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 | { | |
9349d5cc AD |
47 | int requested_index = rdev->pm.requested_power_state_index; |
48 | struct radeon_power_state *ps = &rdev->pm.power_state[requested_index]; | |
49 | struct radeon_voltage *voltage = &ps->clock_info[0].voltage; | |
49e02b73 | 50 | |
9349d5cc AD |
51 | if ((voltage->type == VOLTAGE_SW) && voltage->voltage) |
52 | radeon_atom_set_voltage(rdev, voltage->voltage); | |
49e02b73 | 53 | } |
771fe6b9 JG |
54 | |
55 | /* | |
3ce0a23d | 56 | * GART |
771fe6b9 | 57 | */ |
3ce0a23d | 58 | int rv770_pcie_gart_enable(struct radeon_device *rdev) |
771fe6b9 | 59 | { |
3ce0a23d JG |
60 | u32 tmp; |
61 | int r, i; | |
771fe6b9 | 62 | |
4aac0473 JG |
63 | if (rdev->gart.table.vram.robj == NULL) { |
64 | dev_err(rdev->dev, "No VRAM object for PCIE GART.\n"); | |
65 | return -EINVAL; | |
3ce0a23d | 66 | } |
4aac0473 JG |
67 | r = radeon_gart_table_vram_pin(rdev); |
68 | if (r) | |
3ce0a23d | 69 | return r; |
82568565 | 70 | radeon_gart_restore(rdev); |
3ce0a23d JG |
71 | /* Setup L2 cache */ |
72 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
73 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
74 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
75 | WREG32(VM_L2_CNTL2, 0); | |
76 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
77 | /* Setup TLB control */ | |
78 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
79 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
80 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
81 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
82 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
83 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
84 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
85 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
86 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
87 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
88 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
89 | WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12); | |
1a029b76 | 90 | WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12); |
3ce0a23d JG |
91 | WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12); |
92 | WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) | | |
93 | RANGE_PROTECTION_FAULT_ENABLE_DEFAULT); | |
94 | WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, | |
95 | (u32)(rdev->dummy_page.addr >> 12)); | |
96 | for (i = 1; i < 7; i++) | |
97 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
771fe6b9 | 98 | |
3ce0a23d JG |
99 | r600_pcie_gart_tlb_flush(rdev); |
100 | rdev->gart.ready = true; | |
771fe6b9 JG |
101 | return 0; |
102 | } | |
103 | ||
3ce0a23d | 104 | void rv770_pcie_gart_disable(struct radeon_device *rdev) |
771fe6b9 | 105 | { |
3ce0a23d | 106 | u32 tmp; |
4c788679 | 107 | int i, r; |
3ce0a23d | 108 | |
3ce0a23d JG |
109 | /* Disable all tables */ |
110 | for (i = 0; i < 7; i++) | |
111 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
112 | ||
113 | /* Setup L2 cache */ | |
114 | WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING | | |
115 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
116 | WREG32(VM_L2_CNTL2, 0); | |
117 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
118 | /* Setup TLB control */ | |
119 | tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
120 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
121 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
122 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
123 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
124 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
125 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
126 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
4aac0473 | 127 | if (rdev->gart.table.vram.robj) { |
4c788679 JG |
128 | r = radeon_bo_reserve(rdev->gart.table.vram.robj, false); |
129 | if (likely(r == 0)) { | |
130 | radeon_bo_kunmap(rdev->gart.table.vram.robj); | |
131 | radeon_bo_unpin(rdev->gart.table.vram.robj); | |
132 | radeon_bo_unreserve(rdev->gart.table.vram.robj); | |
133 | } | |
4aac0473 JG |
134 | } |
135 | } | |
136 | ||
137 | void rv770_pcie_gart_fini(struct radeon_device *rdev) | |
138 | { | |
f9274562 | 139 | radeon_gart_fini(rdev); |
4aac0473 JG |
140 | rv770_pcie_gart_disable(rdev); |
141 | radeon_gart_table_vram_free(rdev); | |
771fe6b9 JG |
142 | } |
143 | ||
144 | ||
1a029b76 JG |
145 | void rv770_agp_enable(struct radeon_device *rdev) |
146 | { | |
147 | u32 tmp; | |
148 | int i; | |
149 | ||
150 | /* Setup L2 cache */ | |
151 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
152 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
153 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
154 | WREG32(VM_L2_CNTL2, 0); | |
155 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
156 | /* Setup TLB control */ | |
157 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
158 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
159 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
160 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
161 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
162 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
163 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
164 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
165 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
