drm/radeon/kms: fix bandwidth computation on avivo hardware
[safe/jmp/linux-2.6] / drivers / gpu / drm / radeon / rs690.c
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  */
28 #include "drmP.h"
29 #include "radeon_reg.h"
30 #include "radeon.h"
31 #include "rs690r.h"
32 #include "atom.h"
33 #include "atom-bits.h"
34
35 /* rs690,rs740 depends on : */
36 void r100_hdp_reset(struct radeon_device *rdev);
37 int r300_mc_wait_for_idle(struct radeon_device *rdev);
38 void r420_pipes_init(struct radeon_device *rdev);
39 void rs400_gart_disable(struct radeon_device *rdev);
40 int rs400_gart_enable(struct radeon_device *rdev);
41 void rs400_gart_adjust_size(struct radeon_device *rdev);
42 void rs600_mc_disable_clients(struct radeon_device *rdev);
43 void rs600_disable_vga(struct radeon_device *rdev);
44
45 /* This files gather functions specifics to :
46  * rs690,rs740
47  *
48  * Some of these functions might be used by newer ASICs.
49  */
50 void rs690_gpu_init(struct radeon_device *rdev);
51 int rs690_mc_wait_for_idle(struct radeon_device *rdev);
52
53
54 /*
55  * MC functions.
56  */
57 int rs690_mc_init(struct radeon_device *rdev)
58 {
59         uint32_t tmp;
60         int r;
61
62         if (r100_debugfs_rbbm_init(rdev)) {
63                 DRM_ERROR("Failed to register debugfs file for RBBM !\n");
64         }
65
66         rs690_gpu_init(rdev);
67         rs400_gart_disable(rdev);
68
69         /* Setup GPU memory space */
70         rdev->mc.gtt_location = rdev->mc.vram_size;
71         rdev->mc.gtt_location += (rdev->mc.gtt_size - 1);
72         rdev->mc.gtt_location &= ~(rdev->mc.gtt_size - 1);
73         rdev->mc.vram_location = 0xFFFFFFFFUL;
74         r = radeon_mc_setup(rdev);
75         if (r) {
76                 return r;
77         }
78
79         /* Program GPU memory space */
80         rs600_mc_disable_clients(rdev);
81         if (rs690_mc_wait_for_idle(rdev)) {
82                 printk(KERN_WARNING "Failed to wait MC idle while "
83                        "programming pipes. Bad things might happen.\n");
84         }
85         tmp = rdev->mc.vram_location + rdev->mc.vram_size - 1;
86         tmp = REG_SET(RS690_MC_FB_TOP, tmp >> 16);
87         tmp |= REG_SET(RS690_MC_FB_START, rdev->mc.vram_location >> 16);
88         WREG32_MC(RS690_MCCFG_FB_LOCATION, tmp);
89         /* FIXME: Does this reg exist on RS480,RS740 ? */
90         WREG32(0x310, rdev->mc.vram_location);
91         WREG32(RS690_HDP_FB_LOCATION, rdev->mc.vram_location >> 16);
92         return 0;
93 }
94
95 void rs690_mc_fini(struct radeon_device *rdev)
96 {
97         rs400_gart_disable(rdev);
98         radeon_gart_table_ram_free(rdev);
99         radeon_gart_fini(rdev);
100 }
101
102
103 /*
104  * Global GPU functions
105  */
106 int rs690_mc_wait_for_idle(struct radeon_device *rdev)
107 {
108         unsigned i;
109         uint32_t tmp;
110
111         for (i = 0; i < rdev->usec_timeout; i++) {
112                 /* read MC_STATUS */
113                 tmp = RREG32_MC(RS690_MC_STATUS);
114                 if (tmp & RS690_MC_STATUS_IDLE) {
115                         return 0;
116                 }
117                 DRM_UDELAY(1);
118         }
119         return -1;
120 }
121
122 void rs690_errata(struct radeon_device *rdev)
123 {
124         rdev->pll_errata = 0;
125 }
126
127 void rs690_gpu_init(struct radeon_device *rdev)
128 {
129         /* FIXME: HDP same place on rs690 ? */
130         r100_hdp_reset(rdev);
131         rs600_disable_vga(rdev);
132         /* FIXME: is this correct ? */
133         r420_pipes_init(rdev);
134         if (rs690_mc_wait_for_idle(rdev)) {
135                 printk(KERN_WARNING "Failed to wait MC idle while "
136                        "programming pipes. Bad things might happen.\n");
137         }
138 }
139
140
141 /*
142  * VRAM info.
