Rev 1413 | Rev 1963 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 1246 | serge | 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 | */ |
||
| 1413 | serge | 28 | #include |
| 1246 | serge | 29 | //#include |
| 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; |
||
| 1430 | serge | 59 | radeon_gart_restore(rdev); |
| 1246 | serge | 60 | /* Setup L2 cache */ |
| 61 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | |
||
| 62 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | |
||
| 63 | EFFECTIVE_L2_QUEUE_SIZE(7)); |
||
| 64 | WREG32(VM_L2_CNTL2, 0); |
||
| 65 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); |
||
| 66 | /* Setup TLB control */ |
||
| 67 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | |
||
| 68 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | |
||
| 69 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | |
||
| 70 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); |
||
| 71 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); |
||
| 72 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); |
||
| 73 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); |
||
| 74 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); |
||
| 75 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); |
||
| 76 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); |
||
| 77 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); |
||
| 78 | WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12); |
||
| 79 | WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12); |
||
| 80 | WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12); |
||
| 81 | WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) | |
||
| 82 | RANGE_PROTECTION_FAULT_ENABLE_DEFAULT); |
||
| 83 | WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, |
||
| 84 | (u32)(rdev->dummy_page.addr >> 12)); |
||
| 85 | for (i = 1; i < 7; i++) |
||
| 86 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); |
||
| 87 | |||
| 88 | r600_pcie_gart_tlb_flush(rdev); |
||
| 89 | rdev->gart.ready = true; |
||
| 90 | return 0; |
||
| 91 | } |
||
| 92 | |||
| 93 | void rv770_pcie_gart_disable(struct radeon_device *rdev) |
||
| 94 | { |
||
| 95 | u32 tmp; |
||
| 1403 | serge | 96 | int i, r; |
| 1246 | serge | 97 | |
| 98 | /* Disable all tables */ |
||
| 99 | for (i = 0; i < 7; i++) |
||
| 100 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); |
||
| 101 | |||
| 102 | /* Setup L2 cache */ |
||
| 103 | WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING | |
||
| 104 | EFFECTIVE_L2_QUEUE_SIZE(7)); |
||
| 105 | WREG32(VM_L2_CNTL2, 0); |
||
| 106 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); |
||
| 107 | /* Setup TLB control */ |
||
| 108 | tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); |
||
| 109 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); |
||
| 110 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); |
||
| 111 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); |
||
| 112 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); |
||
| 113 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); |
||
| 114 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); |
||
| 115 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); |
||
| 116 | if (rdev->gart.table.vram.robj) { |
||
| 1404 | serge | 117 | r = radeon_bo_reserve(rdev->gart.table.vram.robj, false); |
| 118 | if (likely(r == 0)) { |
||
| 119 | radeon_bo_kunmap(rdev->gart.table.vram.robj); |
||
| 120 | radeon_bo_unpin(rdev->gart.table.vram.robj); |
||
| 121 | radeon_bo_unreserve(rdev->gart.table.vram.robj); |
||
| 122 | } |
||
| 1246 | serge | 123 | } |
| 124 | } |
||
| 125 | |||
| 126 | void rv770_pcie_gart_fini(struct radeon_device *rdev) |
||
| 127 | { |
||
| 128 | rv770_pcie_gart_disable(rdev); |
||
| 1404 | serge | 129 | radeon_gart_table_vram_free(rdev); |
| 1246 | serge | 130 | radeon_gart_fini(rdev); |
| 131 | } |
||
| 132 | |||
| 133 | |||
| 134 | void rv770_agp_enable(struct radeon_device *rdev) |
||
| 135 | { |
||
| 136 | u32 tmp; |
||
| 137 | int i; |
||
| 138 | |||
| 139 | /* Setup L2 cache */ |
||
| 140 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | |
||
| 141 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | |
||
| 142 | EFFECTIVE_L2_QUEUE_SIZE(7)); |
||
| 143 | WREG32(VM_L2_CNTL2, 0); |
||
| 144 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); |
||
| 145 | /* Setup TLB control */ |
||
| 146 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | |
||
| 147 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | |
||
| 148 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | |
||
| 149 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); |
||
| 150 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); |
||
| 151 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); |
||
| 152 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); |
||
| 153 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); |
||
| 154 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); |
||
| 155 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); |
||
| 156 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); |
||
| 157 | for (i = 0; i < 7; i++) |
||
