]>
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> | |
5a0e3ad6 | 30 | #include <linux/slab.h> |
771fe6b9 | 31 | #include "drmP.h" |
771fe6b9 | 32 | #include "radeon.h" |
e6990375 | 33 | #include "radeon_asic.h" |
4153e584 | 34 | #include "radeon_drm.h" |
3ce0a23d | 35 | #include "rv770d.h" |
3ce0a23d | 36 | #include "atom.h" |
d39c3b89 | 37 | #include "avivod.h" |
771fe6b9 | 38 | |
3ce0a23d JG |
39 | #define R700_PFP_UCODE_SIZE 848 |
40 | #define R700_PM4_UCODE_SIZE 1360 | |
771fe6b9 | 41 | |
3ce0a23d JG |
42 | static void rv770_gpu_init(struct radeon_device *rdev); |
43 | void rv770_fini(struct radeon_device *rdev); | |
771fe6b9 | 44 | |
49e02b73 AD |
45 | void rv770_pm_misc(struct radeon_device *rdev) |
46 | { | |
47 | ||
48 | } | |
771fe6b9 JG |
49 | |
50 | /* | |
3ce0a23d | 51 | * GART |
771fe6b9 | 52 | */ |
3ce0a23d | 53 | int rv770_pcie_gart_enable(struct radeon_device *rdev) |
771fe6b9 | 54 | { |
3ce0a23d JG |
55 | u32 tmp; |
56 | int r, i; | |
771fe6b9 | 57 | |
4aac0473 JG |
58 | if (rdev->gart.table.vram.robj == NULL) { |
59 | dev_err(rdev->dev, "No VRAM object for PCIE GART.\n"); | |
60 | return -EINVAL; | |
3ce0a23d | 61 | } |
4aac0473 JG |
62 | r = radeon_gart_table_vram_pin(rdev); |
63 | if (r) | |
3ce0a23d | 64 | return r; |
82568565 | 65 | radeon_gart_restore(rdev); |
3ce0a23d JG |
66 | /* Setup L2 cache */ |
67 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
68 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
69 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
70 | WREG32(VM_L2_CNTL2, 0); | |
71 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
72 | /* Setup TLB control */ | |
73 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
74 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
75 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
76 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
77 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
78 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
79 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
80 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
81 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
82 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
83 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
84 | WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12); | |
1a029b76 | 85 | WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12); |
3ce0a23d JG |
86 | WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12); |
87 | WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) | | |
88 | RANGE_PROTECTION_FAULT_ENABLE_DEFAULT); | |
89 | WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, | |
90 | (u32)(rdev->dummy_page.addr >> 12)); | |
91 | for (i = 1; i < 7; i++) | |
92 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
771fe6b9 | 93 | |
3ce0a23d JG |
94 | r600_pcie_gart_tlb_flush(rdev); |
95 | rdev->gart.ready = true; | |
771fe6b9 JG |
96 | return 0; |
97 | } | |
98 | ||
3ce0a23d | 99 | void rv770_pcie_gart_disable(struct radeon_device *rdev) |
771fe6b9 | 100 | { |
3ce0a23d | 101 | u32 tmp; |
4c788679 | 102 | int i, r; |
3ce0a23d | 103 | |
3ce0a23d JG |
104 | /* Disable all tables */ |
105 | for (i = 0; i < 7; i++) | |
106 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
107 | ||
108 | /* Setup L2 cache */ | |
109 | WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING | | |
110 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
111 | WREG32(VM_L2_CNTL2, 0); | |
112 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
113 | /* Setup TLB control */ | |
114 | tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
115 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
116 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
117 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
118 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
119 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
120 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
121 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
4aac0473 | 122 | if (rdev->gart.table.vram.robj) { |
4c788679 JG |
123 | r = radeon_bo_reserve(rdev->gart.table.vram.robj, false); |
124 | if (likely(r == 0)) { | |
125 | radeon_bo_kunmap(rdev->gart.table.vram.robj); | |
126 | radeon_bo_unpin(rdev->gart.table.vram.robj); | |
127 | radeon_bo_unreserve(rdev->gart.table.vram.robj); | |
128 | } | |
4aac0473 JG |
129 | } |
130 | } | |
131 | ||
132 | void rv770_pcie_gart_fini(struct radeon_device *rdev) | |
133 | { | |
f9274562 | 134 | radeon_gart_fini(rdev); |
4aac0473 JG |
135 | rv770_pcie_gart_disable(rdev); |
136 | radeon_gart_table_vram_free(rdev); | |
771fe6b9 JG |
137 | } |
138 | ||
139 | ||
1a029b76 JG |
140 | void rv770_agp_enable(struct radeon_device *rdev) |
141 | { | |
142 | u32 tmp; | |
143 | int i; | |
144 | ||
145 | /* Setup L2 cache */ | |
146 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
147 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
148 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
149 | WREG32(VM_L2_CNTL2, 0); | |
150 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
151 | /* Setup TLB control */ | |
152 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
153 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
154 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
155 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
156 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