143  */
144 void rs690_pm_info(struct radeon_device *rdev)
145 {
146         int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
147         struct _ATOM_INTEGRATED_SYSTEM_INFO *info;
148         struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *info_v2;
149         void *ptr;
150         uint16_t data_offset;
151         uint8_t frev, crev;
152         fixed20_12 tmp;
153
154         atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
155                                &frev, &crev, &data_offset);
156         ptr = rdev->mode_info.atom_context->bios + data_offset;
157         info = (struct _ATOM_INTEGRATED_SYSTEM_INFO *)ptr;
158         info_v2 = (struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *)ptr;
159         /* Get various system informations from bios */
160         switch (crev) {
161         case 1:
162                 tmp.full = rfixed_const(100);
163                 rdev->pm.igp_sideport_mclk.full = rfixed_const(info->ulBootUpMemoryClock);
164                 rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
165                 rdev->pm.igp_system_mclk.full = rfixed_const(le16_to_cpu(info->usK8MemoryClock));
166                 rdev->pm.igp_ht_link_clk.full = rfixed_const(le16_to_cpu(info->usFSBClock));
167                 rdev->pm.igp_ht_link_width.full = rfixed_const(info->ucHTLinkWidth);
168                 break;
169         case 2:
170                 tmp.full = rfixed_const(100);
171                 rdev->pm.igp_sideport_mclk.full = rfixed_const(info_v2->ulBootUpSidePortClock);
172                 rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
173                 rdev->pm.igp_system_mclk.full = rfixed_const(info_v2->ulBootUpUMAClock);
174                 rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
175                 rdev->pm.igp_ht_link_clk.full = rfixed_const(info_v2->ulHTLinkFreq);
176                 rdev->pm.igp_ht_link_clk.full = rfixed_div(rdev->pm.igp_ht_link_clk, tmp);
177                 rdev->pm.igp_ht_link_width.full = rfixed_const(le16_to_cpu(info_v2->usMinHTLinkWidth));
178                 break;
179         default:
180                 tmp.full = rfixed_const(100);
181                 /* We assume the slower possible clock ie worst case */
182                 /* DDR 333Mhz */
183                 rdev->pm.igp_sideport_mclk.full = rfixed_const(333);
184                 /* FIXME: system clock ? */
185                 rdev->pm.igp_system_mclk.full = rfixed_const(100);
186                 rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
187                 rdev->pm.igp_ht_link_clk.full = rfixed_const(200);
188                 rdev->pm.igp_ht_link_width.full = rfixed_const(8);
189                 DRM_ERROR("No integrated system info for your GPU, using safe default\n");
190                 break;
191         }
192         /* Compute various bandwidth */
193         /* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
194         tmp.full = rfixed_const(4);
195         rdev->pm.k8_bandwidth.full = rfixed_mul(rdev->pm.igp_system_mclk, tmp);
196         /* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
197          *              = ht_clk * ht_width / 5
198          */
199         tmp.full = rfixed_const(5);
200         rdev->pm.ht_bandwidth.full = rfixed_mul(rdev->pm.igp_ht_link_clk,
201                                                 rdev->pm.igp_ht_link_width);
202         rdev->pm.ht_bandwidth.full = rfixed_div(rdev->pm.ht_bandwidth, tmp);
203         if (tmp.full < rdev->pm.max_bandwidth.full) {
204                 /* HT link is a limiting factor */
205                 rdev->pm.max_bandwidth.full = tmp.full;
206         }
207         /* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
208          *                    = (sideport_clk * 14) / 10
209          */
210         tmp.full = rfixed_const(14);
211         rdev->pm.sideport_bandwidth.full = rfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
212         tmp.full = rfixed_const(10);
213         rdev->pm.sideport_bandwidth.full = rfixed_div(rdev->pm.sideport_bandwidth, tmp);
214 }
215
216 void rs690_vram_info(struct radeon_device *rdev)
217 {
218         uint32_t tmp;
219         fixed20_12 a;
220
221         rs400_gart_adjust_size(rdev);
222         /* DDR for all card after R300 & IGP */
