Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Fri, Sep 05, 2014 at 12:35:18PM -0700, Rodrigo Vivi wrote: On Fri, Sep 5, 2014 at 11:28 AM, Paulo Zanoni przan...@gmail.com wrote: - The screen is disabled - We're runtime suspended These are another error. Not caused by this patch. Why frontbuffer_flush is being called with screen disabled or on runtime suspend? Because the frontbuffer tracking only tells you when a frontbuffer gets updated. It's the job of the consumers to properly filter these events so that they only act upon those which are relevant. psr does that, fbc (as is) doesn't. And if you have that filtering, adding more filtering as a safeguard really isn't all that useful. - We're touching cursors when we decide to flush FBC. This is not actually a flush. But cleaning the cache to allow compression restart. I believe we could have some variable to indicate when we touched ring flush than on next frontbuffer we do the mmio cache clean. If you filter frontbuffers properly you can ignore all the cursor/sprite events completely. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Fri, Sep 05, 2014 at 02:12:49PM -0700, Rodrigo Vivi wrote: As Paulo told the part 2 of his proposal is still missing and we will get the WARN of reading fbc register at fbc_enabled when on runtime_suspend... But I couldn't find the propper place to check for intel_crtc-active. This was another issue reported by power team that on screen off there was to much activity on our driver. So we need to check it more carefully the proper place to avoid all unecessary inactivities with display off. Do you have any suggestion Daniel? Don't check for intel-active since doing that will lead to hilarious locking inversions. You have to add a bit of tracking (with it's own locking) for which frontbuffer events are interesting for fbc like psr does. Since we only do fbc on one plane ever that's only ever going to be one bit, but setting/clearing it correctly in crtc_enable/disable is the important part here. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
2014-08-04 7:51 GMT-03:00 Rodrigo Vivi rodrigo.v...@intel.com: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. This patch causes lots of WARNs when running igt/pm_rpm/cursor and a few other similar test cases. Now the interesting this is that we're trying to do FBC flushing, but: - FBC is disabled by default - The screen is disabled - We're runtime suspended - We're touching cursors when we decide to flush FBC. This is the first WARN that happens: [ 123.581179] [drm:intel_runtime_suspend] Device suspended [ 123.680666] [ cut here ] [ 123.680697] WARNING: CPU: 1 PID: 1776 at drivers/gpu/drm/i915/intel_uncore.c:47 assert_device_not_suspended.isra.8+0x43/0x50 [i915]() [ 123.680699] Device suspended [ 123.680700] Modules linked in: intel_rapl x86_pkg_temp_thermal serio_raw intel_powerclamp efivars btusb iwlmvm iwlwifi mei_me mei int3403_thermal snd_hda_codec_hdmi snd_hda_intel snd_hda_controller i2c_designware_platform snd_hda_codec i2c_designware_core snd_hwdep snd_pcm_oss snd_mixer_oss acpi_pad snd_pcm snd_timer fuse nls_utf8 nls_cp437 vfat fat sd_mod ahci libahci i915 sdhci_pci e1000e drm_kms_helper drm sdhci_acpi sdhci [ 123.680733] CPU: 1 PID: 1776 Comm: pm_rpm Not tainted 3.17.0-rc2.1409051503pz+ #1100 [ 123.680736] Hardware name: Intel Corporation Broadwell Client platform/Wilson Beach SDS, BIOS BDW-E2R1.86C.0072.R03.1405072127 05/07/2014 [ 123.680738] 0009 88024174fab8 816f6d23 88024174fb00 [ 123.680742] 88024174faf0 8107b368 88003770 00050380 [ 123.680746] 00050380 880037700068 0001 88024174fb50 [ 123.680751] Call Trace: [ 123.680758] [816f6d23] dump_stack+0x4d/0x66 [ 123.680763] [8107b368] warn_slowpath_common+0x78/0xa0 [ 123.680766] [8107b3d7] warn_slowpath_fmt+0x47/0x50 [ 123.680772] [810c0efd] ? trace_hardirqs_on_caller+0x15d/0x200 [ 123.680791] [a0118c63] assert_device_not_suspended.isra.8+0x43/0x50 [i915] [ 123.680809] [a011cbd5] gen8_write32+0x35/0x180 [i915] [ 123.680821] [a00dbb49] gen8_fbc_sw_flush+0x29/0x30 [i915] [ 123.680840] [a013150d] intel_frontbuffer_flush+0x7d/0x90 [i915] [ 123.680859] [a0135f6b] intel_cursor_plane_update+0x13b/0x150 [i915] [ 123.680876] [a00276b0] setplane_internal+0x260/0x2a0 [drm] [ 123.680889] [a002780e] drm_mode_cursor_common+0x11e/0x310 [drm] [ 123.680904] [a002ae1c] drm_mode_cursor_ioctl+0x3c/0x40 [drm] [ 123.680914] [a001cc9f] drm_ioctl+0x1df/0x6a0 [drm] [ 123.680920] [816fe769] ? mutex_unlock+0x9/0x10 [ 123.680924] [811f3856] ? seq_read+0xb6/0x3e0 [ 123.680929] [811e2540] do_vfs_ioctl+0x2e0/0x4e0 [ 123.680934] [81701077] ? sysret_check+0x1b/0x56 [ 123.680939] [810c0efd] ? trace_hardirqs_on_caller+0x15d/0x200 [ 123.680943] [811e27c1] SyS_ioctl+0x81/0xa0 [ 123.680947] [81701052] system_call_fastpath+0x16/0x1b [ 123.680949] ---[ end trace 9f389682639d2eb9 ]--- [ 123.680953] [ cut here ] So, what we can/could do: 1 - Change gen8_fbc_sw_flush() do to nothing when FBC is disabled. 2 - At some point of the stack above, avoid calling everything else if the screen and/or plane and/or crtc is disabled. But when? At intel_cursor_plane_update? At intel_frontbuffer_flush? 