Commit | Line | Data |
---|---|---|
771fe6b9 JG |
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 | */ | |
3ce0a23d JG |
28 | #include <linux/firmware.h> |
29 | #include <linux/platform_device.h> | |
771fe6b9 | 30 | #include "drmP.h" |
771fe6b9 | 31 | #include "radeon.h" |
e6990375 | 32 | #include "radeon_asic.h" |
4153e584 | 33 | #include "radeon_drm.h" |
3ce0a23d | 34 | #include "rv770d.h" |
3ce0a23d | 35 | #include "atom.h" |
d39c3b89 | 36 | #include "avivod.h" |
771fe6b9 | 37 | |
3ce0a23d JG |
38 | #define R700_PFP_UCODE_SIZE 848 |
39 | #define R700_PM4_UCODE_SIZE 1360 | |
771fe6b9 | 40 | |
3ce0a23d JG |
41 | static void rv770_gpu_init(struct radeon_device *rdev); |
42 | void rv770_fini(struct radeon_device *rdev); | |
771fe6b9 JG |
43 | |
44 | ||
45 | /* | |
3ce0a23d | 46 | * GART |
771fe6b9 | 47 | */ |
3ce0a23d | 48 | int rv770_pcie_gart_enable(struct radeon_device *rdev) |
771fe6b9 | 49 | { |
3ce0a23d JG |
50 | u32 tmp; |
51 | int r, i; | |
771fe6b9 | 52 | |
4aac0473 JG |
53 | if (rdev->gart.table.vram.robj == NULL) { |
54 | dev_err(rdev->dev, "No VRAM object for PCIE GART.\n"); | |
55 | return -EINVAL; | |
3ce0a23d | 56 | } |
4aac0473 JG |
57 | r = radeon_gart_table_vram_pin(rdev); |
58 | if (r) | |
3ce0a23d | 59 | return r; |
82568565 | 60 | radeon_gart_restore(rdev); |
3ce0a23d JG |
61 | /* Setup L2 cache */ |
62 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
63 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
64 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
65 | WREG32(VM_L2_CNTL2, 0); | |
66 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
67 | /* Setup TLB control */ | |
68 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
69 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
70 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
71 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
72 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
73 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
74 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
75 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
76 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
77 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
78 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
79 | WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12); | |
1a029b76 | 80 | WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12); |
3ce0a23d JG |
81 | WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12); |
82 | WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) | | |
83 | RANGE_PROTECTION_FAULT_ENABLE_DEFAULT); | |
84 | WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, | |
85 | (u32)(rdev->dummy_page.addr >> 12)); | |
86 | for (i = 1; i < 7; i++) | |
87 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
771fe6b9 | 88 | |
3ce0a23d JG |
89 | r600_pcie_gart_tlb_flush(rdev); |
90 | rdev->gart.ready = true; | |
771fe6b9 JG |
91 | return 0; |
92 | } | |
93 | ||
3ce0a23d | 94 | void rv770_pcie_gart_disable(struct radeon_device *rdev) |
771fe6b9 | 95 | { |
3ce0a23d | 96 | u32 tmp; |
4c788679 | 97 | int i, r; |
3ce0a23d | 98 | |
3ce0a23d JG |
99 | /* Disable all tables */ |
100 | for (i = 0; i < 7; i++) | |
101 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
102 | ||
103 | /* Setup L2 cache */ | |
104 | WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING | | |
105 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
106 | WREG32(VM_L2_CNTL2, 0); | |
107 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
108 | /* Setup TLB control */ | |
109 | tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
110 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
111 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
112 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
113 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
114 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
115 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
116 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
4aac0473 | 117 | if (rdev->gart.table.vram.robj) { |
4c788679 JG |
118 | r = radeon_bo_reserve(rdev->gart.table.vram.robj, false); |
119 | if (likely(r == 0)) { | |
120 | radeon_bo_kunmap(rdev->gart.table.vram.robj); | |
121 | radeon_bo_unpin(rdev->gart.table.vram.robj); | |
122 | radeon_bo_unreserve(rdev->gart.table.vram.robj); | |
123 | } | |
4aac0473 JG |
124 | } |
125 | } | |
126 | ||
127 | void rv770_pcie_gart_fini(struct radeon_device *rdev) | |
128 | { | |
f9274562 | 129 | radeon_gart_fini(rdev); |
4aac0473 JG |
130 | rv770_pcie_gart_disable(rdev); |
131 | radeon_gart_table_vram_free(rdev); | |
771fe6b9 JG |
132 | } |
133 | ||
134 | ||
1a029b76 JG |
135 | void rv770_agp_enable(struct radeon_device *rdev) |
136 | { | |
137 | u32 tmp; | |
138 | int i; | |
139 | ||
140 | /* Setup L2 cache */ | |
141 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
142 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
143 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
144 | WREG32(VM_L2_CNTL2, 0); | |
145 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
146 | /* Setup TLB control */ | |
147 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
148 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
149 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
150 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
151 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
152 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