223         rdev->mc.vram_is_ddr = true;
224         /* FIXME: is this correct for RS690/RS740 ? */
225         tmp = RREG32(RADEON_MEM_CNTL);
226         if (tmp & R300_MEM_NUM_CHANNELS_MASK) {
227                 rdev->mc.vram_width = 128;
228         } else {
229                 rdev->mc.vram_width = 64;
230         }
231         rdev->mc.vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
232
233         rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
234         rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
235         rs690_pm_info(rdev);
236         /* FIXME: we should enforce default clock in case GPU is not in
237          * default setup
238          */
239         a.full = rfixed_const(100);
240         rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
241         rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
242         a.full = rfixed_const(16);
243         /* core_bandwidth = sclk(Mhz) * 16 */
244         rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
245 }
246
247 void rs690_line_buffer_adjust(struct radeon_device *rdev,
248                               struct drm_display_mode *mode1,
249                               struct drm_display_mode *mode2)
250 {
251         u32 tmp;
252
253         /*
254          * Line Buffer Setup
255          * There is a single line buffer shared by both display controllers.
256          * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
257          * the display controllers.  The paritioning can either be done
258          * manually or via one of four preset allocations specified in bits 1:0:
259          *  0 - line buffer is divided in half and shared between crtc
260          *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
261          *  2 - D1 gets the whole buffer
262          *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
263          * Setting bit 2 of DC_LB_MEMORY_SPLIT controls switches to manual
264          * allocation mode. In manual allocation mode, D1 always starts at 0,
265          * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
266          */
267         tmp = RREG32(DC_LB_MEMORY_SPLIT) & ~DC_LB_MEMORY_SPLIT_MASK;
268         tmp &= ~DC_LB_MEMORY_SPLIT_SHIFT_MODE;
269         /* auto */
270         if (mode1 && mode2) {
271                 if (mode1->hdisplay > mode2->hdisplay) {
272                         if (mode1->hdisplay > 2560)
273                                 tmp |= DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
274                         else
275                                 tmp |= DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
276                 } else if (mode2->hdisplay > mode1->hdisplay) {
277                         if (mode2->hdisplay > 2560)
278                                 tmp |= DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
279                         else
280                                 tmp |= DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
281                 } else
282                         tmp |= AVIVO_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
283         } else if (mode1) {
284                 tmp |= DC_LB_MEMORY_SPLIT_D1_ONLY;
285         } else if (mode2) {
286                 tmp |= DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
287         }
288         WREG32(DC_LB_MEMORY_SPLIT, tmp);
289 }
290
291 struct rs690_watermark {
292         u32        lb_request_fifo_depth;
293         fixed20_12 num_line_pair;
294         fixed20_12 estimated_width;
295         fixed20_12 worst_case_latency;
296         fixed20_12 consumption_rate;
297         fixed20_12 active_time;
298         fixed20_12 dbpp;
299         fixed20_12 priority_mark_max;
300         fixed20_12 priority_mark;
301         fixed20_12 sclk;
302 };
303
304 void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
305                                   struct radeon_crtc *crtc,
306                                   struct rs690_watermark *wm)
307 {
308         struct drm_display_mode *mode = &crtc->base.mode;
309         fixed20_12 a, b, c;
310         fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
311         fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
312         /* FIXME: detect IGP with sideport memory, i don't think there is any
