drivers/gpu/drm/radeon/rv770.c

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/firmware.h>
  29 #include <linux/platform_device.h>
  30 #include <linux/slab.h>
  31 #include "drmP.h"
  32 #include "radeon.h"
  33 #include "radeon_asic.h"
  34 #include "radeon_drm.h"
  35 #include "rv770d.h"
  36 #include "atom.h"
  37 #include "avivod.h"
  38
  39 #define R700_PFP_UCODE_SIZE 848
  40 #define R700_PM4_UCODE_SIZE 1360
  41
  42 static void rv770_gpu_init(struct radeon_device *rdev);
  43 void rv770_fini(struct radeon_device *rdev);
  44 static void rv770_pcie_gen2_enable(struct radeon_device *rdev);
  45
  46 u32 rv770_page_flip(struct radeon_device *rdev, int crtc_id, u64 crtc_base)
  47 {
  48         struct radeon_crtc *radeon_crtc = rdev->mode_info.crtcs[crtc_id];
  49         u32 tmp = RREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset);
  50
  51         /* Lock the graphics update lock */
  52         tmp |= AVIVO_D1GRPH_UPDATE_LOCK;
  53         WREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset, tmp);
  54
  55         /* update the scanout addresses */
  56         if (radeon_crtc->crtc_id) {
  57                 WREG32(D2GRPH_SECONDARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  58                 WREG32(D2GRPH_PRIMARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  59         } else {
  60                 WREG32(D1GRPH_SECONDARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  61                 WREG32(D1GRPH_PRIMARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  62         }
  63         WREG32(D1GRPH_SECONDARY_SURFACE_ADDRESS + radeon_crtc->crtc_offset,
  64                (u32)crtc_base);
  65         WREG32(D1GRPH_PRIMARY_SURFACE_ADDRESS + radeon_crtc->crtc_offset,
  66                (u32)crtc_base);
  67
  68         /* Wait for update_pending to go high. */
  69         while (!(RREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset) & AVIVO_D1GRPH_SURFACE_UPDATE_PENDING));
  70         DRM_DEBUG("Update pending now high. Unlocking vupdate_lock.\n");
  71
  72         /* Unlock the lock, so double-buffering can take place inside vblank */
  73         tmp &= ~AVIVO_D1GRPH_UPDATE_LOCK;
  74         WREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset, tmp);
  75
  76         /* Return current update_pending status: */
  77         return RREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset) & AVIVO_D1GRPH_SURFACE_UPDATE_PENDING;
  78 }
  79
  80 /* get temperature in millidegrees */
  81 int rv770_get_temp(struct radeon_device *rdev)
  82 {
  83         u32 temp = (RREG32(CG_MULT_THERMAL_STATUS) & ASIC_T_MASK) >>
  84                 ASIC_T_SHIFT;
  85         int actual_temp;
  86
  87         if (temp & 0x400)
  88                 actual_temp = -256;
  89         else if (temp & 0x200)
  90                 actual_temp = 255;
  91         else if (temp & 0x100) {
  92                 actual_temp = temp & 0x1ff;
  93                 actual_temp |= ~0x1ff;
  94         } else
  95                 actual_temp = temp & 0xff;
  96
  97         return (actual_temp * 1000) / 2;
  98 }
  99
 100 void rv770_pm_misc(struct radeon_device *rdev)
 101 {
 102         int req_ps_idx = rdev->pm.requested_power_state_index;
 103         int req_cm_idx = rdev->pm.requested_clock_mode_index;
 104         struct radeon_power_state *ps = &rdev->pm.power_state[req_ps_idx];
 105         struct radeon_voltage *voltage = &ps->clock_info[req_cm_idx].voltage;
 106
 107         if ((voltage->type == VOLTAGE_SW) && voltage->voltage) {
 108                 /* 0xff01 is a flag rather then an actual voltage */
 109                 if (voltage->voltage == 0xff01)
 110                         return;
 111                 if (voltage->voltage != rdev->pm.current_vddc) {
 112                         radeon_atom_set_voltage(rdev, voltage->voltage, SET_VOLTAGE_TYPE_ASIC_VDDC);
 113                         rdev->pm.current_vddc = voltage->voltage;
 114                         DRM_DEBUG("Setting: v: %d\n", voltage->voltage);
 115                 }
 116         }
 117 }
 118
 119 /*
 120  * GART
 121  */
 122 int rv770_pcie_gart_enable(struct radeon_device *rdev)
 123 {
 124         u32 tmp;
 125         int r, i;
 126
 127         if (rdev->gart.table.vram.robj == NULL) {
 128                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
 129                 return -EINVAL;
 130         }
 131         r = radeon_gart_table_vram_pin(rdev);
 132         if (r)
 133                 return r;
 134         radeon_gart_restore(rdev);
 135         /* Setup L2 cache */
 136         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 137                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 138                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 139         WREG32(VM_L2_CNTL2, 0);
 140         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 141         /* Setup TLB control */
 142         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 143                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 144                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 145                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 146         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 147         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 148         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 149         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 150         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 151         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 152         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 153         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
 154         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
 155         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
 156         WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
 157                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
 158         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
 159                         (u32)(rdev->dummy_page.addr >> 12));
 160         for (i = 1; i < 7; i++)
 161                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 162
 163         r600_pcie_gart_tlb_flush(rdev);
 164         rdev->gart.ready = true;
 165         return 0;
 166 }
 167
 168 void rv770_pcie_gart_disable(struct radeon_device *rdev)
 169 {
 170         u32 tmp;
 171         int i, r;
 172
 173         /* Disable all tables */
 174         for (i = 0; i < 7; i++)
 175                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 176
 177         /* Setup L2 cache */
 178         WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING |
 179                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 180         WREG32(VM_L2_CNTL2, 0);
 181         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 182         /* Setup TLB control */
 183         tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 184         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 185         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 186         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 187         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 188         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 189         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 190         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 191         if (rdev->gart.table.vram.robj) {