157 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
158 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
159 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
160 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
161 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
162 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
163 | for (i = 0; i < 7; i++) | |
164 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
165 | } | |
166 | ||
a3c1945a | 167 | static void rv770_mc_program(struct radeon_device *rdev) |
771fe6b9 | 168 | { |
a3c1945a | 169 | struct rv515_mc_save save; |
3ce0a23d JG |
170 | u32 tmp; |
171 | int i, j; | |
172 | ||
173 | /* Initialize HDP */ | |
174 | for (i = 0, j = 0; i < 32; i++, j += 0x18) { | |
175 | WREG32((0x2c14 + j), 0x00000000); | |
176 | WREG32((0x2c18 + j), 0x00000000); | |
177 | WREG32((0x2c1c + j), 0x00000000); | |
178 | WREG32((0x2c20 + j), 0x00000000); | |
179 | WREG32((0x2c24 + j), 0x00000000); | |
180 | } | |
181 | WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0); | |
182 | ||
a3c1945a | 183 | rv515_mc_stop(rdev, &save); |
3ce0a23d | 184 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 185 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 186 | } |
3ce0a23d JG |
187 | /* Lockout access through VGA aperture*/ |
188 | WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE); | |
3ce0a23d | 189 | /* Update configuration */ |
1a029b76 JG |
190 | if (rdev->flags & RADEON_IS_AGP) { |
191 | if (rdev->mc.vram_start < rdev->mc.gtt_start) { | |
192 | /* VRAM before AGP */ | |
193 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
194 | rdev->mc.vram_start >> 12); | |
195 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
196 | rdev->mc.gtt_end >> 12); | |
197 | } else { | |
198 | /* VRAM after AGP */ | |
199 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
200 | rdev->mc.gtt_start >> 12); | |
201 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
202 | rdev->mc.vram_end >> 12); | |
203 | } | |
204 | } else { | |
205 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
206 | rdev->mc.vram_start >> 12); | |
207 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
208 | rdev->mc.vram_end >> 12); | |
209 | } | |
3ce0a23d | 210 | WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0); |
1a029b76 | 211 | tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16; |
3ce0a23d JG |
212 | tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF); |
213 | WREG32(MC_VM_FB_LOCATION, tmp); | |
214 | WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8)); | |
215 | WREG32(HDP_NONSURFACE_INFO, (2 << 7)); | |
216 | WREG32(HDP_NONSURFACE_SIZE, (rdev->mc.mc_vram_size - 1) | 0x3FF); | |
217 | if (rdev->flags & RADEON_IS_AGP) { | |
1a029b76 | 218 | WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16); |
3ce0a23d JG |
219 | WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16); |
220 | WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22); | |
221 | } else { | |
222 | WREG32(MC_VM_AGP_BASE, 0); | |
223 | WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF); | |
224 | WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF); | |
225 | } | |
3ce0a23d | 226 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 227 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 228 | } |
a3c1945a | 229 | rv515_mc_resume(rdev, &save); |
698443d9 DA |
230 | /* we need to own VRAM, so turn off the VGA renderer here |
231 | * to stop it overwriting our objects */ | |
d39c3b89 | 232 | rv515_vga_render_disable(rdev); |
771fe6b9 JG |
233 | } |
234 | ||
3ce0a23d JG |
235 | |
236 | /* | |
237 | * CP. | |
238 | */ | |
239 | void r700_cp_stop(struct radeon_device *rdev) | |
771fe6b9 | 240 | { |
3ce0a23d | 241 | WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT)); |
771fe6b9 JG |
242 | } |
243 | ||
3ce0a23d | 244 | static int rv770_cp_load_microcode(struct radeon_device *rdev) |
771fe6b9 | 245 | { |
3ce0a23d JG |
246 | const __be32 *fw_data; |
247 | int i; | |
248 | ||
249 | if (!rdev->me_fw || !rdev->pfp_fw) | |
250 | return -EINVAL; | |
251 | ||
252 | r700_cp_stop(rdev); | |
253 | WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0)); | |
254 | ||
255 | /* Reset cp */ | |
256 | WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP); | |
257 | RREG32(GRBM_SOFT_RESET); | |
258 | mdelay(15); | |
259 | WREG32(GRBM_SOFT_RESET, 0); | |
260 | ||
261 | fw_data = (const __be32 *)rdev->pfp_fw->data; | |
262 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
263 | for (i = 0; i < R700_PFP_UCODE_SIZE; i++) | |
264 | WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++)); | |
265 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
266 | ||
267 | fw_data = (const __be32 *)rdev->me_fw->data; | |
268 | WREG32(CP_ME_RAM_WADDR, 0); | |
269 | for (i = 0; i < R700_PM4_UCODE_SIZE; i++) | |
270 | WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++)); | |
271 | ||
272 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
273 | WREG32(CP_ME_RAM_WADDR, 0); | |
274 | WREG32(CP_ME_RAM_RADDR, 0); | |
275 | return 0; | |
771fe6b9 JG |
276 | } |
277 | ||
fe251e2f AD |
278 | void r700_cp_fini(struct radeon_device *rdev) |
279 | { | |
280 | r700_cp_stop(rdev); | |
281 | radeon_ring_fini(rdev); | |
282 | } | |
771fe6b9 JG |
283 | |
284 | /* | |
3ce0a23d | 285 | * Core functions |
771fe6b9 | 286 | */ |
d03f5d59 AD |
287 | static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev, |
288 | u32 num_tile_pipes, | |