313          * such product available
314          */
315         bool sideport = false;
316
317         if (!crtc->base.enabled) {
318                 /* FIXME: wouldn't it better to set priority mark to maximum */
319                 wm->lb_request_fifo_depth = 4;
320                 return;
321         }
322
323         if (crtc->vsc.full > rfixed_const(2))
324                 wm->num_line_pair.full = rfixed_const(2);
325         else
326                 wm->num_line_pair.full = rfixed_const(1);
327
328         b.full = rfixed_const(mode->crtc_hdisplay);
329         c.full = rfixed_const(256);
330         a.full = rfixed_mul(wm->num_line_pair, b);
331         request_fifo_depth.full = rfixed_div(a, c);
332         if (a.full < rfixed_const(4)) {
333                 wm->lb_request_fifo_depth = 4;
334         } else {
335                 wm->lb_request_fifo_depth = rfixed_trunc(request_fifo_depth);
336         }
337
338         /* Determine consumption rate
339          *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
340          *  vtaps = number of vertical taps,
341          *  vsc = vertical scaling ratio, defined as source/destination
342          *  hsc = horizontal scaling ration, defined as source/destination
343          */
344         a.full = rfixed_const(mode->clock);
345         b.full = rfixed_const(1000);
346         a.full = rfixed_div(a, b);
347         pclk.full = rfixed_div(b, a);
348         if (crtc->rmx_type != RMX_OFF) {
349                 b.full = rfixed_const(2);
350                 if (crtc->vsc.full > b.full)
351                         b.full = crtc->vsc.full;
352                 b.full = rfixed_mul(b, crtc->hsc);
353                 c.full = rfixed_const(2);
354                 b.full = rfixed_div(b, c);
355                 consumption_time.full = rfixed_div(pclk, b);
356         } else {
357                 consumption_time.full = pclk.full;
358         }
359         a.full = rfixed_const(1);
360         wm->consumption_rate.full = rfixed_div(a, consumption_time);
361
362
363         /* Determine line time
364          *  LineTime = total time for one line of displayhtotal
365          *  LineTime = total number of horizontal pixels
366          *  pclk = pixel clock period(ns)
367          */
368         a.full = rfixed_const(crtc->base.mode.crtc_htotal);
369         line_time.full = rfixed_mul(a, pclk);
370
371         /* Determine active time
372          *  ActiveTime = time of active region of display within one line,
373          *  hactive = total number of horizontal active pixels
374          *  htotal = total number of horizontal pixels
375          */
376         a.full = rfixed_const(crtc->base.mode.crtc_htotal);
377         b.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
378         wm->active_time.full = rfixed_mul(line_time, b);
379         wm->active_time.full = rfixed_div(wm->active_time, a);
380
381         /* Maximun bandwidth is the minimun bandwidth of all component */
382         rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
383         if (sideport) {
384                 if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
385                         rdev->pm.sideport_bandwidth.full)
386                         rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
387                 read_delay_latency.full = rfixed_const(370 * 800 * 1000);
388                 read_delay_latency.full = rfixed_div(read_delay_latency,
389                         rdev->pm.igp_sideport_mclk);
390         } else {
391                 if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
392                         rdev->pm.k8_bandwidth.full)
393                         rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
394                 if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
395                         rdev->pm.ht_bandwidth.full)
396                         rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
397                 read_delay_latency.full = rfixed_const(5000);
398         }
399
400         /* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
401         a.full = rfixed_const(16);
402         rdev->pm.sclk.full = rfixed_mul(rdev->pm.max_bandwidth, a);
403         a.full = rfixed_const(1000);
404         rdev->pm.sclk.full = rfixed_div(a, rdev->pm.sclk);