166 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
167 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
168 | for (i = 0; i < 7; i++) | |
169 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
170 | } | |
171 | ||
a3c1945a | 172 | static void rv770_mc_program(struct radeon_device *rdev) |
771fe6b9 | 173 | { |
a3c1945a | 174 | struct rv515_mc_save save; |
3ce0a23d JG |
175 | u32 tmp; |
176 | int i, j; | |
177 | ||
178 | /* Initialize HDP */ | |
179 | for (i = 0, j = 0; i < 32; i++, j += 0x18) { | |
180 | WREG32((0x2c14 + j), 0x00000000); | |
181 | WREG32((0x2c18 + j), 0x00000000); | |
182 | WREG32((0x2c1c + j), 0x00000000); | |
183 | WREG32((0x2c20 + j), 0x00000000); | |
184 | WREG32((0x2c24 + j), 0x00000000); | |
185 | } | |
186 | WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0); | |
187 | ||
a3c1945a | 188 | rv515_mc_stop(rdev, &save); |
3ce0a23d | 189 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 190 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 191 | } |
3ce0a23d JG |
192 | /* Lockout access through VGA aperture*/ |
193 | WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE); | |
3ce0a23d | 194 | /* Update configuration */ |
1a029b76 JG |
195 | if (rdev->flags & RADEON_IS_AGP) { |
196 | if (rdev->mc.vram_start < rdev->mc.gtt_start) { | |
197 | /* VRAM before AGP */ | |
198 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
199 | rdev->mc.vram_start >> 12); | |
200 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
201 | rdev->mc.gtt_end >> 12); | |
202 | } else { | |
203 | /* VRAM after AGP */ | |
204 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
205 | rdev->mc.gtt_start >> 12); | |
206 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
207 | rdev->mc.vram_end >> 12); | |
208 | } | |
209 | } else { | |
210 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
211 | rdev->mc.vram_start >> 12); | |
212 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
213 | rdev->mc.vram_end >> 12); | |
214 | } | |
3ce0a23d | 215 | WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0); |
1a029b76 | 216 | tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16; |
3ce0a23d JG |
217 | tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF); |
218 | WREG32(MC_VM_FB_LOCATION, tmp); | |
219 | WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8)); | |
220 | WREG32(HDP_NONSURFACE_INFO, (2 << 7)); | |
221 | WREG32(HDP_NONSURFACE_SIZE, (rdev->mc.mc_vram_size - 1) | 0x3FF); | |
222 | if (rdev->flags & RADEON_IS_AGP) { | |
1a029b76 | 223 | WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16); |
3ce0a23d JG |
224 | WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16); |
225 | WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22); | |
226 | } else { | |
227 | WREG32(MC_VM_AGP_BASE, 0); | |
228 | WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF); | |
229 | WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF); | |
230 | } | |
3ce0a23d | 231 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 232 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 233 | } |
a3c1945a | 234 | rv515_mc_resume(rdev, &save); |
698443d9 DA |
235 | /* we need to own VRAM, so turn off the VGA renderer here |
236 | * to stop it overwriting our objects */ | |
d39c3b89 | 237 | rv515_vga_render_disable(rdev); |
771fe6b9 JG |
238 | } |
239 | ||
3ce0a23d JG |
240 | |
241 | /* | |
242 | * CP. | |
243 | */ | |
244 | void r700_cp_stop(struct radeon_device *rdev) | |
771fe6b9 | 245 | { |
3ce0a23d | 246 | WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT)); |
771fe6b9 JG |
247 | } |
248 | ||
3ce0a23d | 249 | static int rv770_cp_load_microcode(struct radeon_device *rdev) |
771fe6b9 | 250 | { |
3ce0a23d JG |
251 | const __be32 *fw_data; |
252 | int i; | |
253 | ||
254 | if (!rdev->me_fw || !rdev->pfp_fw) | |
255 | return -EINVAL; | |
256 | ||
257 | r700_cp_stop(rdev); | |
258 | WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0)); | |
259 | ||
260 | /* Reset cp */ | |
261 | WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP); | |
262 | RREG32(GRBM_SOFT_RESET); | |
263 | mdelay(15); | |
264 | WREG32(GRBM_SOFT_RESET, 0); | |
265 | ||
266 | fw_data = (const __be32 *)rdev->pfp_fw->data; | |
267 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
268 | for (i = 0; i < R700_PFP_UCODE_SIZE; i++) | |
269 | WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++)); | |
270 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
271 | ||
272 | fw_data = (const __be32 *)rdev->me_fw->data; | |
273 | WREG32(CP_ME_RAM_WADDR, 0); | |
274 | for (i = 0; i < R700_PM4_UCODE_SIZE; i++) | |
275 | WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++)); | |
276 | ||
277 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
278 | WREG32(CP_ME_RAM_WADDR, 0); | |
279 | WREG32(CP_ME_RAM_RADDR, 0); | |
280 | return 0; | |
771fe6b9 JG |
281 | } |
282 | ||
fe251e2f AD |
283 | void r700_cp_fini(struct radeon_device *rdev) |
284 | { | |
285 | r700_cp_stop(rdev); | |
286 | radeon_ring_fini(rdev); | |
287 | } | |
771fe6b9 JG |
288 | |
289 | /* | |
3ce0a23d | 290 | * Core functions |
771fe6b9 | 291 | */ |
d03f5d59 AD |
292 | static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev, |
293 | u32 num_tile_pipes, | |
294 | u32 num_backends, | |
295 | u32 backend_disable_mask) | |
771fe6b9 | 296 | { |
3ce0a23d JG |
297 | u32 backend_map = 0; |
298 | u32 enabled_backends_mask; | |