289 | u32 num_backends, | |
290 | u32 backend_disable_mask) | |
771fe6b9 | 291 | { |
3ce0a23d JG |
292 | u32 backend_map = 0; |
293 | u32 enabled_backends_mask; | |
294 | u32 enabled_backends_count; | |
295 | u32 cur_pipe; | |
296 | u32 swizzle_pipe[R7XX_MAX_PIPES]; | |
297 | u32 cur_backend; | |
298 | u32 i; | |
d03f5d59 | 299 | bool force_no_swizzle; |
3ce0a23d JG |
300 | |
301 | if (num_tile_pipes > R7XX_MAX_PIPES) | |
302 | num_tile_pipes = R7XX_MAX_PIPES; | |
303 | if (num_tile_pipes < 1) | |
304 | num_tile_pipes = 1; | |
305 | if (num_backends > R7XX_MAX_BACKENDS) | |
306 | num_backends = R7XX_MAX_BACKENDS; | |
307 | if (num_backends < 1) | |
308 | num_backends = 1; | |
309 | ||
310 | enabled_backends_mask = 0; | |
311 | enabled_backends_count = 0; | |
312 | for (i = 0; i < R7XX_MAX_BACKENDS; ++i) { | |
313 | if (((backend_disable_mask >> i) & 1) == 0) { | |
314 | enabled_backends_mask |= (1 << i); | |
315 | ++enabled_backends_count; | |
316 | } | |
317 | if (enabled_backends_count == num_backends) | |
318 | break; | |
319 | } | |
320 | ||
321 | if (enabled_backends_count == 0) { | |
322 | enabled_backends_mask = 1; | |
323 | enabled_backends_count = 1; | |
324 | } | |
325 | ||
326 | if (enabled_backends_count != num_backends) | |
327 | num_backends = enabled_backends_count; | |
328 | ||
d03f5d59 AD |
329 | switch (rdev->family) { |
330 | case CHIP_RV770: | |
331 | case CHIP_RV730: | |
332 | force_no_swizzle = false; | |
333 | break; | |
334 | case CHIP_RV710: | |
335 | case CHIP_RV740: | |
336 | default: | |
337 | force_no_swizzle = true; | |
338 | break; | |
339 | } | |
340 | ||
3ce0a23d JG |
341 | memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES); |
342 | switch (num_tile_pipes) { | |
343 | case 1: | |
344 | swizzle_pipe[0] = 0; | |
345 | break; | |
346 | case 2: | |
347 | swizzle_pipe[0] = 0; | |
348 | swizzle_pipe[1] = 1; | |
349 | break; | |
350 | case 3: | |
d03f5d59 AD |
351 | if (force_no_swizzle) { |
352 | swizzle_pipe[0] = 0; | |
353 | swizzle_pipe[1] = 1; | |
354 | swizzle_pipe[2] = 2; | |
355 | } else { | |
356 | swizzle_pipe[0] = 0; | |
357 | swizzle_pipe[1] = 2; | |
358 | swizzle_pipe[2] = 1; | |
359 | } | |
3ce0a23d JG |
360 | break; |
361 | case 4: | |
d03f5d59 AD |
362 | if (force_no_swizzle) { |
363 | swizzle_pipe[0] = 0; | |
364 | swizzle_pipe[1] = 1; | |
365 | swizzle_pipe[2] = 2; | |
366 | swizzle_pipe[3] = 3; | |
367 | } else { | |
368 | swizzle_pipe[0] = 0; | |
369 | swizzle_pipe[1] = 2; | |
370 | swizzle_pipe[2] = 3; | |
371 | swizzle_pipe[3] = 1; | |
372 | } | |
3ce0a23d JG |
373 | break; |
374 | case 5: | |
d03f5d59 AD |
375 | if (force_no_swizzle) { |
376 | swizzle_pipe[0] = 0; | |
377 | swizzle_pipe[1] = 1; | |
378 | swizzle_pipe[2] = 2; | |
379 | swizzle_pipe[3] = 3; | |
380 | swizzle_pipe[4] = 4; | |
381 | } else { | |
382 | swizzle_pipe[0] = 0; | |
383 | swizzle_pipe[1] = 2; | |
384 | swizzle_pipe[2] = 4; | |
385 | swizzle_pipe[3] = 1; | |
386 | swizzle_pipe[4] = 3; | |
387 | } | |
3ce0a23d JG |
388 | break; |
389 | case 6: | |
d03f5d59 AD |
390 | if (force_no_swizzle) { |
391 | swizzle_pipe[0] = 0; | |
392 | swizzle_pipe[1] = 1; | |
393 | swizzle_pipe[2] = 2; | |
394 | swizzle_pipe[3] = 3; | |
395 | swizzle_pipe[4] = 4; | |
396 | swizzle_pipe[5] = 5; | |
397 | } else { | |
398 | swizzle_pipe[0] = 0; | |
399 | swizzle_pipe[1] = 2; | |
400 | swizzle_pipe[2] = 4; | |
401 | swizzle_pipe[3] = 5; | |
402 | swizzle_pipe[4] = 3; | |
403 | swizzle_pipe[5] = 1; | |
404 | } | |
3ce0a23d JG |
405 | break; |
406 | case 7: | |
d03f5d59 AD |
407 | if (force_no_swizzle) { |
408 | swizzle_pipe[0] = 0; | |
409 | swizzle_pipe[1] = 1; | |
410 | swizzle_pipe[2] = 2; | |
411 | swizzle_pipe[3] = 3; | |
412 | swizzle_pipe[4] = 4; | |
413 | swizzle_pipe[5] = 5; | |
414 | swizzle_pipe[6] = 6; | |
415 | } else { | |
416 | swizzle_pipe[0] = 0; | |
417 | swizzle_pipe[1] = 2; | |
418 | swizzle_pipe[2] = 4; | |
419 | swizzle_pipe[3] = 6; | |
420 | swizzle_pipe[4] = 3; | |
421 | swizzle_pipe[5] = 1; | |
422 | swizzle_pipe[6] = 5; | |
423 | } | |
3ce0a23d JG |
424 | break; |
425 | case 8: | |
d03f5d59 AD |
426 | if (force_no_swizzle) { |
427 | swizzle_pipe[0] = 0; | |
428 | swizzle_pipe[1] = 1; | |
429 | swizzle_pipe[2] = 2; | |
430 | swizzle_pipe[3] = 3; | |
431 | swizzle_pipe[4] = 4; | |
432 | swizzle_pipe[5] = 5; | |
433 | swizzle_pipe[6] = 6; | |
434 | swizzle_pipe[7] = 7; | |
435 | } else { | |
436 | swizzle_pipe[0] = 0; | |
437 | swizzle_pipe[1] = 2; | |
438 | swizzle_pipe[2] = 4; | |
439 | swizzle_pipe[3] = 6; | |
440 | swizzle_pipe[4] = 3; | |
441 | swizzle_pipe[5] = 1; | |
442 | swizzle_pipe[6] = 7; | |
443 | swizzle_pipe[7] = 5; | |
444 | } | |
3ce0a23d JG |
445 | break; |
446 | } | |
447 | ||
448 | cur_backend = 0; | |
449 | for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) { | |
450 | while (((1 << cur_backend) & enabled_backends_mask) == 0) | |
451 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
452 | ||
453 | backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2))); | |
454 | ||
455 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
456 | } | |
457 | ||
458 | return backend_map; | |
771fe6b9 JG |
459 | } |
460 | ||
3ce0a23d | 461 | static void rv770_gpu_init(struct radeon_device *rdev) |
771fe6b9 | 462 | { |
3ce0a23d | 463 | int i, j, num_qd_pipes; |
d03f5d59 | 464 | u32 ta_aux_cntl; |
3ce0a23d JG |
465 | u32 sx_debug_1; |
466 | u32 smx_dc_ctl0; | |
d03f5d59 | 467 | u32 db_debug3; |
