }
 854         rdev->mc.vram_start = rdev->mc.vram_location;
 855         rdev->mc.vram_end = rdev->mc.vram_location + rdev->mc.mc_vram_size - 1;
 856         rdev->mc.gtt_start = rdev->mc.gtt_location;
 857         rdev->mc.gtt_end = rdev->mc.gtt_location + rdev->mc.gtt_size - 1;
 858         /* FIXME: we should enforce default clock in case GPU is not in
 859          * default setup
 860          */
 861         a.full = rfixed_const(100);
 862         rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
 863         rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
 864         return 0;
 865 }
 866 int rv770_gpu_reset(struct radeon_device *rdev)
 867 {
 868         /* FIXME: implement any rv770 specific bits */
 869         return r600_gpu_reset(rdev);
 870 }
 871
 872 static int rv770_startup(struct radeon_device *rdev)
 873 {
 874         int r;
 875
 876         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
 877                 r = r600_init_microcode(rdev);
 878                 if (r) {
 879                         DRM_ERROR("Failed to load firmware!\n");
 880                         return r;
 881                 }
 882         }
 883
 884         rv770_mc_program(rdev);
 885         if (rdev->flags & RADEON_IS_AGP) {
 886                 rv770_agp_enable(rdev);
 887         } else {
 888                 r = rv770_pcie_gart_enable(rdev);
 889                 if (r)
 890                         return r;
 891         }
 892         rv770_gpu_init(rdev);
 893         r = r600_blit_init(rdev);
 894         if (r) {
 895                 r600_blit_fini(rdev);
 896                 rdev->asic->copy = NULL;
 897                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
 898         }
 899         /* pin copy shader into vram */
 900         if (rdev->r600_blit.shader_obj) {
 901                 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
 902                 if (unlikely(r != 0))
 903                         return r;
 904                 r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM,
 905                                 &rdev->r600_blit.shader_gpu_addr);
 906                 radeon_bo_unreserve(rdev->r600_blit.shader_obj);
 907                 if (r) {
 908                         DRM_ERROR("failed to pin blit object %d\n", r);
 909                         return r;
 910                 }
 911         }
 912         /* Enable IRQ */
 913         r = r600_irq_init(rdev);
 914         if (r) {
 915                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
 916                 radeon_irq_kms_fini(rdev);
 917                 return r;
 918         }
 919         r600_irq_set(rdev);
 920
 921         r = radeon_ring_init(rdev, rdev->cp.ring_size);
 922         if (r)
 923                 return r;
 924         r = rv770_cp_load_microcode(rdev);
 925         if (r)
 926                 return r;
 927         r = r600_cp_resume(rdev);
 928         if (r)
 929                 return r;
 930         /* write back buffer are not vital so don't worry about failure */
 931         r600_wb_enable(rdev);
 932         return 0;
 933 }
 934
 935 int rv770_resume(struct radeon_device *rdev)
 936 {
 937         int r;
 938
 939         /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
 940          * posting will perform necessary task to bring back GPU into good
 941          * shape.
 942          */
 943         /* post card */
 944         atom_asic_init(rdev->mode_info.atom_context);
 945         /* Initialize clocks */
 946         r = radeon_clocks_init(rdev);
 947         if (r) {
 948                 return r;
 949         }
 950
 951         r = rv770_startup(rdev);
 952         if (r) {
 953                 DRM_ERROR("r600 startup failed on resume\n");
 954                 return r;
 955         }
 956
 957         r = r600_ib_test(rdev);
 958         if (r) {
 959                 DRM_ERROR("radeon: failled testing IB (%d).\n", r);
 960                 return r;
 961         }
 962         return r;
 963
 964 }
 965
 966 int rv770_suspend(struct radeon_device *rdev)
 967 {
 968         int r;
 969
 970         /* FIXME: we should wait for ring to be empty */
 971         r700_cp_stop(rdev);
 972         rdev->cp.ready = false;
 973         r600_irq_suspend(rdev);
 974         r600_wb_disable(rdev);
 975         rv770_pcie_gart_disable(rdev);
 976         /* unpin shaders bo */
 977         if (rdev->r600_blit.shader_obj) {
 978                 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
 979                 if (likely(r == 0)) {
 980                         radeon_bo_unpin(rdev->r600_blit.shader_obj);
 981                         radeon_bo_unreserve(rdev->r600_blit.shader_obj);
 982                 }
 983         }
 984         return 0;
 985 }
 986
 987 /* Plan is to move initialization in that function and use
 988  * helper function so that radeon_device_init pretty much
 989  * do nothing more than calling asic specific function. This
 990  * should also allow to remove a bunch of callback function
 991  * like vram_info.