| 158 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); |
||
| 159 | } |
||
| 160 | |||
| 161 | static void rv770_mc_program(struct radeon_device *rdev) |
||
| 162 | { |
||
| 163 | struct rv515_mc_save save; |
||
| 164 | u32 tmp; |
||
| 165 | int i, j; |
||
| 166 | |||
| 167 | /* Initialize HDP */ |
||
| 168 | for (i = 0, j = 0; i < 32; i++, j += 0x18) { |
||
| 169 | WREG32((0x2c14 + j), 0x00000000); |
||
| 170 | WREG32((0x2c18 + j), 0x00000000); |
||
| 171 | WREG32((0x2c1c + j), 0x00000000); |
||
| 172 | WREG32((0x2c20 + j), 0x00000000); |
||
| 173 | WREG32((0x2c24 + j), 0x00000000); |
||
| 174 | } |
||
| 175 | WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0); |
||
| 176 | |||
| 177 | rv515_mc_stop(rdev, &save); |
||
| 178 | if (r600_mc_wait_for_idle(rdev)) { |
||
| 179 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
||
| 180 | } |
||
| 181 | /* Lockout access through VGA aperture*/ |
||
| 182 | WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE); |
||
| 183 | /* Update configuration */ |
||
| 184 | if (rdev->flags & RADEON_IS_AGP) { |
||
| 185 | if (rdev->mc.vram_start < rdev->mc.gtt_start) { |
||
| 186 | /* VRAM before AGP */ |
||
| 187 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, |
||
| 188 | rdev->mc.vram_start >> 12); |
||
| 189 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, |
||
| 190 | rdev->mc.gtt_end >> 12); |
||
| 191 | } else { |
||
| 192 | /* VRAM after AGP */ |
||
| 193 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, |
||
| 194 | rdev->mc.gtt_start >> 12); |
||
| 195 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, |
||
| 196 | rdev->mc.vram_end >> 12); |
||
| 197 | } |
||
| 198 | } else { |
||
| 199 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, |
||
| 200 | rdev->mc.vram_start >> 12); |
||
| 201 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, |
||
| 202 | rdev->mc.vram_end >> 12); |
||
| 203 | } |
||
| 204 | WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0); |
||
| 205 | tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16; |
||
| 206 | tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF); |
||
| 207 | WREG32(MC_VM_FB_LOCATION, tmp); |
||
| 208 | WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8)); |
||
| 209 | WREG32(HDP_NONSURFACE_INFO, (2 << 7)); |
||
| 210 | WREG32(HDP_NONSURFACE_SIZE, (rdev->mc.mc_vram_size - 1) | 0x3FF); |
||
| 211 | if (rdev->flags & RADEON_IS_AGP) { |
||
| 212 | WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16); |
||
| 213 | WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16); |
||
| 214 | WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22); |
||
| 215 | } else { |
||
| 216 | WREG32(MC_VM_AGP_BASE, 0); |
||
| 217 | WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF); |
||
| 218 | WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF); |
||
| 219 | } |
||
| 220 | if (r600_mc_wait_for_idle(rdev)) { |
||
| 221 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
||
| 222 | } |
||
| 223 | rv515_mc_resume(rdev, &save); |
||
| 224 | /* we need to own VRAM, so turn off the VGA renderer here |
||
| 225 | * to stop it overwriting our objects */ |
||
| 226 | rv515_vga_render_disable(rdev); |
||
| 227 | } |
||
| 228 | |||
| 229 | |||
| 230 | /* |
||
| 231 | * CP. |
||
| 232 | */ |
||
| 233 | void r700_cp_stop(struct radeon_device *rdev) |
||
| 234 | { |
||
| 235 | WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT)); |
||
| 236 | } |
||
| 237 | |||
| 1413 | serge | 238 | |
| 1246 | serge | 239 | static int rv770_cp_load_microcode(struct radeon_device *rdev) |
| 240 | { |
||
| 241 | const __be32 *fw_data; |
||
| 242 | int i; |
||
| 243 | |||
| 244 | if (!rdev->me_fw || !rdev->pfp_fw) |
||
| 245 | return -EINVAL; |
||
| 246 | |||
| 247 | r700_cp_stop(rdev); |
||
| 248 | WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0)); |
||
| 249 | |||
| 250 | /* Reset cp */ |
||
| 251 | WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP); |
||
| 252 | RREG32(GRBM_SOFT_RESET); |
||
| 253 | mdelay(15); |
||
| 254 | WREG32(GRBM_SOFT_RESET, 0); |
||
| 255 | |||
| 256 | fw_data = (const __be32 *)rdev->pfp_fw->data; |
||
| 257 | WREG32(CP_PFP_UCODE_ADDR, 0); |
||
| 258 | for (i = 0; i < R700_PFP_UCODE_SIZE; i++) |
||
| 259 | WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++)); |
||
| 260 | WREG32(CP_PFP_UCODE_ADDR, 0); |
||
| 261 | |||
| 262 | fw_data = (const __be32 *)rdev->me_fw->data; |
||
| 263 | WREG32(CP_ME_RAM_WADDR, 0); |
||
| 264 | for (i = 0; i < R700_PM4_UCODE_SIZE; i++) |
||
| 265 | WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++)); |
||
| 266 | |||
| 267 | WREG32(CP_PFP_UCODE_ADDR, 0); |
||
| 268 | WREG32(CP_ME_RAM_WADDR, 0); |
||
| 269 | WREG32(CP_ME_RAM_RADDR, 0); |
||
| 270 | return 0; |
||
| 271 | } |
||
| 272 | |||
| 273 | |||
| 274 | /* |
||
| 275 | * Core functions |
||
| 276 | */ |
||
| 1430 | serge | 277 | static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev, |
| 278 | u32 num_tile_pipes, |
||
| 1246 | serge | 279 | u32 num_backends, |
| 280 | u32 backend_disable_mask) |
||
| 281 | { |
||
| 282 | u32 backend_map = 0; |
||
| 283 | u32 enabled_backends_mask; |
||
| 284 | u32 enabled_backends_count; |
||
| 285 | u32 cur_pipe; |
