Merge branch 'tracing-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[pandora-kernel.git] / drivers / gpu / drm / radeon / rs690.c
index 79ba850..8798825 100644 (file)
@@ -28,6 +28,9 @@
 #include "drmP.h"
 #include "radeon_reg.h"
 #include "radeon.h"
+#include "rs690r.h"
+#include "atom.h"
+#include "atom-bits.h"
 
 /* rs690,rs740 depends on : */
 void r100_hdp_reset(struct radeon_device *rdev);
@@ -64,7 +67,7 @@ int rs690_mc_init(struct radeon_device *rdev)
        rs400_gart_disable(rdev);
 
        /* Setup GPU memory space */
-       rdev->mc.gtt_location = rdev->mc.vram_size;
+       rdev->mc.gtt_location = rdev->mc.mc_vram_size;
        rdev->mc.gtt_location += (rdev->mc.gtt_size - 1);
        rdev->mc.gtt_location &= ~(rdev->mc.gtt_size - 1);
        rdev->mc.vram_location = 0xFFFFFFFFUL;
@@ -79,7 +82,7 @@ int rs690_mc_init(struct radeon_device *rdev)
                printk(KERN_WARNING "Failed to wait MC idle while "
                       "programming pipes. Bad things might happen.\n");
        }
-       tmp = rdev->mc.vram_location + rdev->mc.vram_size - 1;
+       tmp = rdev->mc.vram_location + rdev->mc.mc_vram_size - 1;
        tmp = REG_SET(RS690_MC_FB_TOP, tmp >> 16);
        tmp |= REG_SET(RS690_MC_FB_START, rdev->mc.vram_location >> 16);
        WREG32_MC(RS690_MCCFG_FB_LOCATION, tmp);
@@ -138,9 +141,82 @@ void rs690_gpu_init(struct radeon_device *rdev)
 /*
  * VRAM info.
  */
+void rs690_pm_info(struct radeon_device *rdev)
+{
+       int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
+       struct _ATOM_INTEGRATED_SYSTEM_INFO *info;
+       struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *info_v2;
+       void *ptr;
+       uint16_t data_offset;
+       uint8_t frev, crev;
+       fixed20_12 tmp;
+
+       atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
+                              &frev, &crev, &data_offset);
+       ptr = rdev->mode_info.atom_context->bios + data_offset;
+       info = (struct _ATOM_INTEGRATED_SYSTEM_INFO *)ptr;
+       info_v2 = (struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *)ptr;
+       /* Get various system informations from bios */
+       switch (crev) {
+       case 1:
+               tmp.full = rfixed_const(100);
+               rdev->pm.igp_sideport_mclk.full = rfixed_const(info->ulBootUpMemoryClock);
+               rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
+               rdev->pm.igp_system_mclk.full = rfixed_const(le16_to_cpu(info->usK8MemoryClock));
+               rdev->pm.igp_ht_link_clk.full = rfixed_const(le16_to_cpu(info->usFSBClock));
+               rdev->pm.igp_ht_link_width.full = rfixed_const(info->ucHTLinkWidth);
+               break;
+       case 2:
+               tmp.full = rfixed_const(100);
+               rdev->pm.igp_sideport_mclk.full = rfixed_const(info_v2->ulBootUpSidePortClock);
+               rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
+               rdev->pm.igp_system_mclk.full = rfixed_const(info_v2->ulBootUpUMAClock);
+               rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
+               rdev->pm.igp_ht_link_clk.full = rfixed_const(info_v2->ulHTLinkFreq);
+               rdev->pm.igp_ht_link_clk.full = rfixed_div(rdev->pm.igp_ht_link_clk, tmp);
+               rdev->pm.igp_ht_link_width.full = rfixed_const(le16_to_cpu(info_v2->usMinHTLinkWidth));
+               break;
+       default:
+               tmp.full = rfixed_const(100);
+               /* We assume the slower possible clock ie worst case */
+               /* DDR 333Mhz */
+               rdev->pm.igp_sideport_mclk.full = rfixed_const(333);
+               /* FIXME: system clock ? */
+               rdev->pm.igp_system_mclk.full = rfixed_const(100);
+               rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
+               rdev->pm.igp_ht_link_clk.full = rfixed_const(200);
+               rdev->pm.igp_ht_link_width.full = rfixed_const(8);
+               DRM_ERROR("No integrated system info for your GPU, using safe default\n");
+               break;
+       }
+       /* Compute various bandwidth */
+       /* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
+       tmp.full = rfixed_const(4);
+       rdev->pm.k8_bandwidth.full = rfixed_mul(rdev->pm.igp_system_mclk, tmp);
+       /* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
+        *              = ht_clk * ht_width / 5
+        */
+       tmp.full = rfixed_const(5);
+       rdev->pm.ht_bandwidth.full = rfixed_mul(rdev->pm.igp_ht_link_clk,
+                                               rdev->pm.igp_ht_link_width);
+       rdev->pm.ht_bandwidth.full = rfixed_div(rdev->pm.ht_bandwidth, tmp);
+       if (tmp.full < rdev->pm.max_bandwidth.full) {
+               /* HT link is a limiting factor */
+               rdev->pm.max_bandwidth.full = tmp.full;
+       }
+       /* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
+        *                    = (sideport_clk * 14) / 10
+        */
+       tmp.full = rfixed_const(14);
