rs690.c revision 4aac047323e3082d0866b8ad3784236632105af4
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 : */
36void r100_hdp_reset(struct radeon_device *rdev);
37int r300_mc_wait_for_idle(struct radeon_device *rdev);
38void r420_pipes_init(struct radeon_device *rdev);
39void rs400_gart_disable(struct radeon_device *rdev);
40int rs400_gart_enable(struct radeon_device *rdev);
41void rs400_gart_adjust_size(struct radeon_device *rdev);
42void rs600_mc_disable_clients(struct radeon_device *rdev);
43void 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 */
50void rs690_gpu_init(struct radeon_device *rdev);
51int rs690_mc_wait_for_idle(struct radeon_device *rdev);
52
53
54/*
55 * MC functions.
56 */
57int 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.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.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
95void rs690_mc_fini(struct radeon_device *rdev)
96{
97}
98
99
100/*
101 * Global GPU functions
102 */
103int rs690_mc_wait_for_idle(struct radeon_device *rdev)
104{
105	unsigned i;
106	uint32_t tmp;
107
108	for (i = 0; i < rdev->usec_timeout; i++) {
109		/* read MC_STATUS */
110		tmp = RREG32_MC(RS690_MC_STATUS);
111		if (tmp & RS690_MC_STATUS_IDLE) {
112			return 0;
113		}
114		DRM_UDELAY(1);
115	}
116	return -1;
117}
118
119void rs690_errata(struct radeon_device *rdev)
120{
121	rdev->pll_errata = 0;
122}
123
124void rs690_gpu_init(struct radeon_device *rdev)
125{
126	/* FIXME: HDP same place on rs690 ? */
127	r100_hdp_reset(rdev);
128	rs600_disable_vga(rdev);
129	/* FIXME: is this correct ? */
130	r420_pipes_init(rdev);
131	if (rs690_mc_wait_for_idle(rdev)) {
132		printk(KERN_WARNING "Failed to wait MC idle while "
133		       "programming pipes. Bad things might happen.\n");
134	}
135}
136
137
138/*
139 * VRAM info.
140 */
141void rs690_pm_info(struct radeon_device *rdev)
142{
143	int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
144	struct _ATOM_INTEGRATED_SYSTEM_INFO *info;
145	struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *info_v2;
146	void *ptr;
147	uint16_t data_offset;
148	uint8_t frev, crev;
149	fixed20_12 tmp;
150
151	atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
152			       &frev, &crev, &data_offset);
153	ptr = rdev->mode_info.atom_context->bios + data_offset;
154	info = (struct _ATOM_INTEGRATED_SYSTEM_INFO *)ptr;
155	info_v2 = (struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *)ptr;
156	/* Get various system informations from bios */
157	switch (crev) {
158	case 1:
159		tmp.full = rfixed_const(100);
160		rdev->pm.igp_sideport_mclk.full = rfixed_const(info->ulBootUpMemoryClock);
161		rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
162		rdev->pm.igp_system_mclk.full = rfixed_const(le16_to_cpu(info->usK8MemoryClock));
163		rdev->pm.igp_ht_link_clk.full = rfixed_const(le16_to_cpu(info->usFSBClock));
164		rdev->pm.igp_ht_link_width.full = rfixed_const(info->ucHTLinkWidth);
165		break;
166	case 2:
167		tmp.full = rfixed_const(100);
168		rdev->pm.igp_sideport_mclk.full = rfixed_const(info_v2->ulBootUpSidePortClock);
169		rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
170		rdev->pm.igp_system_mclk.full = rfixed_const(info_v2->ulBootUpUMAClock);