||
| 286 | u32 swizzle_pipe[R7XX_MAX_PIPES]; |
||
| 287 | u32 cur_backend; |
||
| 288 | u32 i; |
||
| 1430 | serge | 289 | bool force_no_swizzle; |
| 1246 | serge | 290 | |
| 291 | if (num_tile_pipes > R7XX_MAX_PIPES) |
||
| 292 | num_tile_pipes = R7XX_MAX_PIPES; |
||
| 293 | if (num_tile_pipes < 1) |
||
| 294 | num_tile_pipes = 1; |
||
| 295 | if (num_backends > R7XX_MAX_BACKENDS) |
||
| 296 | num_backends = R7XX_MAX_BACKENDS; |
||
| 297 | if (num_backends < 1) |
||
| 298 | num_backends = 1; |
||
| 299 | |||
| 300 | enabled_backends_mask = 0; |
||
| 301 | enabled_backends_count = 0; |
||
| 302 | for (i = 0; i < R7XX_MAX_BACKENDS; ++i) { |
||
| 303 | if (((backend_disable_mask >> i) & 1) == 0) { |
||
| 304 | enabled_backends_mask |= (1 << i); |
||
| 305 | ++enabled_backends_count; |
||
| 306 | } |
||
| 307 | if (enabled_backends_count == num_backends) |
||
| 308 | break; |
||
| 309 | } |
||
| 310 | |||
| 311 | if (enabled_backends_count == 0) { |
||
| 312 | enabled_backends_mask = 1; |
||
| 313 | enabled_backends_count = 1; |
||
| 314 | } |
||
| 315 | |||
| 316 | if (enabled_backends_count != num_backends) |
||
| 317 | num_backends = enabled_backends_count; |
||
| 318 | |||
| 1430 | serge | 319 | switch (rdev->family) { |
| 320 | case CHIP_RV770: |
||
| 321 | case CHIP_RV730: |
||
| 322 | force_no_swizzle = false; |
||
| 323 | break; |
||
| 324 | case CHIP_RV710: |
||
| 325 | case CHIP_RV740: |
||
| 326 | default: |
||
| 327 | force_no_swizzle = true; |
||
| 328 | break; |
||
| 329 | } |
||
| 330 | |||
| 1246 | serge | 331 | memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES); |
| 332 | switch (num_tile_pipes) { |
||
| 333 | case 1: |
||
| 334 | swizzle_pipe[0] = 0; |
||
| 335 | break; |
||
| 336 | case 2: |
||
| 337 | swizzle_pipe[0] = 0; |
||
| 338 | swizzle_pipe[1] = 1; |
||
| 339 | break; |
||
| 340 | case 3: |
||
| 1430 | serge | 341 | if (force_no_swizzle) { |
| 342 | swizzle_pipe[0] = 0; |
||
| 343 | swizzle_pipe[1] = 1; |
||
| 344 | swizzle_pipe[2] = 2; |
||
| 345 | } else { |
||
| 1246 | serge | 346 | swizzle_pipe[0] = 0; |
| 347 | swizzle_pipe[1] = 2; |
||
| 348 | swizzle_pipe[2] = 1; |
||
| 1430 | serge | 349 | } |
| 1246 | serge | 350 | break; |
| 351 | case 4: |
||
| 1430 | serge | 352 | if (force_no_swizzle) { |
| 353 | swizzle_pipe[0] = 0; |
||
| 354 | swizzle_pipe[1] = 1; |
||
| 355 | swizzle_pipe[2] = 2; |
||
| 356 | swizzle_pipe[3] = 3; |
||
| 357 | } else { |
||
| 1246 | serge | 358 | swizzle_pipe[0] = 0; |
| 359 | swizzle_pipe[1] = 2; |
||
| 360 | swizzle_pipe[2] = 3; |
||
| 361 | swizzle_pipe[3] = 1; |
||
| 1430 | serge | 362 | } |
| 1246 | serge | 363 | break; |
| 364 | case 5: |
||
| 1430 | serge | 365 | if (force_no_swizzle) { |
| 366 | swizzle_pipe[0] = 0; |
||
| 367 | swizzle_pipe[1] = 1; |
||
| 368 | swizzle_pipe[2] = 2; |
||
| 369 | swizzle_pipe[3] = 3; |
||
| 370 | swizzle_pipe[4] = 4; |
||
| 371 | } else { |
||
| 1246 | serge | 372 | swizzle_pipe[0] = 0; |
| 373 | swizzle_pipe[1] = 2; |
||
| 374 | swizzle_pipe[2] = 4; |
||
| 375 | swizzle_pipe[3] = 1; |
||
| 376 | swizzle_pipe[4] = 3; |
||
| 1430 | serge | 377 | } |
| 1246 | serge | 378 | break; |
| 379 | case 6: |
||
| 1430 | serge | 380 | if (force_no_swizzle) { |
| 381 | swizzle_pipe[0] = 0; |
||
| 382 | swizzle_pipe[1] = 1; |
||
| 383 | swizzle_pipe[2] = 2; |
||
| 384 | swizzle_pipe[3] = 3; |
||
| 385 | swizzle_pipe[4] = 4; |
||
| 386 | swizzle_pipe[5] = 5; |
||
| 387 | } else { |
||
| 1246 | serge | 388 | swizzle_pipe[0] = 0; |
| 389 | swizzle_pipe[1] = 2; |
||
| 390 | swizzle_pipe[2] = 4; |
||
| 391 | swizzle_pipe[3] = 5; |
||
| 392 | swizzle_pipe[4] = 3; |
||
| 393 | swizzle_pipe[5] = 1; |
||
| 1430 | serge | 394 | } |
| 1246 | serge | 395 | break; |
| 396 | case 7: |
||
| 1430 | serge | 397 | if (force_no_swizzle) { |
| 398 | swizzle_pipe[0] = 0; |
||
| 399 | swizzle_pipe[1] = 1; |
||
| 400 | swizzle_pipe[2] = 2; |
||
| 401 | swizzle_pipe[3] = 3; |
||
| 402 | swizzle_pipe[4] = 4; |
||
| 403 | swizzle_pipe[5] = 5; |
||
| 404 | swizzle_pipe[6] = 6; |
||
| 405 | } else { |
||
| 1246 | serge | 406 | swizzle_pipe[0] = 0; |
| 407 | swizzle_pipe[1] = 2; |
||
| 408 | swizzle_pipe[2] = 4; |
||
| 409 | swizzle_pipe[3] = 6; |
||
| 410 | swizzle_pipe[4] = 3; |
||
| 411 | swizzle_pipe[5] = 1; |
||
| 412 | swizzle_pipe[6] = 5; |
||
| 1430 | serge | 413 | } |
| 1246 | serge | 414 | break; |
| 415 | case 8: |
||
| 1430 | serge | 416 | if (force_no_swizzle) { |
| 417 | swizzle_pipe[0] = 0; |
||
| 418 | swizzle_pipe[1] = 1; |
||
| 419 | swizzle_pipe[2] = 2; |
||
| 420 | swizzle_pipe[3] = 3; |
||
| 421 | swizzle_pipe[4] = 4; |
||
| 422 | swizzle_pipe[5] = 5; |
||
| 423 | swizzle_pipe[6] = 6; |
||
| 424 | swizzle_pipe[7] = 7; |
||
| 425 | } else { |
||
| 1246 | serge | 426 | swizzle_pipe[0] = 0; |
| 427 | swizzle_pipe[1] = 2; |
||
| 428 | swizzle_pipe[2] = 4; |
||
| 429 | swizzle_pipe[3] = 6; |
||
| 430 | swizzle_pipe[4] = 3; |
||
| 431 | swizzle_pipe[5] = 1; |
||
| 432 | swizzle_pipe[6] = 7; |
||
| 433 | swizzle_pipe[7] = 5; |
||
| 1430 | serge | 434 | } |
| 1246 | serge | 435 | break; |
| 436 | } |
||
| 437 | |||
| 438 | cur_backend = 0; |
||
| 439 | for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) { |
||