299 | u32 enabled_backends_count; | |
300 | u32 cur_pipe; | |
301 | u32 swizzle_pipe[R7XX_MAX_PIPES]; | |
302 | u32 cur_backend; | |
303 | u32 i; | |
d03f5d59 | 304 | bool force_no_swizzle; |
3ce0a23d JG |
305 | |
306 | if (num_tile_pipes > R7XX_MAX_PIPES) | |
307 | num_tile_pipes = R7XX_MAX_PIPES; | |
308 | if (num_tile_pipes < 1) | |
309 | num_tile_pipes = 1; | |
310 | if (num_backends > R7XX_MAX_BACKENDS) | |
311 | num_backends = R7XX_MAX_BACKENDS; | |
312 | if (num_backends < 1) | |
313 | num_backends = 1; | |
314 | ||
315 | enabled_backends_mask = 0; | |
316 | enabled_backends_count = 0; | |
317 | for (i = 0; i < R7XX_MAX_BACKENDS; ++i) { | |
318 | if (((backend_disable_mask >> i) & 1) == 0) { | |
319 | enabled_backends_mask |= (1 << i); | |
320 | ++enabled_backends_count; | |
321 | } | |
322 | if (enabled_backends_count == num_backends) | |
323 | break; | |
324 | } | |
325 | ||
326 | if (enabled_backends_count == 0) { | |
327 | enabled_backends_mask = 1; | |
328 | enabled_backends_count = 1; | |
329 | } | |
330 | ||
331 | if (enabled_backends_count != num_backends) | |
332 | num_backends = enabled_backends_count; | |
333 | ||
d03f5d59 AD |
334 | switch (rdev->family) { |
335 | case CHIP_RV770: | |
336 | case CHIP_RV730: | |
337 | force_no_swizzle = false; | |
338 | break; | |
339 | case CHIP_RV710: | |
340 | case CHIP_RV740: | |
341 | default: | |
342 | force_no_swizzle = true; | |
343 | break; | |
344 | } | |
345 | ||
3ce0a23d JG |
346 | memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES); |
347 | switch (num_tile_pipes) { | |
348 | case 1: | |
349 | swizzle_pipe[0] = 0; | |
350 | break; | |
351 | case 2: | |
352 | swizzle_pipe[0] = 0; | |
353 | swizzle_pipe[1] = 1; | |
354 | break; | |
355 | case 3: | |
d03f5d59 AD |
356 | if (force_no_swizzle) { |
357 | swizzle_pipe[0] = 0; | |
358 | swizzle_pipe[1] = 1; | |
359 | swizzle_pipe[2] = 2; | |
360 | } else { | |
361 | swizzle_pipe[0] = 0; | |
362 | swizzle_pipe[1] = 2; | |
363 | swizzle_pipe[2] = 1; | |
364 | } | |
3ce0a23d JG |
365 | break; |
366 | case 4: | |
d03f5d59 AD |
367 | if (force_no_swizzle) { |
368 | swizzle_pipe[0] = 0; | |
369 | swizzle_pipe[1] = 1; | |
370 | swizzle_pipe[2] = 2; | |
371 | swizzle_pipe[3] = 3; | |
372 | } else { | |
373 | swizzle_pipe[0] = 0; | |
374 | swizzle_pipe[1] = 2; | |
375 | swizzle_pipe[2] = 3; | |
376 | swizzle_pipe[3] = 1; | |
377 | } | |
3ce0a23d JG |
378 | break; |
379 | case 5: | |
d03f5d59 AD |
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 | } else { | |
387 | swizzle_pipe[0] = 0; | |
388 | swizzle_pipe[1] = 2; | |
389 | swizzle_pipe[2] = 4; | |
390 | swizzle_pipe[3] = 1; | |
391 | swizzle_pipe[4] = 3; | |
392 | } | |
3ce0a23d JG |
393 | break; |
394 | case 6: | |
d03f5d59 AD |
395 | if (force_no_swizzle) { |
396 | swizzle_pipe[0] = 0; | |
397 | swizzle_pipe[1] = 1; | |
398 | swizzle_pipe[2] = 2; | |
399 | swizzle_pipe[3] = 3; | |
400 | swizzle_pipe[4] = 4; | |
401 | swizzle_pipe[5] = 5; | |
402 | } else { | |
403 | swizzle_pipe[0] = 0; | |
404 | swizzle_pipe[1] = 2; | |
405 | swizzle_pipe[2] = 4; | |
406 | swizzle_pipe[3] = 5; | |
407 | swizzle_pipe[4] = 3; | |
408 | swizzle_pipe[5] = 1; | |
409 | } | |
3ce0a23d JG |
410 | break; |
411 | case 7: | |
d03f5d59 AD |
412 | if (force_no_swizzle) { |
413 | swizzle_pipe[0] = 0; | |
414 | swizzle_pipe[1] = 1; | |
415 | swizzle_pipe[2] = 2; | |
416 | swizzle_pipe[3] = 3; | |
417 | swizzle_pipe[4] = 4; | |
418 | swizzle_pipe[5] = 5; | |
419 | swizzle_pipe[6] = 6; | |
420 | } else { | |
421 | swizzle_pipe[0] = 0; | |
422 | swizzle_pipe[1] = 2; | |
423 | swizzle_pipe[2] = 4; | |
424 | swizzle_pipe[3] = 6; | |
425 | swizzle_pipe[4] = 3; | |
426 | swizzle_pipe[5] = 1; | |
427 | swizzle_pipe[6] = 5; | |
428 | } | |
3ce0a23d JG |
429 | break; |
430 | case 8: | |
d03f5d59 AD |
431 | if (force_no_swizzle) { |
432 | swizzle_pipe[0] = 0; | |
433 | swizzle_pipe[1] = 1; | |
434 | swizzle_pipe[2] = 2; | |
435 | swizzle_pipe[3] = 3; | |
436 | swizzle_pipe[4] = 4; | |
437 | swizzle_pipe[5] = 5; | |
438 | swizzle_pipe[6] = 6; | |
439 | swizzle_pipe[7] = 7; | |
440 | } else { | |
441 | swizzle_pipe[0] = 0; | |
442 | swizzle_pipe[1] = 2; | |
443 | swizzle_pipe[2] = 4; | |
444 | swizzle_pipe[3] = 6; | |
445 | swizzle_pipe[4] = 3; | |
446 | swizzle_pipe[5] = 1; | |
447 | swizzle_pipe[6] = 7; | |
448 | swizzle_pipe[7] = 5; | |
449 | } | |
3ce0a23d JG |
450 | break; |
451 | } | |
452 | ||
453 | cur_backend = 0; | |
454 | for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) { | |
455 | while (((1 << cur_backend) & enabled_backends_mask) == 0) | |
456 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
457 | ||
458 | backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2))); | |
459 | ||
460 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
461 | } | |
462 | ||
463 | return backend_map; | |
771fe6b9 JG |
464 | } |
465 | ||
3ce0a23d | 466 | static void rv770_gpu_init(struct radeon_device *rdev) |
771fe6b9 | 467 | { |
3ce0a23d | 468 | int i, j, num_qd_pipes; |
d03f5d59 | 469 | u32 ta_aux_cntl; |
3ce0a23d JG |
470 | u32 sx_debug_1; |
471 | u32 smx_dc_ctl0; | |
d03f5d59 | 472 | u32 db_debug3; |
3ce0a23d JG |
473 | u32 num_gs_verts_per_thread; |
474 | u32 vgt_gs_per_es; | |
475 | u32 gs_prim_buffer_depth = 0; | |
476 | u32 sq_ms_fifo_sizes; | |
477 | u32 sq_config; | |
478 | u32 sq_thread_resource_mgmt; | |
479 | u32 hdp_host_path_cntl; | |