405         /* Determine chunk time
406          * ChunkTime = the time it takes the DCP to send one chunk of data
407          * to the LB which consists of pipeline delay and inter chunk gap
408          * sclk = system clock(ns)
409          */
410         a.full = rfixed_const(256 * 13);
411         chunk_time.full = rfixed_mul(rdev->pm.sclk, a);
412         a.full = rfixed_const(10);
413         chunk_time.full = rfixed_div(chunk_time, a);
414
415         /* Determine the worst case latency
416          * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
417          * WorstCaseLatency = worst case time from urgent to when the MC starts
418          *                    to return data
419          * READ_DELAY_IDLE_MAX = constant of 1us
420          * ChunkTime = time it takes the DCP to send one chunk of data to the LB
421          *             which consists of pipeline delay and inter chunk gap
422          */
423         if (rfixed_trunc(wm->num_line_pair) > 1) {
424                 a.full = rfixed_const(3);
425                 wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
426                 wm->worst_case_latency.full += read_delay_latency.full;
427         } else {
428                 a.full = rfixed_const(2);
429                 wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
430                 wm->worst_case_latency.full += read_delay_latency.full;
431         }
432
433         /* Determine the tolerable latency
434          * TolerableLatency = Any given request has only 1 line time
435          *                    for the data to be returned
436          * LBRequestFifoDepth = Number of chunk requests the LB can
437          *                      put into the request FIFO for a display
438          *  LineTime = total time for one line of display
439          *  ChunkTime = the time it takes the DCP to send one chunk
440          *              of data to the LB which consists of
441          *  pipeline delay and inter chunk gap
442          */
443         if ((2+wm->lb_request_fifo_depth) >= rfixed_trunc(request_fifo_depth)) {
444                 tolerable_latency.full = line_time.full;
445         } else {
446                 tolerable_latency.full = rfixed_const(wm->lb_request_fifo_depth - 2);
447                 tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
448                 tolerable_latency.full = rfixed_mul(tolerable_latency, chunk_time);
449                 tolerable_latency.full = line_time.full - tolerable_latency.full;
450         }
451         /* We assume worst case 32bits (4 bytes) */
452         wm->dbpp.full = rfixed_const(4 * 8);
453
454         /* Determine the maximum priority mark
455          *  width = viewport width in pixels
456          */
457         a.full = rfixed_const(16);
458         wm->priority_mark_max.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
459         wm->priority_mark_max.full = rfixed_div(wm->priority_mark_max, a);
460
461         /* Determine estimated width */
462         estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
463         estimated_width.full = rfixed_div(estimated_width, consumption_time);
464         if (rfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
465                 wm->priority_mark.full = rfixed_const(10);
466         } else {
467                 a.full = rfixed_const(16);
468                 wm->priority_mark.full = rfixed_div(estimated_width, a);
469                 wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
470         }
471 }
472
473 void rs690_bandwidth_update(struct radeon_device *rdev)
474 {
475         struct drm_display_mode *mode0 = NULL;
476         struct drm_display_mode *mode1 = NULL;
477         struct rs690_watermark wm0;
478         struct rs690_watermark wm1;
479         u32 tmp;
480         fixed20_12 priority_mark02, priority_mark12, fill_rate;
481         fixed20_12 a, b;
482
483         if (rdev->mode_info.crtcs[0]->base.enabled)
484                 mode0 = &rdev->mode_info.crtcs[0]->base.mode;
485         if (rdev->mode_info.crtcs[1]->base.enabled)
486                 mode1 = &rdev->mode_info.crtcs[1]->base.mode;
487         /*
488          * Set display0/1 priority up in the memory controller for
489          * modes if the user specifies HIGH for displaypriority
490          * option.