3 - Don't call gen8_fbc_sw_flush if the frontbuffer_bits don't include the primary plane bits (since, AFAIR, FBC ignores the sprite and cursors, right?) Is anybody against any of the 3 items above? Also notice that the cursor subtest is not the only one that causes these WARNs, so I expect similar backtraces with other similar solutions for equivalent problems with the other planes. Thanks, Paulo Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Fri, Sep 5, 2014 at 11:28 AM, Paulo Zanoni przan...@gmail.com wrote: 2014-08-04 7:51 GMT-03:00 Rodrigo Vivi rodrigo.v...@intel.com: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. This patch causes lots of WARNs when running igt/pm_rpm/cursor and a few other similar test cases. Now the interesting this is that we're trying to do FBC flushing, but: As I told you on the IRC already, thank you for finding this. It is good to get reasons and arguments. - FBC is disabled by default Yeah, we need to fix it although it just matter much - The screen is disabled - We're runtime suspended These are another error. Not caused by this patch. Why frontbuffer_flush is being called with screen disabled or on runtime suspend? - We're touching cursors when we decide to flush FBC. This is not actually a flush. But cleaning the cache to allow compression restart. I believe we could have some variable to indicate when we touched ring flush than on next frontbuffer we do the mmio cache clean. This is the first WARN that happens: [ 123.581179] [drm:intel_runtime_suspend] Device suspended [ 123.680666] [ cut here ] [ 123.680697] WARNING: CPU: 1 PID: 1776 at drivers/gpu/drm/i915/intel_uncore.c:47 assert_device_not_suspended.isra.8+0x43/0x50 [i915]() [ 123.680699] Device suspended [ 123.680700] Modules linked in: intel_rapl x86_pkg_temp_thermal serio_raw intel_powerclamp efivars btusb iwlmvm iwlwifi mei_me mei int3403_thermal snd_hda_codec_hdmi snd_hda_intel snd_hda_controller i2c_designware_platform snd_hda_codec i2c_designware_core snd_hwdep snd_pcm_oss snd_mixer_oss acpi_pad snd_pcm snd_timer fuse nls_utf8 nls_cp437 vfat fat sd_mod ahci libahci i915 sdhci_pci e1000e drm_kms_helper drm sdhci_acpi sdhci [ 123.680733] CPU: 1 PID: 1776 Comm: pm_rpm Not tainted 3.17.0-rc2.1409051503pz+ #1100 [ 123.680736] Hardware name: Intel Corporation Broadwell Client platform/Wilson Beach SDS, BIOS BDW-E2R1.86C.0072.R03.1405072127 05/07/2014 [ 123.680738] 0009 88024174fab8 816f6d23 88024174fb00 [ 123.680742] 88024174faf0 8107b368 88003770 00050380 [ 123.680746] 00050380 880037700068 0001 88024174fb50 [ 123.680751] Call Trace: [ 123.680758] [816f6d23] dump_stack+0x4d/0x66 [ 123.680763] [8107b368] warn_slowpath_common+0x78/0xa0 [ 123.680766] [8107b3d7] warn_slowpath_fmt+0x47/0x50 [ 123.680772] [810c0efd] ? trace_hardirqs_on_caller+0x15d/0x200 [ 123.680791] [a0118c63] assert_device_not_suspended.isra.8+0x43/0x50 [i915] [ 123.680809] [a011cbd5] gen8_write32+0x35/0x180 [i915] [ 123.680821] [a00dbb49] gen8_fbc_sw_flush+0x29/0x30 [i915] [ 123.680840] [a013150d] intel_frontbuffer_flush+0x7d/0x90 [i915] [ 123.680859] [a0135f6b] intel_cursor_plane_update+0x13b/0x150 [i915] [ 123.680876] [a00276b0] setplane_internal+0x260/0x2a0 [drm] [ 123.680889] [a002780e] drm_mode_cursor_common+0x11e/0x310 [drm] [ 123.680904] [a002ae1c] drm_mode_cursor_ioctl+0x3c/0x40 [drm] [ 123.680914] [a001cc9f] drm_ioctl+0x1df/0x6a0 [drm] [ 123.680920] [816fe769] ? mutex_unlock+0x9/0x10 [ 123.680924] [811f3856] ? seq_read+0xb6/0x3e0 [ 123.680929] [811e2540] do_vfs_ioctl+0x2e0/0x4e0 [ 123.680934] [81701077] ? sysret_check+0x1b/0x56 [ 123.680939] [810c0efd] ? trace_hardirqs_on_caller+0x15d/0x200 [ 123.680943] [811e27c1] SyS_ioctl+0x81/0xa0 [ 123.680947] [81701052] system_call_fastpath+0x16/0x1b [ 123.680949] ---[ end trace 9f389682639d2eb9 ]--- [ 123.680953] [ cut here ] So, what we can/could do: 1 - Change gen8_fbc_sw_flush() do to nothing when FBC is disabled. Agreed! 2 - At some point of the stack above, avoid calling everything else if the screen and/or plane and/or crtc is disabled. But when? At intel_cursor_plane_update? At intel_frontbuffer_flush? Good question! 3 - Don't call gen8_fbc_sw_flush if the frontbuffer_bits don't include the primary plane bits (since, AFAIR, FBC ignores the sprite and cursors, right?) not sure. but we could try check primary bits + do only when we
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