3ce0a23d JG |
468 | u32 num_gs_verts_per_thread; |
469 | u32 vgt_gs_per_es; | |
470 | u32 gs_prim_buffer_depth = 0; | |
471 | u32 sq_ms_fifo_sizes; | |
472 | u32 sq_config; | |
473 | u32 sq_thread_resource_mgmt; | |
474 | u32 hdp_host_path_cntl; | |
475 | u32 sq_dyn_gpr_size_simd_ab_0; | |
476 | u32 backend_map; | |
477 | u32 gb_tiling_config = 0; | |
478 | u32 cc_rb_backend_disable = 0; | |
479 | u32 cc_gc_shader_pipe_config = 0; | |
480 | u32 mc_arb_ramcfg; | |
481 | u32 db_debug4; | |
771fe6b9 | 482 | |
3ce0a23d JG |
483 | /* setup chip specs */ |
484 | switch (rdev->family) { | |
485 | case CHIP_RV770: | |
486 | rdev->config.rv770.max_pipes = 4; | |
487 | rdev->config.rv770.max_tile_pipes = 8; | |
488 | rdev->config.rv770.max_simds = 10; | |
489 | rdev->config.rv770.max_backends = 4; | |
490 | rdev->config.rv770.max_gprs = 256; | |
491 | rdev->config.rv770.max_threads = 248; | |
492 | rdev->config.rv770.max_stack_entries = 512; | |
493 | rdev->config.rv770.max_hw_contexts = 8; | |
494 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
495 | rdev->config.rv770.sx_max_export_size = 128; | |
496 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
497 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
498 | rdev->config.rv770.sq_num_cf_insts = 2; | |
499 | ||
500 | rdev->config.rv770.sx_num_of_sets = 7; | |
501 | rdev->config.rv770.sc_prim_fifo_size = 0xF9; | |
502 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
503 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
504 | break; | |
505 | case CHIP_RV730: | |
506 | rdev->config.rv770.max_pipes = 2; | |
507 | rdev->config.rv770.max_tile_pipes = 4; | |
508 | rdev->config.rv770.max_simds = 8; | |
509 | rdev->config.rv770.max_backends = 2; | |
510 | rdev->config.rv770.max_gprs = 128; | |
511 | rdev->config.rv770.max_threads = 248; | |
512 | rdev->config.rv770.max_stack_entries = 256; | |
513 | rdev->config.rv770.max_hw_contexts = 8; | |
514 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
515 | rdev->config.rv770.sx_max_export_size = 256; | |
516 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
517 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
518 | rdev->config.rv770.sq_num_cf_insts = 2; | |
519 | ||
520 | rdev->config.rv770.sx_num_of_sets = 7; | |
521 | rdev->config.rv770.sc_prim_fifo_size = 0xf9; | |
522 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
523 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
524 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
525 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
526 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
527 | } | |
528 | break; | |
529 | case CHIP_RV710: | |
530 | rdev->config.rv770.max_pipes = 2; | |
531 | rdev->config.rv770.max_tile_pipes = 2; | |
532 | rdev->config.rv770.max_simds = 2; | |
533 | rdev->config.rv770.max_backends = 1; | |
534 | rdev->config.rv770.max_gprs = 256; | |
535 | rdev->config.rv770.max_threads = 192; | |
536 | rdev->config.rv770.max_stack_entries = 256; | |
537 | rdev->config.rv770.max_hw_contexts = 4; | |
538 | rdev->config.rv770.max_gs_threads = 8 * 2; | |
539 | rdev->config.rv770.sx_max_export_size = 128; | |
540 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
541 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
542 | rdev->config.rv770.sq_num_cf_insts = 1; | |
543 | ||
544 | rdev->config.rv770.sx_num_of_sets = 7; | |
545 | rdev->config.rv770.sc_prim_fifo_size = 0x40; | |
546 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
547 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
548 | break; | |
549 | case CHIP_RV740: | |
550 | rdev->config.rv770.max_pipes = 4; | |
551 | rdev->config.rv770.max_tile_pipes = 4; | |
552 | rdev->config.rv770.max_simds = 8; | |
553 | rdev->config.rv770.max_backends = 4; | |
554 | rdev->config.rv770.max_gprs = 256; | |
555 | rdev->config.rv770.max_threads = 248; | |
556 | rdev->config.rv770.max_stack_entries = 512; | |
557 | rdev->config.rv770.max_hw_contexts = 8; | |
558 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
559 | rdev->config.rv770.sx_max_export_size = 256; | |
560 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
561 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
562 | rdev->config.rv770.sq_num_cf_insts = 2; | |
563 | ||
564 | rdev->config.rv770.sx_num_of_sets = 7; | |
565 | rdev->config.rv770.sc_prim_fifo_size = 0x100; | |
566 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
567 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
568 | ||
569 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
570 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
571 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
572 | } | |
573 | break; | |
574 | default: | |
575 | break; | |
576 | } | |
577 | ||
578 | /* Initialize HDP */ | |
579 | j = 0; | |
580 | for (i = 0; i < 32; i++) { | |
581 | WREG32((0x2c14 + j), 0x00000000); | |
582 | WREG32((0x2c18 + j), 0x00000000); | |
583 | WREG32((0x2c1c + j), 0x00000000); | |
584 | WREG32((0x2c20 + j), 0x00000000); | |
585 | WREG32((0x2c24 + j), 0x00000000); | |
586 | j += 0x18; | |
587 | } | |
588 | ||
589 | WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff)); | |
590 | ||
591 | /* setup tiling, simd, pipe config */ | |
592 | mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG); | |
593 | ||
594 | switch (rdev->config.rv770.max_tile_pipes) { | |
595 | case 1: | |
d03f5d59 | 596 | default: |
3ce0a23d JG |
597 | gb_tiling_config |= PIPE_TILING(0); |