 192                 r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
 193                 if (likely(r == 0)) {
 194                         radeon_bo_kunmap(rdev->gart.table.vram.robj);
 195                         radeon_bo_unpin(rdev->gart.table.vram.robj);
 196                         radeon_bo_unreserve(rdev->gart.table.vram.robj);
 197                 }
 198         }
 199 }
 200
 201 void rv770_pcie_gart_fini(struct radeon_device *rdev)
 202 {
 203         radeon_gart_fini(rdev);
 204         rv770_pcie_gart_disable(rdev);
 205         radeon_gart_table_vram_free(rdev);
 206 }
 207
 208
 209 void rv770_agp_enable(struct radeon_device *rdev)
 210 {
 211         u32 tmp;
 212         int i;
 213
 214         /* Setup L2 cache */
 215         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 216                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 217                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 218         WREG32(VM_L2_CNTL2, 0);
 219         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 220         /* Setup TLB control */
 221         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 222                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 223                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 224                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 225         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 226         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 227         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 228         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 229         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 230         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 231         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 232         for (i = 0; i < 7; i++)
 233                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 234 }
 235
 236 static void rv770_mc_program(struct radeon_device *rdev)
 237 {
 238         struct rv515_mc_save save;
 239         u32 tmp;
 240         int i, j;
 241
 242         /* Initialize HDP */
 243         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
 244                 WREG32((0x2c14 + j), 0x00000000);
 245                 WREG32((0x2c18 + j), 0x00000000);
 246                 WREG32((0x2c1c + j), 0x00000000);
 247                 WREG32((0x2c20 + j), 0x00000000);
 248                 WREG32((0x2c24 + j), 0x00000000);
 249         }
 250         /* r7xx hw bug.  Read from HDP_DEBUG1 rather
 251          * than writing to HDP_REG_COHERENCY_FLUSH_CNTL
 252          */
 253         tmp = RREG32(HDP_DEBUG1);
 254
 255         rv515_mc_stop(rdev, &save);
 256         if (r600_mc_wait_for_idle(rdev)) {
 257                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 258         }
 259         /* Lockout access through VGA aperture*/
 260         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
 261         /* Update configuration */
 262         if (rdev->flags & RADEON_IS_AGP) {
 263                 if (rdev->mc.vram_start < rdev->mc.gtt_start) {
 264                         /* VRAM before AGP */
 265                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 266                                 rdev->mc.vram_start >> 12);
 267                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 268                                 rdev->mc.gtt_end >> 12);
 269                 } else {
 270                         /* VRAM after AGP */
 271                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 272                                 rdev->mc.gtt_start >> 12);
 273                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 274                                 rdev->mc.vram_end >> 12);
 275                 }
 276         } else {
 277                 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 278                         rdev->mc.vram_start >> 12);
 279                 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 280                         rdev->mc.vram_end >> 12);
 281         }
 282         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 283         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
 284         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
 285         WREG32(MC_VM_FB_LOCATION, tmp);
 286         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
 287         WREG32(HDP_NONSURFACE_INFO, (2 << 7));
 288         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
 289         if (rdev->flags & RADEON_IS_AGP) {
 290                 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16);
 291                 WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16);
 292                 WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22);
 293         } else {
 294                 WREG32(MC_VM_AGP_BASE, 0);
 295                 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
 296                 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
 297         }
 298         if (r600_mc_wait_for_idle(rdev)) {
 299                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 300         }
 301         rv515_mc_resume(rdev, &save);
 302         /* we need to own VRAM, so turn off the VGA renderer here
 303          * to stop it overwriting our objects */
 304         rv515_vga_render_disable(rdev);
 305 }
 306
 307
 308 /*
 309  * CP.
 310  */
 311 void r700_cp_stop(struct radeon_device *rdev)
 312 {
 313         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 314         WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
 315         WREG32(SCRATCH_UMSK, 0);
 316 }
 317
 318 static int rv770_cp_load_microcode(struct radeon_device *rdev)
 319 {
 320         const __be32 *fw_data;
 321         int i;
 322
 323         if (!rdev->me_fw || !rdev->pfp_fw)
 324                 return -EINVAL;
 325
 326         r700_cp_stop(rdev);
 327         WREG32(CP_RB_CNTL,
 328 #ifdef __BIG_ENDIAN
 329                BUF_SWAP_32BIT |
 330 #endif
 331                RB_NO_UPDATE | RB_BLKSZ(15) | RB_BUFSZ(3));
 332
 333         /* Reset cp */
 334         WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
 335         RREG32(GRBM_SOFT_RESET);
 336         mdelay(15);
 337         WREG32(GRBM_SOFT_RESET, 0);
 338
 339         fw_data = (const __be32 *)rdev->pfp_fw->data;
 340         WREG32(CP_PFP_UCODE_ADDR, 0);
 341         for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 342                 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
 343         WREG32(CP_PFP_UCODE_ADDR, 0);
 344
 345         fw_data = (const __be32 *)rdev->me_fw->data;
 346         WREG32(CP_ME_RAM_WADDR, 0);
 347         for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 348                 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
 349
 350         WREG32(CP_PFP_UCODE_ADDR, 0);
 351         WREG32(CP_ME_RAM_WADDR, 0);
 352         WREG32(CP_ME_RAM_RADDR, 0);
 353         return 0;
 354 }
 355
 356 void r700_cp_fini(struct radeon_device *rdev)
 357 {
 358         r700_cp_stop(rdev);
 359         radeon_ring_fini(rdev);
 360 }
 361
 362 /*
 363  * Core functions
 364  */
 365 static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev,
 366                                              u32 num_tile_pipes,
 367                                              u32 num_backends,