| 440 | while (((1 << cur_backend) & enabled_backends_mask) == 0) |
||
| 441 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; |
||
| 442 | |||
| 443 | backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2))); |
||
| 444 | |||
| 445 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; |
||
| 446 | } |
||
| 447 | |||
| 448 | return backend_map; |
||
| 449 | } |
||
| 450 | |||
| 451 | static void rv770_gpu_init(struct radeon_device *rdev) |
||
| 452 | { |
||
| 453 | int i, j, num_qd_pipes; |
||
| 1430 | serge | 454 | u32 ta_aux_cntl; |
| 1246 | serge | 455 | u32 sx_debug_1; |
| 456 | u32 smx_dc_ctl0; |
||
| 1430 | serge | 457 | u32 db_debug3; |
| 1246 | serge | 458 | u32 num_gs_verts_per_thread; |
| 459 | u32 vgt_gs_per_es; |
||
| 460 | u32 gs_prim_buffer_depth = 0; |
||
| 461 | u32 sq_ms_fifo_sizes; |
||
| 462 | u32 sq_config; |
||
| 463 | u32 sq_thread_resource_mgmt; |
||
| 464 | u32 hdp_host_path_cntl; |
||
| 465 | u32 sq_dyn_gpr_size_simd_ab_0; |
||
| 466 | u32 backend_map; |
||
| 467 | u32 gb_tiling_config = 0; |
||
| 468 | u32 cc_rb_backend_disable = 0; |
||
| 469 | u32 cc_gc_shader_pipe_config = 0; |
||
| 470 | u32 mc_arb_ramcfg; |
||
| 471 | u32 db_debug4; |
||
| 472 | |||
| 473 | /* setup chip specs */ |
||
| 474 | switch (rdev->family) { |
||
| 475 | case CHIP_RV770: |
||
| 476 | rdev->config.rv770.max_pipes = 4; |
||
| 477 | rdev->config.rv770.max_tile_pipes = 8; |
||
| 478 | rdev->config.rv770.max_simds = 10; |
||
| 479 | rdev->config.rv770.max_backends = 4; |
||
| 480 | rdev->config.rv770.max_gprs = 256; |
||
| 481 | rdev->config.rv770.max_threads = 248; |
||
| 482 | rdev->config.rv770.max_stack_entries = 512; |
||
| 483 | rdev->config.rv770.max_hw_contexts = 8; |
||
| 484 | rdev->config.rv770.max_gs_threads = 16 * 2; |
||
| 485 | rdev->config.rv770.sx_max_export_size = 128; |
||
| 486 | rdev->config.rv770.sx_max_export_pos_size = 16; |
||
| 487 | rdev->config.rv770.sx_max_export_smx_size = 112; |
||
| 488 | rdev->config.rv770.sq_num_cf_insts = 2; |
||
| 489 | |||
| 490 | rdev->config.rv770.sx_num_of_sets = 7; |
||
| 491 | rdev->config.rv770.sc_prim_fifo_size = 0xF9; |
||
| 492 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; |
||
| 493 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; |
||
| 494 | break; |
||
| 495 | case CHIP_RV730: |
||
| 496 | rdev->config.rv770.max_pipes = 2; |
||
| 497 | rdev->config.rv770.max_tile_pipes = 4; |
||
| 498 | rdev->config.rv770.max_simds = 8; |
||
| 499 | rdev->config.rv770.max_backends = 2; |
||
| 500 | rdev->config.rv770.max_gprs = 128; |
||
| 501 | rdev->config.rv770.max_threads = 248; |
||
| 502 | rdev->config.rv770.max_stack_entries = 256; |
||
| 503 | rdev->config.rv770.max_hw_contexts = 8; |
||
| 504 | rdev->config.rv770.max_gs_threads = 16 * 2; |
||
| 505 | rdev->config.rv770.sx_max_export_size = 256; |
||
| 506 | rdev->config.rv770.sx_max_export_pos_size = 32; |
||
| 507 | rdev->config.rv770.sx_max_export_smx_size = 224; |
||
| 508 | rdev->config.rv770.sq_num_cf_insts = 2; |
||
| 509 | |||
| 510 | rdev->config.rv770.sx_num_of_sets = 7; |
||
| 511 | rdev->config.rv770.sc_prim_fifo_size = 0xf9; |
||
| 512 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; |
||
| 513 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; |
||
| 514 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { |
||
| 515 | rdev->config.rv770.sx_max_export_pos_size -= 16; |
||
| 516 | rdev->config.rv770.sx_max_export_smx_size += 16; |
||
| 517 | } |
||
| 518 | break; |
||
| 519 | case CHIP_RV710: |
||
| 520 | rdev->config.rv770.max_pipes = 2; |
||
| 521 | rdev->config.rv770.max_tile_pipes = 2; |
||
| 522 | rdev->config.rv770.max_simds = 2; |
||
| 523 | rdev->config.rv770.max_backends = 1; |
||
| 524 | rdev->config.rv770.max_gprs = 256; |
||
| 525 | rdev->config.rv770.max_threads = 192; |
||
| 526 | rdev->config.rv770.max_stack_entries = 256; |
||
| 527 | rdev->config.rv770.max_hw_contexts = 4; |
||
| 528 | rdev->config.rv770.max_gs_threads = 8 * 2; |
||
| 529 | rdev->config.rv770.sx_max_export_size = 128; |
||
| 530 | rdev->config.rv770.sx_max_export_pos_size = 16; |
||
| 531 | rdev->config.rv770.sx_max_export_smx_size = 112; |
||
| 532 | rdev->config.rv770.sq_num_cf_insts = 1; |
||
| 533 | |||
| 534 | rdev->config.rv770.sx_num_of_sets = 7; |
||
| 535 | rdev->config.rv770.sc_prim_fifo_size = 0x40; |
||
| 536 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; |
||
| 537 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; |
||
| 538 | break; |
||
| 539 | case CHIP_RV740: |
||
| 540 | rdev->config.rv770.max_pipes = 4; |
||
| 541 | rdev->config.rv770.max_tile_pipes = 4; |
||
| 542 | rdev->config.rv770.max_simds = 8; |
||
| 543 | rdev->config.rv770.max_backends = 4; |
||
| 544 | rdev->config.rv770.max_gprs = 256; |
||
| 545 | rdev->config.rv770.max_threads = 248; |
||
| 546 | rdev->config.rv770.max_stack_entries = 512; |
||
| 547 | rdev->config.rv770.max_hw_contexts = 8; |
||
| 548 | rdev->config.rv770.max_gs_threads = 16 * 2; |
||
| 549 | rdev->config.rv770.sx_max_export_size = 256; |
||
| 550 | rdev->config.rv770.sx_max_export_pos_size = 32; |
||
| 551 | rdev->config.rv770.sx_max_export_smx_size = 224; |
||
| 552 | rdev->config.rv770.sq_num_cf_insts = 2; |
