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 | |
21a8122a AD |
45 | /* get temperature in millidegrees */ |
46 | u32 rv770_get_temp(struct radeon_device *rdev) | |
47 | { | |
48 | u32 temp = (RREG32(CG_MULT_THERMAL_STATUS) & ASIC_T_MASK) >> | |
49 | ASIC_T_SHIFT; | |
50 | u32 actual_temp = 0; | |
51 | ||
52 | if ((temp >> 9) & 1) | |
53 | actual_temp = 0; | |
54 | else | |
55 | actual_temp = (temp >> 1) & 0xff; | |
56 | ||
57 | return actual_temp * 1000; | |
58 | } | |
59 | ||
49e02b73 AD |
60 | void rv770_pm_misc(struct radeon_device *rdev) |
61 | { | |
a081a9d6 RM |
62 | int req_ps_idx = rdev->pm.requested_power_state_index; |
63 | int req_cm_idx = rdev->pm.requested_clock_mode_index; | |
64 | struct radeon_power_state *ps = &rdev->pm.power_state[req_ps_idx]; | |
65 | struct radeon_voltage *voltage = &ps->clock_info[req_cm_idx].voltage; | |
4d60173f AD |
66 | |
67 | if ((voltage->type == VOLTAGE_SW) && voltage->voltage) { | |
68 | if (voltage->voltage != rdev->pm.current_vddc) { | |
69 | radeon_atom_set_voltage(rdev, voltage->voltage); | |
70 | rdev->pm.current_vddc = voltage->voltage; | |
0fcbe947 | 71 | DRM_DEBUG("Setting: v: %d\n", voltage->voltage); |
4d60173f AD |
72 | } |
73 | } | |
49e02b73 | 74 | } |
771fe6b9 JG |
75 | |
76 | /* | |
3ce0a23d | 77 | * GART |
771fe6b9 | 78 | */ |
3ce0a23d | 79 | int rv770_pcie_gart_enable(struct radeon_device *rdev) |
771fe6b9 | 80 | { |
3ce0a23d JG |
81 | u32 tmp; |
82 | int r, i; | |
771fe6b9 | 83 | |
4aac0473 JG |
84 | if (rdev->gart.table.vram.robj == NULL) { |
85 | dev_err(rdev->dev, "No VRAM object for PCIE GART.\n"); | |
86 | return -EINVAL; | |
3ce0a23d | 87 | } |
4aac0473 JG |
88 | r = radeon_gart_table_vram_pin(rdev); |
89 | if (r) | |
3ce0a23d | 90 | return r; |
82568565 | 91 | radeon_gart_restore(rdev); |
3ce0a23d JG |
92 | /* Setup L2 cache */ |
93 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
94 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
95 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
96 | WREG32(VM_L2_CNTL2, 0); | |
97 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
98 | /* Setup TLB control */ | |
99 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
100 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
101 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
102 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
103 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
104 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
105 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
106 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
107 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
108 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
109 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
110 | WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12); | |
1a029b76 | 111 | WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12); |
3ce0a23d JG |
112 | WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12); |
113 | WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) | | |
114 | RANGE_PROTECTION_FAULT_ENABLE_DEFAULT); | |
115 | WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, | |
116 | (u32)(rdev->dummy_page.addr >> 12)); | |
117 | for (i = 1; i < 7; i++) | |
118 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
771fe6b9 | 119 | |
3ce0a23d JG |
120 | r600_pcie_gart_tlb_flush(rdev); |
121 | rdev->gart.ready = true; | |
771fe6b9 JG |
122 | return 0; |
123 | } | |
124 | ||
3ce0a23d | 125 | void rv770_pcie_gart_disable(struct radeon_device *rdev) |
771fe6b9 | 126 | { |
3ce0a23d | 127 | u32 tmp; |
4c788679 | 128 | int i, r; |
3ce0a23d | 129 | |
3ce0a23d JG |
130 | /* Disable all tables */ |
131 | for (i = 0; i < 7; i++) | |
132 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
133 | ||
134 | /* Setup L2 cache */ | |
135 | WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING | | |
136 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
137 | WREG32(VM_L2_CNTL2, 0); | |
138 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
139 | /* Setup TLB control */ | |
140 | tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
141 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
142 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
143 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
144 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
145 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
146 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
147 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
4aac0473 | 148 | if (rdev->gart.table.vram.robj) { |
4c788679 JG |
149 | r = radeon_bo_reserve(rdev->gart.table.vram.robj, false); |
150 | if (likely(r == 0)) { | |
151 | radeon_bo_kunmap(rdev->gart.table.vram.robj); | |
152 | radeon_bo_unpin(rdev->gart.table.vram.robj); | |
153 | radeon_bo_unreserve(rdev->gart.table.vram.robj); | |
154 | } | |
4aac0473 JG |
155 | } |
156 | } | |
157 | ||
158 | void rv770_pcie_gart_fini(struct radeon_device *rdev) | |
159 | { | |
f9274562 | 160 | radeon_gart_fini(rdev); |
4aac0473 JG |
161 | rv770_pcie_gart_disable(rdev); |
162 | radeon_gart_table_vram_free(rdev); | |
771fe6b9 JG |
163 | } |
164 | ||
165 | ||
1a029b76 JG |
166 | void rv770_agp_enable(struct radeon_device *rdev) |
167 | { | |
168 | u32 tmp; | |
169 | int i; | |
170 | ||
171 | /* Setup L2 cache */ | |
172 | WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING | | |
173 | ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE | | |
174 | EFFECTIVE_L2_QUEUE_SIZE(7)); | |
175 | WREG32(VM_L2_CNTL2, 0); | |
