From c93bb85b5cba3e3a06f2cad8e9bc5c23d3d10aac Mon Sep 17 00:00:00 2001 From: Jerome Glisse Date: Mon, 13 Jul 2009 21:04:08 +0200 Subject: drm/radeon/kms: fix bandwidth computation on avivo hardware Fix bandwidth computation and crtc priority in memory controller so that crtc memory request are fullfill in time to avoid display artifact. Signed-off-by: Jerome Glisse Signed-off-by: Dave Airlie --- drivers/gpu/drm/radeon/rs690.c | 472 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 472 insertions(+) (limited to 'drivers/gpu/drm/radeon/rs690.c') diff --git a/drivers/gpu/drm/radeon/rs690.c b/drivers/gpu/drm/radeon/rs690.c index 79ba85042b5..97eaee3d28b 100644 --- a/drivers/gpu/drm/radeon/rs690.c +++ b/drivers/gpu/drm/radeon/rs690.c @@ -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); @@ -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 */ @@ -156,8 +232,404 @@ void rs690_vram_info(struct radeon_device *rdev) 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 -- cgit v1.2.3