480 | u32 sq_dyn_gpr_size_simd_ab_0; | |
481 | u32 backend_map; | |
482 | u32 gb_tiling_config = 0; | |
483 | u32 cc_rb_backend_disable = 0; | |
484 | u32 cc_gc_shader_pipe_config = 0; | |
485 | u32 mc_arb_ramcfg; | |
486 | u32 db_debug4; | |
771fe6b9 | 487 | |
3ce0a23d JG |
488 | /* setup chip specs */ |
489 | switch (rdev->family) { | |
490 | case CHIP_RV770: | |
491 | rdev->config.rv770.max_pipes = 4; | |
492 | rdev->config.rv770.max_tile_pipes = 8; | |
493 | rdev->config.rv770.max_simds = 10; | |
494 | rdev->config.rv770.max_backends = 4; | |
495 | rdev->config.rv770.max_gprs = 256; | |
496 | rdev->config.rv770.max_threads = 248; | |
497 | rdev->config.rv770.max_stack_entries = 512; | |
498 | rdev->config.rv770.max_hw_contexts = 8; | |
499 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
500 | rdev->config.rv770.sx_max_export_size = 128; | |
501 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
502 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
503 | rdev->config.rv770.sq_num_cf_insts = 2; | |
504 | ||
505 | rdev->config.rv770.sx_num_of_sets = 7; | |
506 | rdev->config.rv770.sc_prim_fifo_size = 0xF9; | |
507 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
508 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
509 | break; | |
510 | case CHIP_RV730: | |
511 | rdev->config.rv770.max_pipes = 2; | |
512 | rdev->config.rv770.max_tile_pipes = 4; | |
513 | rdev->config.rv770.max_simds = 8; | |
514 | rdev->config.rv770.max_backends = 2; | |
515 | rdev->config.rv770.max_gprs = 128; | |
516 | rdev->config.rv770.max_threads = 248; | |
517 | rdev->config.rv770.max_stack_entries = 256; | |
518 | rdev->config.rv770.max_hw_contexts = 8; | |
519 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
520 | rdev->config.rv770.sx_max_export_size = 256; | |
521 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
522 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
523 | rdev->config.rv770.sq_num_cf_insts = 2; | |
524 | ||
525 | rdev->config.rv770.sx_num_of_sets = 7; | |
526 | rdev->config.rv770.sc_prim_fifo_size = 0xf9; | |
527 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
528 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
529 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
530 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
531 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
532 | } | |
533 | break; | |
534 | case CHIP_RV710: | |
535 | rdev->config.rv770.max_pipes = 2; | |
536 | rdev->config.rv770.max_tile_pipes = 2; | |
537 | rdev->config.rv770.max_simds = 2; | |
538 | rdev->config.rv770.max_backends = 1; | |
539 | rdev->config.rv770.max_gprs = 256; | |
540 | rdev->config.rv770.max_threads = 192; | |
541 | rdev->config.rv770.max_stack_entries = 256; | |
542 | rdev->config.rv770.max_hw_contexts = 4; | |
543 | rdev->config.rv770.max_gs_threads = 8 * 2; | |
544 | rdev->config.rv770.sx_max_export_size = 128; | |
545 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
546 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
547 | rdev->config.rv770.sq_num_cf_insts = 1; | |
548 | ||
549 | rdev->config.rv770.sx_num_of_sets = 7; | |
550 | rdev->config.rv770.sc_prim_fifo_size = 0x40; | |
551 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
552 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
553 | break; | |
554 | case CHIP_RV740: | |
555 | rdev->config.rv770.max_pipes = 4; | |
556 | rdev->config.rv770.max_tile_pipes = 4; | |
557 | rdev->config.rv770.max_simds = 8; | |
558 | rdev->config.rv770.max_backends = 4; | |
559 | rdev->config.rv770.max_gprs = 256; | |
560 | rdev->config.rv770.max_threads = 248; | |
561 | rdev->config.rv770.max_stack_entries = 512; | |
562 | rdev->config.rv770.max_hw_contexts = 8; | |
563 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
564 | rdev->config.rv770.sx_max_export_size = 256; | |
565 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
566 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
567 | rdev->config.rv770.sq_num_cf_insts = 2; | |
568 | ||
569 | rdev->config.rv770.sx_num_of_sets = 7; | |
570 | rdev->config.rv770.sc_prim_fifo_size = 0x100; | |
571 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
572 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
573 | ||
574 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
575 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
576 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
577 | } | |
578 | break; | |
579 | default: | |
580 | break; | |
581 | } | |
582 | ||
583 | /* Initialize HDP */ | |
584 | j = 0; | |
585 | for (i = 0; i < 32; i++) { | |
586 | WREG32((0x2c14 + j), 0x00000000); | |
587 | WREG32((0x2c18 + j), 0x00000000); | |
588 | WREG32((0x2c1c + j), 0x00000000); | |
589 | WREG32((0x2c20 + j), 0x00000000); | |
590 | WREG32((0x2c24 + j), 0x00000000); | |
591 | j += 0x18; | |
592 | } | |
593 | ||
594 | WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff)); | |
595 | ||
596 | /* setup tiling, simd, pipe config */ | |
597 | mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG); | |
598 | ||
599 | switch (rdev->config.rv770.max_tile_pipes) { | |
600 | case 1: | |
d03f5d59 | 601 | default: |
3ce0a23d JG |
602 | gb_tiling_config |= PIPE_TILING(0); |
603 | break; | |
604 | case 2: | |
605 | gb_tiling_config |= PIPE_TILING(1); | |
606 | break; | |
607 | case 4: | |
608 | gb_tiling_config |= PIPE_TILING(2); | |
609 | break; | |
610 | case 8: | |