176 | WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2)); | |
177 | /* Setup TLB control */ | |
178 | tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING | | |
179 | SYSTEM_ACCESS_MODE_NOT_IN_SYS | | |
180 | SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | | |
181 | EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5); | |
182 | WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp); | |
183 | WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp); | |
184 | WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp); | |
185 | WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp); | |
186 | WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp); | |
187 | WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp); | |
188 | WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp); | |
189 | for (i = 0; i < 7; i++) | |
190 | WREG32(VM_CONTEXT0_CNTL + (i * 4), 0); | |
191 | } | |
192 | ||
a3c1945a | 193 | static void rv770_mc_program(struct radeon_device *rdev) |
771fe6b9 | 194 | { |
a3c1945a | 195 | struct rv515_mc_save save; |
3ce0a23d JG |
196 | u32 tmp; |
197 | int i, j; | |
198 | ||
199 | /* Initialize HDP */ | |
200 | for (i = 0, j = 0; i < 32; i++, j += 0x18) { | |
201 | WREG32((0x2c14 + j), 0x00000000); | |
202 | WREG32((0x2c18 + j), 0x00000000); | |
203 | WREG32((0x2c1c + j), 0x00000000); | |
204 | WREG32((0x2c20 + j), 0x00000000); | |
205 | WREG32((0x2c24 + j), 0x00000000); | |
206 | } | |
207 | WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0); | |
208 | ||
a3c1945a | 209 | rv515_mc_stop(rdev, &save); |
3ce0a23d | 210 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 211 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 212 | } |
3ce0a23d JG |
213 | /* Lockout access through VGA aperture*/ |
214 | WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE); | |
3ce0a23d | 215 | /* Update configuration */ |
1a029b76 JG |
216 | if (rdev->flags & RADEON_IS_AGP) { |
217 | if (rdev->mc.vram_start < rdev->mc.gtt_start) { | |
218 | /* VRAM before AGP */ | |
219 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
220 | rdev->mc.vram_start >> 12); | |
221 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
222 | rdev->mc.gtt_end >> 12); | |
223 | } else { | |
224 | /* VRAM after AGP */ | |
225 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
226 | rdev->mc.gtt_start >> 12); | |
227 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
228 | rdev->mc.vram_end >> 12); | |
229 | } | |
230 | } else { | |
231 | WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, | |
232 | rdev->mc.vram_start >> 12); | |
233 | WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, | |
234 | rdev->mc.vram_end >> 12); | |
235 | } | |
3ce0a23d | 236 | WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0); |
1a029b76 | 237 | tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16; |
3ce0a23d JG |
238 | tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF); |
239 | WREG32(MC_VM_FB_LOCATION, tmp); | |
240 | WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8)); | |
241 | WREG32(HDP_NONSURFACE_INFO, (2 << 7)); | |
46fcd2b3 | 242 | WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF); |
3ce0a23d | 243 | if (rdev->flags & RADEON_IS_AGP) { |
1a029b76 | 244 | WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16); |
3ce0a23d JG |
245 | WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16); |
246 | WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22); | |
247 | } else { | |
248 | WREG32(MC_VM_AGP_BASE, 0); | |
249 | WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF); | |
250 | WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF); | |
251 | } | |
3ce0a23d | 252 | if (r600_mc_wait_for_idle(rdev)) { |
a3c1945a | 253 | dev_warn(rdev->dev, "Wait for MC idle timedout !\n"); |
3ce0a23d | 254 | } |
a3c1945a | 255 | rv515_mc_resume(rdev, &save); |
698443d9 DA |
256 | /* we need to own VRAM, so turn off the VGA renderer here |
257 | * to stop it overwriting our objects */ | |
d39c3b89 | 258 | rv515_vga_render_disable(rdev); |
771fe6b9 JG |
259 | } |
260 | ||
3ce0a23d JG |
261 | |
262 | /* | |
263 | * CP. | |
264 | */ | |
265 | void r700_cp_stop(struct radeon_device *rdev) | |
771fe6b9 | 266 | { |
3ce0a23d | 267 | WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT)); |
771fe6b9 JG |
268 | } |
269 | ||
3ce0a23d | 270 | static int rv770_cp_load_microcode(struct radeon_device *rdev) |
771fe6b9 | 271 | { |
3ce0a23d JG |
272 | const __be32 *fw_data; |
273 | int i; | |
274 | ||
275 | if (!rdev->me_fw || !rdev->pfp_fw) | |
276 | return -EINVAL; | |
277 | ||
278 | r700_cp_stop(rdev); | |
279 | WREG32(CP_RB_CNTL, RB_NO_UPDATE | (15 << 8) | (3 << 0)); | |
280 | ||
281 | /* Reset cp */ | |
282 | WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP); | |
283 | RREG32(GRBM_SOFT_RESET); | |
284 | mdelay(15); | |
285 | WREG32(GRBM_SOFT_RESET, 0); | |
286 | ||
287 | fw_data = (const __be32 *)rdev->pfp_fw->data; | |
288 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
289 | for (i = 0; i < R700_PFP_UCODE_SIZE; i++) | |
290 | WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++)); | |
291 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
292 | ||
293 | fw_data = (const __be32 *)rdev->me_fw->data; | |
294 | WREG32(CP_ME_RAM_WADDR, 0); | |
295 | for (i = 0; i < R700_PM4_UCODE_SIZE; i++) | |
296 | WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++)); | |
297 | ||
298 | WREG32(CP_PFP_UCODE_ADDR, 0); | |
299 | WREG32(CP_ME_RAM_WADDR, 0); | |
300 | WREG32(CP_ME_RAM_RADDR, 0); | |
301 | return 0; | |
771fe6b9 JG |
302 | } |
303 | ||
fe251e2f AD |