||
| 553 | |||
| 554 | rdev->config.rv770.sx_num_of_sets = 7; |
||
| 555 | rdev->config.rv770.sc_prim_fifo_size = 0x100; |
||
| 556 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; |
||
| 557 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; |
||
| 558 | |||
| 559 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { |
||
| 560 | rdev->config.rv770.sx_max_export_pos_size -= 16; |
||
| 561 | rdev->config.rv770.sx_max_export_smx_size += 16; |
||
| 562 | } |
||
| 563 | break; |
||
| 564 | default: |
||
| 565 | break; |
||
| 566 | } |
||
| 567 | |||
| 568 | /* Initialize HDP */ |
||
| 569 | j = 0; |
||
| 570 | for (i = 0; i < 32; i++) { |
||
| 571 | WREG32((0x2c14 + j), 0x00000000); |
||
| 572 | WREG32((0x2c18 + j), 0x00000000); |
||
| 573 | WREG32((0x2c1c + j), 0x00000000); |
||
| 574 | WREG32((0x2c20 + j), 0x00000000); |
||
| 575 | WREG32((0x2c24 + j), 0x00000000); |
||
| 576 | j += 0x18; |
||
| 577 | } |
||
| 578 | |||
| 579 | WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff)); |
||
| 580 | |||
| 581 | /* setup tiling, simd, pipe config */ |
||
| 582 | mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG); |
||
| 583 | |||
| 584 | switch (rdev->config.rv770.max_tile_pipes) { |
||
| 585 | case 1: |
||
| 1430 | serge | 586 | default: |
| 1246 | serge | 587 | gb_tiling_config |= PIPE_TILING(0); |
| 588 | break; |
||
| 589 | case 2: |
||
| 590 | gb_tiling_config |= PIPE_TILING(1); |
||
| 591 | break; |
||
| 592 | case 4: |
||
| 593 | gb_tiling_config |= PIPE_TILING(2); |
||
| 594 | break; |
||
| 595 | case 8: |
||
| 596 | gb_tiling_config |= PIPE_TILING(3); |
||
| 597 | break; |
||
| 598 | } |
||
| 1430 | serge | 599 | rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes; |
| 1246 | serge | 600 | |
| 601 | if (rdev->family == CHIP_RV770) |
||
| 602 | gb_tiling_config |= BANK_TILING(1); |
||
| 603 | else |
||
| 1268 | serge | 604 | gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT); |
| 1430 | serge | 605 | rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3); |
| 1246 | serge | 606 | |
| 607 | gb_tiling_config |= GROUP_SIZE(0); |
||
| 1430 | serge | 608 | rdev->config.rv770.tiling_group_size = 256; |
| 1246 | serge | 609 | |
| 1268 | serge | 610 | if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) { |
| 1246 | serge | 611 | gb_tiling_config |= ROW_TILING(3); |
| 612 | gb_tiling_config |= SAMPLE_SPLIT(3); |
||
| 613 | } else { |
||
| 614 | gb_tiling_config |= |
||
| 615 | ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); |
||
| 616 | gb_tiling_config |= |
||
| 617 | SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); |
||
| 618 | } |
||
| 619 | |||
| 620 | gb_tiling_config |= BANK_SWAPS(1); |
||
| 621 | |||
| 1430 | serge | 622 | cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000; |
| 623 | cc_rb_backend_disable |= |
||
| 624 | BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK); |
||
| 1246 | serge | 625 | |
| 1430 | serge | 626 | cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00; |
| 627 | cc_gc_shader_pipe_config |= |
||
| 1246 | serge | 628 | INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK); |
| 629 | cc_gc_shader_pipe_config |= |
||
| 630 | INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK); |
||
| 631 | |||
| 1430 | serge | 632 | if (rdev->family == CHIP_RV740) |
| 633 | backend_map = 0x28; |
||
| 634 | else |
||
| 635 | backend_map = r700_get_tile_pipe_to_backend_map(rdev, |
||
| 636 | rdev->config.rv770.max_tile_pipes, |
||
| 637 | (R7XX_MAX_BACKENDS - |
||
| 638 | r600_count_pipe_bits((cc_rb_backend_disable & |
||
| 639 | R7XX_MAX_BACKENDS_MASK) >> 16)), |
||
| 640 | (cc_rb_backend_disable >> 16)); |
||
| 641 | gb_tiling_config |= BACKEND_MAP(backend_map); |
||
| 1246 | serge | 642 | |
| 1430 | serge | 643 | |
| 1246 | serge | 644 | WREG32(GB_TILING_CONFIG, gb_tiling_config); |
| 645 | WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff)); |
||
| 646 | WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff)); |
||
| 647 | |||
| 648 | WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable); |
||
| 649 | WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); |
||
| 1430 | serge | 650 | WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable); |
| 1246 | serge | 651 | |
| 652 | WREG32(CGTS_SYS_TCC_DISABLE, 0); |
||
| 653 | WREG32(CGTS_TCC_DISABLE, 0); |
||
| 654 | |||
| 655 | num_qd_pipes = |
||
| 1430 | serge | 656 | R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8); |
| 1246 | serge | 657 | WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK); |
| 658 | WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK); |
||
| 659 | |||
| 660 | /* set HW defaults for 3D engine */ |
||
| 661 | WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) | |
||
| 662 | ROQ_IB2_START(0x2b))); |
||
| 663 | |||
| 664 | WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30)); |
||
| 665 | |||
| 1430 | serge | 666 | ta_aux_cntl = RREG32(TA_CNTL_AUX); |
| 667 | WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO); |
||
| 1246 | serge | 668 | |
| 669 | sx_debug_1 = RREG32(SX_DEBUG_1); |
||
| 670 | sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS; |
||
| 671 | WREG32(SX_DEBUG_1, sx_debug_1); |