153 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
154 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
155 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
156 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
157 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
158 | for (i = 0; i < 7; i++) | |
159 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
160 | } | |
161 | ||
a3c1945a | 162 | static void rv770_mc_program(struct radeon_device *rdev) |
771fe6b9 | 163 | { |
a3c1945a | 164 | struct rv515_mc_save save; |
3ce0a23d JG |
165 | u32 tmp; |
166 | int i, j; | |
167 | ||
168 | /* Initialize HDP */ | |
169 | for (i = 0, j = 0; i < 32; i++, j += 0x18) { | |
170 | WREG32((0x2c14 + j), 0x00000000); | |
171 | WREG32((0x2c18 + j), 0x00000000); | |
172 | WREG32((0x2c1c + j), 0x00000000); | |
173 | WREG32((0x2c20 + j), 0x00000000); | |
174 | WREG32((0x2c24 + j), 0x00000000); | |
175 | } | |
176 | WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0); | |
177 | ||
a3c1945a | 178 | rv515_mc_stop(rdev, &save); |
3ce0a23d | 179 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 180 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 181 | } |
3ce0a23d JG |
182 | /* Lockout access through VGA aperture*/ |
183 | WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE); | |
3ce0a23d | 184 | /* Update configuration */ |
1a029b76 JG |
185 | if (rdev->flags & RADEON_IS_AGP) { |
186 | if (rdev->mc.vram_start < rdev->mc.gtt_start) { | |
187 | /* VRAM before AGP */ | |
188 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
189 | rdev->mc.vram_start >> 12); | |
190 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
191 | rdev->mc.gtt_end >> 12); | |
192 | } else { | |
193 | /* VRAM after AGP */ | |
194 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
195 | rdev->mc.gtt_start >> 12); | |
196 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
197 | rdev->mc.vram_end >> 12); | |
198 | } | |
199 | } else { | |
200 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
201 | rdev->mc.vram_start >> 12); | |
202 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
203 | rdev->mc.vram_end >> 12); | |
204 | } | |
3ce0a23d | 205 | WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0); |
1a029b76 | 206 | tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16; |
3ce0a23d JG |
207 | tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF); |
208 | WREG32(MC_VM_FB_LOCATION, tmp); | |
209 | WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8)); | |
210 | WREG32(HDP_NONSURFACE_INFO, (2 << 7)); | |
211 | WREG32(HDP_NONSURFACE_SIZE, (rdev->mc.mc_vram_size - 1) | 0x3FF); | |
212 | if (rdev->flags & RADEON_IS_AGP) { | |
1a029b76 | 213 | WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16); |
3ce0a23d JG |
214 | WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16); |
215 | WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22); | |
216 | } else { | |
217 | WREG32(MC_VM_AGP_BASE, 0); | |
218 | WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF); | |
219 | WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF); | |
220 | } | |
3ce0a23d | 221 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 222 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 223 | } |
a3c1945a | 224 | rv515_mc_resume(rdev, &save); |
698443d9 DA |
225 | /* we need to own VRAM, so turn off the VGA renderer here |
226 | * to stop it overwriting our objects */ | |
d39c3b89 | 227 | rv515_vga_render_disable(rdev); |
771fe6b9 JG |
228 | } |
229 | ||
3ce0a23d JG |
230 | |
231 | /* | |
232 | * CP. | |
233 | */ | |
234 | void r700_cp_stop(struct radeon_device *rdev) | |
771fe6b9 | 235 | { |
3ce0a23d | 236 | WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT)); |
771fe6b9 JG |
237 | } |
238 | ||
3ce0a23d JG |
239 | |
240 | static int rv770_cp_load_microcode(struct radeon_device *rdev) | |
771fe6b9 | 241 | { |
3ce0a23d JG |
242 | const __be32 *fw_data; |
243 | int i; | |
244 | ||
245 | if (!rdev->me_fw || !rdev->pfp_fw) | |
246 | return -EINVAL; | |
247 | ||
248 | r700_cp_stop(rdev); | |
249 | WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0)); | |
250 | ||
251 | /* Reset cp */ | |
252 | WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP); | |
253 | RREG32(GRBM_SOFT_RESET); | |
254 | mdelay(15); | |
255 | WREG32(GRBM_SOFT_RESET, 0); | |
256 | ||
257 | fw_data = (const __be32 *)rdev->pfp_fw->data; | |
258 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
259 | for (i = 0; i < R700_PFP_UCODE_SIZE; i++) | |
260 | WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++)); | |
261 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
262 | ||
263 | fw_data = (const __be32 *)rdev->me_fw->data; | |
264 | WREG32(CP_ME_RAM_WADDR, 0); | |
265 | for (i = 0; i < R700_PM4_UCODE_SIZE; i++) | |
266 | WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++)); | |
267 | ||
268 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
269 | WREG32(CP_ME_RAM_WADDR, 0); | |
270 | WREG32(CP_ME_RAM_RADDR, 0); | |
271 | return 0; | |
771fe6b9 JG |
272 | } |
273 | ||
274 | ||
275 | /* | |
3ce0a23d | 276 | * Core functions |
771fe6b9 | 277 | */ |
d03f5d59 AD |
278 | static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev, |
279 | u32 num_tile_pipes, | |
280 | u32 num_backends, | |
281 | u32 backend_disable_mask) | |
771fe6b9 | 282 | { |
3ce0a23d JG |
283 | u32 backend_map = 0; |
284 | u32 enabled_backends_mask; | |
285 | u32 enabled_backends_count; | |
286 | u32 cur_pipe; | |