611 | gb_tiling_config |= PIPE_TILING(3); | |
3ce0a23d JG |
612 | break; |
613 | } | |
d03f5d59 | 614 | rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes; |
3ce0a23d JG |
615 | |
616 | if (rdev->family == CHIP_RV770) | |
617 | gb_tiling_config |= BANK_TILING(1); | |
618 | else | |
e29649db | 619 | gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT); |
961fb597 | 620 | rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3); |
3ce0a23d JG |
621 | |
622 | gb_tiling_config |= GROUP_SIZE(0); | |
961fb597 | 623 | rdev->config.rv770.tiling_group_size = 256; |
3ce0a23d | 624 | |
e29649db | 625 | if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) { |
3ce0a23d JG |
626 | gb_tiling_config |= ROW_TILING(3); |
627 | gb_tiling_config |= SAMPLE_SPLIT(3); | |
628 | } else { | |
629 | gb_tiling_config |= | |
630 | ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
631 | gb_tiling_config |= | |
632 | SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
633 | } | |
634 | ||
635 | gb_tiling_config |= BANK_SWAPS(1); | |
636 | ||
d03f5d59 AD |
637 | cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000; |
638 | cc_rb_backend_disable |= | |
639 | BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK); | |
3ce0a23d | 640 | |
d03f5d59 AD |
641 | cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00; |
642 | cc_gc_shader_pipe_config |= | |
3ce0a23d JG |
643 | INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK); |
644 | cc_gc_shader_pipe_config |= | |
645 | INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK); | |
646 | ||
d03f5d59 AD |
647 | if (rdev->family == CHIP_RV740) |
648 | backend_map = 0x28; | |
649 | else | |
650 | backend_map = r700_get_tile_pipe_to_backend_map(rdev, | |
651 | rdev->config.rv770.max_tile_pipes, | |
652 | (R7XX_MAX_BACKENDS - | |
653 | r600_count_pipe_bits((cc_rb_backend_disable & | |
654 | R7XX_MAX_BACKENDS_MASK) >> 16)), | |
655 | (cc_rb_backend_disable >> 16)); | |
656 | gb_tiling_config |= BACKEND_MAP(backend_map); | |
657 | ||
3ce0a23d JG |
658 | |
659 | WREG32(GB_TILING_CONFIG, gb_tiling_config); | |
660 | WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
661 | WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
662 | ||
663 | WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable); | |
664 | WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); | |
f867c60d | 665 | WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); |
d03f5d59 | 666 | WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable); |
3ce0a23d | 667 | |
3ce0a23d JG |
668 | WREG32(CGTS_SYS_TCC_DISABLE, 0); |
669 | WREG32(CGTS_TCC_DISABLE, 0); | |
f867c60d AD |
670 | WREG32(CGTS_USER_SYS_TCC_DISABLE, 0); |
671 | WREG32(CGTS_USER_TCC_DISABLE, 0); | |
3ce0a23d JG |
672 | |
673 | num_qd_pipes = | |
d03f5d59 | 674 | R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8); |
3ce0a23d JG |
675 | WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK); |
676 | WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK); | |
677 | ||
678 | /* set HW defaults for 3D engine */ | |
679 | WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) | | |
e29649db | 680 | ROQ_IB2_START(0x2b))); |
3ce0a23d JG |
681 | |
682 | WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30)); | |
683 | ||
d03f5d59 AD |
684 | ta_aux_cntl = RREG32(TA_CNTL_AUX); |
685 | WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO); | |
3ce0a23d JG |
686 | |
687 | sx_debug_1 = RREG32(SX_DEBUG_1); | |
688 | sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS; | |
689 | WREG32(SX_DEBUG_1, sx_debug_1); | |
690 | ||
691 | smx_dc_ctl0 = RREG32(SMX_DC_CTL0); | |
692 | smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff); | |
693 | smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1); | |
694 | WREG32(SMX_DC_CTL0, smx_dc_ctl0); | |
695 | ||
d03f5d59 AD |
696 | if (rdev->family != CHIP_RV740) |
697 | WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) | | |
698 | GS_FLUSH_CTL(4) | | |
699 | ACK_FLUSH_CTL(3) | | |
700 | SYNC_FLUSH_CTL)); | |
3ce0a23d | 701 | |
d03f5d59 AD |
702 | db_debug3 = RREG32(DB_DEBUG3); |
703 | db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f); | |
704 | switch (rdev->family) { | |
705 | case CHIP_RV770: | |
706 | case CHIP_RV740: | |
707 | db_debug3 |= DB_CLK_OFF_DELAY(0x1f); | |
708 | break; | |
709 | case CHIP_RV710: | |
710 | case CHIP_RV730: | |
711 | default: | |
712 | db_debug3 |= DB_CLK_OFF_DELAY(2); | |
713 | break; | |
714 | } | |
715 | WREG32(DB_DEBUG3, db_debug3); | |
716 | ||
717 | if (rdev->family != CHIP_RV770) { | |
3ce0a23d JG |
718 | db_debug4 = RREG32(DB_DEBUG4); |
719 | db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER; | |
720 | WREG32(DB_DEBUG4, db_debug4); | |
721 | } | |
722 | ||
723 | WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) | | |
e29649db AD |
724 | POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) | |
725 | SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1))); | |
3ce0a23d JG |
726 | |
727 | WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) | | |
e29649db AD |
728 | SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) | |
729 | SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize))); | |
3ce0a23d JG |
730 | |
731 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
732 | ||
733 | WREG32(VGT_NUM_INSTANCES, 1); | |
734 | ||
735 | WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0)); | |