 992  */
 993 int rv770_init(struct radeon_device *rdev)
 994 {
 995         int r;
 996
 997         r = radeon_dummy_page_init(rdev);
 998         if (r)
 999                 return r;
1000         /* This don't do much */
1001         r = radeon_gem_init(rdev);
1002         if (r)
1003                 return r;
1004         /* Read BIOS */
1005         if (!radeon_get_bios(rdev)) {
1006                 if (ASIC_IS_AVIVO(rdev))
1007                         return -EINVAL;
1008         }
1009         /* Must be an ATOMBIOS */
1010         if (!rdev->is_atom_bios) {
1011                 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
1012                 return -EINVAL;
1013         }
1014         r = radeon_atombios_init(rdev);
1015         if (r)
1016                 return r;
1017         /* Post card if necessary */
1018         if (!r600_card_posted(rdev)) {
1019                 if (!rdev->bios) {
1020                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1021                         return -EINVAL;
1022                 }
1023                 DRM_INFO("GPU not posted. posting now...\n");
1024                 atom_asic_init(rdev->mode_info.atom_context);
1025         }
1026         /* Initialize scratch registers */
1027         r600_scratch_init(rdev);
1028         /* Initialize surface registers */
1029         radeon_surface_init(rdev);
1030         /* Initialize clocks */
1031         radeon_get_clock_info(rdev->ddev);
1032         r = radeon_clocks_init(rdev);
1033         if (r)
1034                 return r;
1035         /* Initialize power management */
1036         radeon_pm_init(rdev);
1037         /* Fence driver */
1038         r = radeon_fence_driver_init(rdev);
1039         if (r)
1040                 return r;
1041         if (rdev->flags & RADEON_IS_AGP) {
1042                 r = radeon_agp_init(rdev);
1043                 if (r)
1044                         radeon_agp_disable(rdev);
1045         }
1046         r = rv770_mc_init(rdev);
1047         if (r)
1048                 return r;
1049         /* Memory manager */
1050         r = radeon_bo_init(rdev);
1051         if (r)
1052                 return r;
1053
1054         r = radeon_irq_kms_init(rdev);
1055         if (r)
1056                 return r;
1057
1058         rdev->cp.ring_obj = NULL;
1059         r600_ring_init(rdev, 1024 * 1024);
1060
1061         rdev->ih.ring_obj = NULL;
1062         r600_ih_ring_init(rdev, 64 * 1024);
1063
1064         r = r600_pcie_gart_init(rdev);
1065         if (r)
1066                 return r;
1067
1068         rdev->accel_working = true;
1069         r = rv770_startup(rdev);
1070         if (r) {
1071                 dev_err(rdev->dev, "disabling GPU acceleration\n");
1072                 r600_cp_fini(rdev);
1073                 r600_wb_fini(rdev);
1074                 r600_irq_fini(rdev);
1075                 radeon_irq_kms_fini(rdev);
1076                 rv770_pcie_gart_fini(rdev);
1077                 rdev->accel_working = false;
1078         }
1079         if (rdev->accel_working) {
1080                 r = radeon_ib_pool_init(rdev);
1081                 if (r) {
1082                         dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
1083                         rdev->accel_working = false;
1084                 } else {
1085                         r = r600_ib_test(rdev);
1086                         if (r) {
1087                                 dev_err(rdev->dev, "IB test failed (%d).\n", r);
1088                                 rdev->accel_working = false;
1089                         }
1090                 }
1091         }
1092         return 0;
1093 }
1094
1095 void rv770_fini(struct radeon_device *rdev)
1096 {
1097         r600_blit_fini(rdev);
1098         r600_cp_fini(rdev);
1099         r600_wb_fini(rdev);
1100         r600_irq_fini(rdev);
1101         radeon_irq_kms_fini(rdev);
1102         rv770_pcie_gart_fini(rdev);
1103         radeon_gem_fini(rdev);
1104         radeon_fence_driver_fini(rdev);
1105         radeon_clocks_fini(rdev);
1106         radeon_agp_fini(rdev);
1107         radeon_bo_fini(rdev);
1108         radeon_atombios_fini(rdev);
1109         kfree(rdev->bios);
1110         rdev->bios = NULL;
1111         radeon_dummy_page_fini(rdev);
1112 }