287 | u32 swizzle_pipe[R7XX_MAX_PIPES]; | |
288 | u32 cur_backend; | |
289 | u32 i; | |
d03f5d59 | 290 | bool force_no_swizzle; |
3ce0a23d JG |
291 | |
292 | if (num_tile_pipes > R7XX_MAX_PIPES) | |
293 | num_tile_pipes = R7XX_MAX_PIPES; | |
294 | if (num_tile_pipes < 1) | |
295 | num_tile_pipes = 1; | |
296 | if (num_backends > R7XX_MAX_BACKENDS) | |
297 | num_backends = R7XX_MAX_BACKENDS; | |
298 | if (num_backends < 1) | |
299 | num_backends = 1; | |
300 | ||
301 | enabled_backends_mask = 0; | |
302 | enabled_backends_count = 0; | |
303 | for (i = 0; i < R7XX_MAX_BACKENDS; ++i) { | |
304 | if (((backend_disable_mask >> i) & 1) == 0) { | |
305 | enabled_backends_mask |= (1 << i); | |
306 | ++enabled_backends_count; | |
307 | } | |
308 | if (enabled_backends_count == num_backends) | |
309 | break; | |
310 | } | |
311 | ||
312 | if (enabled_backends_count == 0) { | |
313 | enabled_backends_mask = 1; | |
314 | enabled_backends_count = 1; | |
315 | } | |
316 | ||
317 | if (enabled_backends_count != num_backends) | |
318 | num_backends = enabled_backends_count; | |
319 | ||
d03f5d59 AD |
320 | switch (rdev->family) { |
321 | case CHIP_RV770: | |
322 | case CHIP_RV730: | |
323 | force_no_swizzle = false; | |
324 | break; | |
325 | case CHIP_RV710: | |
326 | case CHIP_RV740: | |
327 | default: | |
328 | force_no_swizzle = true; | |
329 | break; | |
330 | } | |
331 | ||
3ce0a23d JG |
332 | memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES); |
333 | switch (num_tile_pipes) { | |
334 | case 1: | |
335 | swizzle_pipe[0] = 0; | |
336 | break; | |
337 | case 2: | |
338 | swizzle_pipe[0] = 0; | |
339 | swizzle_pipe[1] = 1; | |
340 | break; | |
341 | case 3: | |
d03f5d59 AD |
342 | if (force_no_swizzle) { |
343 | swizzle_pipe[0] = 0; | |
344 | swizzle_pipe[1] = 1; | |
345 | swizzle_pipe[2] = 2; | |
346 | } else { | |
347 | swizzle_pipe[0] = 0; | |
348 | swizzle_pipe[1] = 2; | |
349 | swizzle_pipe[2] = 1; | |
350 | } | |
3ce0a23d JG |
351 | break; |
352 | case 4: | |
d03f5d59 AD |
353 | if (force_no_swizzle) { |
354 | swizzle_pipe[0] = 0; | |
355 | swizzle_pipe[1] = 1; | |
356 | swizzle_pipe[2] = 2; | |
357 | swizzle_pipe[3] = 3; | |
358 | } else { | |
359 | swizzle_pipe[0] = 0; | |
360 | swizzle_pipe[1] = 2; | |
361 | swizzle_pipe[2] = 3; | |
362 | swizzle_pipe[3] = 1; | |
363 | } | |
3ce0a23d JG |
364 | break; |
365 | case 5: | |
d03f5d59 AD |
366 | if (force_no_swizzle) { |
367 | swizzle_pipe[0] = 0; | |
368 | swizzle_pipe[1] = 1; | |
369 | swizzle_pipe[2] = 2; | |
370 | swizzle_pipe[3] = 3; | |
371 | swizzle_pipe[4] = 4; | |
372 | } else { | |
373 | swizzle_pipe[0] = 0; | |
374 | swizzle_pipe[1] = 2; | |
375 | swizzle_pipe[2] = 4; | |
376 | swizzle_pipe[3] = 1; | |
377 | swizzle_pipe[4] = 3; | |
378 | } | |
3ce0a23d JG |
379 | break; |
380 | case 6: | |
d03f5d59 AD |
381 | if (force_no_swizzle) { |
382 | swizzle_pipe[0] = 0; | |
383 | swizzle_pipe[1] = 1; | |
384 | swizzle_pipe[2] = 2; | |
385 | swizzle_pipe[3] = 3; | |
386 | swizzle_pipe[4] = 4; | |
387 | swizzle_pipe[5] = 5; | |
388 | } else { | |
389 | swizzle_pipe[0] = 0; | |
390 | swizzle_pipe[1] = 2; | |
391 | swizzle_pipe[2] = 4; | |
392 | swizzle_pipe[3] = 5; | |
393 | swizzle_pipe[4] = 3; | |
394 | swizzle_pipe[5] = 1; | |
395 | } | |
3ce0a23d JG |
396 | break; |
397 | case 7: | |
d03f5d59 AD |
398 | if (force_no_swizzle) { |
399 | swizzle_pipe[0] = 0; | |
400 | swizzle_pipe[1] = 1; | |
401 | swizzle_pipe[2] = 2; | |
402 | swizzle_pipe[3] = 3; | |
403 | swizzle_pipe[4] = 4; | |
404 | swizzle_pipe[5] = 5; | |
405 | swizzle_pipe[6] = 6; | |
406 | } else { | |
407 | swizzle_pipe[0] = 0; | |
408 | swizzle_pipe[1] = 2; | |
409 | swizzle_pipe[2] = 4; | |
410 | swizzle_pipe[3] = 6; | |
411 | swizzle_pipe[4] = 3; | |
412 | swizzle_pipe[5] = 1; | |
413 | swizzle_pipe[6] = 5; | |
414 | } | |
3ce0a23d JG |
415 | break; |
416 | case 8: | |
d03f5d59 AD |
417 | if (force_no_swizzle) { |
418 | swizzle_pipe[0] = 0; | |
419 | swizzle_pipe[1] = 1; | |
420 | swizzle_pipe[2] = 2; | |
421 | swizzle_pipe[3] = 3; | |
422 | swizzle_pipe[4] = 4; | |
423 | swizzle_pipe[5] = 5; | |
424 | swizzle_pipe[6] = 6; | |
425 | swizzle_pipe[7] = 7; | |
426 | } else { | |
427 | swizzle_pipe[0] = 0; | |
428 | swizzle_pipe[1] = 2; | |
429 | swizzle_pipe[2] = 4; | |
430 | swizzle_pipe[3] = 6; | |
431 | swizzle_pipe[4] = 3; | |
432 | swizzle_pipe[5] = 1; | |
433 | swizzle_pipe[6] = 7; | |
434 | swizzle_pipe[7] = 5; | |
435 | } | |
3ce0a23d JG |
436 | break; |
437 | } | |
438 | ||
439 | cur_backend = 0; | |
440 | for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) { | |
441 | while (((1 << cur_backend) & enabled_backends_mask) == 0) | |
442 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
443 | ||
444 | backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2))); | |
445 | ||
446 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
447 | } | |
448 | ||
449 | return backend_map; | |
771fe6b9 JG |
450 | } |
451 | ||
3ce0a23d | 452 | static void rv770_gpu_init(struct radeon_device *rdev) |
771fe6b9 | 453 | { |
3ce0a23d | 454 | int i, j, num_qd_pipes; |
d03f5d59 | 455 | u32 ta_aux_cntl; |
3ce0a23d JG |
456 | u32 sx_debug_1; |
457 | u32 smx_dc_ctl0; | |
d03f5d59 | 458 | u32 db_debug3; |
3ce0a23d JG |
459 | u32 num_gs_verts_per_thread; |
460 | u32 vgt_gs_per_es; | |
461 | u32 gs_prim_buffer_depth = 0; | |
462 | u32 sq_ms_fifo_sizes; | |
463 | u32 sq_config; | |
464 | u32 sq_thread_resource_mgmt; | |
465 | u32 hdp_host_path_cntl; | |