 368                                              u32 backend_disable_mask)
 369 {
 370         u32 backend_map = 0;
 371         u32 enabled_backends_mask;
 372         u32 enabled_backends_count;
 373         u32 cur_pipe;
 374         u32 swizzle_pipe[R7XX_MAX_PIPES];
 375         u32 cur_backend;
 376         u32 i;
 377         bool force_no_swizzle;
 378
 379         if (num_tile_pipes > R7XX_MAX_PIPES)
 380                 num_tile_pipes = R7XX_MAX_PIPES;
 381         if (num_tile_pipes < 1)
 382                 num_tile_pipes = 1;
 383         if (num_backends > R7XX_MAX_BACKENDS)
 384                 num_backends = R7XX_MAX_BACKENDS;
 385         if (num_backends < 1)
 386                 num_backends = 1;
 387
 388         enabled_backends_mask = 0;
 389         enabled_backends_count = 0;
 390         for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
 391                 if (((backend_disable_mask >> i) & 1) == 0) {
 392                         enabled_backends_mask |= (1 << i);
 393                         ++enabled_backends_count;
 394                 }
 395                 if (enabled_backends_count == num_backends)
 396                         break;
 397         }
 398
 399         if (enabled_backends_count == 0) {
 400                 enabled_backends_mask = 1;
 401                 enabled_backends_count = 1;
 402         }
 403
 404         if (enabled_backends_count != num_backends)
 405                 num_backends = enabled_backends_count;
 406
 407         switch (rdev->family) {
 408         case CHIP_RV770:
 409         case CHIP_RV730:
 410                 force_no_swizzle = false;
 411                 break;
 412         case CHIP_RV710:
 413         case CHIP_RV740:
 414         default:
 415                 force_no_swizzle = true;
 416                 break;
 417         }
 418
 419         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
 420         switch (num_tile_pipes) {
 421         case 1:
 422                 swizzle_pipe[0] = 0;
 423                 break;
 424         case 2:
 425                 swizzle_pipe[0] = 0;
 426                 swizzle_pipe[1] = 1;
 427                 break;
 428         case 3:
 429                 if (force_no_swizzle) {
 430                         swizzle_pipe[0] = 0;
 431                         swizzle_pipe[1] = 1;
 432                         swizzle_pipe[2] = 2;
 433                 } else {
 434                         swizzle_pipe[0] = 0;
 435                         swizzle_pipe[1] = 2;
 436                         swizzle_pipe[2] = 1;
 437                 }
 438                 break;
 439         case 4:
 440                 if (force_no_swizzle) {
 441                         swizzle_pipe[0] = 0;
 442                         swizzle_pipe[1] = 1;
 443                         swizzle_pipe[2] = 2;
 444                         swizzle_pipe[3] = 3;
 445                 } else {
 446                         swizzle_pipe[0] = 0;
 447                         swizzle_pipe[1] = 2;
 448                         swizzle_pipe[2] = 3;
 449                         swizzle_pipe[3] = 1;
 450                 }
 451                 break;
 452         case 5:
 453                 if (force_no_swizzle) {
 454                         swizzle_pipe[0] = 0;
 455                         swizzle_pipe[1] = 1;
 456                         swizzle_pipe[2] = 2;
 457                         swizzle_pipe[3] = 3;
 458                         swizzle_pipe[4] = 4;
 459                 } else {
 460                         swizzle_pipe[0] = 0;
 461                         swizzle_pipe[1] = 2;
 462                         swizzle_pipe[2] = 4;
 463                         swizzle_pipe[3] = 1;
 464                         swizzle_pipe[4] = 3;
 465                 }
 466                 break;
 467         case 6:
 468                 if (force_no_swizzle) {
 469                         swizzle_pipe[0] = 0;
 470                         swizzle_pipe[1] = 1;
 471                         swizzle_pipe[2] = 2;
 472                         swizzle_pipe[3] = 3;
 473                         swizzle_pipe[4] = 4;
 474                         swizzle_pipe[5] = 5;
 475                 } else {
 476                         swizzle_pipe[0] = 0;
 477                         swizzle_pipe[1] = 2;
 478                         swizzle_pipe[2] = 4;
 479                         swizzle_pipe[3] = 5;
 480                         swizzle_pipe[4] = 3;
 481                         swizzle_pipe[5] = 1;
 482                 }
 483                 break;
 484         case 7:
 485                 if (force_no_swizzle) {
 486                         swizzle_pipe[0] = 0;
 487                         swizzle_pipe[1] = 1;
 488                         swizzle_pipe[2] = 2;
 489                         swizzle_pipe[3] = 3;
 490                         swizzle_pipe[4] = 4;
 491                         swizzle_pipe[5] = 5;
 492                         swizzle_pipe[6] = 6;
 493                 } else {
 494                         swizzle_pipe[0] = 0;
 495                         swizzle_pipe[1] = 2;
 496                         swizzle_pipe[2] = 4;
 497                         swizzle_pipe[3] = 6;
 498                         swizzle_pipe[4] = 3;
 499                         swizzle_pipe[5] = 1;
 500                         swizzle_pipe[6] = 5;
 501                 }
 502                 break;
 503         case 8:
 504                 if (force_no_swizzle) {
 505                         swizzle_pipe[0] = 0;
 506                         swizzle_pipe[1] = 1;
 507                         swizzle_pipe[2] = 2;
 508                         swizzle_pipe[3] = 3;
 509                         swizzle_pipe[4] = 4;
 510                         swizzle_pipe[5] = 5;
 511                         swizzle_pipe[6] = 6;
 512                         swizzle_pipe[7] = 7;
 513                 } else {
 514                         swizzle_pipe[0] = 0;
 515                         swizzle_pipe[1] = 2;
 516                         swizzle_pipe[2] = 4;
 517                         swizzle_pipe[3] = 6;
 518                         swizzle_pipe[4] = 3;
 519                         swizzle_pipe[5] = 1;
 520                         swizzle_pipe[6] = 7;
 521                         swizzle_pipe[7] = 5;
 522                 }
 523                 break;
 524         }
 525
 526         cur_backend = 0;
 527         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
 528                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
 529                         cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
 530
 531                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
 532
 533                 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
 534         }
 535
 536         return backend_map;
 537 }
 538
 539 static void rv770_program_channel_remap(struct radeon_device *rdev)
 540 {
 541         u32 tcp_chan_steer, mc_shared_chremap, tmp;
 542         bool force_no_swizzle;
 543
 544         switch (rdev->family) {
 545         case CHIP_RV770:
 546         case CHIP_RV730:
 547                 force_no_swizzle = false;
 548                 break;
 549         case CHIP_RV710:
 550         case CHIP_RV740:
 551         default:
 552                 force_no_swizzle = true;
 553                 break;
 554         }
 555
 556         tmp = RREG32(MC_SHARED_CHMAP);
 557         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