304 | void r700_cp_fini(struct radeon_device *rdev) |
305 | { | |
306 | r700_cp_stop(rdev); | |
307 | radeon_ring_fini(rdev); | |
308 | } | |
771fe6b9 JG |
309 | |
310 | /* | |
3ce0a23d | 311 | * Core functions |
771fe6b9 | 312 | */ |
d03f5d59 AD |
313 | static u32 r700_get_tile_pipe_to_backend_map(struct radeon_device *rdev, |
314 | u32 num_tile_pipes, | |
315 | u32 num_backends, | |
316 | u32 backend_disable_mask) | |
771fe6b9 | 317 | { |
3ce0a23d JG |
318 | u32 backend_map = 0; |
319 | u32 enabled_backends_mask; | |
320 | u32 enabled_backends_count; | |
321 | u32 cur_pipe; | |
322 | u32 swizzle_pipe[R7XX_MAX_PIPES]; | |
323 | u32 cur_backend; | |
324 | u32 i; | |
d03f5d59 | 325 | bool force_no_swizzle; |
3ce0a23d JG |
326 | |
327 | if (num_tile_pipes > R7XX_MAX_PIPES) | |
328 | num_tile_pipes = R7XX_MAX_PIPES; | |
329 | if (num_tile_pipes < 1) | |
330 | num_tile_pipes = 1; | |
331 | if (num_backends > R7XX_MAX_BACKENDS) | |
332 | num_backends = R7XX_MAX_BACKENDS; | |
333 | if (num_backends < 1) | |
334 | num_backends = 1; | |
335 | ||
336 | enabled_backends_mask = 0; | |
337 | enabled_backends_count = 0; | |
338 | for (i = 0; i < R7XX_MAX_BACKENDS; ++i) { | |
339 | if (((backend_disable_mask >> i) & 1) == 0) { | |
340 | enabled_backends_mask |= (1 << i); | |
341 | ++enabled_backends_count; | |
342 | } | |
343 | if (enabled_backends_count == num_backends) | |
344 | break; | |
345 | } | |
346 | ||
347 | if (enabled_backends_count == 0) { | |
348 | enabled_backends_mask = 1; | |
349 | enabled_backends_count = 1; | |
350 | } | |
351 | ||
352 | if (enabled_backends_count != num_backends) | |
353 | num_backends = enabled_backends_count; | |
354 | ||
d03f5d59 AD |
355 | switch (rdev->family) { |
356 | case CHIP_RV770: | |
357 | case CHIP_RV730: | |
358 | force_no_swizzle = false; | |
359 | break; | |
360 | case CHIP_RV710: | |
361 | case CHIP_RV740: | |
362 | default: | |
363 | force_no_swizzle = true; | |
364 | break; | |
365 | } | |
366 | ||
3ce0a23d JG |
367 | memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES); |
368 | switch (num_tile_pipes) { | |
369 | case 1: | |
370 | swizzle_pipe[0] = 0; | |
371 | break; | |
372 | case 2: | |
373 | swizzle_pipe[0] = 0; | |
374 | swizzle_pipe[1] = 1; | |
375 | break; | |
376 | case 3: | |
d03f5d59 AD |
377 | if (force_no_swizzle) { |
378 | swizzle_pipe[0] = 0; | |
379 | swizzle_pipe[1] = 1; | |
380 | swizzle_pipe[2] = 2; | |
381 | } else { | |
382 | swizzle_pipe[0] = 0; | |
383 | swizzle_pipe[1] = 2; | |
384 | swizzle_pipe[2] = 1; | |
385 | } | |
3ce0a23d JG |
386 | break; |
387 | case 4: | |
d03f5d59 AD |
388 | if (force_no_swizzle) { |
389 | swizzle_pipe[0] = 0; | |
390 | swizzle_pipe[1] = 1; | |
391 | swizzle_pipe[2] = 2; | |
392 | swizzle_pipe[3] = 3; | |
393 | } else { | |
394 | swizzle_pipe[0] = 0; | |
395 | swizzle_pipe[1] = 2; | |
396 | swizzle_pipe[2] = 3; | |
397 | swizzle_pipe[3] = 1; | |
398 | } | |
3ce0a23d JG |
399 | break; |
400 | case 5: | |
d03f5d59 AD |
401 | if (force_no_swizzle) { |
402 | swizzle_pipe[0] = 0; | |
403 | swizzle_pipe[1] = 1; | |
404 | swizzle_pipe[2] = 2; | |
405 | swizzle_pipe[3] = 3; | |
406 | swizzle_pipe[4] = 4; | |
407 | } else { | |
408 | swizzle_pipe[0] = 0; | |
409 | swizzle_pipe[1] = 2; | |
410 | swizzle_pipe[2] = 4; | |
411 | swizzle_pipe[3] = 1; | |
412 | swizzle_pipe[4] = 3; | |
413 | } | |
3ce0a23d JG |
414 | break; |
415 | case 6: | |
d03f5d59 AD |
416 | if (force_no_swizzle) { |
417 | swizzle_pipe[0] = 0; | |
418 | swizzle_pipe[1] = 1; | |
419 | swizzle_pipe[2] = 2; | |
420 | swizzle_pipe[3] = 3; | |
421 | swizzle_pipe[4] = 4; | |
422 | swizzle_pipe[5] = 5; | |
423 | } else { | |
424 | swizzle_pipe[0] = 0; | |
425 | swizzle_pipe[1] = 2; | |
426 | swizzle_pipe[2] = 4; | |
427 | swizzle_pipe[3] = 5; | |
428 | swizzle_pipe[4] = 3; | |
429 | swizzle_pipe[5] = 1; | |
430 | } | |
3ce0a23d JG |
431 | break; |
432 | case 7: | |
d03f5d59 AD |
433 | if (force_no_swizzle) { |
434 | swizzle_pipe[0] = 0; | |
435 | swizzle_pipe[1] = 1; | |
436 | swizzle_pipe[2] = 2; | |
437 | swizzle_pipe[3] = 3; | |
438 | swizzle_pipe[4] = 4; | |
439 | swizzle_pipe[5] = 5; | |
440 | swizzle_pipe[6] = 6; | |
441 | } else { | |
442 | swizzle_pipe[0] = 0; | |
443 | swizzle_pipe[1] = 2; | |
444 | swizzle_pipe[2] = 4; | |
445 | swizzle_pipe[3] = 6; | |
446 | swizzle_pipe[4] = 3; | |
447 | swizzle_pipe[5] = 1; | |
448 | swizzle_pipe[6] = 5; | |
449 | } | |
3ce0a23d JG |
450 | break; |
451 | case 8: | |
d03f5d59 AD |
452 | if (force_no_swizzle) { |
453 | swizzle_pipe[0] = 0; | |
454 | swizzle_pipe[1] = 1; | |
455 | swizzle_pipe[2] = 2; | |
456 | swizzle_pipe[3] = 3; | |
457 | swizzle_pipe[4] = 4; | |
458 | swizzle_pipe[5] = 5; | |
459 | swizzle_pipe[6] = 6; | |
460 | swizzle_pipe[7] = 7; | |
461 | } else { | |
462 | swizzle_pipe[0] = 0; | |
463 | swizzle_pipe[1] = 2; | |
464 | swizzle_pipe[2] = 4; | |
465 | swizzle_pipe[3] = 6; | |
466 | swizzle_pipe[4] = 3; | |
467 | swizzle_pipe[5] = 1; | |
468 | swizzle_pipe[6] = 7; | |
469 | swizzle_pipe[7] = 5; | |
470 | } | |
3ce0a23d JG |
471 | break; |
472 | } | |
473 | ||
474 | cur_backend = 0; | |
475 | for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) { | |
476 | while (((1 << cur_backend) & enabled_backends_mask) == 0) | |
477 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
478 | ||
479 | backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2))); | |
480 | ||
481 | cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS; | |
482 | } | |
483 | ||
484 | return backend_map; | |
771fe6b9 JG |
485 | } |
486 | ||
3ce0a23d | 487 | static void rv770_gpu_init(struct radeon_device *rdev) |