466 | u32 sq_dyn_gpr_size_simd_ab_0; | |
467 | u32 backend_map; | |
468 | u32 gb_tiling_config = 0; | |
469 | u32 cc_rb_backend_disable = 0; | |
470 | u32 cc_gc_shader_pipe_config = 0; | |
471 | u32 mc_arb_ramcfg; | |
472 | u32 db_debug4; | |
771fe6b9 | 473 | |
3ce0a23d JG |
474 | /* setup chip specs */ |
475 | switch (rdev->family) { | |
476 | case CHIP_RV770: | |
477 | rdev->config.rv770.max_pipes = 4; | |
478 | rdev->config.rv770.max_tile_pipes = 8; | |
479 | rdev->config.rv770.max_simds = 10; | |
480 | rdev->config.rv770.max_backends = 4; | |
481 | rdev->config.rv770.max_gprs = 256; | |
482 | rdev->config.rv770.max_threads = 248; | |
483 | rdev->config.rv770.max_stack_entries = 512; | |
484 | rdev->config.rv770.max_hw_contexts = 8; | |
485 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
486 | rdev->config.rv770.sx_max_export_size = 128; | |
487 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
488 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
489 | rdev->config.rv770.sq_num_cf_insts = 2; | |
490 | ||
491 | rdev->config.rv770.sx_num_of_sets = 7; | |
492 | rdev->config.rv770.sc_prim_fifo_size = 0xF9; | |
493 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
494 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
495 | break; | |
496 | case CHIP_RV730: | |
497 | rdev->config.rv770.max_pipes = 2; | |
498 | rdev->config.rv770.max_tile_pipes = 4; | |
499 | rdev->config.rv770.max_simds = 8; | |
500 | rdev->config.rv770.max_backends = 2; | |
501 | rdev->config.rv770.max_gprs = 128; | |
502 | rdev->config.rv770.max_threads = 248; | |
503 | rdev->config.rv770.max_stack_entries = 256; | |
504 | rdev->config.rv770.max_hw_contexts = 8; | |
505 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
506 | rdev->config.rv770.sx_max_export_size = 256; | |
507 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
508 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
509 | rdev->config.rv770.sq_num_cf_insts = 2; | |
510 | ||
511 | rdev->config.rv770.sx_num_of_sets = 7; | |
512 | rdev->config.rv770.sc_prim_fifo_size = 0xf9; | |
513 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
514 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
515 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
516 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
517 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
518 | } | |
519 | break; | |
520 | case CHIP_RV710: | |
521 | rdev->config.rv770.max_pipes = 2; | |
522 | rdev->config.rv770.max_tile_pipes = 2; | |
523 | rdev->config.rv770.max_simds = 2; | |
524 | rdev->config.rv770.max_backends = 1; | |
525 | rdev->config.rv770.max_gprs = 256; | |
526 | rdev->config.rv770.max_threads = 192; | |
527 | rdev->config.rv770.max_stack_entries = 256; | |
528 | rdev->config.rv770.max_hw_contexts = 4; | |
529 | rdev->config.rv770.max_gs_threads = 8 * 2; | |
530 | rdev->config.rv770.sx_max_export_size = 128; | |
531 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
532 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
533 | rdev->config.rv770.sq_num_cf_insts = 1; | |
534 | ||
535 | rdev->config.rv770.sx_num_of_sets = 7; | |
536 | rdev->config.rv770.sc_prim_fifo_size = 0x40; | |
537 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
538 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
539 | break; | |
540 | case CHIP_RV740: | |
541 | rdev->config.rv770.max_pipes = 4; | |
542 | rdev->config.rv770.max_tile_pipes = 4; | |
543 | rdev->config.rv770.max_simds = 8; | |
544 | rdev->config.rv770.max_backends = 4; | |
545 | rdev->config.rv770.max_gprs = 256; | |
546 | rdev->config.rv770.max_threads = 248; | |
547 | rdev->config.rv770.max_stack_entries = 512; | |
548 | rdev->config.rv770.max_hw_contexts = 8; | |
549 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
550 | rdev->config.rv770.sx_max_export_size = 256; | |
551 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
552 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
553 | rdev->config.rv770.sq_num_cf_insts = 2; | |
554 | ||
555 | rdev->config.rv770.sx_num_of_sets = 7; | |
556 | rdev->config.rv770.sc_prim_fifo_size = 0x100; | |
557 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
558 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
559 | ||
560 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
561 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
562 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
563 | } | |
564 | break; | |
565 | default: | |
566 | break; | |
567 | } | |
568 | ||
569 | /* Initialize HDP */ | |
570 | j = 0; | |
571 | for (i = 0; i < 32; i++) { | |
572 | WREG32((0x2c14 + j), 0x00000000); | |
573 | WREG32((0x2c18 + j), 0x00000000); | |
574 | WREG32((0x2c1c + j), 0x00000000); | |
575 | WREG32((0x2c20 + j), 0x00000000); | |
576 | WREG32((0x2c24 + j), 0x00000000); | |
577 | j += 0x18; | |
578 | } | |
579 | ||
580 | WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff)); | |
581 | ||
582 | /* setup tiling, simd, pipe config */ | |
583 | mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG); | |
584 | ||
585 | switch (rdev->config.rv770.max_tile_pipes) { | |
586 | case 1: | |
d03f5d59 | 587 | default: |
3ce0a23d JG |
588 | gb_tiling_config |= PIPE_TILING(0); |
589 | break; | |
590 | case 2: | |
591 | gb_tiling_config |= PIPE_TILING(1); | |
592 | break; | |
593 | case 4: | |
594 | gb_tiling_config |= PIPE_TILING(2); | |
595 | break; | |
596 | case 8: | |
597 | gb_tiling_config |= PIPE_TILING(3); | |
3ce0a23d JG |
598 | break; |
599 | } | |