 558         case 0:
 559         case 1:
 560         default:
 561                 /* default mapping */
 562                 mc_shared_chremap = 0x00fac688;
 563                 break;
 564         case 2:
 565         case 3:
 566                 if (force_no_swizzle)
 567                         mc_shared_chremap = 0x00fac688;
 568                 else
 569                         mc_shared_chremap = 0x00bbc298;
 570                 break;
 571         }
 572
 573         if (rdev->family == CHIP_RV740)
 574                 tcp_chan_steer = 0x00ef2a60;
 575         else
 576                 tcp_chan_steer = 0x00fac688;
 577
 578         WREG32(TCP_CHAN_STEER, tcp_chan_steer);
 579         WREG32(MC_SHARED_CHREMAP, mc_shared_chremap);
 580 }
 581
 582 static void rv770_gpu_init(struct radeon_device *rdev)
 583 {
 584         int i, j, num_qd_pipes;
 585         u32 ta_aux_cntl;
 586         u32 sx_debug_1;
 587         u32 smx_dc_ctl0;
 588         u32 db_debug3;
 589         u32 num_gs_verts_per_thread;
 590         u32 vgt_gs_per_es;
 591         u32 gs_prim_buffer_depth = 0;
 592         u32 sq_ms_fifo_sizes;
 593         u32 sq_config;
 594         u32 sq_thread_resource_mgmt;
 595         u32 hdp_host_path_cntl;
 596         u32 sq_dyn_gpr_size_simd_ab_0;
 597         u32 backend_map;
 598         u32 gb_tiling_config = 0;
 599         u32 cc_rb_backend_disable = 0;
 600         u32 cc_gc_shader_pipe_config = 0;
 601         u32 mc_arb_ramcfg;
 602         u32 db_debug4;
 603
 604         /* setup chip specs */
 605         switch (rdev->family) {
 606         case CHIP_RV770:
 607                 rdev->config.rv770.max_pipes = 4;
 608                 rdev->config.rv770.max_tile_pipes = 8;
 609                 rdev->config.rv770.max_simds = 10;
 610                 rdev->config.rv770.max_backends = 4;
 611                 rdev->config.rv770.max_gprs = 256;
 612                 rdev->config.rv770.max_threads = 248;
 613                 rdev->config.rv770.max_stack_entries = 512;
 614                 rdev->config.rv770.max_hw_contexts = 8;
 615                 rdev->config.rv770.max_gs_threads = 16 * 2;
 616                 rdev->config.rv770.sx_max_export_size = 128;
 617                 rdev->config.rv770.sx_max_export_pos_size = 16;
 618                 rdev->config.rv770.sx_max_export_smx_size = 112;
 619                 rdev->config.rv770.sq_num_cf_insts = 2;
 620
 621                 rdev->config.rv770.sx_num_of_sets = 7;
 622                 rdev->config.rv770.sc_prim_fifo_size = 0xF9;
 623                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 624                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 625                 break;
 626         case CHIP_RV730:
 627                 rdev->config.rv770.max_pipes = 2;
 628                 rdev->config.rv770.max_tile_pipes = 4;
 629                 rdev->config.rv770.max_simds = 8;
 630                 rdev->config.rv770.max_backends = 2;
 631                 rdev->config.rv770.max_gprs = 128;
 632                 rdev->config.rv770.max_threads = 248;
 633                 rdev->config.rv770.max_stack_entries = 256;
 634                 rdev->config.rv770.max_hw_contexts = 8;
 635                 rdev->config.rv770.max_gs_threads = 16 * 2;
 636                 rdev->config.rv770.sx_max_export_size = 256;
 637                 rdev->config.rv770.sx_max_export_pos_size = 32;
 638                 rdev->config.rv770.sx_max_export_smx_size = 224;
 639                 rdev->config.rv770.sq_num_cf_insts = 2;
 640
 641                 rdev->config.rv770.sx_num_of_sets = 7;
 642                 rdev->config.rv770.sc_prim_fifo_size = 0xf9;
 643                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 644                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 645                 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
 646                         rdev->config.rv770.sx_max_export_pos_size -= 16;
 647                         rdev->config.rv770.sx_max_export_smx_size += 16;
 648                 }
 649                 break;
 650         case CHIP_RV710:
 651                 rdev->config.rv770.max_pipes = 2;
 652                 rdev->config.rv770.max_tile_pipes = 2;
 653                 rdev->config.rv770.max_simds = 2;
 654                 rdev->config.rv770.max_backends = 1;
 655                 rdev->config.rv770.max_gprs = 256;
 656                 rdev->config.rv770.max_threads = 192;
 657                 rdev->config.rv770.max_stack_entries = 256;
 658                 rdev->config.rv770.max_hw_contexts = 4;
 659                 rdev->config.rv770.max_gs_threads = 8 * 2;
 660                 rdev->config.rv770.sx_max_export_size = 128;
 661                 rdev->config.rv770.sx_max_export_pos_size = 16;
 662                 rdev->config.rv770.sx_max_export_smx_size = 112;
 663                 rdev->config.rv770.sq_num_cf_insts = 1;
 664
 665                 rdev->config.rv770.sx_num_of_sets = 7;
 666                 rdev->config.rv770.sc_prim_fifo_size = 0x40;
 667                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 668                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 669                 break;
 670         case CHIP_RV740:
 671                 rdev->config.rv770.max_pipes = 4;
 672                 rdev->config.rv770.max_tile_pipes = 4;
 673                 rdev->config.rv770.max_simds = 8;
 674                 rdev->config.rv770.max_backends = 4;
 675                 rdev->config.rv770.max_gprs = 256;
 676                 rdev->config.rv770.max_threads = 248;
 677                 rdev->config.rv770.max_stack_entries = 512;
 678                 rdev->config.rv770.max_hw_contexts = 8;
 679                 rdev->config.rv770.max_gs_threads = 16 * 2;
 680                 rdev->config.rv770.sx_max_export_size = 256;
 681                 rdev->config.rv770.sx_max_export_pos_size = 32;
 682                 rdev->config.rv770.sx_max_export_smx_size = 224;
 683                 rdev->config.rv770.sq_num_cf_insts = 2;
 684
 685                 rdev->config.rv770.sx_num_of_sets = 7;
 686                 rdev->config.rv770.sc_prim_fifo_size = 0x100;
 687                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 688                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 689
 690                 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
 691                         rdev->config.rv770.sx_max_export_pos_size -= 16;
 692                         rdev->config.rv770.sx_max_export_smx_size += 16;
 693                 }
 694                 break;
 695         default:
 696                 break;
 697         }
 698
 699         /* Initialize HDP */
 700         j = 0;
 701         for (i = 0; i < 32; i++) {
 702                 WREG32((0x2c14 + j), 0x00000000);
 703                 WREG32((0x2c18 + j), 0x00000000);
 704                 WREG32((0x2c1c + j), 0x00000000);
 705                 WREG32((0x2c20 + j), 0x00000000);
 706                 WREG32((0x2c24 + j), 0x00000000);
 707                 j += 0x18;
 708         }
 709
 710         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
 711
 712         /* setup tiling, simd, pipe config */
 713         mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
 714
 715         switch (rdev->config.rv770.max_tile_pipes) {
 716         case 1:
 717         default:
 718                 gb_tiling_config |= PIPE_TILING(0);
 719                 break;