||
| 672 | |||
| 673 | smx_dc_ctl0 = RREG32(SMX_DC_CTL0); |
||
| 674 | smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff); |
||
| 675 | smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1); |
||
| 676 | WREG32(SMX_DC_CTL0, smx_dc_ctl0); |
||
| 677 | |||
| 1430 | serge | 678 | if (rdev->family != CHIP_RV740) |
| 1246 | serge | 679 | WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) | |
| 680 | GS_FLUSH_CTL(4) | |
||
| 681 | ACK_FLUSH_CTL(3) | |
||
| 682 | SYNC_FLUSH_CTL)); |
||
| 683 | |||
| 1430 | serge | 684 | db_debug3 = RREG32(DB_DEBUG3); |
| 685 | db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f); |
||
| 686 | switch (rdev->family) { |
||
| 687 | case CHIP_RV770: |
||
| 688 | case CHIP_RV740: |
||
| 689 | db_debug3 |= DB_CLK_OFF_DELAY(0x1f); |
||
| 690 | break; |
||
| 691 | case CHIP_RV710: |
||
| 692 | case CHIP_RV730: |
||
| 693 | default: |
||
| 694 | db_debug3 |= DB_CLK_OFF_DELAY(2); |
||
| 695 | break; |
||
| 696 | } |
||
| 697 | WREG32(DB_DEBUG3, db_debug3); |
||
| 698 | |||
| 699 | if (rdev->family != CHIP_RV770) { |
||
| 1246 | serge | 700 | db_debug4 = RREG32(DB_DEBUG4); |
| 701 | db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER; |
||
| 702 | WREG32(DB_DEBUG4, db_debug4); |
||
| 703 | } |
||
| 704 | |||
| 705 | WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) | |
||
| 706 | POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) | |
||
| 707 | SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1))); |
||
| 708 | |||
| 709 | WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) | |
||
| 710 | SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) | |
||
| 711 | SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize))); |
||
| 712 | |||
| 713 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); |
||
| 714 | |||
| 715 | WREG32(VGT_NUM_INSTANCES, 1); |
||
| 716 | |||
| 717 | WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0)); |
||
| 718 | |||
| 719 | WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4)); |
||
| 720 | |||
| 721 | WREG32(CP_PERFMON_CNTL, 0); |
||
| 722 | |||
| 723 | sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) | |
||
| 724 | DONE_FIFO_HIWATER(0xe0) | |
||
| 725 | ALU_UPDATE_FIFO_HIWATER(0x8)); |
||
| 726 | switch (rdev->family) { |
||
| 727 | case CHIP_RV770: |
||
| 1430 | serge | 728 | case CHIP_RV730: |
| 729 | case CHIP_RV710: |
||
| 1246 | serge | 730 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1); |
| 731 | break; |
||
| 732 | case CHIP_RV740: |
||
| 733 | default: |
||
| 734 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4); |
||
| 735 | break; |
||
| 736 | } |
||
| 737 | WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes); |
||
| 738 | |||
| 739 | /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT |
||
| 740 | * should be adjusted as needed by the 2D/3D drivers. This just sets default values |
||
| 741 | */ |
||
| 742 | sq_config = RREG32(SQ_CONFIG); |
||
| 743 | sq_config &= ~(PS_PRIO(3) | |
||
| 744 | VS_PRIO(3) | |
||
| 745 | GS_PRIO(3) | |
||
| 746 | ES_PRIO(3)); |
||
| 747 | sq_config |= (DX9_CONSTS | |
||
| 748 | VC_ENABLE | |
||
| 749 | EXPORT_SRC_C | |
||
| 750 | PS_PRIO(0) | |
||
| 751 | VS_PRIO(1) | |
||
| 752 | GS_PRIO(2) | |
||
| 753 | ES_PRIO(3)); |
||
| 754 | if (rdev->family == CHIP_RV710) |
||
| 755 | /* no vertex cache */ |
||
| 756 | sq_config &= ~VC_ENABLE; |
||
| 757 | |||
| 758 | WREG32(SQ_CONFIG, sq_config); |
||
| 759 | |||
| 760 | WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | |
||
| 761 | NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | |
||
| 762 | NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2))); |
||
| 763 | |||
| 764 | WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) | |
||
| 765 | NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64))); |
||
| 766 | |||
| 767 | sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) | |
||
| 768 | NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) | |
||
| 769 | NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8)); |
||
| 770 | if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads) |
||
| 771 | sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads); |
||
| 772 | else |
||
| 773 | sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8); |
||
| 774 | WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt); |
||
| 775 | |||
| 776 | WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | |
||
| 777 | NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); |
||
| 778 | |||
| 779 | WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | |
||
| 780 | NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); |
||
| 781 | |||
| 782 | sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) | |
||
| 783 | SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) | |
||
| 784 | SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) | |
||
| 785 | SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64)); |
||
| 786 | |||
| 787 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0); |
||
| 788 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0); |
||
| 789 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0); |
||