d03f5d59 | 600 | rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes; |
3ce0a23d JG |
601 | |
602 | if (rdev->family == CHIP_RV770) | |
603 | gb_tiling_config |= BANK_TILING(1); | |
604 | else | |
e29649db | 605 | gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT); |
961fb597 | 606 | rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3); |
3ce0a23d JG |
607 | |
608 | gb_tiling_config |= GROUP_SIZE(0); | |
961fb597 | 609 | rdev->config.rv770.tiling_group_size = 256; |
3ce0a23d | 610 | |
e29649db | 611 | if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) { |
3ce0a23d JG |
612 | gb_tiling_config |= ROW_TILING(3); |
613 | gb_tiling_config |= SAMPLE_SPLIT(3); | |
614 | } else { | |
615 | gb_tiling_config |= | |
616 | ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
617 | gb_tiling_config |= | |
618 | SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
619 | } | |
620 | ||
621 | gb_tiling_config |= BANK_SWAPS(1); | |
622 | ||
d03f5d59 AD |
623 | cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000; |
624 | cc_rb_backend_disable |= | |
625 | BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK); | |
3ce0a23d | 626 | |
d03f5d59 AD |
627 | cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00; |
628 | cc_gc_shader_pipe_config |= | |
3ce0a23d JG |
629 | INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK); |
630 | cc_gc_shader_pipe_config |= | |
631 | INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK); | |
632 | ||
d03f5d59 AD |
633 | if (rdev->family == CHIP_RV740) |
634 | backend_map = 0x28; | |
635 | else | |
636 | backend_map = r700_get_tile_pipe_to_backend_map(rdev, | |
637 | rdev->config.rv770.max_tile_pipes, | |
638 | (R7XX_MAX_BACKENDS - | |
639 | r600_count_pipe_bits((cc_rb_backend_disable & | |
640 | R7XX_MAX_BACKENDS_MASK) >> 16)), | |
641 | (cc_rb_backend_disable >> 16)); | |
642 | gb_tiling_config |= BACKEND_MAP(backend_map); | |
643 | ||
3ce0a23d JG |
644 | |
645 | WREG32(GB_TILING_CONFIG, gb_tiling_config); | |
646 | WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
647 | WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
648 | ||
649 | WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable); | |
650 | WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); | |
f867c60d | 651 | WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); |
d03f5d59 | 652 | WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable); |
3ce0a23d | 653 | |
3ce0a23d JG |
654 | WREG32(CGTS_SYS_TCC_DISABLE, 0); |
655 | WREG32(CGTS_TCC_DISABLE, 0); | |
f867c60d AD |
656 | WREG32(CGTS_USER_SYS_TCC_DISABLE, 0); |
657 | WREG32(CGTS_USER_TCC_DISABLE, 0); | |
3ce0a23d JG |
658 | |
659 | num_qd_pipes = | |
d03f5d59 | 660 | R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8); |
3ce0a23d JG |
661 | WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK); |
662 | WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK); | |
663 | ||
664 | /* set HW defaults for 3D engine */ | |
665 | WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) | | |
e29649db | 666 | ROQ_IB2_START(0x2b))); |
3ce0a23d JG |
667 | |
668 | WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30)); | |
669 | ||
d03f5d59 AD |
670 | ta_aux_cntl = RREG32(TA_CNTL_AUX); |
671 | WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO); | |
3ce0a23d JG |
672 | |
673 | sx_debug_1 = RREG32(SX_DEBUG_1); | |
674 | sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS; | |
675 | WREG32(SX_DEBUG_1, sx_debug_1); | |
676 | ||
677 | smx_dc_ctl0 = RREG32(SMX_DC_CTL0); | |
678 | smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff); | |
679 | smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1); | |
680 | WREG32(SMX_DC_CTL0, smx_dc_ctl0); | |
681 | ||
d03f5d59 AD |
682 | if (rdev->family != CHIP_RV740) |
683 | WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) | | |
684 | GS_FLUSH_CTL(4) | | |
685 | ACK_FLUSH_CTL(3) | | |
686 | SYNC_FLUSH_CTL)); | |
3ce0a23d | 687 | |
d03f5d59 AD |
688 | db_debug3 = RREG32(DB_DEBUG3); |
689 | db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f); | |
690 | switch (rdev->family) { | |
691 | case CHIP_RV770: | |
692 | case CHIP_RV740: | |
693 | db_debug3 |= DB_CLK_OFF_DELAY(0x1f); | |
694 | break; | |
695 | case CHIP_RV710: | |
696 | case CHIP_RV730: | |
697 | default: | |
698 | db_debug3 |= DB_CLK_OFF_DELAY(2); | |
699 | break; | |
700 | } | |
701 | WREG32(DB_DEBUG3, db_debug3); | |
702 | ||
703 | if (rdev->family != CHIP_RV770) { | |
3ce0a23d JG |
704 | db_debug4 = RREG32(DB_DEBUG4); |
705 | db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER; | |
706 | WREG32(DB_DEBUG4, db_debug4); | |
707 | } | |
708 | ||
709 | WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) | | |
e29649db AD |
710 | POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) | |
711 | SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1))); | |
3ce0a23d JG |
712 | |
713 | WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) | | |
e29649db AD |
714 | SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) | |
715 | SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize))); | |
3ce0a23d JG |
716 | |
717 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
718 | ||
719 | WREG32(VGT_NUM_INSTANCES, 1); | |
720 | ||
721 | WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0)); | |
722 | ||
723 | WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4)); | |
724 | ||
725 | WREG32(CP_PERFMON_CNTL, 0); | |
726 | ||