771fe6b9 | 488 | { |
3ce0a23d | 489 | int i, j, num_qd_pipes; |
d03f5d59 | 490 | u32 ta_aux_cntl; |
3ce0a23d JG |
491 | u32 sx_debug_1; |
492 | u32 smx_dc_ctl0; | |
d03f5d59 | 493 | u32 db_debug3; |
3ce0a23d JG |
494 | u32 num_gs_verts_per_thread; |
495 | u32 vgt_gs_per_es; | |
496 | u32 gs_prim_buffer_depth = 0; | |
497 | u32 sq_ms_fifo_sizes; | |
498 | u32 sq_config; | |
499 | u32 sq_thread_resource_mgmt; | |
500 | u32 hdp_host_path_cntl; | |
501 | u32 sq_dyn_gpr_size_simd_ab_0; | |
502 | u32 backend_map; | |
503 | u32 gb_tiling_config = 0; | |
504 | u32 cc_rb_backend_disable = 0; | |
505 | u32 cc_gc_shader_pipe_config = 0; | |
506 | u32 mc_arb_ramcfg; | |
507 | u32 db_debug4; | |
771fe6b9 | 508 | |
3ce0a23d JG |
509 | /* setup chip specs */ |
510 | switch (rdev->family) { | |
511 | case CHIP_RV770: | |
512 | rdev->config.rv770.max_pipes = 4; | |
513 | rdev->config.rv770.max_tile_pipes = 8; | |
514 | rdev->config.rv770.max_simds = 10; | |
515 | rdev->config.rv770.max_backends = 4; | |
516 | rdev->config.rv770.max_gprs = 256; | |
517 | rdev->config.rv770.max_threads = 248; | |
518 | rdev->config.rv770.max_stack_entries = 512; | |
519 | rdev->config.rv770.max_hw_contexts = 8; | |
520 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
521 | rdev->config.rv770.sx_max_export_size = 128; | |
522 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
523 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
524 | rdev->config.rv770.sq_num_cf_insts = 2; | |
525 | ||
526 | rdev->config.rv770.sx_num_of_sets = 7; | |
527 | rdev->config.rv770.sc_prim_fifo_size = 0xF9; | |
528 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
529 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
530 | break; | |
531 | case CHIP_RV730: | |
532 | rdev->config.rv770.max_pipes = 2; | |
533 | rdev->config.rv770.max_tile_pipes = 4; | |
534 | rdev->config.rv770.max_simds = 8; | |
535 | rdev->config.rv770.max_backends = 2; | |
536 | rdev->config.rv770.max_gprs = 128; | |
537 | rdev->config.rv770.max_threads = 248; | |
538 | rdev->config.rv770.max_stack_entries = 256; | |
539 | rdev->config.rv770.max_hw_contexts = 8; | |
540 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
541 | rdev->config.rv770.sx_max_export_size = 256; | |
542 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
543 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
544 | rdev->config.rv770.sq_num_cf_insts = 2; | |
545 | ||
546 | rdev->config.rv770.sx_num_of_sets = 7; | |
547 | rdev->config.rv770.sc_prim_fifo_size = 0xf9; | |
548 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
549 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
550 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
551 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
552 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
553 | } | |
554 | break; | |
555 | case CHIP_RV710: | |
556 | rdev->config.rv770.max_pipes = 2; | |
557 | rdev->config.rv770.max_tile_pipes = 2; | |
558 | rdev->config.rv770.max_simds = 2; | |
559 | rdev->config.rv770.max_backends = 1; | |
560 | rdev->config.rv770.max_gprs = 256; | |
561 | rdev->config.rv770.max_threads = 192; | |
562 | rdev->config.rv770.max_stack_entries = 256; | |
563 | rdev->config.rv770.max_hw_contexts = 4; | |
564 | rdev->config.rv770.max_gs_threads = 8 * 2; | |
565 | rdev->config.rv770.sx_max_export_size = 128; | |
566 | rdev->config.rv770.sx_max_export_pos_size = 16; | |
567 | rdev->config.rv770.sx_max_export_smx_size = 112; | |
568 | rdev->config.rv770.sq_num_cf_insts = 1; | |
569 | ||
570 | rdev->config.rv770.sx_num_of_sets = 7; | |
571 | rdev->config.rv770.sc_prim_fifo_size = 0x40; | |
572 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
573 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
574 | break; | |
575 | case CHIP_RV740: | |
576 | rdev->config.rv770.max_pipes = 4; | |
577 | rdev->config.rv770.max_tile_pipes = 4; | |
578 | rdev->config.rv770.max_simds = 8; | |
579 | rdev->config.rv770.max_backends = 4; | |
580 | rdev->config.rv770.max_gprs = 256; | |
581 | rdev->config.rv770.max_threads = 248; | |
582 | rdev->config.rv770.max_stack_entries = 512; | |
583 | rdev->config.rv770.max_hw_contexts = 8; | |
584 | rdev->config.rv770.max_gs_threads = 16 * 2; | |
585 | rdev->config.rv770.sx_max_export_size = 256; | |
586 | rdev->config.rv770.sx_max_export_pos_size = 32; | |
587 | rdev->config.rv770.sx_max_export_smx_size = 224; | |
588 | rdev->config.rv770.sq_num_cf_insts = 2; | |
589 | ||
590 | rdev->config.rv770.sx_num_of_sets = 7; | |
591 | rdev->config.rv770.sc_prim_fifo_size = 0x100; | |
592 | rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30; | |
593 | rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130; | |
594 | ||
595 | if (rdev->config.rv770.sx_max_export_pos_size > 16) { | |
596 | rdev->config.rv770.sx_max_export_pos_size -= 16; | |
597 | rdev->config.rv770.sx_max_export_smx_size += 16; | |
598 | } | |
599 | break; | |
600 | default: | |
601 | break; | |
602 | } | |
603 | ||
604 | /* Initialize HDP */ | |
605 | j = 0; | |
606 | for (i = 0; i < 32; i++) { | |
607 | WREG32((0x2c14 + j), 0x00000000); | |
608 | WREG32((0x2c18 + j), 0x00000000); | |
609 | WREG32((0x2c1c + j), 0x00000000); | |
610 | WREG32((0x2c20 + j), 0x00000000); | |
611 | WREG32((0x2c24 + j), 0x00000000); | |
612 | j += 0x18; | |
613 | } | |
614 | ||
615 | WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff)); | |
616 | ||
617 | /* setup tiling, simd, pipe config */ | |
618 | mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG); | |
619 | ||
620 | switch (rdev->config.rv770.max_tile_pipes) { | |