598 | break; | |
599 | case 2: | |
600 | gb_tiling_config |= PIPE_TILING(1); | |
601 | break; | |
602 | case 4: | |
603 | gb_tiling_config |= PIPE_TILING(2); | |
604 | break; | |
605 | case 8: | |
606 | gb_tiling_config |= PIPE_TILING(3); | |
3ce0a23d JG |
607 | break; |
608 | } | |
d03f5d59 | 609 | rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes; |
3ce0a23d JG |
610 | |
611 | if (rdev->family == CHIP_RV770) | |
612 | gb_tiling_config |= BANK_TILING(1); | |
613 | else | |
e29649db | 614 | gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT); |
961fb597 | 615 | rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3); |
3ce0a23d JG |
616 | |
617 | gb_tiling_config |= GROUP_SIZE(0); | |
961fb597 | 618 | rdev->config.rv770.tiling_group_size = 256; |
3ce0a23d | 619 | |
e29649db | 620 | if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) { |
3ce0a23d JG |
621 | gb_tiling_config |= ROW_TILING(3); |
622 | gb_tiling_config |= SAMPLE_SPLIT(3); | |
623 | } else { | |
624 | gb_tiling_config |= | |
625 | ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
626 | gb_tiling_config |= | |
627 | SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
628 | } | |
629 | ||
630 | gb_tiling_config |= BANK_SWAPS(1); | |
631 | ||
d03f5d59 AD |
632 | cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000; |
633 | cc_rb_backend_disable |= | |
634 | BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK); | |
3ce0a23d | 635 | |
d03f5d59 AD |
636 | cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00; |
637 | cc_gc_shader_pipe_config |= | |
3ce0a23d JG |
638 | INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK); |
639 | cc_gc_shader_pipe_config |= | |
640 | INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK); | |
641 | ||
d03f5d59 AD |
642 | if (rdev->family == CHIP_RV740) |
643 | backend_map = 0x28; | |
644 | else | |
645 | backend_map = r700_get_tile_pipe_to_backend_map(rdev, | |
646 | rdev->config.rv770.max_tile_pipes, | |
647 | (R7XX_MAX_BACKENDS - | |
648 | r600_count_pipe_bits((cc_rb_backend_disable & | |
649 | R7XX_MAX_BACKENDS_MASK) >> 16)), | |
650 | (cc_rb_backend_disable >> 16)); | |
651 | gb_tiling_config |= BACKEND_MAP(backend_map); | |
652 | ||
3ce0a23d JG |
653 | |
654 | WREG32(GB_TILING_CONFIG, gb_tiling_config); | |
655 | WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
656 | WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
657 | ||
658 | WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable); | |
659 | WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); | |
f867c60d | 660 | WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); |
d03f5d59 | 661 | WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable); |
3ce0a23d | 662 | |
3ce0a23d JG |
663 | WREG32(CGTS_SYS_TCC_DISABLE, 0); |
664 | WREG32(CGTS_TCC_DISABLE, 0); | |
f867c60d AD |
665 | WREG32(CGTS_USER_SYS_TCC_DISABLE, 0); |
666 | WREG32(CGTS_USER_TCC_DISABLE, 0); | |
3ce0a23d JG |
667 | |
668 | num_qd_pipes = | |
d03f5d59 | 669 | R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8); |
3ce0a23d JG |
670 | WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK); |
671 | WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK); | |
672 | ||
673 | /* set HW defaults for 3D engine */ | |
674 | WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) | | |
e29649db | 675 | ROQ_IB2_START(0x2b))); |
3ce0a23d JG |
676 | |
677 | WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30)); | |
678 | ||
d03f5d59 AD |
679 | ta_aux_cntl = RREG32(TA_CNTL_AUX); |
680 | WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO); | |
3ce0a23d JG |
681 | |
682 | sx_debug_1 = RREG32(SX_DEBUG_1); | |
683 | sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS; | |
684 | WREG32(SX_DEBUG_1, sx_debug_1); | |
685 | ||
686 | smx_dc_ctl0 = RREG32(SMX_DC_CTL0); | |
687 | smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff); | |
688 | smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1); | |
689 | WREG32(SMX_DC_CTL0, smx_dc_ctl0); | |
690 | ||
d03f5d59 AD |
691 | if (rdev->family != CHIP_RV740) |
692 | WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) | | |
693 | GS_FLUSH_CTL(4) | | |
694 | ACK_FLUSH_CTL(3) | | |
695 | SYNC_FLUSH_CTL)); | |
3ce0a23d | 696 | |
d03f5d59 AD |
697 | db_debug3 = RREG32(DB_DEBUG3); |
698 | db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f); | |
699 | switch (rdev->family) { | |
700 | case CHIP_RV770: | |
701 | case CHIP_RV740: | |
702 | db_debug3 |= DB_CLK_OFF_DELAY(0x1f); | |
703 | break; | |
704 | case CHIP_RV710: | |
705 | case CHIP_RV730: | |
706 | default: | |
707 | db_debug3 |= DB_CLK_OFF_DELAY(2); | |
708 | break; | |
709 | } | |
710 | WREG32(DB_DEBUG3, db_debug3); | |
711 | ||
712 | if (rdev->family != CHIP_RV770) { | |
3ce0a23d JG |
713 | db_debug4 = RREG32(DB_DEBUG4); |
714 | db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER; | |
715 | WREG32(DB_DEBUG4, db_debug4); | |
716 | } | |
717 | ||
718 | WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) | | |
e29649db AD |
719 | POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) | |
720 | SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1))); | |
3ce0a23d JG |
721 | |
722 | WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) | | |
e29649db AD |
723 | SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) | |
724 | SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize))); | |
3ce0a23d JG |
725 | |
726 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
727 | ||
728 | WREG32(VGT_NUM_INSTANCES, 1); | |
729 | ||
730 | WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0)); | |
731 | ||
732 | WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4)); | |
733 | ||
734 | WREG32(CP_PERFMON_CNTL, 0); | |
735 | ||
736 | sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) | | |
737 | DONE_FIFO_HIWATER(0xe0) | | |
738 | ALU_UPDATE_FIFO_HIWATER(0x8)); | |
739 | switch (rdev->family) { | |
740 | case CHIP_RV770: | |
3ce0a23d JG |
741 | case CHIP_RV730: |
742 | case CHIP_RV710: | |
d03f5d59 AD |
743 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1); |
744 | break; | |
3ce0a23d JG |
745 | case CHIP_RV740: |
746 | default: | |
747 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4); | |
748 | break; | |
749 | } | |
750 | WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes); | |
751 | ||
752 | /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT | |
753 | * should be adjusted as needed by the 2D/3D drivers. This just sets default values | |
754 | */ | |
755 | sq_config = RREG32(SQ_CONFIG); | |
756 | sq_config &= ~(PS_PRIO(3) | | |
757 | VS_PRIO(3) | | |
758 | GS_PRIO(3) | | |
759 | ES_PRIO(3)); | |
760 | sq_config |= (DX9_CONSTS | | |
761 | VC_ENABLE | | |
762 | EXPORT_SRC_C | | |
763 | PS_PRIO(0) | | |
764 | VS_PRIO(1) | | |
765 | GS_PRIO(2) | | |
766 | ES_PRIO(3)); | |
767 | if (rdev->family == CHIP_RV710) | |
768 | /* no vertex cache */ | |
769 | sq_config &= ~VC_ENABLE; | |
770 | ||
771 | WREG32(SQ_CONFIG, sq_config); | |
772 | ||
773 | WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | | |
fe62e1a4 DA |
774 | NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | |
775 | NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2))); | |
3ce0a23d JG |
776 | |
777 | WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) | | |
fe62e1a4 | 778 | NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64))); |
3ce0a23d JG |
779 | |
780 | sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) | | |
781 | NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) | | |
782 | NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8)); | |
783 | if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads) | |
784 | sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads); | |
785 | else | |
786 | sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8); | |
787 | WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt); | |
788 | ||
789 | WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
790 | NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
791 | ||
792 | WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
793 | NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
794 | ||
795 | sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
796 | SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) | | |
797 | SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
798 | SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64)); | |
799 | ||
800 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0); | |
801 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0); | |
802 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0); | |
803 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0); | |
804 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0); | |
805 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0); | |
806 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0); | |
807 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0); | |
808 | ||
809 | WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) | | |
fe62e1a4 | 810 | FORCE_EOV_MAX_REZ_CNT(255))); |
3ce0a23d JG |
811 | |
812 | if (rdev->family == CHIP_RV710) | |
813 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) | | |
fe62e1a4 | 814 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
815 | else |
816 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) | | |
fe62e1a4 | 817 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
818 | |
819 | switch (rdev->family) { | |
820 | case CHIP_RV770: | |
821 | case CHIP_RV730: | |
822 | case CHIP_RV740: | |
823 | gs_prim_buffer_depth = 384; | |
824 | break; | |
825 | case CHIP_RV710: | |
826 | gs_prim_buffer_depth = 128; | |
827 | break; | |
828 | default: | |
829 | break; | |
830 | } | |
831 | ||
832 | num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16; | |
833 | vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread; | |
834 | /* Max value for this is 256 */ | |
835 | if (vgt_gs_per_es > 256) | |
836 | vgt_gs_per_es = 256; | |
837 | ||
838 | WREG32(VGT_ES_PER_GS, 128); | |
839 | WREG32(VGT_GS_PER_ES, vgt_gs_per_es); | |
840 | WREG32(VGT_GS_PER_VS, 2); | |
841 | ||
842 | /* more default values. 2D/3D driver should adjust as needed */ | |
843 | WREG32(VGT_GS_VERTEX_REUSE, 16); | |
844 | WREG32(PA_SC_LINE_STIPPLE_STATE, 0); | |
845 | WREG32(VGT_STRMOUT_EN, 0); | |
846 | WREG32(SX_MISC, 0); | |
847 | WREG32(PA_SC_MODE_CNTL, 0); | |
848 | WREG32(PA_SC_EDGERULE, 0xaaaaaaaa); | |
849 | WREG32(PA_SC_AA_CONFIG, 0); | |
850 | WREG32(PA_SC_CLIPRECT_RULE, 0xffff); | |
851 | WREG32(PA_SC_LINE_STIPPLE, 0); | |
852 | WREG32(SPI_INPUT_Z, 0); | |
853 | WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2)); | |
854 | WREG32(CB_COLOR7_FRAG, 0); | |
855 | ||
856 | /* clear render buffer base addresses */ | |
857 | WREG32(CB_COLOR0_BASE, 0); | |
858 | WREG32(CB_COLOR1_BASE, 0); | |
859 | WREG32(CB_COLOR2_BASE, 0); | |
860 | WREG32(CB_COLOR3_BASE, 0); | |
861 | WREG32(CB_COLOR4_BASE, 0); | |
862 | WREG32(CB_COLOR5_BASE, 0); | |
863 | WREG32(CB_COLOR6_BASE, 0); | |
864 | WREG32(CB_COLOR7_BASE, 0); | |
865 | ||
866 | WREG32(TCP_CNTL, 0); | |
867 | ||
868 | hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL); | |
869 | WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl); | |
870 | ||
871 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
872 | ||
873 | WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA | | |
874 | NUM_CLIP_SEQ(3))); | |
875 | ||
876 | } | |
877 | ||
878 | int rv770_mc_init(struct radeon_device *rdev) | |
879 | { | |
3ce0a23d | 880 | u32 tmp; |
5885b7a9 | 881 | int chansize, numchan; |
3ce0a23d JG |
882 | |
883 | /* Get VRAM informations */ | |
3ce0a23d | 884 | rdev->mc.vram_is_ddr = true; |
5885b7a9 AD |
885 | tmp = RREG32(MC_ARB_RAMCFG); |
886 | if (tmp & CHANSIZE_OVERRIDE) { | |
887 | chansize = 16; | |
888 | } else if (tmp & CHANSIZE_MASK) { | |
889 | chansize = 64; | |
890 | } else { | |
891 | chansize = 32; | |
892 | } | |
893 | tmp = RREG32(MC_SHARED_CHMAP); | |
894 | switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) { | |
895 | case 0: | |
896 | default: | |
897 | numchan = 1; | |
898 | break; | |
899 | case 1: | |
900 | numchan = 2; | |
901 | break; | |
902 | case 2: | |
903 | numchan = 4; | |
904 | break; | |
905 | case 3: | |
906 | numchan = 8; | |
907 | break; | |
908 | } | |
909 | rdev->mc.vram_width = numchan * chansize; | |
771fe6b9 JG |
910 | /* Could aper size report 0 ? */ |
911 | rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0); | |
912 | rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0); | |
3ce0a23d JG |
913 | /* Setup GPU memory space */ |
914 | rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE); | |
915 | rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE); | |
51e5fcd3 | 916 | rdev->mc.visible_vram_size = rdev->mc.aper_size; |
d594e46a | 917 | r600_vram_gtt_location(rdev, &rdev->mc); |
f47299c5 AD |
918 | radeon_update_bandwidth_info(rdev); |
919 | ||
3ce0a23d JG |
920 | return 0; |
921 | } | |
d594e46a | 922 | |
fc30b8ef | 923 | static int rv770_startup(struct radeon_device *rdev) |
3ce0a23d JG |
924 | { |
925 | int r; | |
926 | ||
779720a3 AD |
927 | if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) { |
928 | r = r600_init_microcode(rdev); | |
929 | if (r) { | |
930 | DRM_ERROR("Failed to load firmware!\n"); | |
931 | return r; | |
932 | } | |
933 | } | |
934 | ||
a3c1945a | 935 | rv770_mc_program(rdev); |
1a029b76 JG |
936 | if (rdev->flags & RADEON_IS_AGP) { |
937 | rv770_agp_enable(rdev); | |
938 | } else { | |
939 | r = rv770_pcie_gart_enable(rdev); | |
940 | if (r) | |
941 | return r; | |
942 | } | |
3ce0a23d | 943 | rv770_gpu_init(rdev); |
c38c7b64 JG |
944 | r = r600_blit_init(rdev); |
945 | if (r) { | |
946 | r600_blit_fini(rdev); | |
947 | rdev->asic->copy = NULL; | |
948 | dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r); | |
949 | } | |
ff82f052 JG |
950 | /* pin copy shader into vram */ |
951 | if (rdev->r600_blit.shader_obj) { | |
952 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
953 | if (unlikely(r != 0)) | |
954 | return r; | |
955 | r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM, | |
956 | &rdev->r600_blit.shader_gpu_addr); | |
957 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
7923c615 | 958 | if (r) { |
ff82f052 | 959 | DRM_ERROR("failed to pin blit object %d\n", r); |
7923c615 AD |
960 | return r; |
961 | } | |
962 | } | |
d8f60cfc | 963 | /* Enable IRQ */ |
d8f60cfc AD |
964 | r = r600_irq_init(rdev); |
965 | if (r) { | |
966 | DRM_ERROR("radeon: IH init failed (%d).\n", r); | |
967 | radeon_irq_kms_fini(rdev); | |
968 | return r; | |
969 | } | |
970 | r600_irq_set(rdev); | |
971 | ||
3ce0a23d JG |
972 | r = radeon_ring_init(rdev, rdev->cp.ring_size); |
973 | if (r) | |
974 | return r; | |
975 | r = rv770_cp_load_microcode(rdev); | |
976 | if (r) | |
977 | return r; | |
978 | r = r600_cp_resume(rdev); | |
979 | if (r) | |
980 | return r; | |
81cc35bf JG |
981 | /* write back buffer are not vital so don't worry about failure */ |
982 | r600_wb_enable(rdev); | |
3ce0a23d JG |
983 | return 0; |
984 | } | |
985 | ||
fc30b8ef DA |
986 | int rv770_resume(struct radeon_device *rdev) |
987 | { | |
988 | int r; | |
989 | ||
1a029b76 JG |
990 | /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw, |
991 | * posting will perform necessary task to bring back GPU into good | |
992 | * shape. | |
993 | */ | |
fc30b8ef | 994 | /* post card */ |
e7d40b9a | 995 | atom_asic_init(rdev->mode_info.atom_context); |
fc30b8ef DA |
996 | /* Initialize clocks */ |
997 | r = radeon_clocks_init(rdev); | |
998 | if (r) { | |
999 | return r; | |
1000 | } | |
1001 | ||
1002 | r = rv770_startup(rdev); | |
1003 | if (r) { | |
1004 | DRM_ERROR("r600 startup failed on resume\n"); | |
1005 | return r; | |
1006 | } | |
1007 | ||
62a8ea3f | 1008 | r = r600_ib_test(rdev); |
fc30b8ef DA |
1009 | if (r) { |
1010 | DRM_ERROR("radeon: failled testing IB (%d).\n", r); | |
1011 | return r; | |
1012 | } | |
8a8c6e7c RM |
1013 | |
1014 | r = r600_audio_init(rdev); | |
1015 | if (r) { | |
1016 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1017 | return r; | |
1018 | } | |
1019 | ||
fc30b8ef DA |
1020 | return r; |
1021 | ||
1022 | } | |
1023 | ||
3ce0a23d JG |
1024 | int rv770_suspend(struct radeon_device *rdev) |
1025 | { | |
4c788679 JG |
1026 | int r; |
1027 | ||
8a8c6e7c | 1028 | r600_audio_fini(rdev); |
3ce0a23d JG |
1029 | /* FIXME: we should wait for ring to be empty */ |
1030 | r700_cp_stop(rdev); | |
4153e584 | 1031 | rdev->cp.ready = false; |
0c45249f | 1032 | r600_irq_suspend(rdev); |
81cc35bf | 1033 | r600_wb_disable(rdev); |
4aac0473 | 1034 | rv770_pcie_gart_disable(rdev); |
4153e584 | 1035 | /* unpin shaders bo */ |
30d2d9a5 JG |
1036 | if (rdev->r600_blit.shader_obj) { |
1037 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
1038 | if (likely(r == 0)) { | |
1039 | radeon_bo_unpin(rdev->r600_blit.shader_obj); | |
1040 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
1041 | } | |
4c788679 | 1042 | } |
3ce0a23d JG |
1043 | return 0; |
1044 | } | |
1045 | ||
1046 | /* Plan is to move initialization in that function and use | |
1047 | * helper function so that radeon_device_init pretty much | |
1048 | * do nothing more than calling asic specific function. This | |
1049 | * should also allow to remove a bunch of callback function | |
1050 | * like vram_info. | |
1051 | */ | |
1052 | int rv770_init(struct radeon_device *rdev) | |
1053 | { | |
1054 | int r; | |
1055 | ||
3ce0a23d JG |
1056 | r = radeon_dummy_page_init(rdev); |
1057 | if (r) | |
1058 | return r; | |
1059 | /* This don't do much */ | |
1060 | r = radeon_gem_init(rdev); | |
1061 | if (r) | |
1062 | return r; | |
1063 | /* Read BIOS */ | |
1064 | if (!radeon_get_bios(rdev)) { | |
1065 | if (ASIC_IS_AVIVO(rdev)) | |
1066 | return -EINVAL; | |
1067 | } | |
1068 | /* Must be an ATOMBIOS */ | |
e7d40b9a JG |
1069 | if (!rdev->is_atom_bios) { |
1070 | dev_err(rdev->dev, "Expecting atombios for R600 GPU\n"); | |
3ce0a23d | 1071 | return -EINVAL; |
e7d40b9a | 1072 | } |
3ce0a23d JG |
1073 | r = radeon_atombios_init(rdev); |
1074 | if (r) | |
1075 | return r; | |
1076 | /* Post card if necessary */ | |
72542d77 DA |
1077 | if (!r600_card_posted(rdev)) { |
1078 | if (!rdev->bios) { | |
1079 | dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n"); | |
1080 | return -EINVAL; | |
1081 | } | |
3ce0a23d JG |
1082 | DRM_INFO("GPU not posted. posting now...\n"); |
1083 | atom_asic_init(rdev->mode_info.atom_context); | |
1084 | } | |
1085 | /* Initialize scratch registers */ | |
1086 | r600_scratch_init(rdev); | |
1087 | /* Initialize surface registers */ | |
1088 | radeon_surface_init(rdev); | |
7433874e | 1089 | /* Initialize clocks */ |
5e6dde7e | 1090 | radeon_get_clock_info(rdev->ddev); |
3ce0a23d JG |
1091 | r = radeon_clocks_init(rdev); |
1092 | if (r) | |
1093 | return r; | |
1094 | /* Fence driver */ | |
1095 | r = radeon_fence_driver_init(rdev); | |
1096 | if (r) | |
1097 | return r; | |
d594e46a | 1098 | /* initialize AGP */ |
700a0cc0 JG |
1099 | if (rdev->flags & RADEON_IS_AGP) { |
1100 | r = radeon_agp_init(rdev); | |
1101 | if (r) | |
1102 | radeon_agp_disable(rdev); | |
1103 | } | |
3ce0a23d | 1104 | r = rv770_mc_init(rdev); |
b574f251 | 1105 | if (r) |
3ce0a23d | 1106 | return r; |
3ce0a23d | 1107 | /* Memory manager */ |
4c788679 | 1108 | r = radeon_bo_init(rdev); |
3ce0a23d JG |
1109 | if (r) |
1110 | return r; | |
d8f60cfc AD |
1111 | |
1112 | r = radeon_irq_kms_init(rdev); | |
1113 | if (r) | |
1114 | return r; | |
1115 | ||
3ce0a23d JG |
1116 | rdev->cp.ring_obj = NULL; |
1117 | r600_ring_init(rdev, 1024 * 1024); | |
1118 | ||
d8f60cfc AD |
1119 | rdev->ih.ring_obj = NULL; |
1120 | r600_ih_ring_init(rdev, 64 * 1024); | |
1121 | ||
4aac0473 JG |
1122 | r = r600_pcie_gart_init(rdev); |
1123 | if (r) | |
1124 | return r; | |
1125 | ||
779720a3 | 1126 | rdev->accel_working = true; |
fc30b8ef | 1127 | r = rv770_startup(rdev); |
3ce0a23d | 1128 | if (r) { |
655efd3d | 1129 | dev_err(rdev->dev, "disabling GPU acceleration\n"); |
fe251e2f | 1130 | r700_cp_fini(rdev); |
75c81298 | 1131 | r600_wb_fini(rdev); |
655efd3d JG |
1132 | r600_irq_fini(rdev); |
1133 | radeon_irq_kms_fini(rdev); | |
75c81298 | 1134 | rv770_pcie_gart_fini(rdev); |
733289c2 | 1135 | rdev->accel_working = false; |
3ce0a23d | 1136 | } |
733289c2 | 1137 | if (rdev->accel_working) { |
733289c2 JG |
1138 | r = radeon_ib_pool_init(rdev); |
1139 | if (r) { | |
db96380e | 1140 | dev_err(rdev->dev, "IB initialization failed (%d).\n", r); |
733289c2 | 1141 | rdev->accel_working = false; |
db96380e JG |
1142 | } else { |
1143 | r = r600_ib_test(rdev); | |
1144 | if (r) { | |
1145 | dev_err(rdev->dev, "IB test failed (%d).\n", r); | |
1146 | rdev->accel_working = false; | |
1147 | } | |
733289c2 | 1148 | } |
3ce0a23d | 1149 | } |
8a8c6e7c RM |
1150 | |
1151 | r = r600_audio_init(rdev); | |
1152 | if (r) { | |
1153 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1154 | return r; | |
1155 | } | |
1156 | ||
3ce0a23d JG |
1157 | return 0; |
1158 | } | |
1159 | ||
1160 | void rv770_fini(struct radeon_device *rdev) | |
1161 | { | |
1162 | r600_blit_fini(rdev); | |
fe251e2f | 1163 | r700_cp_fini(rdev); |
655efd3d | 1164 | r600_wb_fini(rdev); |
d8f60cfc AD |
1165 | r600_irq_fini(rdev); |
1166 | radeon_irq_kms_fini(rdev); | |
4aac0473 | 1167 | rv770_pcie_gart_fini(rdev); |
3ce0a23d JG |
1168 | radeon_gem_fini(rdev); |
1169 | radeon_fence_driver_fini(rdev); | |
1170 | radeon_clocks_fini(rdev); | |
d0269ed8 | 1171 | radeon_agp_fini(rdev); |
4c788679 | 1172 | radeon_bo_fini(rdev); |
e7d40b9a | 1173 | radeon_atombios_fini(rdev); |
3ce0a23d JG |
1174 | kfree(rdev->bios); |
1175 | rdev->bios = NULL; | |
1176 | radeon_dummy_page_fini(rdev); | |
771fe6b9 | 1177 | } |