171		rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
172		rdev->pm.igp_ht_link_clk.full = rfixed_const(info_v2->ulHTLinkFreq);
173		rdev->pm.igp_ht_link_clk.full = rfixed_div(rdev->pm.igp_ht_link_clk, tmp);
174		rdev->pm.igp_ht_link_width.full = rfixed_const(le16_to_cpu(info_v2->usMinHTLinkWidth));
175		break;
176	default:
177		tmp.full = rfixed_const(100);
178		/* We assume the slower possible clock ie worst case */
179		/* DDR 333Mhz */
180		rdev->pm.igp_sideport_mclk.full = rfixed_const(333);
181		/* FIXME: system clock ? */
182		rdev->pm.igp_system_mclk.full = rfixed_const(100);
183		rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
184		rdev->pm.igp_ht_link_clk.full = rfixed_const(200);
185		rdev->pm.igp_ht_link_width.full = rfixed_const(8);
186		DRM_ERROR("No integrated system info for your GPU, using safe default\n");
187		break;
188	}
189	/* Compute various bandwidth */
190	/* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
191	tmp.full = rfixed_const(4);
192	rdev->pm.k8_bandwidth.full = rfixed_mul(rdev->pm.igp_system_mclk, tmp);
193	/* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
194	 *              = ht_clk * ht_width / 5
195	 */
196	tmp.full = rfixed_const(5);
197	rdev->pm.ht_bandwidth.full = rfixed_mul(rdev->pm.igp_ht_link_clk,
198						rdev->pm.igp_ht_link_width);
199	rdev->pm.ht_bandwidth.full = rfixed_div(rdev->pm.ht_bandwidth, tmp);
200	if (tmp.full < rdev->pm.max_bandwidth.full) {
201		/* HT link is a limiting factor */
202		rdev->pm.max_bandwidth.full = tmp.full;
203	}
204	/* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
205	 *                    = (sideport_clk * 14) / 10
206	 */
207	tmp.full = rfixed_const(14);
208	rdev->pm.sideport_bandwidth.full = rfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
209	tmp.full = rfixed_const(10);
210	rdev->pm.sideport_bandwidth.full = rfixed_div(rdev->pm.sideport_bandwidth, tmp);
211}
212
213void rs690_vram_info(struct radeon_device *rdev)
214{
215	uint32_t tmp;
216	fixed20_12 a;
217
218	rs400_gart_adjust_size(rdev);
219	/* DDR for all card after R300 & IGP */
220	rdev->mc.vram_is_ddr = true;
221	/* FIXME: is this correct for RS690/RS740 ? */
222	tmp = RREG32(RADEON_MEM_CNTL);
223	if (tmp & R300_MEM_NUM_CHANNELS_MASK) {
224		rdev->mc.vram_width = 128;
225	} else {
226		rdev->mc.vram_width = 64;
227	}
228	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
229	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
230
231	rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
232	rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
233	rs690_pm_info(rdev);
234	/* FIXME: we should enforce default clock in case GPU is not in
235	 * default setup
236	 */
237	a.full = rfixed_const(100);
238	rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
239	rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
240	a.full = rfixed_const(16);
241	/* core_bandwidth = sclk(Mhz) * 16 */
242	rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
243}
244
245void rs690_line_buffer_adjust(struct radeon_device *rdev,
246			      struct drm_display_mode *mode1,
247			      struct drm_display_mode *mode2)
248{
249	u32 tmp;
250
251	/*
252	 * Line Buffer Setup
253	 * There is a single line buffer shared by both display controllers.