 720         case 2:
 721                 gb_tiling_config |= PIPE_TILING(1);
 722                 break;
 723         case 4:
 724                 gb_tiling_config |= PIPE_TILING(2);
 725                 break;
 726         case 8:
 727                 gb_tiling_config |= PIPE_TILING(3);
 728                 break;
 729         }
 730         rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes;
 731
 732         if (rdev->family == CHIP_RV770)
 733                 gb_tiling_config |= BANK_TILING(1);
 734         else
 735                 gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
 736         rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3);
 737         gb_tiling_config |= GROUP_SIZE((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT);
 738         if ((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT)
 739                 rdev->config.rv770.tiling_group_size = 512;
 740         else
 741                 rdev->config.rv770.tiling_group_size = 256;
 742         if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) {
 743                 gb_tiling_config |= ROW_TILING(3);
 744                 gb_tiling_config |= SAMPLE_SPLIT(3);
 745         } else {
 746                 gb_tiling_config |=
 747                         ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 748                 gb_tiling_config |=
 749                         SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 750         }
 751
 752         gb_tiling_config |= BANK_SWAPS(1);
 753
 754         cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000;
 755         cc_rb_backend_disable |=
 756                 BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK);
 757
 758         cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
 759         cc_gc_shader_pipe_config |=
 760                 INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK);
 761         cc_gc_shader_pipe_config |=
 762                 INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK);
 763
 764         if (rdev->family == CHIP_RV740)
 765                 backend_map = 0x28;
 766         else
 767                 backend_map = r700_get_tile_pipe_to_backend_map(rdev,
 768                                                                 rdev->config.rv770.max_tile_pipes,
 769                                                                 (R7XX_MAX_BACKENDS -
 770                                                                  r600_count_pipe_bits((cc_rb_backend_disable &
 771                                                                                        R7XX_MAX_BACKENDS_MASK) >> 16)),
 772                                                                 (cc_rb_backend_disable >> 16));
 773
 774         rdev->config.rv770.tile_config = gb_tiling_config;
 775         gb_tiling_config |= BACKEND_MAP(backend_map);
 776
 777         WREG32(GB_TILING_CONFIG, gb_tiling_config);
 778         WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 779         WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 780
 781         rv770_program_channel_remap(rdev);
 782
 783         WREG32(CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
 784         WREG32(CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
 785         WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
 786         WREG32(CC_SYS_RB_BACKEND_DISABLE,  cc_rb_backend_disable);
 787
 788         WREG32(CGTS_SYS_TCC_DISABLE, 0);
 789         WREG32(CGTS_TCC_DISABLE, 0);
 790         WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
 791         WREG32(CGTS_USER_TCC_DISABLE, 0);
 792
 793         num_qd_pipes =
 794                 R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
 795         WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK);
 796         WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK);
 797
 798         /* set HW defaults for 3D engine */
 799         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
 800                                      ROQ_IB2_START(0x2b)));
 801
 802         WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
 803
 804         ta_aux_cntl = RREG32(TA_CNTL_AUX);
 805         WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO);
 806
 807         sx_debug_1 = RREG32(SX_DEBUG_1);
 808         sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
 809         WREG32(SX_DEBUG_1, sx_debug_1);
 810
 811         smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
 812         smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff);
 813         smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1);
 814         WREG32(SMX_DC_CTL0, smx_dc_ctl0);
 815
 816         if (rdev->family != CHIP_RV740)
 817                 WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) |
 818                                        GS_FLUSH_CTL(4) |
 819                                        ACK_FLUSH_CTL(3) |
 820                                        SYNC_FLUSH_CTL));
 821
 822         db_debug3 = RREG32(DB_DEBUG3);
 823         db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f);
 824         switch (rdev->family) {
 825         case CHIP_RV770:
 826         case CHIP_RV740:
 827                 db_debug3 |= DB_CLK_OFF_DELAY(0x1f);
 828                 break;
 829         case CHIP_RV710:
 830         case CHIP_RV730:
 831         default:
 832                 db_debug3 |= DB_CLK_OFF_DELAY(2);
 833                 break;
 834         }
 835         WREG32(DB_DEBUG3, db_debug3);
 836
 837         if (rdev->family != CHIP_RV770) {
 838                 db_debug4 = RREG32(DB_DEBUG4);
 839                 db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER;
 840                 WREG32(DB_DEBUG4, db_debug4);
 841         }
 842
 843         WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) |
 844                                         POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) |
 845                                         SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1)));
 846
 847         WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) |
 848                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) |
 849                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize)));
 850
 851         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 852
 853         WREG32(VGT_NUM_INSTANCES, 1);
 854
 855         WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0));
 856
 857         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
 858
 859         WREG32(CP_PERFMON_CNTL, 0);
 860
 861         sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) |
 862                             DONE_FIFO_HIWATER(0xe0) |
 863                             ALU_UPDATE_FIFO_HIWATER(0x8));
 864         switch (rdev->family) {
 865         case CHIP_RV770:
 866         case CHIP_RV730:
 867         case CHIP_RV710:
 868                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1);
 869                 break;
 870         case CHIP_RV740:
 871         default:
 872                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4);
 873                 break;
 874         }
 875         WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 876
 877         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 878          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 879          */
 880         sq_config = RREG32(SQ_CONFIG);
 881         sq_config &= ~(PS_PRIO(3) |