621 | case 1: | |
d03f5d59 | 622 | default: |
3ce0a23d JG |
623 | gb_tiling_config |= PIPE_TILING(0); |
624 | break; | |
625 | case 2: | |
626 | gb_tiling_config |= PIPE_TILING(1); | |
627 | break; | |
628 | case 4: | |
629 | gb_tiling_config |= PIPE_TILING(2); | |
630 | break; | |
631 | case 8: | |
632 | gb_tiling_config |= PIPE_TILING(3); | |
3ce0a23d JG |
633 | break; |
634 | } | |
d03f5d59 | 635 | rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes; |
3ce0a23d JG |
636 | |
637 | if (rdev->family == CHIP_RV770) | |
638 | gb_tiling_config |= BANK_TILING(1); | |
639 | else | |
e29649db | 640 | gb_tiling_config |= BANK_TILING((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT); |
961fb597 | 641 | rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3); |
3ce0a23d JG |
642 | |
643 | gb_tiling_config |= GROUP_SIZE(0); | |
961fb597 | 644 | rdev->config.rv770.tiling_group_size = 256; |
3ce0a23d | 645 | |
e29649db | 646 | if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) { |
3ce0a23d JG |
647 | gb_tiling_config |= ROW_TILING(3); |
648 | gb_tiling_config |= SAMPLE_SPLIT(3); | |
649 | } else { | |
650 | gb_tiling_config |= | |
651 | ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
652 | gb_tiling_config |= | |
653 | SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT)); | |
654 | } | |
655 | ||
656 | gb_tiling_config |= BANK_SWAPS(1); | |
657 | ||
d03f5d59 AD |
658 | cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000; |
659 | cc_rb_backend_disable |= | |
660 | BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << rdev->config.rv770.max_backends) & R7XX_MAX_BACKENDS_MASK); | |
3ce0a23d | 661 | |
d03f5d59 AD |
662 | cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00; |
663 | cc_gc_shader_pipe_config |= | |
3ce0a23d JG |
664 | INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << rdev->config.rv770.max_pipes) & R7XX_MAX_PIPES_MASK); |
665 | cc_gc_shader_pipe_config |= | |
666 | INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << rdev->config.rv770.max_simds) & R7XX_MAX_SIMDS_MASK); | |
667 | ||
d03f5d59 AD |
668 | if (rdev->family == CHIP_RV740) |
669 | backend_map = 0x28; | |
670 | else | |
671 | backend_map = r700_get_tile_pipe_to_backend_map(rdev, | |
672 | rdev->config.rv770.max_tile_pipes, | |
673 | (R7XX_MAX_BACKENDS - | |
674 | r600_count_pipe_bits((cc_rb_backend_disable & | |
675 | R7XX_MAX_BACKENDS_MASK) >> 16)), | |
676 | (cc_rb_backend_disable >> 16)); | |
d03f5d59 | 677 | |
e7aeeba6 AD |
678 | rdev->config.rv770.tile_config = gb_tiling_config; |
679 | gb_tiling_config |= BACKEND_MAP(backend_map); | |
3ce0a23d JG |
680 | |
681 | WREG32(GB_TILING_CONFIG, gb_tiling_config); | |
682 | WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
683 | WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff)); | |
684 | ||
685 | WREG32(CC_RB_BACKEND_DISABLE, cc_rb_backend_disable); | |
686 | WREG32(CC_GC_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); | |
f867c60d | 687 | WREG32(GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config); |
d03f5d59 | 688 | WREG32(CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable); |
3ce0a23d | 689 | |
3ce0a23d JG |
690 | WREG32(CGTS_SYS_TCC_DISABLE, 0); |
691 | WREG32(CGTS_TCC_DISABLE, 0); | |
f867c60d AD |
692 | WREG32(CGTS_USER_SYS_TCC_DISABLE, 0); |
693 | WREG32(CGTS_USER_TCC_DISABLE, 0); | |
3ce0a23d JG |
694 | |
695 | num_qd_pipes = | |
d03f5d59 | 696 | R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8); |
3ce0a23d JG |
697 | WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK); |
698 | WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK); | |
699 | ||
700 | /* set HW defaults for 3D engine */ | |
701 | WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) | | |
e29649db | 702 | ROQ_IB2_START(0x2b))); |
3ce0a23d JG |
703 | |
704 | WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30)); | |
705 | ||
d03f5d59 AD |
706 | ta_aux_cntl = RREG32(TA_CNTL_AUX); |
707 | WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO); | |
3ce0a23d JG |
708 | |
709 | sx_debug_1 = RREG32(SX_DEBUG_1); | |
710 | sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS; | |
711 | WREG32(SX_DEBUG_1, sx_debug_1); | |
712 | ||
713 | smx_dc_ctl0 = RREG32(SMX_DC_CTL0); | |
714 | smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff); | |
715 | smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1); | |
716 | WREG32(SMX_DC_CTL0, smx_dc_ctl0); | |
717 | ||
d03f5d59 AD |
718 | if (rdev->family != CHIP_RV740) |
719 | WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) | | |
720 | GS_FLUSH_CTL(4) | | |
721 | ACK_FLUSH_CTL(3) | | |
722 | SYNC_FLUSH_CTL)); | |
3ce0a23d | 723 | |
d03f5d59 AD |
724 | db_debug3 = RREG32(DB_DEBUG3); |
725 | db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f); | |
726 | switch (rdev->family) { | |
727 | case CHIP_RV770: | |
728 | case CHIP_RV740: | |
729 | db_debug3 |= DB_CLK_OFF_DELAY(0x1f); | |
730 | break; | |
731 | case CHIP_RV710: | |
732 | case CHIP_RV730: | |
733 | default: | |
734 | db_debug3 |= DB_CLK_OFF_DELAY(2); | |
735 | break; | |
736 | } | |
737 | WREG32(DB_DEBUG3, db_debug3); | |
738 | ||
739 | if (rdev->family != CHIP_RV770) { | |
3ce0a23d JG |
740 | db_debug4 = RREG32(DB_DEBUG4); |
741 | db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER; | |
742 | WREG32(DB_DEBUG4, db_debug4); | |
743 | } | |
744 | ||
745 | WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) | | |
e29649db AD |
746 | POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) | |
747 | SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1))); | |