| 790 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0); |
||
| 791 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0); |
||
| 792 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0); |
||
| 793 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0); |
||
| 794 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0); |
||
| 795 | |||
| 796 | WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) | |
||
| 797 | FORCE_EOV_MAX_REZ_CNT(255))); |
||
| 798 | |||
| 799 | if (rdev->family == CHIP_RV710) |
||
| 800 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) | |
||
| 801 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
||
| 802 | else |
||
| 803 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) | |
||
| 804 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
||
| 805 | |||
| 806 | switch (rdev->family) { |
||
| 807 | case CHIP_RV770: |
||
| 808 | case CHIP_RV730: |
||
| 809 | case CHIP_RV740: |
||
| 810 | gs_prim_buffer_depth = 384; |
||
| 811 | break; |
||
| 812 | case CHIP_RV710: |
||
| 813 | gs_prim_buffer_depth = 128; |
||
| 814 | break; |
||
| 815 | default: |
||
| 816 | break; |
||
| 817 | } |
||
| 818 | |||
| 819 | num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16; |
||
| 820 | vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread; |
||
| 821 | /* Max value for this is 256 */ |
||
| 822 | if (vgt_gs_per_es > 256) |
||
| 823 | vgt_gs_per_es = 256; |
||
| 824 | |||
| 825 | WREG32(VGT_ES_PER_GS, 128); |
||
| 826 | WREG32(VGT_GS_PER_ES, vgt_gs_per_es); |
||
| 827 | WREG32(VGT_GS_PER_VS, 2); |
||
| 828 | |||
| 829 | /* more default values. 2D/3D driver should adjust as needed */ |
||
| 830 | WREG32(VGT_GS_VERTEX_REUSE, 16); |
||
| 831 | WREG32(PA_SC_LINE_STIPPLE_STATE, 0); |
||
| 832 | WREG32(VGT_STRMOUT_EN, 0); |
||
| 833 | WREG32(SX_MISC, 0); |
||
| 834 | WREG32(PA_SC_MODE_CNTL, 0); |
||
| 835 | WREG32(PA_SC_EDGERULE, 0xaaaaaaaa); |
||
| 836 | WREG32(PA_SC_AA_CONFIG, 0); |
||
| 837 | WREG32(PA_SC_CLIPRECT_RULE, 0xffff); |
||
| 838 | WREG32(PA_SC_LINE_STIPPLE, 0); |
||
| 839 | WREG32(SPI_INPUT_Z, 0); |
||
| 840 | WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2)); |
||
| 841 | WREG32(CB_COLOR7_FRAG, 0); |
||
| 842 | |||
| 843 | /* clear render buffer base addresses */ |
||
| 844 | WREG32(CB_COLOR0_BASE, 0); |
||
| 845 | WREG32(CB_COLOR1_BASE, 0); |
||
| 846 | WREG32(CB_COLOR2_BASE, 0); |
||
| 847 | WREG32(CB_COLOR3_BASE, 0); |
||
| 848 | WREG32(CB_COLOR4_BASE, 0); |
||
| 849 | WREG32(CB_COLOR5_BASE, 0); |
||
| 850 | WREG32(CB_COLOR6_BASE, 0); |
||
| 851 | WREG32(CB_COLOR7_BASE, 0); |
||
| 852 | |||
| 853 | WREG32(TCP_CNTL, 0); |
||
| 854 | |||
| 855 | hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL); |
||
| 856 | WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl); |
||
| 857 | |||
| 858 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); |
||
| 859 | |||
| 860 | WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA | |
||
| 861 | NUM_CLIP_SEQ(3))); |
||
| 862 | |||
| 863 | } |
||
| 864 | |||
| 865 | int rv770_mc_init(struct radeon_device *rdev) |
||
| 866 | { |
||
| 867 | fixed20_12 a; |
||
| 868 | u32 tmp; |
||
| 1268 | serge | 869 | int chansize, numchan; |
| 1246 | serge | 870 | |
| 871 | /* Get VRAM informations */ |
||
| 872 | rdev->mc.vram_is_ddr = true; |
||
| 1268 | serge | 873 | tmp = RREG32(MC_ARB_RAMCFG); |
| 874 | if (tmp & CHANSIZE_OVERRIDE) { |
||
| 875 | chansize = 16; |
||
| 876 | } else if (tmp & CHANSIZE_MASK) { |
||
| 877 | chansize = 64; |
||
| 878 | } else { |
||
| 879 | chansize = 32; |
||
| 880 | } |
||
| 881 | tmp = RREG32(MC_SHARED_CHMAP); |
||
| 882 | switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) { |
||
| 883 | case 0: |
||
| 884 | default: |
||
| 885 | numchan = 1; |
||
| 886 | break; |
||
| 887 | case 1: |
||
| 888 | numchan = 2; |
||
| 889 | break; |
||
| 890 | case 2: |
||
| 891 | numchan = 4; |
||
| 892 | break; |
||
| 893 | case 3: |
||
| 894 | numchan = 8; |
||
| 895 | break; |
||
| 896 | } |
||
| 897 | rdev->mc.vram_width = numchan * chansize; |
||
| 1246 | serge | 898 | /* Could aper size report 0 ? */ |
| 899 | rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0); |
||
| 900 | rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0); |
||
| 901 | /* Setup GPU memory space */ |
||
| 902 | rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE); |
||
| 903 | rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE); |
||
| 1430 | serge | 904 | rdev->mc.visible_vram_size = rdev->mc.aper_size; |
| 905 | /* FIXME remove this once we support unmappable VRAM */ |
||
| 906 | if (rdev->mc.mc_vram_size > rdev->mc.aper_size) { |
||
| 1246 | serge | 907 | rdev->mc.mc_vram_size = rdev->mc.aper_size; |
| 908 | rdev->mc.real_vram_size = rdev->mc.aper_size; |
||
| 909 | } |
||
| 1430 | serge | 910 | r600_vram_gtt_location(rdev, &rdev->mc); |
| 1246 | serge | 911 | /* FIXME: we should enforce default clock in case GPU is not in |
| 912 | * default setup |
||
| 913 | */ |
||
| 914 | a.full = rfixed_const(100); |
||
| 915 | rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk); |
||
| 916 | rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a); |
||
| 917 | return 0; |
||
| 918 | } |
||
| 1430 | serge | 919 | |