736 | ||
737 | WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4)); | |
738 | ||
739 | WREG32(CP_PERFMON_CNTL, 0); | |
740 | ||
741 | sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) | | |
742 | DONE_FIFO_HIWATER(0xe0) | | |
743 | ALU_UPDATE_FIFO_HIWATER(0x8)); | |
744 | switch (rdev->family) { | |
745 | case CHIP_RV770: | |
3ce0a23d JG |
746 | case CHIP_RV730: |
747 | case CHIP_RV710: | |
d03f5d59 AD |
748 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1); |
749 | break; | |
3ce0a23d JG |
750 | case CHIP_RV740: |
751 | default: | |
752 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4); | |
753 | break; | |
754 | } | |
755 | WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes); | |
756 | ||
757 | /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT | |
758 | * should be adjusted as needed by the 2D/3D drivers. This just sets default values | |
759 | */ | |
760 | sq_config = RREG32(SQ_CONFIG); | |
761 | sq_config &= ~(PS_PRIO(3) | | |
762 | VS_PRIO(3) | | |
763 | GS_PRIO(3) | | |
764 | ES_PRIO(3)); | |
765 | sq_config |= (DX9_CONSTS | | |
766 | VC_ENABLE | | |
767 | EXPORT_SRC_C | | |
768 | PS_PRIO(0) | | |
769 | VS_PRIO(1) | | |
770 | GS_PRIO(2) | | |
771 | ES_PRIO(3)); | |
772 | if (rdev->family == CHIP_RV710) | |
773 | /* no vertex cache */ | |
774 | sq_config &= ~VC_ENABLE; | |
775 | ||
776 | WREG32(SQ_CONFIG, sq_config); | |
777 | ||
778 | WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | | |
fe62e1a4 DA |
779 | NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | |
780 | NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2))); | |
3ce0a23d JG |
781 | |
782 | WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) | | |
fe62e1a4 | 783 | NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64))); |
3ce0a23d JG |
784 | |
785 | sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) | | |
786 | NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) | | |
787 | NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8)); | |
788 | if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads) | |
789 | sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads); | |
790 | else | |
791 | sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8); | |
792 | WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt); | |
793 | ||
794 | WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
795 | NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
796 | ||
797 | WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
798 | NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
799 | ||
800 | sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
801 | SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) | | |
802 | SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
803 | SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64)); | |
804 | ||
805 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0); | |
806 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0); | |
807 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0); | |
808 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0); | |
809 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0); | |
810 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0); | |
811 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0); | |
812 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0); | |
813 | ||
814 | WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) | | |
fe62e1a4 | 815 | FORCE_EOV_MAX_REZ_CNT(255))); |
3ce0a23d JG |
816 | |
817 | if (rdev->family == CHIP_RV710) | |
818 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) | | |
fe62e1a4 | 819 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
820 | else |
821 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) | | |
fe62e1a4 | 822 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
823 | |
824 | switch (rdev->family) { | |
825 | case CHIP_RV770: | |
826 | case CHIP_RV730: | |
827 | case CHIP_RV740: | |
828 | gs_prim_buffer_depth = 384; | |
829 | break; | |
830 | case CHIP_RV710: | |
831 | gs_prim_buffer_depth = 128; | |
832 | break; | |
833 | default: | |
834 | break; | |
835 | } | |
836 | ||
837 | num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16; | |
838 | vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread; | |
839 | /* Max value for this is 256 */ | |
840 | if (vgt_gs_per_es > 256) | |
841 | vgt_gs_per_es = 256; | |
842 | ||
843 | WREG32(VGT_ES_PER_GS, 128); | |
844 | WREG32(VGT_GS_PER_ES, vgt_gs_per_es); | |
845 | WREG32(VGT_GS_PER_VS, 2); | |
846 | ||
847 | /* more default values. 2D/3D driver should adjust as needed */ | |
848 | WREG32(VGT_GS_VERTEX_REUSE, 16); | |
849 | WREG32(PA_SC_LINE_STIPPLE_STATE, 0); | |
850 | WREG32(VGT_STRMOUT_EN, 0); | |
851 | WREG32(SX_MISC, 0); | |
852 | WREG32(PA_SC_MODE_CNTL, 0); | |
853 | WREG32(PA_SC_EDGERULE, 0xaaaaaaaa); | |
854 | WREG32(PA_SC_AA_CONFIG, 0); | |
855 | WREG32(PA_SC_CLIPRECT_RULE, 0xffff); | |
856 | WREG32(PA_SC_LINE_STIPPLE, 0); | |
857 | WREG32(SPI_INPUT_Z, 0); | |
858 | WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2)); | |
859 | WREG32(CB_COLOR7_FRAG, 0); | |
860 | ||
861 | /* clear render buffer base addresses */ | |
862 | WREG32(CB_COLOR0_BASE, 0); | |
863 | WREG32(CB_COLOR1_BASE, 0); | |
864 | WREG32(CB_COLOR2_BASE, 0); | |
865 | WREG32(CB_COLOR3_BASE, 0); | |
866 | WREG32(CB_COLOR4_BASE, 0); | |
867 | WREG32(CB_COLOR5_BASE, 0); | |