+       rdev->pm.sideport_bandwidth.full = rfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
+       tmp.full = rfixed_const(10);
+       rdev->pm.sideport_bandwidth.full = rfixed_div(rdev->pm.sideport_bandwidth, tmp);
+}
+
 void rs690_vram_info(struct radeon_device *rdev)
 {
        uint32_t tmp;
+       fixed20_12 a;
 
        rs400_gart_adjust_size(rdev);
        /* DDR for all card after R300 & IGP */
@@ -152,12 +228,409 @@ void rs690_vram_info(struct radeon_device *rdev)
        } else {
                rdev->mc.vram_width = 64;
        }
-       rdev->mc.vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
+       rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
+       rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
 
        rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
        rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
+       rs690_pm_info(rdev);
+       /* FIXME: we should enforce default clock in case GPU is not in
+        * default setup
+        */
+       a.full = rfixed_const(100);
+       rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
+       rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
+       a.full = rfixed_const(16);
+       /* core_bandwidth = sclk(Mhz) * 16 */
+       rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
+}
+
+void rs690_line_buffer_adjust(struct radeon_device *rdev,
+                             struct drm_display_mode *mode1,
+                             struct drm_display_mode *mode2)
+{
+       u32 tmp;
+
+       /*
+        * Line Buffer Setup
+        * There is a single line buffer shared by both display controllers.
+        * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
+        * the display controllers.  The paritioning can either be done
+        * manually or via one of four preset allocations specified in bits 1:0:
+        *  0 - line buffer is divided in half and shared between crtc
+        *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
+        *  2 - D1 gets the whole buffer
+        *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
+        * Setting bit 2 of DC_LB_MEMORY_SPLIT controls switches to manual
+        * allocation mode. In manual allocation mode, D1 always starts at 0,
+        * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
+        */
+       tmp = RREG32(DC_LB_MEMORY_SPLIT) & ~DC_LB_MEMORY_SPLIT_MASK;
+       tmp &= ~DC_LB_MEMORY_SPLIT_SHIFT_MODE;
+       /* auto */
+       if (mode1 && mode2) {
+               if (mode1->hdisplay > mode2->hdisplay) {
+                       if (mode1->hdisplay > 2560)
+                               tmp |= DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
+                       else
+                               tmp |= DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
+               } else if (mode2->hdisplay > mode1->hdisplay) {
+                       if (mode2->hdisplay > 2560)
+                               tmp |= DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
+                       else
+                               tmp |= DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
+               } else
+                       tmp |= AVIVO_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
+       } else if (mode1) {
+               tmp |= DC_LB_MEMORY_SPLIT_D1_ONLY;
+       } else if (mode2) {
+               tmp |= DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
+       }
+       WREG32(DC_LB_MEMORY_SPLIT, tmp);
+}
+
+struct rs690_watermark {
+       u32        lb_request_fifo_depth;
+       fixed20_12 num_line_pair;
+       fixed20_12 estimated_width;
+       fixed20_12 worst_case_latency;
+       fixed20_12 consumption_rate;
+       fixed20_12 active_time;
+       fixed20_12 dbpp;
+       fixed20_12 priority_mark_max;
+       fixed20_12 priority_mark;
+       fixed20_12 sclk;
+};
+
+void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
+                                 struct radeon_crtc *crtc,
+                                 struct rs690_watermark *wm)
+{
+       struct drm_display_mode *mode = &crtc->base.mode;
+       fixed20_12 a, b, c;
+       fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
+       fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
+       /* FIXME: detect IGP with sideport memory, i don't think there is any
+        * such product available
+        */
+       bool sideport = false;
+
+       if (!crtc->base.enabled) {
+               /* FIXME: wouldn't it better to set priority mark to maximum */
+               wm->lb_request_fifo_depth = 4;
+               return;
+       }
+
+       if (crtc->vsc.full > rfixed_const(2))
+               wm->num_line_pair.full = rfixed_const(2);
+       else
+               wm->num_line_pair.full = rfixed_const(1);
+
+       b.full = rfixed_const(mode->crtc_hdisplay);
+       c.full = rfixed_const(256);
+       a.full = rfixed_mul(wm->num_line_pair, b);
+       request_fifo_depth.full = rfixed_div(a, c);
+       if (a.full < rfixed_const(4)) {