254	 * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
255	 * the display controllers.  The paritioning can either be done
256	 * manually or via one of four preset allocations specified in bits 1:0:
257	 *  0 - line buffer is divided in half and shared between crtc
258	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
259	 *  2 - D1 gets the whole buffer
260	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
261	 * Setting bit 2 of DC_LB_MEMORY_SPLIT controls switches to manual
262	 * allocation mode. In manual allocation mode, D1 always starts at 0,
263	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
264	 */
265	tmp = RREG32(DC_LB_MEMORY_SPLIT) & ~DC_LB_MEMORY_SPLIT_MASK;
266	tmp &= ~DC_LB_MEMORY_SPLIT_SHIFT_MODE;
267	/* auto */
268	if (mode1 && mode2) {
269		if (mode1->hdisplay > mode2->hdisplay) {
270			if (mode1->hdisplay > 2560)
271				tmp |= DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
272			else
273				tmp |= DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
274		} else if (mode2->hdisplay > mode1->hdisplay) {
275			if (mode2->hdisplay > 2560)
276				tmp |= DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
277			else
278				tmp |= DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
279		} else
280			tmp |= AVIVO_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
281	} else if (mode1) {
282		tmp |= DC_LB_MEMORY_SPLIT_D1_ONLY;
283	} else if (mode2) {
284		tmp |= DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
285	}
286	WREG32(DC_LB_MEMORY_SPLIT, tmp);
287}
288
289struct rs690_watermark {
290	u32        lb_request_fifo_depth;
291	fixed20_12 num_line_pair;
292	fixed20_12 estimated_width;
293	fixed20_12 worst_case_latency;
294	fixed20_12 consumption_rate;
295	fixed20_12 active_time;
296	fixed20_12 dbpp;
297	fixed20_12 priority_mark_max;
298	fixed20_12 priority_mark;
299	fixed20_12 sclk;
300};
301
302void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
303				  struct radeon_crtc *crtc,
304				  struct rs690_watermark *wm)
305{
306	struct drm_display_mode *mode = &crtc->base.mode;
307	fixed20_12 a, b, c;
308	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
309	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
310	/* FIXME: detect IGP with sideport memory, i don't think there is any
311	 * such product available
312	 */
313	bool sideport = false;
314
315	if (!crtc->base.enabled) {
316		/* FIXME: wouldn't it better to set priority mark to maximum */
317		wm->lb_request_fifo_depth = 4;
318		return;
319	}
320
321	if (crtc->vsc.full > rfixed_const(2))
322		wm->num_line_pair.full = rfixed_const(2);
323	else
324		wm->num_line_pair.full = rfixed_const(1);
325
326	b.full = rfixed_const(mode->crtc_hdisplay);
327	c.full = rfixed_const(256);
328	a.full = rfixed_mul(wm->num_line_pair, b);
329	request_fifo_depth.full = rfixed_div(a, c);
330	if (a.full < rfixed_const(4)) {
331		wm->lb_request_fifo_depth = 4;
332	} else {
333		wm->lb_request_fifo_depth = rfixed_trunc(request_fifo_depth);
334	}
335
336	/* Determine consumption rate
337	 *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
338	 *  vtaps = number of vertical taps,
339	 *  vsc = vertical scaling ratio, defined as source/destination
340	 *  hsc = horizontal scaling ration, defined as source/destination
341	 */
342	a.full = rfixed_const(mode->clock);
343	b.full = rfixed_const(1000);
344	a.full = rfixed_div(a, b);
345	pclk.full = rfixed_div(b, a);
346	if (crtc->rmx_type != RMX_OFF) {
347		b.full = rfixed_const(2);
348		if (crtc->vsc.full > b.full)
349			b.full = crtc->vsc.full;
350		b.full = rfixed_mul(b, crtc->hsc);
351		c.full = rfixed_const(2);
352		b.full = rfixed_div(b, c);
353		consumption_time.full = rfixed_div(pclk, b);
354	} else {
355		consumption_time.full = pclk.full;
356	}
357	a.full = rfixed_const(1);
358	wm->consumption_rate.full = rfixed_div(a, consumption_time);
359
360
361	/* Determine line time
362	 *  LineTime = total time for one line of displayhtotal
363	 *  LineTime = total number of horizontal pixels
364	 *  pclk = pixel clock period(ns)
365	 */
366	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
367	line_time.full = rfixed_mul(a, pclk);
368
369	/* Determine active time
370	 *  ActiveTime = time of active region of display within one line,
371	 *  hactive = total number of horizontal active pixels
372	 *  htotal = total number of horizontal pixels
373	 */
374	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
375	b.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
376	wm->active_time.full = rfixed_mul(line_time, b);
377	wm->active_time.full = rfixed_div(wm->active_time, a);
378
379	/* Maximun bandwidth is the minimun bandwidth of all component */