491          */
492         if (rdev->disp_priority == 2) {
493                 tmp = RREG32_MC(MC_INIT_MISC_LAT_TIMER);
494                 tmp &= ~MC_DISP1R_INIT_LAT_MASK;
495                 tmp &= ~MC_DISP0R_INIT_LAT_MASK;
496                 if (mode1)
497                         tmp |= (1 << MC_DISP1R_INIT_LAT_SHIFT);
498                 if (mode0)
499                         tmp |= (1 << MC_DISP0R_INIT_LAT_SHIFT);
500                 WREG32_MC(MC_INIT_MISC_LAT_TIMER, tmp);
501         }
502         rs690_line_buffer_adjust(rdev, mode0, mode1);
503
504         if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
505                 WREG32(DCP_CONTROL, 0);
506         if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
507                 WREG32(DCP_CONTROL, 2);
508
509         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
510         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
511
512         tmp = (wm0.lb_request_fifo_depth - 1);
513         tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
514         WREG32(LB_MAX_REQ_OUTSTANDING, tmp);
515
516         if (mode0 && mode1) {
517                 if (rfixed_trunc(wm0.dbpp) > 64)
518                         a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
519                 else
520                         a.full = wm0.num_line_pair.full;
521                 if (rfixed_trunc(wm1.dbpp) > 64)
522                         b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
523                 else
524                         b.full = wm1.num_line_pair.full;
525                 a.full += b.full;
526                 fill_rate.full = rfixed_div(wm0.sclk, a);
527                 if (wm0.consumption_rate.full > fill_rate.full) {
528                         b.full = wm0.consumption_rate.full - fill_rate.full;
529                         b.full = rfixed_mul(b, wm0.active_time);
530                         a.full = rfixed_mul(wm0.worst_case_latency,
531                                                 wm0.consumption_rate);
532                         a.full = a.full + b.full;
533                         b.full = rfixed_const(16 * 1000);
534                         priority_mark02.full = rfixed_div(a, b);
535                 } else {
536                         a.full = rfixed_mul(wm0.worst_case_latency,
537                                                 wm0.consumption_rate);
538                         b.full = rfixed_const(16 * 1000);
539                         priority_mark02.full = rfixed_div(a, b);
540                 }
541                 if (wm1.consumption_rate.full > fill_rate.full) {
542                         b.full = wm1.consumption_rate.full - fill_rate.full;
543                         b.full = rfixed_mul(b, wm1.active_time);
544                         a.full = rfixed_mul(wm1.worst_case_latency,
545                                                 wm1.consumption_rate);
546                         a.full = a.full + b.full;
547                         b.full = rfixed_const(16 * 1000);
548                         priority_mark12.full = rfixed_div(a, b);
549                 } else {
550                         a.full = rfixed_mul(wm1.worst_case_latency,
551                                                 wm1.consumption_rate);
552                         b.full = rfixed_const(16 * 1000);
553                         priority_mark12.full = rfixed_div(a, b);
554                 }
555                 if (wm0.priority_mark.full > priority_mark02.full)
556                         priority_mark02.full = wm0.priority_mark.full;
557                 if (rfixed_trunc(priority_mark02) < 0)
558                         priority_mark02.full = 0;
559                 if (wm0.priority_mark_max.full > priority_mark02.full)
560                         priority_mark02.full = wm0.priority_mark_max.full;
561                 if (wm1.priority_mark.full > priority_mark12.full)
562                         priority_mark12.full = wm1.priority_mark.full;
563                 if (rfixed_trunc(priority_mark12) < 0)
564                         priority_mark12.full = 0;
565                 if (wm1.priority_mark_max.full > priority_mark12.full)
566                         priority_mark12.full = wm1.priority_mark_max.full;
567                 WREG32(D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
568                 WREG32(D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
569                 WREG32(D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