+               wm->lb_request_fifo_depth = 4;
+       } else {
+               wm->lb_request_fifo_depth = rfixed_trunc(request_fifo_depth);
+       }
+
+       /* Determine consumption rate
+        *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
+        *  vtaps = number of vertical taps,
+        *  vsc = vertical scaling ratio, defined as source/destination
+        *  hsc = horizontal scaling ration, defined as source/destination
+        */
+       a.full = rfixed_const(mode->clock);
+       b.full = rfixed_const(1000);
+       a.full = rfixed_div(a, b);
+       pclk.full = rfixed_div(b, a);
+       if (crtc->rmx_type != RMX_OFF) {
+               b.full = rfixed_const(2);
+               if (crtc->vsc.full > b.full)
+                       b.full = crtc->vsc.full;
+               b.full = rfixed_mul(b, crtc->hsc);
+               c.full = rfixed_const(2);
+               b.full = rfixed_div(b, c);
+               consumption_time.full = rfixed_div(pclk, b);
+       } else {
+               consumption_time.full = pclk.full;
+       }
+       a.full = rfixed_const(1);
+       wm->consumption_rate.full = rfixed_div(a, consumption_time);
+
+
+       /* Determine line time
+        *  LineTime = total time for one line of displayhtotal
+        *  LineTime = total number of horizontal pixels
+        *  pclk = pixel clock period(ns)
+        */
+       a.full = rfixed_const(crtc->base.mode.crtc_htotal);
+       line_time.full = rfixed_mul(a, pclk);
+
+       /* Determine active time
+        *  ActiveTime = time of active region of display within one line,
+        *  hactive = total number of horizontal active pixels
+        *  htotal = total number of horizontal pixels
+        */
+       a.full = rfixed_const(crtc->base.mode.crtc_htotal);
+       b.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
+       wm->active_time.full = rfixed_mul(line_time, b);
+       wm->active_time.full = rfixed_div(wm->active_time, a);
+
+       /* Maximun bandwidth is the minimun bandwidth of all component */
+       rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
+       if (sideport) {
+               if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
+                       rdev->pm.sideport_bandwidth.full)
+                       rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
+               read_delay_latency.full = rfixed_const(370 * 800 * 1000);
+               read_delay_latency.full = rfixed_div(read_delay_latency,
+                       rdev->pm.igp_sideport_mclk);
+       } else {
+               if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
+                       rdev->pm.k8_bandwidth.full)
+                       rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
+               if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
+                       rdev->pm.ht_bandwidth.full)
+                       rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
+               read_delay_latency.full = rfixed_const(5000);
+       }
+
+       /* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
+       a.full = rfixed_const(16);
+       rdev->pm.sclk.full = rfixed_mul(rdev->pm.max_bandwidth, a);
+       a.full = rfixed_const(1000);
+       rdev->pm.sclk.full = rfixed_div(a, rdev->pm.sclk);
+       /* Determine chunk time
+        * ChunkTime = the time it takes the DCP to send one chunk of data
+        * to the LB which consists of pipeline delay and inter chunk gap
+        * sclk = system clock(ns)
+        */
+       a.full = rfixed_const(256 * 13);
+       chunk_time.full = rfixed_mul(rdev->pm.sclk, a);
+       a.full = rfixed_const(10);
+       chunk_time.full = rfixed_div(chunk_time, a);
+
+       /* Determine the worst case latency
+        * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
+        * WorstCaseLatency = worst case time from urgent to when the MC starts
+        *                    to return data
+        * READ_DELAY_IDLE_MAX = constant of 1us
+        * ChunkTime = time it takes the DCP to send one chunk of data to the LB
+        *             which consists of pipeline delay and inter chunk gap
+        */
+       if (rfixed_trunc(wm->num_line_pair) > 1) {
+               a.full = rfixed_const(3);
+               wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
+               wm->worst_case_latency.full += read_delay_latency.full;
+       } else {
+               a.full = rfixed_const(2);
+               wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
+               wm->worst_case_latency.full += read_delay_latency.full;
+       }
+
+       /* Determine the tolerable latency
+        * TolerableLatency = Any given request has only 1 line time
+        *                    for the data to be returned
+        * LBRequestFifoDepth = Number of chunk requests the LB can
+        *                      put into the request FIFO for a display
+        *  LineTime = total time for one line of display
+        *  ChunkTime = the time it takes the DCP to send one chunk