As Paulo told the part 2 of his proposal is still missing and we will get the WARN of reading fbc register at fbc_enabled when on runtime_suspend... But I couldn't find the propper place to check for intel_crtc-active. This was another issue reported by power team that on screen off there was to much activity on our driver. So we need to check it more carefully the proper place to avoid all unecessary inactivities with display off. Do you have any suggestion Daniel? On Fri, Sep 5, 2014 at 12:35 PM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: On Fri, Sep 5, 2014 at 11:28 AM, Paulo Zanoni przan...@gmail.com wrote: 2014-08-04 7:51 GMT-03:00 Rodrigo Vivi rodrigo.v...@intel.com: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. This patch causes lots of WARNs when running igt/pm_rpm/cursor and a few other similar test cases. Now the interesting this is that we're trying to do FBC flushing, but: As I told you on the IRC already, thank you for finding this. It is good to get reasons and arguments. - FBC is disabled by default Yeah, we need to fix it although it just matter much - The screen is disabled - We're runtime suspended These are another error. Not caused by this patch. Why frontbuffer_flush is being called with screen disabled or on runtime suspend? - We're touching cursors when we decide to flush FBC. This is not actually a flush. But cleaning the cache to allow compression restart. I believe we could have some variable to indicate when we touched ring flush than on next frontbuffer we do the mmio cache clean. This is the first WARN that happens: [ 123.581179] [drm:intel_runtime_suspend] Device suspended [ 123.680666] [ cut here ] [ 123.680697] WARNING: CPU: 1 PID: 1776 at drivers/gpu/drm/i915/intel_uncore.c:47 assert_device_not_suspended.isra.8+0x43/0x50 [i915]() [ 123.680699] Device suspended [ 123.680700] Modules linked in: intel_rapl x86_pkg_temp_thermal serio_raw intel_powerclamp efivars btusb iwlmvm iwlwifi mei_me mei int3403_thermal snd_hda_codec_hdmi snd_hda_intel snd_hda_controller i2c_designware_platform snd_hda_codec i2c_designware_core snd_hwdep snd_pcm_oss snd_mixer_oss acpi_pad snd_pcm snd_timer fuse nls_utf8 nls_cp437 vfat fat sd_mod ahci libahci i915 sdhci_pci e1000e drm_kms_helper drm sdhci_acpi sdhci [ 123.680733] CPU: 1 PID: 1776 Comm: pm_rpm Not tainted 3.17.0-rc2.1409051503pz+ #1100 [ 123.680736] Hardware name: Intel Corporation Broadwell Client platform/Wilson Beach SDS, BIOS BDW-E2R1.86C.0072.R03.1405072127 05/07/2014 [ 123.680738] 0009 88024174fab8 816f6d23 88024174fb00 [ 123.680742] 88024174faf0 8107b368 88003770 00050380 [ 123.680746] 00050380 880037700068 0001 88024174fb50 [ 123.680751] Call Trace: [ 123.680758] [816f6d23] dump_stack+0x4d/0x66 [ 123.680763] [8107b368] warn_slowpath_common+0x78/0xa0 [ 123.680766] [8107b3d7] warn_slowpath_fmt+0x47/0x50 [ 123.680772] [810c0efd] ? trace_hardirqs_on_caller+0x15d/0x200 [ 123.680791] [a0118c63] assert_device_not_suspended.isra.8+0x43/0x50 [i915] [ 123.680809] [a011cbd5] gen8_write32+0x35/0x180 [i915] [ 123.680821] [a00dbb49] gen8_fbc_sw_flush+0x29/0x30 [i915] [ 123.680840] [a013150d] intel_frontbuffer_flush+0x7d/0x90 [i915] [ 123.680859] [a0135f6b] intel_cursor_plane_update+0x13b/0x150 [i915] [ 123.680876] [a00276b0] setplane_internal+0x260/0x2a0 [drm] [ 123.680889] [a002780e] drm_mode_cursor_common+0x11e/0x310 [drm] [ 123.680904] [a002ae1c] drm_mode_cursor_ioctl+0x3c/0x40 [drm] [ 123.680914] [a001cc9f] drm_ioctl+0x1df/0x6a0 [drm] [ 123.680920] [816fe769] ? mutex_unlock+0x9/0x10 [ 123.680924] [811f3856] ? seq_read+0xb6/0x3e0 [ 123.680929] [811e2540] do_vfs_ioctl+0x2e0/0x4e0 [ 123.680934] [81701077] ? sysret_check+0x1b/0x56 [ 123.680939] [810c0efd] ? trace_hardirqs_on_caller+0x15d/0x200 [ 123.680943] [811e27c1] SyS_ioctl+0x81/0xa0 [ 123.680947] [81701052] system_call_fastpath+0x16/0x1b [ 123.680949] ---[ end trace 9f389682639d2eb9 ]--- [ 123.680953]
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Tue, Aug 26, 2014 at 2:39 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: So I prefer to continue using the HW/ring version we have already working for HSW and merge this v3 to get FBC working at BDW. Well for that we first need to fix up the psr testcase. I really want that. And then I also want fbc enabled by default, which means you need to rebase/review Ville's patch series to make that work. Also I really think the fb frontbuffer tracking can be made to work for fbc - if you do it right you should actually end up with fewer frontbuffer flushes than what we currently do by submitting them through rings. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Tue, Aug 26, 2014 at 12:54 AM, Daniel Vetter dan...@ffwll.ch wrote: On Tue, Aug 26, 2014 at 2:39 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: So I prefer to continue using the HW/ring version we have already working for HSW and merge this v3 to get FBC working at BDW. Well for that we first need to fix up the psr testcase. I really want that. fbc and psr are independend features and tasks prioritization should come from managers and program managers, right?! And then I also want fbc enabled by default, which means you need to rebase/review Ville's patch series to make that work. We have FBC working with issues and protected by parameter on all platforms. On BDW there is no fbc at all. This patch makes FBC state at least go to the same level as we already have FBC working on all other platforms. I don't see a dependency here between this fix and the big FBC rework-fix. This patch isn't enabling FBC by default. But allowing people that want and need to use FBC. Also I really think the fb frontbuffer tracking can be made to work for fbc - if you do it right you should actually end up with fewer frontbuffer flushes than what we currently do by submitting them through rings. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch -- Rodrigo Vivi Blog: http://blog.vivi.eng.br ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Tue, Aug 26, 2014 at 8:38 PM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: On Tue, Aug 26, 2014 at 12:54 AM, Daniel Vetter dan...@ffwll.ch wrote: On Tue, Aug 26, 2014 at 2:39 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: So I prefer to continue using the HW/ring version we have already working for HSW and merge this v3 to get FBC working at BDW. Well for that we first need to fix up the psr testcase. I really want that. fbc and psr are independend features and tasks prioritization should come from managers and program managers, right?! And then I also want fbc enabled by default, which means you need to rebase/review Ville's patch series to make that work. We have FBC working with issues and protected by parameter on all platforms. On BDW there is no fbc at all. This patch makes FBC state at least go to the same level as we already have FBC working on all other platforms. I don't see a dependency here between this fix and the big FBC rework-fix. This patch isn't enabling FBC by default. But allowing people that want and need to use FBC. It's going to be the same answer for both parts - I don't really like if we add features but don't complete them (so testcases and enabled by default). And in our long meeting today a lot of people asked my why I'm reluctant to merge patches so that we can clean them up in-tree (which I agree is often the much more efficient approach). Reactions like yours here are pretty much the reason for that - getting something in to make an internal customer happy or check of a box in our tracking or some other requirement and then move on right away. And I know that you yourself are in a very bad spot trenched between me and requests from management and project tracking. And this is also not to single out you at all, it's just what I see all over. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
I tried your suggestion for nuke fbc less... so I was just nuking along with psr_exit and cleaning the cache along with psr_do_enable and it got really better... I could see false colors working nice enough... However it got worst on PC7 residency... PC7 flutuated much more with the sw version... So I prefer to continue using the HW/ring version we have already working for HSW and merge this v3 to get FBC working at BDW. On Thu, Aug 21, 2014 at 9:44 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: List was accidentally drop. I didn't mean it. Sorry. And yes, you are right, we need a way to reduce nukes and cleans similar that we have for psr. I'll try it. On Tue, Aug 19, 2014 at 11:58 AM, Daniel Vetter dan...@ffwll.ch wrote: Readding intel-gfx. Please don't drop mailing lists cc's without telling me. Thanks, Daniel On Tue, Aug 19, 2014 at 8:57 PM, Daniel Vetter dan...@ffwll.ch wrote: Yeah, that does a lot too much flushing - you need to track relevant dirty bits like psr does, and then only flush when there has been a preceeding invalidate with the primary plane frontbuffer bit for the pipe that's using fbc. On top of that there's room for more improvements (filtering out pageflips and optimizing that more, atm we just disable fbc over a pageflip which is a bit meh), and we should also be able to ditch all the existing fbc nuking we do from the cmd streamer. -Daniel On Tue, Aug 19, 2014 at 12:09 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=fbc-sw-nuke-hswid=71875d3331aa3baef4f6f6bd297cc70dd94df1b6 On Fri, Aug 8, 2014 at 12:06 AM, Daniel Vetter dan...@ffwll.ch wrote: On Thu, Aug 07, 2014 at 01:04:19PM -0700, Rodrigo Vivi wrote: I tested here on HSW a full sw nuke/cache clean and I didn't liked the result. It seems to compress less than the hw one and to recompress everything a lot and stay less time compressed. That is really unexpected. For a modern desktop (i.e. anything that pageflips) there should be zero difference. And for actual frontbuffer rendering there should only be a difference when doing tiny cpu rendering to the frontbuffer. So something didn't work out as expected. Can you please push the code somewhere, or just submit patches to intel-gfx? Thanks, Daniel So, imho v3 is the way to go. On Mon, Aug 4, 2014 at 3:51 AM, Rodrigo Vivi rodrigo.v...@intel.com wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -2713,6 +2713,7 @@ extern void intel_modeset_setup_hw_state(struct drm_device *dev, extern void i915_redisable_vga(struct drm_device *dev); extern void i915_redisable_vga_power_on(struct drm_device *dev); extern bool intel_fbc_enabled(struct drm_device *dev); +extern void gen8_fbc_sw_flush(struct drm_device *dev, u32 value); extern void intel_disable_fbc(struct drm_device *dev); extern bool ironlake_set_drps(struct drm_device *dev, u8 val); extern void intel_init_pch_refclk(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 883af0b..c8421cd 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -9044,6 +9044,9 @@ void intel_frontbuffer_flush(struct drm_device *dev, intel_mark_fb_busy(dev, frontbuffer_bits, NULL); intel_edp_psr_flush(dev, frontbuffer_bits); + + if (IS_GEN8(dev)) + gen8_fbc_sw_flush(dev, FBC_REND_CACHE_CLEAN); } /** diff --git a/drivers/gpu/drm/i915/intel_pm.c