 882                        VS_PRIO(3) |
 883                        GS_PRIO(3) |
 884                        ES_PRIO(3));
 885         sq_config |= (DX9_CONSTS |
 886                       VC_ENABLE |
 887                       EXPORT_SRC_C |
 888                       PS_PRIO(0) |
 889                       VS_PRIO(1) |
 890                       GS_PRIO(2) |
 891                       ES_PRIO(3));
 892         if (rdev->family == CHIP_RV710)
 893                 /* no vertex cache */
 894                 sq_config &= ~VC_ENABLE;
 895
 896         WREG32(SQ_CONFIG, sq_config);
 897
 898         WREG32(SQ_GPR_RESOURCE_MGMT_1,  (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 899                                          NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 900                                          NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2)));
 901
 902         WREG32(SQ_GPR_RESOURCE_MGMT_2,  (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) |
 903                                          NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64)));
 904
 905         sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) |
 906                                    NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) |
 907                                    NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8));
 908         if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads)
 909                 sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads);
 910         else
 911                 sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8);
 912         WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
 913
 914         WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 915                                                      NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 916
 917         WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 918                                                      NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 919
 920         sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 921                                      SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) |
 922                                      SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 923                                      SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64));
 924
 925         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
 926         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
 927         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
 928         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
 929         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
 930         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
 931         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
 932         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
 933
 934         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
 935                                           FORCE_EOV_MAX_REZ_CNT(255)));
 936
 937         if (rdev->family == CHIP_RV710)
 938                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) |
 939                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 940         else
 941                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) |
 942                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 943
 944         switch (rdev->family) {
 945         case CHIP_RV770:
 946         case CHIP_RV730:
 947         case CHIP_RV740:
 948                 gs_prim_buffer_depth = 384;
 949                 break;
 950         case CHIP_RV710:
 951                 gs_prim_buffer_depth = 128;
 952                 break;
 953         default:
 954                 break;
 955         }
 956
 957         num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16;
 958         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
 959         /* Max value for this is 256 */
 960         if (vgt_gs_per_es > 256)
 961                 vgt_gs_per_es = 256;
 962
 963         WREG32(VGT_ES_PER_GS, 128);
 964         WREG32(VGT_GS_PER_ES, vgt_gs_per_es);
 965         WREG32(VGT_GS_PER_VS, 2);
 966
 967         /* more default values. 2D/3D driver should adjust as needed */
 968         WREG32(VGT_GS_VERTEX_REUSE, 16);
 969         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
 970         WREG32(VGT_STRMOUT_EN, 0);
 971         WREG32(SX_MISC, 0);
 972         WREG32(PA_SC_MODE_CNTL, 0);
 973         WREG32(PA_SC_EDGERULE, 0xaaaaaaaa);
 974         WREG32(PA_SC_AA_CONFIG, 0);
 975         WREG32(PA_SC_CLIPRECT_RULE, 0xffff);
 976         WREG32(PA_SC_LINE_STIPPLE, 0);
 977         WREG32(SPI_INPUT_Z, 0);
 978         WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
 979         WREG32(CB_COLOR7_FRAG, 0);
 980
 981         /* clear render buffer base addresses */
 982         WREG32(CB_COLOR0_BASE, 0);
 983         WREG32(CB_COLOR1_BASE, 0);
 984         WREG32(CB_COLOR2_BASE, 0);
 985         WREG32(CB_COLOR3_BASE, 0);
 986         WREG32(CB_COLOR4_BASE, 0);
 987         WREG32(CB_COLOR5_BASE, 0);
 988         WREG32(CB_COLOR6_BASE, 0);
 989         WREG32(CB_COLOR7_BASE, 0);
 990
 991         WREG32(TCP_CNTL, 0);
 992
 993         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
 994         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
 995
 996         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 997
 998         WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
 999                                           NUM_CLIP_SEQ(3)));
1000
1001 }
1002
1003 static int rv770_vram_scratch_init(struct radeon_device *rdev)
1004 {
1005         int r;
1006         u64 gpu_addr;
1007
1008         if (rdev->vram_scratch.robj == NULL) {
1009                 r = radeon_bo_create(rdev, RADEON_GPU_PAGE_SIZE,
1010                                      PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
1011                                      &rdev->vram_scratch.robj);
1012                 if (r) {
1013                         return r;
1014                 }
1015         }
1016
1017         r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
1018         if (unlikely(r != 0))
1019                 return r;
1020         r = radeon_bo_pin(rdev->vram_scratch.robj,
1021                           RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
1022         if (r) {
1023                 radeon_bo_unreserve(rdev->vram_scratch.robj);
1024                 return r;
1025         }
1026         r = radeon_bo_kmap(rdev->vram_scratch.robj,
1027                                 (void **)&rdev->vram_scratch.ptr);
1028         if (r)
1029                 radeon_bo_unpin(rdev->vram_scratch.robj);
1030         radeon_bo_unreserve(rdev->vram_scratch.robj);
1031
1032         return r;
1033 }
1034
1035 static void rv770_vram_scratch_fini(struct radeon_device *rdev)
1036 {
1037         int r;
1038
1039         if (rdev->vram_scratch.robj == NULL) {
1040                 return;
1041         }
1042         r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
1043         if (likely(r == 0)) {
1044                 radeon_bo_kunmap(rdev->vram_scratch.robj);
1045                 radeon_bo_unpin(rdev->vram_scratch.robj);
1046                 radeon_bo_unreserve(rdev->vram_scratch.robj);
1047         }
1048         radeon_bo_unref(&rdev->vram_scratch.robj);