3ce0a23d JG |
748 | |
749 | WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) | | |
e29649db AD |
750 | SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) | |
751 | SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize))); | |
3ce0a23d JG |
752 | |
753 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
754 | ||
755 | WREG32(VGT_NUM_INSTANCES, 1); | |
756 | ||
757 | WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0)); | |
758 | ||
759 | WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4)); | |
760 | ||
761 | WREG32(CP_PERFMON_CNTL, 0); | |
762 | ||
763 | sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) | | |
764 | DONE_FIFO_HIWATER(0xe0) | | |
765 | ALU_UPDATE_FIFO_HIWATER(0x8)); | |
766 | switch (rdev->family) { | |
767 | case CHIP_RV770: | |
3ce0a23d JG |
768 | case CHIP_RV730: |
769 | case CHIP_RV710: | |
d03f5d59 AD |
770 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1); |
771 | break; | |
3ce0a23d JG |
772 | case CHIP_RV740: |
773 | default: | |
774 | sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4); | |
775 | break; | |
776 | } | |
777 | WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes); | |
778 | ||
779 | /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT | |
780 | * should be adjusted as needed by the 2D/3D drivers. This just sets default values | |
781 | */ | |
782 | sq_config = RREG32(SQ_CONFIG); | |
783 | sq_config &= ~(PS_PRIO(3) | | |
784 | VS_PRIO(3) | | |
785 | GS_PRIO(3) | | |
786 | ES_PRIO(3)); | |
787 | sq_config |= (DX9_CONSTS | | |
788 | VC_ENABLE | | |
789 | EXPORT_SRC_C | | |
790 | PS_PRIO(0) | | |
791 | VS_PRIO(1) | | |
792 | GS_PRIO(2) | | |
793 | ES_PRIO(3)); | |
794 | if (rdev->family == CHIP_RV710) | |
795 | /* no vertex cache */ | |
796 | sq_config &= ~VC_ENABLE; | |
797 | ||
798 | WREG32(SQ_CONFIG, sq_config); | |
799 | ||
800 | WREG32(SQ_GPR_RESOURCE_MGMT_1, (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | | |
fe62e1a4 DA |
801 | NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) | |
802 | NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2))); | |
3ce0a23d JG |
803 | |
804 | WREG32(SQ_GPR_RESOURCE_MGMT_2, (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) | | |
fe62e1a4 | 805 | NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64))); |
3ce0a23d JG |
806 | |
807 | sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) | | |
808 | NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) | | |
809 | NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8)); | |
810 | if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads) | |
811 | sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads); | |
812 | else | |
813 | sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8); | |
814 | WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt); | |
815 | ||
816 | WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
817 | NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
818 | ||
819 | WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) | | |
820 | NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4))); | |
821 | ||
822 | sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
823 | SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) | | |
824 | SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) | | |
825 | SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64)); | |
826 | ||
827 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0); | |
828 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0); | |
829 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0); | |
830 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0); | |
831 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0); | |
832 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0); | |
833 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0); | |
834 | WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0); | |
835 | ||
836 | WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) | | |
fe62e1a4 | 837 | FORCE_EOV_MAX_REZ_CNT(255))); |
3ce0a23d JG |
838 | |
839 | if (rdev->family == CHIP_RV710) | |
840 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) | | |
fe62e1a4 | 841 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
842 | else |
843 | WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) | | |
fe62e1a4 | 844 | AUTO_INVLD_EN(ES_AND_GS_AUTO))); |
3ce0a23d JG |
845 | |
846 | switch (rdev->family) { | |
847 | case CHIP_RV770: | |
848 | case CHIP_RV730: | |
849 | case CHIP_RV740: | |
850 | gs_prim_buffer_depth = 384; | |
851 | break; | |
852 | case CHIP_RV710: | |
853 | gs_prim_buffer_depth = 128; | |
854 | break; | |
855 | default: | |
856 | break; | |
857 | } | |
858 | ||
859 | num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16; | |
860 | vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread; | |
861 | /* Max value for this is 256 */ | |
862 | if (vgt_gs_per_es > 256) | |
863 | vgt_gs_per_es = 256; | |
864 | ||
865 | WREG32(VGT_ES_PER_GS, 128); | |
866 | WREG32(VGT_GS_PER_ES, vgt_gs_per_es); | |
867 | WREG32(VGT_GS_PER_VS, 2); | |
868 | ||
869 | /* more default values. 2D/3D driver should adjust as needed */ | |
870 | WREG32(VGT_GS_VERTEX_REUSE, 16); | |
871 | WREG32(PA_SC_LINE_STIPPLE_STATE, 0); | |
872 | WREG32(VGT_STRMOUT_EN, 0); | |
873 | WREG32(SX_MISC, 0); | |
874 | WREG32(PA_SC_MODE_CNTL, 0); | |
875 | WREG32(PA_SC_EDGERULE, 0xaaaaaaaa); | |
876 | WREG32(PA_SC_AA_CONFIG, 0); | |
877 | WREG32(PA_SC_CLIPRECT_RULE, 0xffff); | |
878 | WREG32(PA_SC_LINE_STIPPLE, 0); | |
879 | WREG32(SPI_INPUT_Z, 0); | |
880 | WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2)); | |