727 | sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) | | |
728 | DONE_FIFO_HIWATER(0xe0) | | |
729 | ALU_UPDATE_FIFO_HIWATER(0x8)); | |
730 | switch (rdev->family) { | |
731 | case CHIP_RV770: | |
3ce0a23d JG |
732 | case CHIP_RV730: |
733 | case CHIP_RV710: | |
d03f5d59 AD |
734 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1); |
735 | break; | |
3ce0a23d JG |
736 | case CHIP_RV740: |
737 | default: | |
738 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4); | |
739 | break; | |
740 | } | |
741 | WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes); | |
742 | ||
743 | /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT | |
744 | * should be adjusted as needed by the 2D/3D drivers. This just sets default values | |
745 | */ | |
746 | sq_config = RREG32(SQ_CONFIG); | |
747 | sq_config &= ~(PS_PRIO(3) | | |
748 | VS_PRIO(3) | | |
749 | GS_PRIO(3) | | |
750 | ES_PRIO(3)); | |
751 | sq_config |= (DX9_CONSTS | | |
752 | VC_ENABLE | | |
753 | EXPORT_SRC_C | | |
754 | PS_PRIO(0) | | |
755 | VS_PRIO(1) | | |
756 | GS_PRIO(2) | | |
757 | ES_PRIO(3)); | |
758 | if (rdev->family == CHIP_RV710) | |
759 | /* no vertex cache */ | |
760 | sq_config &= ~VC_ENABLE; | |
761 | ||
762 | WREG32(SQ_CONFIG, sq_config); | |
763 | ||
764 | WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | | |
fe62e1a4 DA |
765 | NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | |
766 | NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2))); | |
3ce0a23d JG |
767 | |
768 | WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) | | |
fe62e1a4 | 769 | NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64))); |
3ce0a23d JG |
770 | |
771 | sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) | | |
772 | NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) | | |
773 | NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8)); | |
774 | if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads) | |
775 | sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads); | |
776 | else | |
777 | sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8); | |
778 | WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt); | |
779 | ||
780 | WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
781 | NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
782 | ||
783 | WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
784 | NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
785 | ||
786 | sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
787 | SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) | | |
788 | SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
789 | SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64)); | |
790 | ||
791 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0); | |
792 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0); | |
793 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0); | |
794 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0); | |
795 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0); | |
796 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0); | |
797 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0); | |
798 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0); | |
799 | ||
800 | WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) | | |
fe62e1a4 | 801 | FORCE_EOV_MAX_REZ_CNT(255))); |
3ce0a23d JG |
802 | |
803 | if (rdev->family == CHIP_RV710) | |
804 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) | | |
fe62e1a4 | 805 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
806 | else |
807 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) | | |
fe62e1a4 | 808 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
809 | |
810 | switch (rdev->family) { | |
811 | case CHIP_RV770: | |
812 | case CHIP_RV730: | |
813 | case CHIP_RV740: | |
814 | gs_prim_buffer_depth = 384; | |
815 | break; | |
816 | case CHIP_RV710: | |
817 | gs_prim_buffer_depth = 128; | |
818 | break; | |
819 | default: | |
820 | break; | |
821 | } | |
822 | ||
823 | num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16; | |
824 | vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread; | |
825 | /* Max value for this is 256 */ | |
826 | if (vgt_gs_per_es > 256) | |
827 | vgt_gs_per_es = 256; | |
828 | ||
829 | WREG32(VGT_ES_PER_GS, 128); | |
830 | WREG32(VGT_GS_PER_ES, vgt_gs_per_es); | |
831 | WREG32(VGT_GS_PER_VS, 2); | |
832 | ||
833 | /* more default values. 2D/3D driver should adjust as needed */ | |
834 | WREG32(VGT_GS_VERTEX_REUSE, 16); | |
835 | WREG32(PA_SC_LINE_STIPPLE_STATE, 0); | |
836 | WREG32(VGT_STRMOUT_EN, 0); | |
837 | WREG32(SX_MISC, 0); | |
838 | WREG32(PA_SC_MODE_CNTL, 0); | |
839 | WREG32(PA_SC_EDGERULE, 0xaaaaaaaa); | |
840 | WREG32(PA_SC_AA_CONFIG, 0); | |
841 | WREG32(PA_SC_CLIPRECT_RULE, 0xffff); | |
842 | WREG32(PA_SC_LINE_STIPPLE, 0); | |
843 | WREG32(SPI_INPUT_Z, 0); | |
844 | WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2)); | |
845 | WREG32(CB_COLOR7_FRAG, 0); | |
846 | ||
847 | /* clear render buffer base addresses */ | |
848 | WREG32(CB_COLOR0_BASE, 0); | |
849 | WREG32(CB_COLOR1_BASE, 0); | |
850 | WREG32(CB_COLOR2_BASE, 0); | |
851 | WREG32(CB_COLOR3_BASE, 0); | |
852 | WREG32(CB_COLOR4_BASE, 0); | |
853 | WREG32(CB_COLOR5_BASE, 0); | |
854 | WREG32(CB_COLOR6_BASE, 0); | |
855 | WREG32(CB_COLOR7_BASE, 0); | |
856 | ||
857 | WREG32(TCP_CNTL, 0); | |
858 | ||
859 | hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL); | |
860 | WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl); | |