+        *              of data to the LB which consists of
+        *  pipeline delay and inter chunk gap
+        */
+       if ((2+wm->lb_request_fifo_depth) >= rfixed_trunc(request_fifo_depth)) {
+               tolerable_latency.full = line_time.full;
+       } else {
+               tolerable_latency.full = rfixed_const(wm->lb_request_fifo_depth - 2);
+               tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
+               tolerable_latency.full = rfixed_mul(tolerable_latency, chunk_time);
+               tolerable_latency.full = line_time.full - tolerable_latency.full;
+       }
+       /* We assume worst case 32bits (4 bytes) */
+       wm->dbpp.full = rfixed_const(4 * 8);
+
+       /* Determine the maximum priority mark
+        *  width = viewport width in pixels
+        */
+       a.full = rfixed_const(16);
+       wm->priority_mark_max.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
+       wm->priority_mark_max.full = rfixed_div(wm->priority_mark_max, a);
+
+       /* Determine estimated width */
+       estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
+       estimated_width.full = rfixed_div(estimated_width, consumption_time);
+       if (rfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
+               wm->priority_mark.full = rfixed_const(10);
+       } else {
+               a.full = rfixed_const(16);
+               wm->priority_mark.full = rfixed_div(estimated_width, a);
+               wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
+       }
 }
 
+void rs690_bandwidth_update(struct radeon_device *rdev)
+{
+       struct drm_display_mode *mode0 = NULL;
+       struct drm_display_mode *mode1 = NULL;
+       struct rs690_watermark wm0;
+       struct rs690_watermark wm1;
+       u32 tmp;
+       fixed20_12 priority_mark02, priority_mark12, fill_rate;
+       fixed20_12 a, b;
+
+       if (rdev->mode_info.crtcs[0]->base.enabled)
+               mode0 = &rdev->mode_info.crtcs[0]->base.mode;
+       if (rdev->mode_info.crtcs[1]->base.enabled)
+               mode1 = &rdev->mode_info.crtcs[1]->base.mode;
+       /*
+        * Set display0/1 priority up in the memory controller for
+        * modes if the user specifies HIGH for displaypriority
+        * option.
+        */
+       if (rdev->disp_priority == 2) {
+               tmp = RREG32_MC(MC_INIT_MISC_LAT_TIMER);
+               tmp &= ~MC_DISP1R_INIT_LAT_MASK;
+               tmp &= ~MC_DISP0R_INIT_LAT_MASK;
+               if (mode1)
+                       tmp |= (1 << MC_DISP1R_INIT_LAT_SHIFT);
+               if (mode0)
+                       tmp |= (1 << MC_DISP0R_INIT_LAT_SHIFT);
+               WREG32_MC(MC_INIT_MISC_LAT_TIMER, tmp);
+       }
+       rs690_line_buffer_adjust(rdev, mode0, mode1);
+
+       if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
+               WREG32(DCP_CONTROL, 0);
+       if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
+               WREG32(DCP_CONTROL, 2);
+
+       rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
+       rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
+
+       tmp = (wm0.lb_request_fifo_depth - 1);
+       tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
+       WREG32(LB_MAX_REQ_OUTSTANDING, tmp);
+
+       if (mode0 && mode1) {
+               if (rfixed_trunc(wm0.dbpp) > 64)
+                       a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
+               else
+                       a.full = wm0.num_line_pair.full;
+               if (rfixed_trunc(wm1.dbpp) > 64)
+                       b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
+               else
+                       b.full = wm1.num_line_pair.full;
+               a.full += b.full;
+               fill_rate.full = rfixed_div(wm0.sclk, a);
+               if (wm0.consumption_rate.full > fill_rate.full) {
+                       b.full = wm0.consumption_rate.full - fill_rate.full;
+                       b.full = rfixed_mul(b, wm0.active_time);
+                       a.full = rfixed_mul(wm0.worst_case_latency,
+                                               wm0.consumption_rate);
+                       a.full = a.full + b.full;
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark02.full = rfixed_div(a, b);
+               } else {
+                       a.full = rfixed_mul(wm0.worst_case_latency,
+                                               wm0.consumption_rate);
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark02.full = rfixed_div(a, b);
+               }
+               if (wm1.consumption_rate.full > fill_rate.full) {
+                       b.full = wm1.consumption_rate.full - fill_rate.full;
+                       b.full = rfixed_mul(b, wm1.active_time);
+                       a.full = rfixed_mul(wm1.worst_case_latency,
+                                               wm1.consumption_rate);
+                       a.full = a.full + b.full;
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark12.full = rfixed_div(a, b);