| 1246 | serge | 920 | int rv770_gpu_reset(struct radeon_device *rdev) |
| 921 | { |
||
| 922 | /* FIXME: implement any rv770 specific bits */ |
||
| 923 | return r600_gpu_reset(rdev); |
||
| 924 | } |
||
| 925 | |||
| 926 | static int rv770_startup(struct radeon_device *rdev) |
||
| 927 | { |
||
| 928 | int r; |
||
| 929 | |||
| 1413 | serge | 930 | if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) { |
| 931 | r = r600_init_microcode(rdev); |
||
| 932 | if (r) { |
||
| 933 | DRM_ERROR("Failed to load firmware!\n"); |
||
| 934 | return r; |
||
| 935 | } |
||
| 936 | } |
||
| 937 | |||
| 1246 | serge | 938 | rv770_mc_program(rdev); |
| 939 | if (rdev->flags & RADEON_IS_AGP) { |
||
| 940 | rv770_agp_enable(rdev); |
||
| 941 | } else { |
||
| 942 | r = rv770_pcie_gart_enable(rdev); |
||
| 943 | if (r) |
||
| 944 | return r; |
||
| 945 | } |
||
| 946 | rv770_gpu_init(rdev); |
||
| 1413 | serge | 947 | r = radeon_ring_init(rdev, rdev->cp.ring_size); |
| 948 | if (r) |
||
| 949 | return r; |
||
| 950 | r = rv770_cp_load_microcode(rdev); |
||
| 951 | if (r) |
||
| 952 | return r; |
||
| 953 | r = r600_cp_resume(rdev); |
||
| 954 | if (r) |
||
| 955 | return r; |
||
| 1246 | serge | 956 | /* write back buffer are not vital so don't worry about failure */ |
| 1413 | serge | 957 | // r600_wb_enable(rdev); |
| 1246 | serge | 958 | return 0; |
| 959 | } |
||
| 960 | |||
| 961 | |||
| 1413 | serge | 962 | |
| 963 | |||
| 964 | |||
| 965 | |||
| 966 | |||
| 1246 | serge | 967 | /* Plan is to move initialization in that function and use |
| 968 | * helper function so that radeon_device_init pretty much |
||
| 969 | * do nothing more than calling asic specific function. This |
||
| 970 | * should also allow to remove a bunch of callback function |
||
| 971 | * like vram_info. |
||
| 972 | */ |
||
| 973 | int rv770_init(struct radeon_device *rdev) |
||
| 974 | { |
||
| 975 | int r; |
||
| 976 | |||
| 977 | r = radeon_dummy_page_init(rdev); |
||
| 978 | if (r) |
||
| 979 | return r; |
||
| 980 | /* This don't do much */ |
||
| 981 | r = radeon_gem_init(rdev); |
||
| 982 | if (r) |
||
| 983 | return r; |
||
| 984 | /* Read BIOS */ |
||
| 985 | if (!radeon_get_bios(rdev)) { |
||
| 986 | if (ASIC_IS_AVIVO(rdev)) |
||
| 987 | return -EINVAL; |
||
| 988 | } |
||
| 989 | /* Must be an ATOMBIOS */ |
||
| 990 | if (!rdev->is_atom_bios) { |
||
| 991 | dev_err(rdev->dev, "Expecting atombios for R600 GPU\n"); |
||
| 992 | return -EINVAL; |
||
| 993 | } |
||
| 994 | r = radeon_atombios_init(rdev); |
||
| 995 | if (r) |
||
| 996 | return r; |
||
| 997 | /* Post card if necessary */ |
||
| 1403 | serge | 998 | if (!r600_card_posted(rdev)) { |
| 999 | if (!rdev->bios) { |
||
| 1000 | dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n"); |
||
| 1001 | return -EINVAL; |
||
| 1002 | } |
||
| 1246 | serge | 1003 | DRM_INFO("GPU not posted. posting now...\n"); |
| 1004 | atom_asic_init(rdev->mode_info.atom_context); |
||
| 1005 | } |
||
| 1006 | /* Initialize scratch registers */ |
||
| 1007 | r600_scratch_init(rdev); |
||
| 1008 | /* Initialize surface registers */ |
||
| 1009 | radeon_surface_init(rdev); |
||
| 1268 | serge | 1010 | /* Initialize clocks */ |
| 1246 | serge | 1011 | radeon_get_clock_info(rdev->ddev); |
| 1012 | r = radeon_clocks_init(rdev); |
||
| 1013 | if (r) |
||
| 1014 | return r; |
||
| 1268 | serge | 1015 | /* Initialize power management */ |
| 1016 | radeon_pm_init(rdev); |
||
| 1246 | serge | 1017 | /* Fence driver */ |
| 1018 | // r = radeon_fence_driver_init(rdev); |
||
| 1019 | // if (r) |
||
| 1020 | // return r; |
||
| 1430 | serge | 1021 | /* initialize AGP */ |
| 1403 | serge | 1022 | if (rdev->flags & RADEON_IS_AGP) { |
| 1023 | r = radeon_agp_init(rdev); |
||
| 1024 | if (r) |
||
| 1025 | radeon_agp_disable(rdev); |
||
| 1026 | } |
||
| 1246 | serge | 1027 | r = rv770_mc_init(rdev); |
| 1028 | if (r) |
||
| 1029 | return r; |
||
| 1030 | /* Memory manager */ |
||
| 1403 | serge | 1031 | r = radeon_bo_init(rdev); |
| 1246 | serge | 1032 | if (r) |
| 1033 | return r; |
||
| 1034 | |||
| 1035 | |||
| 1413 | serge | 1036 | rdev->cp.ring_obj = NULL; |
| 1037 | r600_ring_init(rdev, 1024 * 1024); |
||
| 1038 | |||
| 1039 | |||
| 1246 | serge | 1040 | r = r600_pcie_gart_init(rdev); |
| 1041 | if (r) |
||
| 1042 | return r; |
||
| 1043 | |||
| 1044 | rdev->accel_working = true; |
||
| 1045 | r = rv770_startup(rdev); |
||
| 1046 | if (r) { |
||
| 1413 | serge | 1047 | dev_err(rdev->dev, "disabling GPU acceleration\n"); |
| 1048 | |||
| 1246 | serge | 1049 | rv770_pcie_gart_fini(rdev); |
| 1050 | rdev->accel_working = false; |
||
| 1051 | } |
||
| 1052 | if (rdev->accel_working) { |
||
| 1430 | serge | 1053 | // r = radeon_ib_pool_init(rdev); |
| 1054 | // if (r) { |
||
| 1055 | // dev_err(rdev->dev, "IB initialization failed (%d).\n", r); |
||
| 1056 | // rdev->accel_working = false; |
||
| 1057 | // } else { |
||
| 1058 | // r = r600_ib_test(rdev); |
||
| 1059 | // if (r) { |
||
| 1060 | // dev_err(rdev->dev, "IB test failed (%d).\n", r); |
||
| 1061 | // rdev->accel_working = false; |
||
| 1062 | // } |
||
| 1063 | // } |
||
| 1246 | serge | 1064 | } |
| 1065 | return 0; |
||
| 1066 | } |
||
| 1067 |