570                 WREG32(D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
571         } else if (mode0) {
572                 if (rfixed_trunc(wm0.dbpp) > 64)
573                         a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
574                 else
575                         a.full = wm0.num_line_pair.full;
576                 fill_rate.full = rfixed_div(wm0.sclk, a);
577                 if (wm0.consumption_rate.full > fill_rate.full) {
578                         b.full = wm0.consumption_rate.full - fill_rate.full;
579                         b.full = rfixed_mul(b, wm0.active_time);
580                         a.full = rfixed_mul(wm0.worst_case_latency,
581                                                 wm0.consumption_rate);
582                         a.full = a.full + b.full;
583                         b.full = rfixed_const(16 * 1000);
584                         priority_mark02.full = rfixed_div(a, b);
585                 } else {
586                         a.full = rfixed_mul(wm0.worst_case_latency,
587                                                 wm0.consumption_rate);
588                         b.full = rfixed_const(16 * 1000);
589                         priority_mark02.full = rfixed_div(a, b);
590                 }
591                 if (wm0.priority_mark.full > priority_mark02.full)
592                         priority_mark02.full = wm0.priority_mark.full;
593                 if (rfixed_trunc(priority_mark02) < 0)
594                         priority_mark02.full = 0;
595                 if (wm0.priority_mark_max.full > priority_mark02.full)
596                         priority_mark02.full = wm0.priority_mark_max.full;
597                 WREG32(D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
598                 WREG32(D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
599                 WREG32(D2MODE_PRIORITY_A_CNT, MODE_PRIORITY_OFF);
600                 WREG32(D2MODE_PRIORITY_B_CNT, MODE_PRIORITY_OFF);
601         } else {
602                 if (rfixed_trunc(wm1.dbpp) > 64)
603                         a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
604                 else
605                         a.full = wm1.num_line_pair.full;
606                 fill_rate.full = rfixed_div(wm1.sclk, a);
607                 if (wm1.consumption_rate.full > fill_rate.full) {
608                         b.full = wm1.consumption_rate.full - fill_rate.full;
609                         b.full = rfixed_mul(b, wm1.active_time);
610                         a.full = rfixed_mul(wm1.worst_case_latency,
611                                                 wm1.consumption_rate);
612                         a.full = a.full + b.full;
613                         b.full = rfixed_const(16 * 1000);
614                         priority_mark12.full = rfixed_div(a, b);
615                 } else {
616                         a.full = rfixed_mul(wm1.worst_case_latency,
617                                                 wm1.consumption_rate);
618                         b.full = rfixed_const(16 * 1000);
619                         priority_mark12.full = rfixed_div(a, b);
620                 }
621                 if (wm1.priority_mark.full > priority_mark12.full)
622                         priority_mark12.full = wm1.priority_mark.full;
623                 if (rfixed_trunc(priority_mark12) < 0)
624                         priority_mark12.full = 0;
625                 if (wm1.priority_mark_max.full > priority_mark12.full)
626                         priority_mark12.full = wm1.priority_mark_max.full;
627                 WREG32(D1MODE_PRIORITY_A_CNT, MODE_PRIORITY_OFF);
628                 WREG32(D1MODE_PRIORITY_B_CNT, MODE_PRIORITY_OFF);
629                 WREG32(D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
630                 WREG32(D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
631         }
632 }
633
634 /*
635  * Indirect registers accessor
636  */
637 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
638 {
639         uint32_t r;
640
641         WREG32(RS690_MC_INDEX, (reg & RS690_MC_INDEX_MASK));
642         r = RREG32(RS690_MC_DATA);
643         WREG32(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
644         return r;
645 }
646
647 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
648 {
649         WREG32(RS690_MC_INDEX,
650                RS690_MC_INDEX_WR_EN | ((reg) & RS690_MC_INDEX_MASK));
651         WREG32(RS690_MC_DATA, v);
652         WREG32(RS690_MC_INDEX, RS690_MC_INDEX_WR_ACK);
653 }