+               } else {
+                       a.full = rfixed_mul(wm1.worst_case_latency,
+                                               wm1.consumption_rate);
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark12.full = rfixed_div(a, b);
+               }
+               if (wm0.priority_mark.full > priority_mark02.full)
+                       priority_mark02.full = wm0.priority_mark.full;
+               if (rfixed_trunc(priority_mark02) < 0)
+                       priority_mark02.full = 0;
+               if (wm0.priority_mark_max.full > priority_mark02.full)
+                       priority_mark02.full = wm0.priority_mark_max.full;
+               if (wm1.priority_mark.full > priority_mark12.full)
+                       priority_mark12.full = wm1.priority_mark.full;
+               if (rfixed_trunc(priority_mark12) < 0)
+                       priority_mark12.full = 0;
+               if (wm1.priority_mark_max.full > priority_mark12.full)
+                       priority_mark12.full = wm1.priority_mark_max.full;
+               WREG32(D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
+               WREG32(D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
+               WREG32(D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
+               WREG32(D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
+       } else if (mode0) {
+               if (rfixed_trunc(wm0.dbpp) > 64)
+                       a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
+               else
+                       a.full = wm0.num_line_pair.full;
+               fill_rate.full = rfixed_div(wm0.sclk, a);
+               if (wm0.consumption_rate.full > fill_rate.full) {
+                       b.full = wm0.consumption_rate.full - fill_rate.full;
+                       b.full = rfixed_mul(b, wm0.active_time);
+                       a.full = rfixed_mul(wm0.worst_case_latency,
+                                               wm0.consumption_rate);
+                       a.full = a.full + b.full;
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark02.full = rfixed_div(a, b);
+               } else {
+                       a.full = rfixed_mul(wm0.worst_case_latency,
+                                               wm0.consumption_rate);
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark02.full = rfixed_div(a, b);
+               }
+               if (wm0.priority_mark.full > priority_mark02.full)
+                       priority_mark02.full = wm0.priority_mark.full;
+               if (rfixed_trunc(priority_mark02) < 0)
+                       priority_mark02.full = 0;
+               if (wm0.priority_mark_max.full > priority_mark02.full)
+                       priority_mark02.full = wm0.priority_mark_max.full;
+               WREG32(D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
+               WREG32(D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
+               WREG32(D2MODE_PRIORITY_A_CNT, MODE_PRIORITY_OFF);
+               WREG32(D2MODE_PRIORITY_B_CNT, MODE_PRIORITY_OFF);
+       } else {
+               if (rfixed_trunc(wm1.dbpp) > 64)
+                       a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
+               else
+                       a.full = wm1.num_line_pair.full;
+               fill_rate.full = rfixed_div(wm1.sclk, a);
+               if (wm1.consumption_rate.full > fill_rate.full) {
+                       b.full = wm1.consumption_rate.full - fill_rate.full;
+                       b.full = rfixed_mul(b, wm1.active_time);
+                       a.full = rfixed_mul(wm1.worst_case_latency,
+                                               wm1.consumption_rate);
+                       a.full = a.full + b.full;
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark12.full = rfixed_div(a, b);
+               } else {
+                       a.full = rfixed_mul(wm1.worst_case_latency,
+                                               wm1.consumption_rate);
+                       b.full = rfixed_const(16 * 1000);
+                       priority_mark12.full = rfixed_div(a, b);
+               }
+               if (wm1.priority_mark.full > priority_mark12.full)
+                       priority_mark12.full = wm1.priority_mark.full;
+               if (rfixed_trunc(priority_mark12) < 0)
+                       priority_mark12.full = 0;
+               if (wm1.priority_mark_max.full > priority_mark12.full)
+                       priority_mark12.full = wm1.priority_mark_max.full;
+               WREG32(D1MODE_PRIORITY_A_CNT, MODE_PRIORITY_OFF);
+               WREG32(D1MODE_PRIORITY_B_CNT, MODE_PRIORITY_OFF);
+               WREG32(D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
+               WREG32(D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
+       }
+}
 
 /*
  * Indirect registers accessor
@@ -179,3 +652,4 @@ void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
        WREG32(RS690_MC_DATA, v);
        WREG32(RS690_MC_INDEX, RS690_MC_INDEX_WR_ACK);
 }
+