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
List was accidentally drop. I didn't mean it. Sorry. And yes, you are right, we need a way to reduce nukes and cleans similar that we have for psr. I'll try it. On Tue, Aug 19, 2014 at 11:58 AM, Daniel Vetter dan...@ffwll.ch wrote: Readding intel-gfx. Please don't drop mailing lists cc's without telling me. Thanks, Daniel On Tue, Aug 19, 2014 at 8:57 PM, Daniel Vetter dan...@ffwll.ch wrote: Yeah, that does a lot too much flushing - you need to track relevant dirty bits like psr does, and then only flush when there has been a preceeding invalidate with the primary plane frontbuffer bit for the pipe that's using fbc. On top of that there's room for more improvements (filtering out pageflips and optimizing that more, atm we just disable fbc over a pageflip which is a bit meh), and we should also be able to ditch all the existing fbc nuking we do from the cmd streamer. -Daniel On Tue, Aug 19, 2014 at 12:09 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=fbc-sw-nuke-hswid=71875d3331aa3baef4f6f6bd297cc70dd94df1b6 On Fri, Aug 8, 2014 at 12:06 AM, Daniel Vetter dan...@ffwll.ch wrote: On Thu, Aug 07, 2014 at 01:04:19PM -0700, Rodrigo Vivi wrote: I tested here on HSW a full sw nuke/cache clean and I didn't liked the result. It seems to compress less than the hw one and to recompress everything a lot and stay less time compressed. That is really unexpected. For a modern desktop (i.e. anything that pageflips) there should be zero difference. And for actual frontbuffer rendering there should only be a difference when doing tiny cpu rendering to the frontbuffer. So something didn't work out as expected. Can you please push the code somewhere, or just submit patches to intel-gfx? Thanks, Daniel So, imho v3 is the way to go. On Mon, Aug 4, 2014 at 3:51 AM, Rodrigo Vivi rodrigo.v...@intel.com wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -2713,6 +2713,7 @@ extern void intel_modeset_setup_hw_state(struct drm_device *dev, extern void i915_redisable_vga(struct drm_device *dev); extern void i915_redisable_vga_power_on(struct drm_device *dev); extern bool intel_fbc_enabled(struct drm_device *dev); +extern void gen8_fbc_sw_flush(struct drm_device *dev, u32 value); extern void intel_disable_fbc(struct drm_device *dev); extern bool ironlake_set_drps(struct drm_device *dev, u8 val); extern void intel_init_pch_refclk(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 883af0b..c8421cd 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -9044,6 +9044,9 @@ void intel_frontbuffer_flush(struct drm_device *dev, intel_mark_fb_busy(dev, frontbuffer_bits, NULL); intel_edp_psr_flush(dev, frontbuffer_bits); + + if (IS_GEN8(dev)) + gen8_fbc_sw_flush(dev, FBC_REND_CACHE_CLEAN); } /** diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 684dc5f..de07d3e 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -345,6 +345,16 @@ bool intel_fbc_enabled(struct drm_device *dev) return dev_priv-display.fbc_enabled(dev); } +void gen8_fbc_sw_flush(struct drm_device *dev, u32 value) +{ + struct drm_i915_private *dev_priv = dev-dev_private; + + if (!IS_GEN8(dev)) + return; + + I915_WRITE(MSG_FBC_REND_STATE,
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
Readding intel-gfx. Please don't drop mailing lists cc's without telling me. Thanks, Daniel On Tue, Aug 19, 2014 at 8:57 PM, Daniel Vetter dan...@ffwll.ch wrote: Yeah, that does a lot too much flushing - you need to track relevant dirty bits like psr does, and then only flush when there has been a preceeding invalidate with the primary plane frontbuffer bit for the pipe that's using fbc. On top of that there's room for more improvements (filtering out pageflips and optimizing that more, atm we just disable fbc over a pageflip which is a bit meh), and we should also be able to ditch all the existing fbc nuking we do from the cmd streamer. -Daniel On Tue, Aug 19, 2014 at 12:09 AM, Rodrigo Vivi rodrigo.v...@gmail.com wrote: http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=fbc-sw-nuke-hswid=71875d3331aa3baef4f6f6bd297cc70dd94df1b6 On Fri, Aug 8, 2014 at 12:06 AM, Daniel Vetter dan...@ffwll.ch wrote: On Thu, Aug 07, 2014 at 01:04:19PM -0700, Rodrigo Vivi wrote: I tested here on HSW a full sw nuke/cache clean and I didn't liked the result. It seems to compress less than the hw one and to recompress everything a lot and stay less time compressed. That is really unexpected. For a modern desktop (i.e. anything that pageflips) there should be zero difference. And for actual frontbuffer rendering there should only be a difference when doing tiny cpu rendering to the frontbuffer. So something didn't work out as expected. Can you please push the code somewhere, or just submit patches to intel-gfx? Thanks, Daniel So, imho v3 is the way to go. On Mon, Aug 4, 2014 at 3:51 AM, Rodrigo Vivi rodrigo.v...