380	rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
381	if (sideport) {
382		if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
383			rdev->pm.sideport_bandwidth.full)
384			rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
385		read_delay_latency.full = rfixed_const(370 * 800 * 1000);
386		read_delay_latency.full = rfixed_div(read_delay_latency,
387			rdev->pm.igp_sideport_mclk);
388	} else {
389		if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
390			rdev->pm.k8_bandwidth.full)
391			rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
392		if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
393			rdev->pm.ht_bandwidth.full)
394			rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
395		read_delay_latency.full = rfixed_const(5000);
396	}
397
398	/* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
399	a.full = rfixed_const(16);
400	rdev->pm.sclk.full = rfixed_mul(rdev->pm.max_bandwidth, a);
401	a.full = rfixed_const(1000);
402	rdev->pm.sclk.full = rfixed_div(a, rdev->pm.sclk);
403	/* Determine chunk time
404	 * ChunkTime = the time it takes the DCP to send one chunk of data
405	 * to the LB which consists of pipeline delay and inter chunk gap
406	 * sclk = system clock(ns)
407	 */
408	a.full = rfixed_const(256 * 13);
409	chunk_time.full = rfixed_mul(rdev->pm.sclk, a);
410	a.full = rfixed_const(10);
411	chunk_time.full = rfixed_div(chunk_time, a);
412
413	/* Determine the worst case latency
414	 * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
415	 * WorstCaseLatency = worst case time from urgent to when the MC starts
416	 *                    to return data
417	 * READ_DELAY_IDLE_MAX = constant of 1us
418	 * ChunkTime = time it takes the DCP to send one chunk of data to the LB
419	 *             which consists of pipeline delay and inter chunk gap
420	 */
421	if (rfixed_trunc(wm->num_line_pair) > 1) {
422		a.full = rfixed_const(3);
423		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
424		wm->worst_case_latency.full += read_delay_latency.full;
425	} else {
426		a.full = rfixed_const(2);
427		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
428		wm->worst_case_latency.full += read_delay_latency.full;
429	}
430
431	/* Determine the tolerable latency
432	 * TolerableLatency = Any given request has only 1 line time
433	 *                    for the data to be returned
434	 * LBRequestFifoDepth = Number of chunk requests the LB can
435	 *                      put into the request FIFO for a display
436	 *  LineTime = total time for one line of display
437	 *  ChunkTime = the time it takes the DCP to send one chunk
438	 *              of data to the LB which consists of
439	 *  pipeline delay and inter chunk gap
440	 */
441	if ((2+wm->lb_request_fifo_depth) >= rfixed_trunc(request_fifo_depth)) {
442		tolerable_latency.full = line_time.full;
443	} else {
444		tolerable_latency.full = rfixed_const(wm->lb_request_fifo_depth - 2);
445		tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
446		tolerable_latency.full = rfixed_mul(tolerable_latency, chunk_time);
447		tolerable_latency.full = line_time.full - tolerable_latency.full;
448	}
449	/* We assume worst case 32bits (4 bytes) */
450	wm->dbpp.full = rfixed_const(4 * 8);
451
452	/* Determine the maximum priority mark
453	 *  width = viewport width in pixels
454	 */
455	a.full = rfixed_const(16);
456	wm->priority_mark_max.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
457	wm->priority_mark_max.full = rfixed_div(wm->priority_mark_max, a);
458
459	/* Determine estimated width */
460	estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
461	estimated_width.full = rfixed_div(estimated_width, consumption_time);
462	if (rfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
463		wm->priority_mark.full = rfixed_const(10);
464	} else {
465		a.full = rfixed_const(16);
466		wm->priority_mark.full = rfixed_div(estimated_width, a);
467		wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
468	}
469}
470
471void rs690_bandwidth_update(struct radeon_device *rdev)
472{
473	struct drm_display_mode *mode0 = NULL;
474	struct drm_display_mode *mode1 = NULL;
475	struct rs690_watermark wm0;
476	struct rs690_watermark wm1;
477	u32 tmp;
478	fixed20_12 priority_mark02, priority_mark12, fill_rate;
479	fixed20_12 a, b;
480
481	if (rdev->mode_info.crtcs[0]->base.enabled)
482		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
483	if (rdev->mode_info.crtcs[1]->base.enabled)
484		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
485	/*
486	 * Set display0/1 priority up in the memory controller for
487	 * modes if the user specifies HIGH for displaypriority
488	 * option.