881 | WREG32(CB_COLOR7_FRAG, 0); | |
882 | ||
883 | /* clear render buffer base addresses */ | |
884 | WREG32(CB_COLOR0_BASE, 0); | |
885 | WREG32(CB_COLOR1_BASE, 0); | |
886 | WREG32(CB_COLOR2_BASE, 0); | |
887 | WREG32(CB_COLOR3_BASE, 0); | |
888 | WREG32(CB_COLOR4_BASE, 0); | |
889 | WREG32(CB_COLOR5_BASE, 0); | |
890 | WREG32(CB_COLOR6_BASE, 0); | |
891 | WREG32(CB_COLOR7_BASE, 0); | |
892 | ||
893 | WREG32(TCP_CNTL, 0); | |
894 | ||
895 | hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL); | |
896 | WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl); | |
897 | ||
898 | WREG32(PA_SC_MULTI_CHIP_CNTL, 0); | |
899 | ||
900 | WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA | | |
901 | NUM_CLIP_SEQ(3))); | |
902 | ||
903 | } | |
904 | ||
905 | int rv770_mc_init(struct radeon_device *rdev) | |
906 | { | |
3ce0a23d | 907 | u32 tmp; |
5885b7a9 | 908 | int chansize, numchan; |
3ce0a23d JG |
909 | |
910 | /* Get VRAM informations */ | |
3ce0a23d | 911 | rdev->mc.vram_is_ddr = true; |
5885b7a9 AD |
912 | tmp = RREG32(MC_ARB_RAMCFG); |
913 | if (tmp & CHANSIZE_OVERRIDE) { | |
914 | chansize = 16; | |
915 | } else if (tmp & CHANSIZE_MASK) { | |
916 | chansize = 64; | |
917 | } else { | |
918 | chansize = 32; | |
919 | } | |
920 | tmp = RREG32(MC_SHARED_CHMAP); | |
921 | switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) { | |
922 | case 0: | |
923 | default: | |
924 | numchan = 1; | |
925 | break; | |
926 | case 1: | |
927 | numchan = 2; | |
928 | break; | |
929 | case 2: | |
930 | numchan = 4; | |
931 | break; | |
932 | case 3: | |
933 | numchan = 8; | |
934 | break; | |
935 | } | |
936 | rdev->mc.vram_width = numchan * chansize; | |
771fe6b9 | 937 | /* Could aper size report 0 ? */ |
01d73a69 JC |
938 | rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0); |
939 | rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0); | |
3ce0a23d JG |
940 | /* Setup GPU memory space */ |
941 | rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE); | |
942 | rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE); | |
51e5fcd3 | 943 | rdev->mc.visible_vram_size = rdev->mc.aper_size; |
d594e46a | 944 | r600_vram_gtt_location(rdev, &rdev->mc); |
f47299c5 AD |
945 | radeon_update_bandwidth_info(rdev); |
946 | ||
3ce0a23d JG |
947 | return 0; |
948 | } | |
d594e46a | 949 | |
fc30b8ef | 950 | static int rv770_startup(struct radeon_device *rdev) |
3ce0a23d JG |
951 | { |
952 | int r; | |
953 | ||
779720a3 AD |
954 | if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) { |
955 | r = r600_init_microcode(rdev); | |
956 | if (r) { | |
957 | DRM_ERROR("Failed to load firmware!\n"); | |
958 | return r; | |
959 | } | |
960 | } | |
961 | ||
a3c1945a | 962 | rv770_mc_program(rdev); |
1a029b76 JG |
963 | if (rdev->flags & RADEON_IS_AGP) { |
964 | rv770_agp_enable(rdev); | |
965 | } else { | |
966 | r = rv770_pcie_gart_enable(rdev); | |
967 | if (r) | |
968 | return r; | |
969 | } | |
3ce0a23d | 970 | rv770_gpu_init(rdev); |
c38c7b64 JG |
971 | r = r600_blit_init(rdev); |
972 | if (r) { | |
973 | r600_blit_fini(rdev); | |
974 | rdev->asic->copy = NULL; | |
975 | dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r); | |
976 | } | |
ff82f052 JG |
977 | /* pin copy shader into vram */ |
978 | if (rdev->r600_blit.shader_obj) { | |
979 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
980 | if (unlikely(r != 0)) | |
981 | return r; | |
982 | r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM, | |
983 | &rdev->r600_blit.shader_gpu_addr); | |
984 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
7923c615 | 985 | if (r) { |
ff82f052 | 986 | DRM_ERROR("failed to pin blit object %d\n", r); |
7923c615 AD |
987 | return r; |
988 | } | |
989 | } | |
d8f60cfc | 990 | /* Enable IRQ */ |
d8f60cfc AD |
991 | r = r600_irq_init(rdev); |
992 | if (r) { | |
993 | DRM_ERROR("radeon: IH init failed (%d).\n", r); | |
994 | radeon_irq_kms_fini(rdev); | |
995 | return r; | |
996 | } | |
997 | r600_irq_set(rdev); | |
998 | ||
3ce0a23d JG |
999 | r = radeon_ring_init(rdev, rdev->cp.ring_size); |
1000 | if (r) | |
1001 | return r; | |
1002 | r = rv770_cp_load_microcode(rdev); | |
1003 | if (r) | |
1004 | return r; | |
1005 | r = r600_cp_resume(rdev); | |
1006 | if (r) | |
1007 | return r; | |
81cc35bf JG |
1008 | /* write back buffer are not vital so don't worry about failure */ |
1009 | r600_wb_enable(rdev); | |
3ce0a23d JG |
1010 | return 0; |
1011 | } | |
1012 | ||
fc30b8ef DA |
1013 | int rv770_resume(struct radeon_device *rdev) |
1014 | { | |
1015 | int r; | |
1016 | ||
1a029b76 JG |
1017 | /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw, |
1018 | * posting will perform necessary task to bring back GPU into good | |
1019 | * shape. | |
1020 | */ | |
fc30b8ef | 1021 | /* post card */ |
e7d40b9a | 1022 | atom_asic_init(rdev->mode_info.atom_context); |
fc30b8ef DA |
1023 | /* Initialize clocks */ |
1024 | r = radeon_clocks_init(rdev); | |
1025 | if (r) { | |
1026 | return r; | |
1027 | } | |
1028 | ||
1029 | r = rv770_startup(rdev); | |
1030 | if (r) { | |
1031 | DRM_ERROR("r600 startup failed on resume\n"); | |
1032 | return r; | |
1033 | } | |
1034 | ||
62a8ea3f | 1035 | r = r600_ib_test(rdev); |
fc30b8ef DA |
1036 | if (r) { |
1037 | DRM_ERROR("radeon: failled testing IB (%d).\n", r); | |
1038 | return r; | |
1039 | } | |
8a8c6e7c RM |
1040 | |
1041 | r = r600_audio_init(rdev); | |
1042 | if (r) { | |
1043 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1044 | return r; | |
1045 | } | |
1046 | ||
fc30b8ef DA |
1047 | return r; |
1048 | ||
1049 | } | |