861 | ||
862 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
863 | ||
864 | WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA | | |
865 | NUM_CLIP_SEQ(3))); | |
866 | ||
867 | } | |
868 | ||
869 | int rv770_mc_init(struct radeon_device *rdev) | |
870 | { | |
3ce0a23d | 871 | u32 tmp; |
5885b7a9 | 872 | int chansize, numchan; |
3ce0a23d JG |
873 | |
874 | /* Get VRAM informations */ | |
3ce0a23d | 875 | rdev->mc.vram_is_ddr = true; |
5885b7a9 AD |
876 | tmp = RREG32(MC_ARB_RAMCFG); |
877 | if (tmp & CHANSIZE_OVERRIDE) { | |
878 | chansize = 16; | |
879 | } else if (tmp & CHANSIZE_MASK) { | |
880 | chansize = 64; | |
881 | } else { | |
882 | chansize = 32; | |
883 | } | |
884 | tmp = RREG32(MC_SHARED_CHMAP); | |
885 | switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) { | |
886 | case 0: | |
887 | default: | |
888 | numchan = 1; | |
889 | break; | |
890 | case 1: | |
891 | numchan = 2; | |
892 | break; | |
893 | case 2: | |
894 | numchan = 4; | |
895 | break; | |
896 | case 3: | |
897 | numchan = 8; | |
898 | break; | |
899 | } | |
900 | rdev->mc.vram_width = numchan * chansize; | |
771fe6b9 JG |
901 | /* Could aper size report 0 ? */ |
902 | rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0); | |
903 | rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0); | |
3ce0a23d JG |
904 | /* Setup GPU memory space */ |
905 | rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE); | |
906 | rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE); | |
51e5fcd3 | 907 | rdev->mc.visible_vram_size = rdev->mc.aper_size; |
d594e46a JG |
908 | /* FIXME remove this once we support unmappable VRAM */ |
909 | if (rdev->mc.mc_vram_size > rdev->mc.aper_size) { | |
974b16e3 | 910 | rdev->mc.mc_vram_size = rdev->mc.aper_size; |
974b16e3 | 911 | rdev->mc.real_vram_size = rdev->mc.aper_size; |
3ce0a23d | 912 | } |
d594e46a | 913 | r600_vram_gtt_location(rdev, &rdev->mc); |
f47299c5 AD |
914 | radeon_update_bandwidth_info(rdev); |
915 | ||
3ce0a23d JG |
916 | return 0; |
917 | } | |
d594e46a | 918 | |
fc30b8ef | 919 | static int rv770_startup(struct radeon_device *rdev) |
3ce0a23d JG |
920 | { |
921 | int r; | |
922 | ||
779720a3 AD |
923 | if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) { |
924 | r = r600_init_microcode(rdev); | |
925 | if (r) { | |
926 | DRM_ERROR("Failed to load firmware!\n"); | |
927 | return r; | |
928 | } | |
929 | } | |
930 | ||
a3c1945a | 931 | rv770_mc_program(rdev); |
1a029b76 JG |
932 | if (rdev->flags & RADEON_IS_AGP) { |
933 | rv770_agp_enable(rdev); | |
934 | } else { | |
935 | r = rv770_pcie_gart_enable(rdev); | |
936 | if (r) | |
937 | return r; | |
938 | } | |
3ce0a23d | 939 | rv770_gpu_init(rdev); |
c38c7b64 JG |
940 | r = r600_blit_init(rdev); |
941 | if (r) { | |
942 | r600_blit_fini(rdev); | |
943 | rdev->asic->copy = NULL; | |
944 | dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r); | |
945 | } | |
ff82f052 JG |
946 | /* pin copy shader into vram */ |
947 | if (rdev->r600_blit.shader_obj) { | |
948 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
949 | if (unlikely(r != 0)) | |
950 | return r; | |
951 | r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM, | |
952 | &rdev->r600_blit.shader_gpu_addr); | |
953 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
7923c615 | 954 | if (r) { |
ff82f052 | 955 | DRM_ERROR("failed to pin blit object %d\n", r); |
7923c615 AD |
956 | return r; |
957 | } | |
958 | } | |
d8f60cfc | 959 | /* Enable IRQ */ |
d8f60cfc AD |
960 | r = r600_irq_init(rdev); |
961 | if (r) { | |
962 | DRM_ERROR("radeon: IH init failed (%d).\n", r); | |
963 | radeon_irq_kms_fini(rdev); | |
964 | return r; | |
965 | } | |
966 | r600_irq_set(rdev); | |
967 | ||
3ce0a23d JG |
968 | r = radeon_ring_init(rdev, rdev->cp.ring_size); |
969 | if (r) | |
970 | return r; | |
971 | r = rv770_cp_load_microcode(rdev); | |
972 | if (r) | |
973 | return r; | |
974 | r = r600_cp_resume(rdev); | |
975 | if (r) | |
976 | return r; | |
81cc35bf JG |
977 | /* write back buffer are not vital so don't worry about failure */ |
978 | r600_wb_enable(rdev); | |
3ce0a23d JG |
979 | return 0; |
980 | } | |
981 | ||
fc30b8ef DA |
982 | int rv770_resume(struct radeon_device *rdev) |
983 | { | |
984 | int r; | |
985 | ||
1a029b76 JG |
986 | /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw, |
987 | * posting will perform necessary task to bring back GPU into good | |
988 | * shape. | |
989 | */ | |
fc30b8ef | 990 | /* post card */ |
e7d40b9a | 991 | atom_asic_init(rdev->mode_info.atom_context); |
fc30b8ef DA |
992 | /* Initialize clocks */ |
993 | r = radeon_clocks_init(rdev); | |
994 | if (r) { | |
995 | return r; | |
996 | } | |
997 | ||
998 | r = rv770_startup(rdev); | |
999 | if (r) { | |
1000 | DRM_ERROR("r600 startup failed on resume\n"); | |
1001 | return r; | |
1002 | } | |
1003 | ||
62a8ea3f | 1004 | r = r600_ib_test(rdev); |
fc30b8ef DA |
1005 | if (r) { |
1006 | DRM_ERROR("radeon: failled testing IB (%d).\n", r); | |
1007 | return r; | |
1008 | } | |
8a8c6e7c RM |
1009 | |
1010 | r = r600_audio_init(rdev); | |
1011 | if (r) { | |
1012 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1013 | return r; | |
1014 | } | |
1015 | ||
fc30b8ef DA |
1016 | return r; |
1017 | ||
1018 | } | |
1019 | ||
3ce0a23d JG |
1020 | int rv770_suspend(struct radeon_device *rdev) |
1021 | { | |
4c788679 JG |
1022 | int r; |
1023 | ||
8a8c6e7c | 1024 | r600_audio_fini(rdev); |
3ce0a23d JG |
1025 | /* FIXME: we should wait for ring to be empty */ |