1049 }
1050
1051 void r700_vram_gtt_location(struct radeon_device *rdev, struct radeon_mc *mc)
1052 {
1053         u64 size_bf, size_af;
1054
1055         if (mc->mc_vram_size > 0xE0000000) {
1056                 /* leave room for at least 512M GTT */
1057                 dev_warn(rdev->dev, "limiting VRAM\n");
1058                 mc->real_vram_size = 0xE0000000;
1059                 mc->mc_vram_size = 0xE0000000;
1060         }
1061         if (rdev->flags & RADEON_IS_AGP) {
1062                 size_bf = mc->gtt_start;
1063                 size_af = 0xFFFFFFFF - mc->gtt_end + 1;
1064                 if (size_bf > size_af) {
1065                         if (mc->mc_vram_size > size_bf) {
1066                                 dev_warn(rdev->dev, "limiting VRAM\n");
1067                                 mc->real_vram_size = size_bf;
1068                                 mc->mc_vram_size = size_bf;
1069                         }
1070                         mc->vram_start = mc->gtt_start - mc->mc_vram_size;
1071                 } else {
1072                         if (mc->mc_vram_size > size_af) {
1073                                 dev_warn(rdev->dev, "limiting VRAM\n");
1074                                 mc->real_vram_size = size_af;
1075                                 mc->mc_vram_size = size_af;
1076                         }
1077                         mc->vram_start = mc->gtt_end;
1078                 }
1079                 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
1080                 dev_info(rdev->dev, "VRAM: %lluM 0x%08llX - 0x%08llX (%lluM used)\n",
1081                                 mc->mc_vram_size >> 20, mc->vram_start,
1082                                 mc->vram_end, mc->real_vram_size >> 20);
1083         } else {
1084                 radeon_vram_location(rdev, &rdev->mc, 0);
1085                 rdev->mc.gtt_base_align = 0;
1086                 radeon_gtt_location(rdev, mc);
1087         }
1088 }
1089
1090 int rv770_mc_init(struct radeon_device *rdev)
1091 {
1092         u32 tmp;
1093         int chansize, numchan;
1094
1095         /* Get VRAM informations */
1096         rdev->mc.vram_is_ddr = true;
1097         tmp = RREG32(MC_ARB_RAMCFG);
1098         if (tmp & CHANSIZE_OVERRIDE) {
1099                 chansize = 16;
1100         } else if (tmp & CHANSIZE_MASK) {
1101                 chansize = 64;
1102         } else {
1103                 chansize = 32;
1104         }
1105         tmp = RREG32(MC_SHARED_CHMAP);
1106         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
1107         case 0:
1108         default:
1109                 numchan = 1;
1110                 break;
1111         case 1:
1112                 numchan = 2;
1113                 break;
1114         case 2:
1115                 numchan = 4;
1116                 break;
1117         case 3:
1118                 numchan = 8;
1119                 break;
1120         }
1121         rdev->mc.vram_width = numchan * chansize;
1122         /* Could aper size report 0 ? */
1123         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
1124         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
1125         /* Setup GPU memory space */
1126         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
1127         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
1128         rdev->mc.visible_vram_size = rdev->mc.aper_size;
1129         r700_vram_gtt_location(rdev, &rdev->mc);
1130         radeon_update_bandwidth_info(rdev);
1131
1132         return 0;
1133 }
1134
1135 static int rv770_startup(struct radeon_device *rdev)
1136 {
1137         int r;
1138
1139         /* enable pcie gen2 link */
1140         rv770_pcie_gen2_enable(rdev);
1141
1142         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
1143                 r = r600_init_microcode(rdev);
1144                 if (r) {
1145                         DRM_ERROR("Failed to load firmware!\n");
1146                         return r;
1147                 }
1148         }
1149
1150         rv770_mc_program(rdev);
1151         if (rdev->flags & RADEON_IS_AGP) {
1152                 rv770_agp_enable(rdev);
1153         } else {
1154                 r = rv770_pcie_gart_enable(rdev);
1155                 if (r)
1156                         return r;
1157         }
1158         r = rv770_vram_scratch_init(rdev);
1159         if (r)
1160                 return r;
1161         rv770_gpu_init(rdev);
1162         r = r600_blit_init(rdev);
1163         if (r) {
1164                 r600_blit_fini(rdev);
1165                 rdev->asic->copy = NULL;
1166                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
1167         }
1168
1169         /* allocate wb buffer */
1170         r = radeon_wb_init(rdev);
1171         if (r)
1172                 return r;
1173
1174         /* Enable IRQ */
1175         r = r600_irq_init(rdev);
1176         if (r) {
1177                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
1178                 radeon_irq_kms_fini(rdev);
1179                 return r;
1180         }
1181         r600_irq_set(rdev);
1182
1183         r = radeon_ring_init(rdev, rdev->cp.ring_size);
1184         if (r)
1185                 return r;
1186         r = rv770_cp_load_microcode(rdev);
1187         if (r)
1188                 return r;
1189         r = r600_cp_resume(rdev);
1190         if (r)
1191                 return r;
1192
1193         return 0;
1194 }
1195
1196 int rv770_resume(struct radeon_device *rdev)
1197 {
1198         int r;
1199
1200         /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
1201          * posting will perform necessary task to bring back GPU into good
1202          * shape.
1203          */
1204         /* post card */
1205         atom_asic_init(rdev->mode_info.atom_context);
1206
1207         r = rv770_startup(rdev);
1208         if (r) {
1209                 DRM_ERROR("r600 startup failed on resume\n");
1210                 return r;
1211         }
1212
1213         r = r600_ib_test(rdev);
1214         if (r) {
1215                 DRM_ERROR("radeon: failed testing IB (%d).\n", r);
1216                 return r;
1217         }
1218
1219         r = r600_audio_init(rdev);
1220         if (r) {
1221                 dev_err(rdev->dev, "radeon: audio init failed\n");
1222                 return r;
1223         }
1224
1225         return r;
1226
1227 }
1228
1229 int rv770_suspend(struct radeon_device *rdev)
1230 {
1231         int r;
1232
1233         r600_audio_fini(rdev);
1234         /* FIXME: we should wait for ring to be empty */
1235         r700_cp_stop(rdev);
1236         rdev->cp.ready = false;
1237         r600_irq_suspend(rdev);
1238         radeon_wb_disable(rdev);
1239         rv770_pcie_gart_disable(rdev);
1240         /* unpin shaders bo */
1241         if (rdev->r600_blit.shader_obj) {
1242                 r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
1243                 if (likely(r == 0)) {
1244                         radeon_bo_unpin(rdev->r600_blit.shader_obj);
1245                         radeon_bo_unreserve(rdev->r600_blit.shader_obj);
1246                 }
1247         }
1248         return 0;
1249 }
1250
1251 /* Plan is to move initialization in that function and use
1252  * helper function so that radeon_device_init pretty much
1253  * do nothing more than calling asic specific function. This
1254  * should also allow to remove a bunch of callback function
1255  * like vram_info.