1050 | ||
3ce0a23d JG |
1051 | int rv770_suspend(struct radeon_device *rdev) |
1052 | { | |
4c788679 JG |
1053 | int r; |
1054 | ||
8a8c6e7c | 1055 | r600_audio_fini(rdev); |
3ce0a23d JG |
1056 | /* FIXME: we should wait for ring to be empty */ |
1057 | r700_cp_stop(rdev); | |
4153e584 | 1058 | rdev->cp.ready = false; |
0c45249f | 1059 | r600_irq_suspend(rdev); |
81cc35bf | 1060 | r600_wb_disable(rdev); |
4aac0473 | 1061 | rv770_pcie_gart_disable(rdev); |
4153e584 | 1062 | /* unpin shaders bo */ |
30d2d9a5 JG |
1063 | if (rdev->r600_blit.shader_obj) { |
1064 | r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false); | |
1065 | if (likely(r == 0)) { | |
1066 | radeon_bo_unpin(rdev->r600_blit.shader_obj); | |
1067 | radeon_bo_unreserve(rdev->r600_blit.shader_obj); | |
1068 | } | |
4c788679 | 1069 | } |
3ce0a23d JG |
1070 | return 0; |
1071 | } | |
1072 | ||
1073 | /* Plan is to move initialization in that function and use | |
1074 | * helper function so that radeon_device_init pretty much | |
1075 | * do nothing more than calling asic specific function. This | |
1076 | * should also allow to remove a bunch of callback function | |
1077 | * like vram_info. | |
1078 | */ | |
1079 | int rv770_init(struct radeon_device *rdev) | |
1080 | { | |
1081 | int r; | |
1082 | ||
3ce0a23d JG |
1083 | r = radeon_dummy_page_init(rdev); |
1084 | if (r) | |
1085 | return r; | |
1086 | /* This don't do much */ | |
1087 | r = radeon_gem_init(rdev); | |
1088 | if (r) | |
1089 | return r; | |
1090 | /* Read BIOS */ | |
1091 | if (!radeon_get_bios(rdev)) { | |
1092 | if (ASIC_IS_AVIVO(rdev)) | |
1093 | return -EINVAL; | |
1094 | } | |
1095 | /* Must be an ATOMBIOS */ | |
e7d40b9a JG |
1096 | if (!rdev->is_atom_bios) { |
1097 | dev_err(rdev->dev, "Expecting atombios for R600 GPU\n"); | |
3ce0a23d | 1098 | return -EINVAL; |
e7d40b9a | 1099 | } |
3ce0a23d JG |
1100 | r = radeon_atombios_init(rdev); |
1101 | if (r) | |
1102 | return r; | |
1103 | /* Post card if necessary */ | |
72542d77 DA |
1104 | if (!r600_card_posted(rdev)) { |
1105 | if (!rdev->bios) { | |
1106 | dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n"); | |
1107 | return -EINVAL; | |
1108 | } | |
3ce0a23d JG |
1109 | DRM_INFO("GPU not posted. posting now...\n"); |
1110 | atom_asic_init(rdev->mode_info.atom_context); | |
1111 | } | |
1112 | /* Initialize scratch registers */ | |
1113 | r600_scratch_init(rdev); | |
1114 | /* Initialize surface registers */ | |
1115 | radeon_surface_init(rdev); | |
7433874e | 1116 | /* Initialize clocks */ |
5e6dde7e | 1117 | radeon_get_clock_info(rdev->ddev); |
3ce0a23d JG |
1118 | r = radeon_clocks_init(rdev); |
1119 | if (r) | |
1120 | return r; | |
1121 | /* Fence driver */ | |
1122 | r = radeon_fence_driver_init(rdev); | |
1123 | if (r) | |
1124 | return r; | |
d594e46a | 1125 | /* initialize AGP */ |
700a0cc0 JG |
1126 | if (rdev->flags & RADEON_IS_AGP) { |
1127 | r = radeon_agp_init(rdev); | |
1128 | if (r) | |
1129 | radeon_agp_disable(rdev); | |
1130 | } | |
3ce0a23d | 1131 | r = rv770_mc_init(rdev); |
b574f251 | 1132 | if (r) |
3ce0a23d | 1133 | return r; |
3ce0a23d | 1134 | /* Memory manager */ |
4c788679 | 1135 | r = radeon_bo_init(rdev); |
3ce0a23d JG |
1136 | if (r) |
1137 | return r; | |
d8f60cfc AD |
1138 | |
1139 | r = radeon_irq_kms_init(rdev); | |
1140 | if (r) | |
1141 | return r; | |
1142 | ||
3ce0a23d JG |
1143 | rdev->cp.ring_obj = NULL; |
1144 | r600_ring_init(rdev, 1024 * 1024); | |
1145 | ||
d8f60cfc AD |
1146 | rdev->ih.ring_obj = NULL; |
1147 | r600_ih_ring_init(rdev, 64 * 1024); | |
1148 | ||
4aac0473 JG |
1149 | r = r600_pcie_gart_init(rdev); |
1150 | if (r) | |
1151 | return r; | |
1152 | ||
779720a3 | 1153 | rdev->accel_working = true; |
fc30b8ef | 1154 | r = rv770_startup(rdev); |
3ce0a23d | 1155 | if (r) { |
655efd3d | 1156 | dev_err(rdev->dev, "disabling GPU acceleration\n"); |
fe251e2f | 1157 | r700_cp_fini(rdev); |
75c81298 | 1158 | r600_wb_fini(rdev); |
655efd3d JG |
1159 | r600_irq_fini(rdev); |
1160 | radeon_irq_kms_fini(rdev); | |
75c81298 | 1161 | rv770_pcie_gart_fini(rdev); |
733289c2 | 1162 | rdev->accel_working = false; |
3ce0a23d | 1163 | } |
733289c2 | 1164 | if (rdev->accel_working) { |
733289c2 JG |
1165 | r = radeon_ib_pool_init(rdev); |
1166 | if (r) { | |
db96380e | 1167 | dev_err(rdev->dev, "IB initialization failed (%d).\n", r); |
733289c2 | 1168 | rdev->accel_working = false; |
db96380e JG |
1169 | } else { |
1170 | r = r600_ib_test(rdev); | |
1171 | if (r) { | |
1172 | dev_err(rdev->dev, "IB test failed (%d).\n", r); | |
1173 | rdev->accel_working = false; | |
1174 | } | |
733289c2 | 1175 | } |
3ce0a23d | 1176 | } |
8a8c6e7c RM |
1177 | |
1178 | r = r600_audio_init(rdev); | |
1179 | if (r) { | |
1180 | dev_err(rdev->dev, "radeon: audio init failed\n"); | |
1181 | return r; | |
1182 | } | |
1183 | ||
3ce0a23d JG |
1184 | return 0; |
1185 | } | |
1186 | ||
1187 | void rv770_fini(struct radeon_device *rdev) | |
1188 | { | |
1189 | r600_blit_fini(rdev); | |
fe251e2f | 1190 | r700_cp_fini(rdev); |
655efd3d | 1191 | r600_wb_fini(rdev); |
d8f60cfc AD |
1192 | r600_irq_fini(rdev); |
1193 | radeon_irq_kms_fini(rdev); | |
4aac0473 | 1194 | rv770_pcie_gart_fini(rdev); |
3ce0a23d JG |
1195 | radeon_gem_fini(rdev); |
1196 | radeon_fence_driver_fini(rdev); | |
1197 | radeon_clocks_fini(rdev); | |
d0269ed8 | 1198 | radeon_agp_fini(rdev); |
4c788679 | 1199 | radeon_bo_fini(rdev); |
e7d40b9a | 1200 | radeon_atombios_fini(rdev); |
3ce0a23d JG |
1201 | kfree(rdev->bios); |
1202 | rdev->bios = NULL; | |
1203 | radeon_dummy_page_fini(rdev); | |
771fe6b9 | 1204 | } |