868 | WREG32(CB_COLOR6_BASE, 0); | |
869 | WREG32(CB_COLOR7_BASE, 0); | |
870 | ||
871 | WREG32(TCP_CNTL, 0); | |
872 | ||
873 | hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL); | |
874 | WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl); | |
875 | ||
876 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
877 | ||
878 | WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA | | |
879 | NUM_CLIP_SEQ(3))); | |
880 | ||
881 | } | |
882 | ||
883 | int rv770_mc_init(struct radeon_device *rdev) | |
884 | { | |
3ce0a23d | 885 | u32 tmp; |
5885b7a9 | 886 | int chansize, numchan; |
3ce0a23d JG |
887 | |
888 | /* Get VRAM informations */ | |
3ce0a23d | 889 | rdev->mc.vram_is_ddr = true; |
5885b7a9 AD |
890 | tmp = RREG32(MC_ARB_RAMCFG); |
891 | if (tmp & CHANSIZE_OVERRIDE) { | |
892 | chansize = 16; | |
893 | } else if (tmp & CHANSIZE_MASK) { | |
894 | chansize = 64; | |
895 | } else { | |
896 | chansize = 32; | |
897 | } | |
898 | tmp = RREG32(MC_SHARED_CHMAP); | |
899 | switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) { | |
900 | case 0: | |
901 | default: | |
902 | numchan = 1; | |
903 | break; | |
904 | case 1: | |
905 | numchan = 2; | |
906 | break; | |
907 | case 2: | |
908 | numchan = 4; | |
909 | break; | |
910 | case 3: | |
911 | numchan = 8; | |
912 | break; | |
913 | } | |
914 | rdev->mc.vram_width = numchan * chansize; | |
771fe6b9 JG |
915 | /* Could aper size report 0 ? */ |
916 | rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0); | |
917 | rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0); | |
3ce0a23d JG |
918 | /* Setup GPU memory space */ |
919 | rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE); | |
920 | rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE); | |
51e5fcd3 | 921 | rdev->mc.visible_vram_size = rdev->mc.aper_size; |
d594e46a | 922 | r600_vram_gtt_location(rdev, &rdev->mc); |
f47299c5 AD |
923 | radeon_update_bandwidth_info(rdev); |
924 | ||
3ce0a23d JG |
925 | return 0; |
926 | } | |
d594e46a | 927 | |
fc30b8ef | 928 | static int rv770_startup(struct radeon_device *rdev) |
3ce0a23d JG |
929 | { |
930 | int r; | |
931 | ||
779720a3 AD |
932 | if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) { |
933 | r = r600_init_microcode(rdev); | |
934 | if (r) { | |
935 | DRM_ERROR("Failed to load firmware!\n"); | |
936 | return r; | |
937 | } | |
938 | } | |
939 | ||
a3c1945a | 940 | rv770_mc_program(rdev); |
1a029b76 JG |
941 | if (rdev->flags & RADEON_IS_AGP) { |
942 | rv770_agp_enable(rdev); | |
943 | } else { | |
944 | r = rv770_pcie_gart_enable(rdev); | |
945 | if (r) | |
946 | return r; | |
947 | } | |
3ce0a23d | 948 | rv770_gpu_init(rdev); |
c38c7b64 JG |
949 | r = r600_blit_init(rdev); |
950 | if (r) { | |
951 | r600_blit_fini(rdev); | |
952 | rdev->asic->copy = NULL; | |
953 | dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r); | |
954 | } | |
ff82f052 JG |
955 | /* pin copy shader into vram */ |
956 | if (rdev->r600_blit.shader_obj) { | |
957 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
958 | if (unlikely(r != 0)) | |
959 | return r; | |
960 | r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM, | |
961 | &rdev->r600_blit.shader_gpu_addr); | |
962 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
7923c615 | 963 | if (r) { |
ff82f052 | 964 | DRM_ERROR("failed to pin blit object %d\n", r); |
7923c615 AD |
965 | return r; |
966 | } | |
967 | } | |
d8f60cfc | 968 | /* Enable IRQ */ |
d8f60cfc AD |
969 | r = r600_irq_init(rdev); |
970 | if (r) { | |
971 | DRM_ERROR("radeon: IH init failed (%d).\n", r); | |
972 | radeon_irq_kms_fini(rdev); | |
973 | return r; | |
974 | } | |
975 | r600_irq_set(rdev); | |
976 | ||
3ce0a23d JG |
977 | r = radeon_ring_init(rdev, rdev->cp.ring_size); |
978 | if (r) | |
979 | return r; | |
980 | r = rv770_cp_load_microcode(rdev); | |
981 | if (r) | |
982 | return r; | |
983 | r = r600_cp_resume(rdev); | |
984 | if (r) | |
985 | return r; | |
81cc35bf JG |
986 | /* write back buffer are not vital so don't worry about failure */ |
987 | r600_wb_enable(rdev); | |
3ce0a23d JG |
988 | return 0; |
989 | } | |
990 | ||
fc30b8ef DA |
991 | int rv770_resume(struct radeon_device *rdev) |
992 | { | |
993 | int r; | |
994 | ||
1a029b76 JG |
995 | /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw, |
996 | * posting will perform necessary task to bring back GPU into good | |
997 | * shape. | |
998 | */ | |
fc30b8ef | 999 | /* post card */ |
e7d40b9a | 1000 | atom_asic_init(rdev->mode_info.atom_context); |
fc30b8ef DA |
1001 | /* Initialize clocks */ |
1002 | r = radeon_clocks_init(rdev); | |
1003 | if (r) { | |
1004 | return r; | |
1005 | } | |
1006 | ||
1007 | r = rv770_startup(rdev); | |
1008 | if (r) { | |
1009 | DRM_ERROR("r600 startup failed on resume\n"); | |
1010 | return r; | |
1011 | } | |
1012 | ||
62a8ea3f | 1013 | r = r600_ib_test(rdev); |
fc30b8ef DA |
1014 | if (r) { |
1015 | DRM_ERROR("radeon: failled testing IB (%d).\n", r); | |
1016 | return r; | |
1017 | } | |
8a8c6e7c RM |
1018 | |
1019 | r = r600_audio_init(rdev); | |
1020 | if (r) { | |
1021 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1022 | return r; | |
1023 | } | |
1024 | ||
fc30b8ef DA |
1025 | return r; |
1026 | ||
1027 | } | |
1028 | ||
3ce0a23d JG |
1029 | int rv770_suspend(struct radeon_device *rdev) |
1030 | { | |
4c788679 JG |
1031 | int r; |
1032 | ||
8a8c6e7c | 1033 | r600_audio_fini(rdev); |