489	 */
490	if (rdev->disp_priority == 2) {
491		tmp = RREG32_MC(MC_INIT_MISC_LAT_TIMER);
492		tmp &= ~MC_DISP1R_INIT_LAT_MASK;
493		tmp &= ~MC_DISP0R_INIT_LAT_MASK;
494		if (mode1)
495			tmp |= (1 << MC_DISP1R_INIT_LAT_SHIFT);
496		if (mode0)
497			tmp |= (1 << MC_DISP0R_INIT_LAT_SHIFT);
498		WREG32_MC(MC_INIT_MISC_LAT_TIMER, tmp);
499	}
500	rs690_line_buffer_adjust(rdev, mode0, mode1);
501
502	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
503		WREG32(DCP_CONTROL, 0);
504	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
505		WREG32(DCP_CONTROL, 2);
506
507	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
508	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
509
510	tmp = (wm0.lb_request_fifo_depth - 1);
511	tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
512	WREG32(LB_MAX_REQ_OUTSTANDING, tmp);
513
514	if (mode0 && mode1) {
515		if (rfixed_trunc(wm0.dbpp) > 64)
516			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
517		else
518			a.full = wm0.num_line_pair.full;
519		if (rfixed_trunc(wm1.dbpp) > 64)
520			b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
521		else
522			b.full = wm1.num_line_pair.full;
523		a.full += b.full;
524		fill_rate.full = rfixed_div(wm0.sclk, a);
525		if (wm0.consumption_rate.full > fill_rate.full) {
526			b.full = wm0.consumption_rate.full - fill_rate.full;
527			b.full = rfixed_mul(b, wm0.active_time);
528			a.full = rfixed_mul(wm0.worst_case_latency,
529						wm0.consumption_rate);
530			a.full = a.full + b.full;
531			b.full = rfixed_const(16 * 1000);
532			priority_mark02.full = rfixed_div(a, b);
533		} else {
534			a.full = rfixed_mul(wm0.worst_case_latency,
535						wm0.consumption_rate);
536			b.full = rfixed_const(16 * 1000);
537			priority_mark02.full = rfixed_div(a, b);
538		}
539		if (wm1.consumption_rate.full > fill_rate.full) {
540			b.full = wm1.consumption_rate.full - fill_rate.full;
541			b.full = rfixed_mul(b, wm1.active_time);
542			a.full = rfixed_mul(wm1.worst_case_latency,
543						wm1.consumption_rate);
544			a.full = a.full + b.full;
545			b.full = rfixed_const(16 * 1000);
546			priority_mark12.full = rfixed_div(a, b);
547		} else {
548			a.full = rfixed_mul(wm1.worst_case_latency,
549						wm1.consumption_rate);
550			b.full = rfixed_const(16 * 1000);
551			priority_mark12.full = rfixed_div(a, b);
552		}
553		if (wm0.priority_mark.full > priority_mark02.full)
554			priority_mark02.full = wm0.priority_mark.full;
555		if (rfixed_trunc(priority_mark02) < 0)
556			priority_mark02.full = 0;
557		if (wm0.priority_mark_max.full > priority_mark02.full)
558			priority_mark02.full = wm0.priority_mark_max.full;
559		if (wm1.priority_mark.full > priority_mark12.full)
560			priority_mark12.full = wm1.priority_mark.full;
561		if (rfixed_trunc(priority_mark12) < 0)
562			priority_mark12.full = 0;
563		if (wm1.priority_mark_max.full > priority_mark12.full)
564			priority_mark12.full = wm1.priority_mark_max.full;
565		WREG32(D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
566		WREG32(D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
567		WREG32(D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
568		WREG32(D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