@intel.com wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -2713,6 +2713,7 @@ extern void intel_modeset_setup_hw_state(struct drm_device *dev, extern void i915_redisable_vga(struct drm_device *dev); extern void i915_redisable_vga_power_on(struct drm_device *dev); extern bool intel_fbc_enabled(struct drm_device *dev); +extern void gen8_fbc_sw_flush(struct drm_device *dev, u32 value); extern void intel_disable_fbc(struct drm_device *dev); extern bool ironlake_set_drps(struct drm_device *dev, u8 val); extern void intel_init_pch_refclk(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 883af0b..c8421cd 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -9044,6 +9044,9 @@ void intel_frontbuffer_flush(struct drm_device *dev, intel_mark_fb_busy(dev, frontbuffer_bits, NULL); intel_edp_psr_flush(dev, frontbuffer_bits); + + if (IS_GEN8(dev)) + gen8_fbc_sw_flush(dev, FBC_REND_CACHE_CLEAN); } /** diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 684dc5f..de07d3e 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -345,6 +345,16 @@ bool intel_fbc_enabled(struct drm_device *dev) return dev_priv-display.fbc_enabled(dev); } +void gen8_fbc_sw_flush(struct drm_device *dev, u32 value) +{ + struct drm_i915_private *dev_priv = dev-dev_private; + + if (!IS_GEN8(dev)) + return; + + I915_WRITE(MSG_FBC_REND_STATE, value); +} + static void intel_fbc_work_fn(struct work_struct *__work) { struct intel_fbc_work *work = diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..2fe871c 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Thu, Aug 07, 2014 at 01:04:19PM -0700, Rodrigo Vivi wrote: I tested here on HSW a full sw nuke/cache clean and I didn't liked the result. It seems to compress less than the hw one and to recompress everything a lot and stay less time compressed. That is really unexpected. For a modern desktop (i.e. anything that pageflips) there should be zero difference. And for actual frontbuffer rendering there should only be a difference when doing tiny cpu rendering to the frontbuffer. So something didn't work out as expected. Can you please push the code somewhere, or just submit patches to intel-gfx? Thanks, Daniel So, imho v3 is the way to go. On Mon, Aug 4, 2014 at 3:51 AM, Rodrigo Vivi rodrigo.v...@intel.com wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -2713,6 +2713,7 @@ extern void intel_modeset_setup_hw_state(struct drm_device *dev, extern void i915_redisable_vga(struct drm_device *dev); extern void i915_redisable_vga_power_on(struct drm_device *dev); extern bool intel_fbc_enabled(struct drm_device *dev); +extern void gen8_fbc_sw_flush(struct drm_device *dev, u32 value); extern void intel_disable_fbc(struct drm_device *dev); extern bool ironlake_set_drps(struct drm_device *dev, u8 val); extern void intel_init_pch_refclk(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 883af0b..c8421cd 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -9044,6 +9044,9 @@ void intel_frontbuffer_flush(struct drm_device *dev, intel_mark_fb_busy(dev, frontbuffer_bits, NULL); intel_edp_psr_flush(dev, frontbuffer_bits); + + if (IS_GEN8(dev)) + gen8_fbc_sw_flush(dev, FBC_REND_CACHE_CLEAN); } /** diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 684dc5f..de07d3e 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -345,6 +345,16 @@ bool intel_fbc_enabled(struct drm_device *dev) return dev_priv-display.fbc_enabled(dev); } +void gen8_fbc_sw_flush(struct drm_device *dev, u32 value) +{ + struct drm_i915_private *dev_priv = dev-dev_private; + + if (!IS_GEN8(dev)) + return; + + I915_WRITE(MSG_FBC_REND_STATE, value); +} + static void intel_fbc_work_fn(struct work_struct *__work) { struct intel_fbc_work *work = diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..2fe871c 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -406,6 +406,7 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, { u32 flags = 0; u32 scratch_addr = ring-scratch.gtt_offset + 2 * CACHELINE_BYTES; + int ret; flags |= PIPE_CONTROL_CS_STALL; @@ -424,7 +425,14 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; } - return gen8_emit_pipe_control(ring, flags, scratch_addr); + ret = gen8_emit_pipe_control(ring, flags, scratch_addr); + if (ret) + return ret; + + if (!invalidate_domains flush_domains) + return gen7_ring_fbc_flush(ring, FBC_REND_NUKE); + + return 0; } static void ring_write_tail(struct intel_engine_cs *ring, -- 1.9.3 ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx -- Rodrigo Vivi Blog: http://blog.vivi.eng.br
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