3ce0a23d JG |
1034 | /* FIXME: we should wait for ring to be empty */ |
1035 | r700_cp_stop(rdev); | |
4153e584 | 1036 | rdev->cp.ready = false; |
0c45249f | 1037 | r600_irq_suspend(rdev); |
81cc35bf | 1038 | r600_wb_disable(rdev); |
4aac0473 | 1039 | rv770_pcie_gart_disable(rdev); |
4153e584 | 1040 | /* unpin shaders bo */ |
30d2d9a5 JG |
1041 | if (rdev->r600_blit.shader_obj) { |
1042 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
1043 | if (likely(r == 0)) { | |
1044 | radeon_bo_unpin(rdev->r600_blit.shader_obj); | |
1045 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
1046 | } | |
4c788679 | 1047 | } |
3ce0a23d JG |
1048 | return 0; |
1049 | } | |
1050 | ||
1051 | /* Plan is to move initialization in that function and use | |
1052 | * helper function so that radeon_device_init pretty much | |
1053 | * do nothing more than calling asic specific function. This | |
1054 | * should also allow to remove a bunch of callback function | |
1055 | * like vram_info. | |
1056 | */ | |
1057 | int rv770_init(struct radeon_device *rdev) | |
1058 | { | |
1059 | int r; | |
1060 | ||
3ce0a23d JG |
1061 | r = radeon_dummy_page_init(rdev); |
1062 | if (r) | |
1063 | return r; | |
1064 | /* This don't do much */ | |
1065 | r = radeon_gem_init(rdev); | |
1066 | if (r) | |
1067 | return r; | |
1068 | /* Read BIOS */ | |
1069 | if (!radeon_get_bios(rdev)) { | |
1070 | if (ASIC_IS_AVIVO(rdev)) | |
1071 | return -EINVAL; | |
1072 | } | |
1073 | /* Must be an ATOMBIOS */ | |
e7d40b9a JG |
1074 | if (!rdev->is_atom_bios) { |
1075 | dev_err(rdev->dev, "Expecting atombios for R600 GPU\n"); | |
3ce0a23d | 1076 | return -EINVAL; |
e7d40b9a | 1077 | } |
3ce0a23d JG |
1078 | r = radeon_atombios_init(rdev); |
1079 | if (r) | |
1080 | return r; | |
1081 | /* Post card if necessary */ | |
72542d77 DA |
1082 | if (!r600_card_posted(rdev)) { |
1083 | if (!rdev->bios) { | |
1084 | dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n"); | |
1085 | return -EINVAL; | |
1086 | } | |
3ce0a23d JG |
1087 | DRM_INFO("GPU not posted. posting now...\n"); |
1088 | atom_asic_init(rdev->mode_info.atom_context); | |
1089 | } | |
1090 | /* Initialize scratch registers */ | |
1091 | r600_scratch_init(rdev); | |
1092 | /* Initialize surface registers */ | |
1093 | radeon_surface_init(rdev); | |
7433874e | 1094 | /* Initialize clocks */ |
5e6dde7e | 1095 | radeon_get_clock_info(rdev->ddev); |
3ce0a23d JG |
1096 | r = radeon_clocks_init(rdev); |
1097 | if (r) | |
1098 | return r; | |
1099 | /* Fence driver */ | |
1100 | r = radeon_fence_driver_init(rdev); | |
1101 | if (r) | |
1102 | return r; | |
d594e46a | 1103 | /* initialize AGP */ |
700a0cc0 JG |
1104 | if (rdev->flags & RADEON_IS_AGP) { |
1105 | r = radeon_agp_init(rdev); | |
1106 | if (r) | |
1107 | radeon_agp_disable(rdev); | |
1108 | } | |
3ce0a23d | 1109 | r = rv770_mc_init(rdev); |
b574f251 | 1110 | if (r) |
3ce0a23d | 1111 | return r; |
3ce0a23d | 1112 | /* Memory manager */ |
4c788679 | 1113 | r = radeon_bo_init(rdev); |
3ce0a23d JG |
1114 | if (r) |
1115 | return r; | |
d8f60cfc AD |
1116 | |
1117 | r = radeon_irq_kms_init(rdev); | |
1118 | if (r) | |
1119 | return r; | |
1120 | ||
3ce0a23d JG |
1121 | rdev->cp.ring_obj = NULL; |
1122 | r600_ring_init(rdev, 1024 * 1024); | |
1123 | ||
d8f60cfc AD |
1124 | rdev->ih.ring_obj = NULL; |
1125 | r600_ih_ring_init(rdev, 64 * 1024); | |
1126 | ||
4aac0473 JG |
1127 | r = r600_pcie_gart_init(rdev); |
1128 | if (r) | |
1129 | return r; | |
1130 | ||
779720a3 | 1131 | rdev->accel_working = true; |
fc30b8ef | 1132 | r = rv770_startup(rdev); |
3ce0a23d | 1133 | if (r) { |
655efd3d | 1134 | dev_err(rdev->dev, "disabling GPU acceleration\n"); |
fe251e2f | 1135 | r700_cp_fini(rdev); |
75c81298 | 1136 | r600_wb_fini(rdev); |
655efd3d JG |
1137 | r600_irq_fini(rdev); |
1138 | radeon_irq_kms_fini(rdev); | |
75c81298 | 1139 | rv770_pcie_gart_fini(rdev); |
733289c2 | 1140 | rdev->accel_working = false; |
3ce0a23d | 1141 | } |
733289c2 | 1142 | if (rdev->accel_working) { |
733289c2 JG |
1143 | r = radeon_ib_pool_init(rdev); |
1144 | if (r) { | |
db96380e | 1145 | dev_err(rdev->dev, "IB initialization failed (%d).\n", r); |
733289c2 | 1146 | rdev->accel_working = false; |
db96380e JG |
1147 | } else { |
1148 | r = r600_ib_test(rdev); | |
1149 | if (r) { | |
1150 | dev_err(rdev->dev, "IB test failed (%d).\n", r); | |
1151 | rdev->accel_working = false; | |
1152 | } | |
733289c2 | 1153 | } |
3ce0a23d | 1154 | } |
8a8c6e7c RM |
1155 | |
1156 | r = r600_audio_init(rdev); | |
1157 | if (r) { | |
1158 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1159 | return r; | |
1160 | } | |
1161 | ||
3ce0a23d JG |
1162 | return 0; |
1163 | } | |
1164 | ||
1165 | void rv770_fini(struct radeon_device *rdev) | |
1166 | { | |
1167 | r600_blit_fini(rdev); | |
fe251e2f | 1168 | r700_cp_fini(rdev); |
655efd3d | 1169 | r600_wb_fini(rdev); |
d8f60cfc AD |
1170 | r600_irq_fini(rdev); |
1171 | radeon_irq_kms_fini(rdev); | |
4aac0473 | 1172 | rv770_pcie_gart_fini(rdev); |
3ce0a23d JG |
1173 | radeon_gem_fini(rdev); |
1174 | radeon_fence_driver_fini(rdev); | |
1175 | radeon_clocks_fini(rdev); | |
d0269ed8 | 1176 | radeon_agp_fini(rdev); |
4c788679 | 1177 | radeon_bo_fini(rdev); |
e7d40b9a | 1178 | radeon_atombios_fini(rdev); |
3ce0a23d JG |
1179 | kfree(rdev->bios); |
1180 | rdev->bios = NULL; | |
1181 | radeon_dummy_page_fini(rdev); | |
771fe6b9 | 1182 | } |