569	} else if (mode0) {
570		if (rfixed_trunc(wm0.dbpp) > 64)
571			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
572		else
573			a.full = wm0.num_line_pair.full;
574		fill_rate.full = rfixed_div(wm0.sclk, a);
575		if (wm0.consumption_rate.full > fill_rate.full) {
576			b.full = wm0.consumption_rate.full - fill_rate.full;
577			b.full = rfixed_mul(b, wm0.active_time);
578			a.full = rfixed_mul(wm0.worst_case_latency,
579						wm0.consumption_rate);
580			a.full = a.full + b.full;
581			b.full = rfixed_const(16 * 1000);
582			priority_mark02.full = rfixed_div(a, b);
583		} else {
584			a.full = rfixed_mul(wm0.worst_case_latency,
585						wm0.consumption_rate);
586			b.full = rfixed_const(16 * 1000);
587			priority_mark02.full = rfixed_div(a, b);
588		}
589		if (wm0.priority_mark.full > priority_mark02.full)
590			priority_mark02.full = wm0.priority_mark.full;
591		if (rfixed_trunc(priority_mark02) < 0)
592			priority_mark02.full = 0;
593		if (wm0.priority_mark_max.full > priority_mark02.full)
594			priority_mark02.full = wm0.priority_mark_max.full;
595		WREG32(D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
596		WREG32(D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
597		WREG32(D2MODE_PRIORITY_A_CNT, MODE_PRIORITY_OFF);
598		WREG32(D2MODE_PRIORITY_B_CNT, MODE_PRIORITY_OFF);
599	} else {
600		if (rfixed_trunc(wm1.dbpp) > 64)
601			a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
602		else
603			a.full = wm1.num_line_pair.full;
604		fill_rate.full = rfixed_div(wm1.sclk, a);
605		if (wm1.consumption_rate.full > fill_rate.full) {
606			b.full = wm1.consumption_rate.full - fill_rate.full;
607			b.full = rfixed_mul(b, wm1.active_time);
608			a.full = rfixed_mul(wm1.worst_case_latency,
609						wm1.consumption_rate);
610			a.full = a.full + b.full;
611			b.full = rfixed_const(16 * 1000);
612			priority_mark12.full = rfixed_div(a, b);
613		} else {
614			a.full = rfixed_mul(wm1.worst_case_latency,
615						wm1.consumption_rate);
616			b.full = rfixed_const(16 * 1000);
617			priority_mark12.full = rfixed_div(a, b);
618		}
619		if (wm1.priority_mark.full > priority_mark12.full)
620			priority_mark12.full = wm1.priority_mark.full;
621		if (rfixed_trunc(priority_mark12) < 0)
622			priority_mark12.full = 0;
623		if (wm1.priority_mark_max.full > priority_mark12.full)
624			priority_mark12.full = wm1.priority_mark_max.full;
625		WREG32(D1MODE_PRIORITY_A_CNT, MODE_PRIORITY_OFF);
626		WREG32(D1MODE_PRIORITY_B_CNT, MODE_PRIORITY_OFF);
627		WREG32(D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
628		WREG32(D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
629	}
630}
631
632/*
633 * Indirect registers accessor
634 */
635uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
636{
637	uint32_t r;
638
639	WREG32(RS690_MC_INDEX, (reg & RS690_MC_INDEX_MASK));
640	r = RREG32(RS690_MC_DATA);
641	WREG32(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
642	return r;
643}
644
645void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
646{
647	WREG32(RS690_MC_INDEX,
648	       RS690_MC_INDEX_WR_EN | ((reg) & RS690_MC_INDEX_MASK));
649	WREG32(RS690_MC_DATA, v);
650	WREG32(RS690_MC_INDEX, RS690_MC_INDEX_WR_ACK);
651}
652