1256  */
1257 int rv770_init(struct radeon_device *rdev)
1258 {
1259         int r;
1260
1261         /* This don't do much */
1262         r = radeon_gem_init(rdev);
1263         if (r)
1264                 return r;
1265         /* Read BIOS */
1266         if (!radeon_get_bios(rdev)) {
1267                 if (ASIC_IS_AVIVO(rdev))
1268                         return -EINVAL;
1269         }
1270         /* Must be an ATOMBIOS */
1271         if (!rdev->is_atom_bios) {
1272                 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
1273                 return -EINVAL;
1274         }
1275         r = radeon_atombios_init(rdev);
1276         if (r)
1277                 return r;
1278         /* Post card if necessary */
1279         if (!radeon_card_posted(rdev)) {
1280                 if (!rdev->bios) {
1281                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1282                         return -EINVAL;
1283                 }
1284                 DRM_INFO("GPU not posted. posting now...\n");
1285                 atom_asic_init(rdev->mode_info.atom_context);
1286         }
1287         /* Initialize scratch registers */
1288         r600_scratch_init(rdev);
1289         /* Initialize surface registers */
1290         radeon_surface_init(rdev);
1291         /* Initialize clocks */
1292         radeon_get_clock_info(rdev->ddev);
1293         /* Fence driver */
1294         r = radeon_fence_driver_init(rdev);
1295         if (r)
1296                 return r;
1297         /* initialize AGP */
1298         if (rdev->flags & RADEON_IS_AGP) {
1299                 r = radeon_agp_init(rdev);
1300                 if (r)
1301                         radeon_agp_disable(rdev);
1302         }
1303         r = rv770_mc_init(rdev);
1304         if (r)
1305                 return r;
1306         /* Memory manager */
1307         r = radeon_bo_init(rdev);
1308         if (r)
1309                 return r;
1310
1311         r = radeon_irq_kms_init(rdev);
1312         if (r)
1313                 return r;
1314
1315         rdev->cp.ring_obj = NULL;
1316         r600_ring_init(rdev, 1024 * 1024);
1317
1318         rdev->ih.ring_obj = NULL;
1319         r600_ih_ring_init(rdev, 64 * 1024);
1320
1321         r = r600_pcie_gart_init(rdev);
1322         if (r)
1323                 return r;
1324
1325         rdev->accel_working = true;
1326         r = rv770_startup(rdev);
1327         if (r) {
1328                 dev_err(rdev->dev, "disabling GPU acceleration\n");
1329                 r700_cp_fini(rdev);
1330                 r600_irq_fini(rdev);
1331                 radeon_wb_fini(rdev);
1332                 radeon_irq_kms_fini(rdev);
1333                 rv770_pcie_gart_fini(rdev);
1334                 rdev->accel_working = false;
1335         }
1336         if (rdev->accel_working) {
1337                 r = radeon_ib_pool_init(rdev);
1338                 if (r) {
1339                         dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
1340                         rdev->accel_working = false;
1341                 } else {
1342                         r = r600_ib_test(rdev);
1343                         if (r) {
1344                                 dev_err(rdev->dev, "IB test failed (%d).\n", r);
1345                                 rdev->accel_working = false;
1346                         }
1347                 }
1348         }
1349
1350         r = r600_audio_init(rdev);
1351         if (r) {
1352                 dev_err(rdev->dev, "radeon: audio init failed\n");
1353                 return r;
1354         }
1355
1356         return 0;
1357 }
1358
1359 void rv770_fini(struct radeon_device *rdev)
1360 {
1361         r600_blit_fini(rdev);
1362         r700_cp_fini(rdev);
1363         r600_irq_fini(rdev);
1364         radeon_wb_fini(rdev);
1365         radeon_irq_kms_fini(rdev);
1366         rv770_pcie_gart_fini(rdev);
1367         rv770_vram_scratch_fini(rdev);
1368         radeon_gem_fini(rdev);
1369         radeon_fence_driver_fini(rdev);
1370         radeon_agp_fini(rdev);
1371         radeon_bo_fini(rdev);
1372         radeon_atombios_fini(rdev);
1373         kfree(rdev->bios);
1374         rdev->bios = NULL;
1375 }
1376
1377 static void rv770_pcie_gen2_enable(struct radeon_device *rdev)
1378 {
1379         u32 link_width_cntl, lanes, speed_cntl, tmp;
1380         u16 link_cntl2;
1381
1382         if (radeon_pcie_gen2 == 0)
1383                 return;
1384
1385         if (rdev->flags & RADEON_IS_IGP)
1386                 return;
1387
1388         if (!(rdev->flags & RADEON_IS_PCIE))
1389                 return;
1390
1391         /* x2 cards have a special sequence */
1392         if (ASIC_IS_X2(rdev))
1393                 return;
1394
1395         /* advertise upconfig capability */
1396         link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
1397         link_width_cntl &= ~LC_UPCONFIGURE_DIS;
1398         WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1399         link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
1400         if (link_width_cntl & LC_RENEGOTIATION_SUPPORT) {
1401                 lanes = (link_width_cntl & LC_LINK_WIDTH_RD_MASK) >> LC_LINK_WIDTH_RD_SHIFT;
1402                 link_width_cntl &= ~(LC_LINK_WIDTH_MASK |
1403                                      LC_RECONFIG_ARC_MISSING_ESCAPE);
1404                 link_width_cntl |= lanes | LC_RECONFIG_NOW |
1405                         LC_RENEGOTIATE_EN | LC_UPCONFIGURE_SUPPORT;
1406                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1407         } else {
1408                 link_width_cntl |= LC_UPCONFIGURE_DIS;
1409                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1410         }
1411
1412         speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1413         if ((speed_cntl & LC_OTHER_SIDE_EVER_SENT_GEN2) &&
1414             (speed_cntl & LC_OTHER_SIDE_SUPPORTS_GEN2)) {
1415
1416                 tmp = RREG32(0x541c);
1417                 WREG32(0x541c, tmp | 0x8);
1418                 WREG32(MM_CFGREGS_CNTL, MM_WR_TO_CFG_EN);
1419                 link_cntl2 = RREG16(0x4088);
1420                 link_cntl2 &= ~TARGET_LINK_SPEED_MASK;
1421                 link_cntl2 |= 0x2;
1422                 WREG16(0x4088, link_cntl2);
1423                 WREG32(MM_CFGREGS_CNTL, 0);
1424
1425                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1426                 speed_cntl &= ~LC_TARGET_LINK_SPEED_OVERRIDE_EN;
1427                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1428
1429                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1430                 speed_cntl |= LC_CLR_FAILED_SPD_CHANGE_CNT;
1431                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1432
1433                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1434                 speed_cntl &= ~LC_CLR_FAILED_SPD_CHANGE_CNT;
1435                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1436
1437                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1438                 speed_cntl |= LC_GEN2_EN_STRAP;
1439                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1440
1441         } else {
1442                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
1443                 /* XXX: only disable it if gen1 bridge vendor == 0x111d or 0x1106 */
1444                 if (1)
1445                         link_width_cntl |= LC_UPCONFIGURE_DIS;
1446                 else
1447                         link_width_cntl &= ~LC_UPCONFIGURE_DIS;
1448                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1449         }
1450 }