1026 | r700_cp_stop(rdev); | |
4153e584 | 1027 | rdev->cp.ready = false; |
0c45249f | 1028 | r600_irq_suspend(rdev); |
81cc35bf | 1029 | r600_wb_disable(rdev); |
4aac0473 | 1030 | rv770_pcie_gart_disable(rdev); |
4153e584 | 1031 | /* unpin shaders bo */ |
30d2d9a5 JG |
1032 | if (rdev->r600_blit.shader_obj) { |
1033 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
1034 | if (likely(r == 0)) { | |
1035 | radeon_bo_unpin(rdev->r600_blit.shader_obj); | |
1036 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
1037 | } | |
4c788679 | 1038 | } |
3ce0a23d JG |
1039 | return 0; |
1040 | } | |
1041 | ||
1042 | /* Plan is to move initialization in that function and use | |
1043 | * helper function so that radeon_device_init pretty much | |
1044 | * do nothing more than calling asic specific function. This | |
1045 | * should also allow to remove a bunch of callback function | |
1046 | * like vram_info. | |
1047 | */ | |
1048 | int rv770_init(struct radeon_device *rdev) | |
1049 | { | |
1050 | int r; | |
1051 | ||
3ce0a23d JG |
1052 | r = radeon_dummy_page_init(rdev); |
1053 | if (r) | |
1054 | return r; | |
1055 | /* This don't do much */ | |
1056 | r = radeon_gem_init(rdev); | |
1057 | if (r) | |
1058 | return r; | |
1059 | /* Read BIOS */ | |
1060 | if (!radeon_get_bios(rdev)) { | |
1061 | if (ASIC_IS_AVIVO(rdev)) | |
1062 | return -EINVAL; | |
1063 | } | |
1064 | /* Must be an ATOMBIOS */ | |
e7d40b9a JG |
1065 | if (!rdev->is_atom_bios) { |
1066 | dev_err(rdev->dev, "Expecting atombios for R600 GPU\n"); | |
3ce0a23d | 1067 | return -EINVAL; |
e7d40b9a | 1068 | } |
3ce0a23d JG |
1069 | r = radeon_atombios_init(rdev); |
1070 | if (r) | |
1071 | return r; | |
1072 | /* Post card if necessary */ | |
72542d77 DA |
1073 | if (!r600_card_posted(rdev)) { |
1074 | if (!rdev->bios) { | |
1075 | dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n"); | |
1076 | return -EINVAL; | |
1077 | } | |
3ce0a23d JG |
1078 | DRM_INFO("GPU not posted. posting now...\n"); |
1079 | atom_asic_init(rdev->mode_info.atom_context); | |
1080 | } | |
1081 | /* Initialize scratch registers */ | |
1082 | r600_scratch_init(rdev); | |
1083 | /* Initialize surface registers */ | |
1084 | radeon_surface_init(rdev); | |
7433874e | 1085 | /* Initialize clocks */ |
5e6dde7e | 1086 | radeon_get_clock_info(rdev->ddev); |
3ce0a23d JG |
1087 | r = radeon_clocks_init(rdev); |
1088 | if (r) | |
1089 | return r; | |
7433874e RM |
1090 | /* Initialize power management */ |
1091 | radeon_pm_init(rdev); | |
3ce0a23d JG |
1092 | /* Fence driver */ |
1093 | r = radeon_fence_driver_init(rdev); | |
1094 | if (r) | |
1095 | return r; | |
d594e46a | 1096 | /* initialize AGP */ |
700a0cc0 JG |
1097 | if (rdev->flags & RADEON_IS_AGP) { |
1098 | r = radeon_agp_init(rdev); | |
1099 | if (r) | |
1100 | radeon_agp_disable(rdev); | |
1101 | } | |
3ce0a23d | 1102 | r = rv770_mc_init(rdev); |
b574f251 | 1103 | if (r) |
3ce0a23d | 1104 | return r; |
3ce0a23d | 1105 | /* Memory manager */ |
4c788679 | 1106 | r = radeon_bo_init(rdev); |
3ce0a23d JG |
1107 | if (r) |
1108 | return r; | |
d8f60cfc AD |
1109 | |
1110 | r = radeon_irq_kms_init(rdev); | |
1111 | if (r) | |
1112 | return r; | |
1113 | ||
3ce0a23d JG |
1114 | rdev->cp.ring_obj = NULL; |
1115 | r600_ring_init(rdev, 1024 * 1024); | |
1116 | ||
d8f60cfc AD |
1117 | rdev->ih.ring_obj = NULL; |
1118 | r600_ih_ring_init(rdev, 64 * 1024); | |
1119 | ||
4aac0473 JG |
1120 | r = r600_pcie_gart_init(rdev); |
1121 | if (r) | |
1122 | return r; | |
1123 | ||
779720a3 | 1124 | rdev->accel_working = true; |
fc30b8ef | 1125 | r = rv770_startup(rdev); |
3ce0a23d | 1126 | if (r) { |
655efd3d JG |
1127 | dev_err(rdev->dev, "disabling GPU acceleration\n"); |
1128 | r600_cp_fini(rdev); | |
75c81298 | 1129 | r600_wb_fini(rdev); |
655efd3d JG |
1130 | r600_irq_fini(rdev); |
1131 | radeon_irq_kms_fini(rdev); | |
75c81298 | 1132 | rv770_pcie_gart_fini(rdev); |
733289c2 | 1133 | rdev->accel_working = false; |
3ce0a23d | 1134 | } |
733289c2 | 1135 | if (rdev->accel_working) { |
733289c2 JG |
1136 | r = radeon_ib_pool_init(rdev); |
1137 | if (r) { | |
db96380e | 1138 | dev_err(rdev->dev, "IB initialization failed (%d).\n", r); |
733289c2 | 1139 | rdev->accel_working = false; |
db96380e JG |
1140 | } else { |
1141 | r = r600_ib_test(rdev); | |
1142 | if (r) { | |
1143 | dev_err(rdev->dev, "IB test failed (%d).\n", r); | |
1144 | rdev->accel_working = false; | |
1145 | } | |
733289c2 | 1146 | } |
3ce0a23d | 1147 | } |
8a8c6e7c RM |
1148 | |
1149 | r = r600_audio_init(rdev); | |
1150 | if (r) { | |
1151 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1152 | return r; | |
1153 | } | |
1154 | ||
3ce0a23d JG |
1155 | return 0; |
1156 | } | |
1157 | ||
1158 | void rv770_fini(struct radeon_device *rdev) | |
1159 | { | |
29fb52ca | 1160 | radeon_pm_fini(rdev); |
3ce0a23d | 1161 | r600_blit_fini(rdev); |
655efd3d JG |
1162 | r600_cp_fini(rdev); |
1163 | r600_wb_fini(rdev); | |
d8f60cfc AD |
1164 | r600_irq_fini(rdev); |
1165 | radeon_irq_kms_fini(rdev); | |
4aac0473 | 1166 | rv770_pcie_gart_fini(rdev); |
3ce0a23d JG |
1167 | radeon_gem_fini(rdev); |
1168 | radeon_fence_driver_fini(rdev); | |
1169 | radeon_clocks_fini(rdev); | |
d0269ed8 | 1170 | radeon_agp_fini(rdev); |
4c788679 | 1171 | radeon_bo_fini(rdev); |
e7d40b9a | 1172 | radeon_atombios_fini(rdev); |
3ce0a23d JG |
1173 | kfree(rdev->bios); |
1174 | rdev->bios = NULL; | |
1175 | radeon_dummy_page_fini(rdev); | |
771fe6b9 | 1176 | } |