I tested here on HSW a full sw nuke/cache clean and I didn't liked the result. It seems to compress less than the hw one and to recompress everything a lot and stay less time compressed. So, imho v3 is the way to go. On Mon, Aug 4, 2014 at 3:51 AM, Rodrigo Vivi rodrigo.v...@intel.com wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -2713,6 +2713,7 @@ extern void intel_modeset_setup_hw_state(struct drm_device *dev, extern void i915_redisable_vga(struct drm_device *dev); extern void i915_redisable_vga_power_on(struct drm_device *dev); extern bool intel_fbc_enabled(struct drm_device *dev); +extern void gen8_fbc_sw_flush(struct drm_device *dev, u32 value); extern void intel_disable_fbc(struct drm_device *dev); extern bool ironlake_set_drps(struct drm_device *dev, u8 val); extern void intel_init_pch_refclk(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 883af0b..c8421cd 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -9044,6 +9044,9 @@ void intel_frontbuffer_flush(struct drm_device *dev, intel_mark_fb_busy(dev, frontbuffer_bits, NULL); intel_edp_psr_flush(dev, frontbuffer_bits); + + if (IS_GEN8(dev)) + gen8_fbc_sw_flush(dev, FBC_REND_CACHE_CLEAN); } /** diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 684dc5f..de07d3e 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -345,6 +345,16 @@ bool intel_fbc_enabled(struct drm_device *dev) return dev_priv-display.fbc_enabled(dev); } +void gen8_fbc_sw_flush(struct drm_device *dev, u32 value) +{ + struct drm_i915_private *dev_priv = dev-dev_private; + + if (!IS_GEN8(dev)) + return; + + I915_WRITE(MSG_FBC_REND_STATE, value); +} + static void intel_fbc_work_fn(struct work_struct *__work) { struct intel_fbc_work *work = diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..2fe871c 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -406,6 +406,7 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, { u32 flags = 0; u32 scratch_addr = ring-scratch.gtt_offset + 2 * CACHELINE_BYTES; + int ret; flags |= PIPE_CONTROL_CS_STALL; @@ -424,7 +425,14 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; } - return gen8_emit_pipe_control(ring, flags, scratch_addr); + ret = gen8_emit_pipe_control(ring, flags, scratch_addr); + if (ret) + return ret; + + if (!invalidate_domains flush_domains) + return gen7_ring_fbc_flush(ring, FBC_REND_NUKE); + + return 0; } static void ring_write_tail(struct intel_engine_cs *ring, -- 1.9.3 ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx -- Rodrigo Vivi Blog: http://blog.vivi.eng.br ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Fri, Aug 01, 2014 at 05:07:30PM +0300, Ville Syrjälä wrote: On Thu, Jul 31, 2014 at 12:07:44PM -0700, Rodrigo Vivi wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. v2: Fix rebase conflict. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/intel_ringbuffer.c | 12 ++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..4ba3db1 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -406,6 +406,7 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, { u32 flags = 0; u32 scratch_addr = ring-scratch.gtt_offset + 2 * CACHELINE_BYTES; + int ret; flags |= PIPE_CONTROL_CS_STALL; @@ -424,7 +425,14 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; } - return gen8_emit_pipe_control(ring, flags, scratch_addr); + ret = gen8_emit_pipe_control(ring, flags, scratch_addr); + if (ret) + return ret; + + if (!invalidate_domains flush_domains) + return gen7_ring_fbc_flush(ring, FBC_REND_NUKE); + + return 0; } static void ring_write_tail(struct intel_engine_cs *ring, @@ -2065,7 +2073,7 @@ static int gen6_ring_flush(struct intel_engine_cs *ring, } intel_ring_advance(ring); - if (IS_GEN7(dev) !invalidate flush) + if (INTEL_INFO(dev)-gen = 7 !invalidate flush) return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN); Apparently BDW has problems with LRIs to certain register offsets on !RCS and this here would hit that. The suggested workaround is to emit such LRIs only from RCS. Doing that would also involve tossing in a semaphore, so it feels like something that ought to be handled somewhere a bit higher up. *cough* sw frontbuffer tracking *cough* At least with fbc we have this nice igt test already. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
[Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. Unfortunatelly on some rings like BCS on BDW we have to avoid Bits 22:18 on LRIs due to a high risk of hung. So, when using Blt ring for frontbuffer rend cache would never been cleaned and FBC would stop compressing buffer. One alternative is to cache clean on software frontbuffer tracking. v2: Fix rebase conflict. v3: Do not clean cache on BCS ring. Instead use sw frontbuffer tracking. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/intel_display.c| 3 +++ drivers/gpu/drm/i915/intel_pm.c | 10 ++ drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +- 4 files changed, 23 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 2a372f2..25d7365 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -2713,6 +2713,7 @@ extern void intel_modeset_setup_hw_state(struct drm_device *dev, extern void i915_redisable_vga(struct drm_device *dev); extern void i915_redisable_vga_power_on(struct drm_device *dev); extern bool intel_fbc_enabled(struct drm_device *dev); +extern void gen8_fbc_sw_flush(struct drm_device *dev, u32 value); extern void intel_disable_fbc(struct drm_device *dev); extern bool ironlake_set_drps(struct drm_device *dev, u8 val); extern void intel_init_pch_refclk(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 883af0b..c8421cd 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -9044,6 +9044,9 @@ void intel_frontbuffer_flush(struct drm_device *dev, intel_mark_fb_busy(dev, frontbuffer_bits, NULL); intel_edp_psr_flush(dev, frontbuffer_bits); + + if (IS_GEN8(dev)) + gen8_fbc_sw_flush(dev, FBC_REND_CACHE_CLEAN); } /** diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 684dc5f..de07d3e 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -345,6 +345,16 @@ bool intel_fbc_enabled(struct drm_device *dev) return dev_priv-display.fbc_enabled(dev); } +void gen8_fbc_sw_flush(struct drm_device *dev, u32 value) +{ + struct drm_i915_private *dev_priv = dev-dev_private; + + if (!IS_GEN8(dev)) + return; + + I915_WRITE(MSG_FBC_REND_STATE, value); +} + static void intel_fbc_work_fn(struct work_struct *__work) { struct intel_fbc_work *work = diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..2fe871c 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -406,6 +406,7 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, { u32 flags = 0; u32 scratch_addr = ring-scratch.gtt_offset + 2 * CACHELINE_BYTES; + int ret; flags |= PIPE_CONTROL_CS_STALL; @@ -424,7 +425,14 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; } - return gen8_emit_pipe_control(ring, flags, scratch_addr); + ret = gen8_emit_pipe_control(ring, flags, scratch_addr); + if (ret) + return ret; + + if (!invalidate_domains flush_domains) + return gen7_ring_fbc_flush(ring, FBC_REND_NUKE); + + return 0; } static void ring_write_tail(struct intel_engine_cs *ring, -- 1.9.3 ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
On Thu, Jul 31, 2014 at 12:07:44PM -0700, Rodrigo Vivi wrote: According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. v2: Fix rebase conflict. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/intel_ringbuffer.c | 12 ++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..4ba3db1 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -406,6 +406,7 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, { u32 flags = 0; u32 scratch_addr = ring-scratch.gtt_offset + 2 * CACHELINE_BYTES; + int ret; flags |= PIPE_CONTROL_CS_STALL; @@ -424,7 +425,14 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; } - return gen8_emit_pipe_control(ring, flags, scratch_addr); + ret = gen8_emit_pipe_control(ring, flags, scratch_addr); + if (ret) + return ret; + + if (!invalidate_domains flush_domains) + return gen7_ring_fbc_flush(ring, FBC_REND_NUKE); + + return 0; } static void ring_write_tail(struct intel_engine_cs *ring, @@ -2065,7 +2073,7 @@ static int gen6_ring_flush(struct intel_engine_cs *ring, } intel_ring_advance(ring); - if (IS_GEN7(dev) !invalidate flush) + if (INTEL_INFO(dev)-gen = 7 !invalidate flush) return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN); Apparently BDW has problems with LRIs to certain register offsets on !RCS and this here would hit that. The suggested workaround is to emit such LRIs only from RCS. Doing that would also involve tossing in a semaphore, so it feels like something that ought to be handled somewhere a bit higher up. return 0; -- 1.9.3 ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx -- Ville Syrjälä Intel OTC ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
[Intel-gfx] [PATCH] drm/i915: FBC flush nuke for BDW
According to spec FBC on BDW and HSW are identical without any gaps. So let's copy the nuke and let FBC really start compressing stuff. Without this patch we can verify with false color that nothing is being compressed. With the nuke in place and false color it is possible to see false color debugs. v2: Fix rebase conflict. Signed-off-by: Rodrigo Vivi rodrigo.v...@intel.com --- drivers/gpu/drm/i915/intel_ringbuffer.c | 12 ++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 2908896..4ba3db1 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -406,6 +406,7 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, { u32 flags = 0; u32 scratch_addr = ring-scratch.gtt_offset + 2 * CACHELINE_BYTES; + int ret; flags |= PIPE_CONTROL_CS_STALL; @@ -424,7 +425,14 @@ gen8_render_ring_flush(struct intel_engine_cs *ring, flags |= PIPE_CONTROL_GLOBAL_GTT_IVB; } - return gen8_emit_pipe_control(ring, flags, scratch_addr); + ret = gen8_emit_pipe_control(ring, flags, scratch_addr); + if (ret) + return ret; + + if (!invalidate_domains flush_domains) + return gen7_ring_fbc_flush(ring, FBC_REND_NUKE); + + return 0; } static void ring_write_tail(struct intel_engine_cs *ring, @@ -2065,7 +2073,7 @@ static int gen6_ring_flush(struct intel_engine_cs *ring, } intel_ring_advance(ring); - if (IS_GEN7(dev) !invalidate flush) + if (INTEL_INFO(dev)-gen = 7 !invalidate flush) return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN); return 0; -- 1.9.3 ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx