* [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. @ 2015-01-05 1:05 Zhigang Gong 2015-01-05 3:03 ` Francisco Jerez 2015-01-05 7:05 ` shuang.he 0 siblings, 2 replies; 7+ messages in thread From: Zhigang Gong @ 2015-01-05 1:05 UTC (permalink / raw) To: intel-gfx; +Cc: Zhigang Gong According to bspec, ROW_CHICKEN3's bit 6 which is to disable move of cacheable global atomics to L3 is needed for GT3 D stepping. I enabled it and tested it with HSW GT2 D stepping and GT3 E stepping. The atomics works fine in beignet. And it could get more than 10x performance improvement with some workload, for an example, the "splat" kernel in darktable, without this patch, it consumes 50 seconds in one large raw picture processing. But with this patch, the same process only takes less than 1 second. Signed-off-by: Zhigang Gong <zhigang.gong@intel.com> --- drivers/gpu/drm/i915/intel_pm.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 7d99a9c..8a27802 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -5938,10 +5938,12 @@ static void haswell_init_clock_gating(struct drm_device *dev) ilk_init_lp_watermarks(dev); - /* L3 caching of data atomics doesn't work -- disable it. */ - I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); - I915_WRITE(HSW_ROW_CHICKEN3, - _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); + if (IS_HSW_GT3(dev) && dev->pdev->revision <= 6) { + /* L3 caching of data atomics doesn't work -- disable it. */ + I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); + I915_WRITE(HSW_ROW_CHICKEN3, + _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); + } /* This is required by WaCatErrorRejectionIssue:hsw */ I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG, -- 1.8.3.2 _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. 2015-01-05 1:05 [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings Zhigang Gong @ 2015-01-05 3:03 ` Francisco Jerez 2015-01-05 2:36 ` Zhigang Gong 2015-01-05 7:05 ` shuang.he 1 sibling, 1 reply; 7+ messages in thread From: Francisco Jerez @ 2015-01-05 3:03 UTC (permalink / raw) To: intel-gfx; +Cc: Zhigang Gong [-- Attachment #1.1.1: Type: text/plain, Size: 2552 bytes --] Zhigang Gong <zhigang.gong@intel.com> writes: > According to bspec, ROW_CHICKEN3's bit 6 which is to disable > move of cacheable global atomics to L3 is needed for GT3 D > stepping. > > I enabled it and tested it with HSW GT2 D stepping and GT3 E stepping. > The atomics works fine in beignet. And it could get more than 10x performance > improvement with some workload, for an example, the "splat" kernel in darktable, > without this patch, it consumes 50 seconds in one large raw picture processing. > But with this patch, the same process only takes less than 1 second. > I tried this already (on HSW GT2 D as well) and I don't think it's enough to get L3 atomics working reliably. Even though they did seem to work OK at first glance I observed some corruption issues (e.g. atomic writes not landing in system memory) when doing atomic writes to contiguous (as in within the same cache-line) locations in memory. The "unused" ARB_shader_image_load_store test [1] I sent to the Piglit mailing list some time ago exposes this IIRC, and probably a couple of other tests too. Also this change is going to cause an instant lock-up anytime Mesa uses atomics because Mesa doesn't change the default L3 way allocation for the DC, which turns out to be 0 on HSW. [1] http://lists.freedesktop.org/archives/piglit/2014-December/013571.html > Signed-off-by: Zhigang Gong <zhigang.gong@intel.com> > --- > drivers/gpu/drm/i915/intel_pm.c | 10 ++++++---- > 1 file changed, 6 insertions(+), 4 deletions(-) > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c > index 7d99a9c..8a27802 100644 > --- a/drivers/gpu/drm/i915/intel_pm.c > +++ b/drivers/gpu/drm/i915/intel_pm.c > @@ -5938,10 +5938,12 @@ static void haswell_init_clock_gating(struct drm_device *dev) > > ilk_init_lp_watermarks(dev); > > - /* L3 caching of data atomics doesn't work -- disable it. */ > - I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); > - I915_WRITE(HSW_ROW_CHICKEN3, > - _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); > + if (IS_HSW_GT3(dev) && dev->pdev->revision <= 6) { > + /* L3 caching of data atomics doesn't work -- disable it. */ > + I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); > + I915_WRITE(HSW_ROW_CHICKEN3, > + _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); > + } > > /* This is required by WaCatErrorRejectionIssue:hsw */ > I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG, > -- > 1.8.3.2 [-- Attachment #1.2: Type: application/pgp-signature, Size: 212 bytes --] [-- Attachment #2: Type: text/plain, Size: 159 bytes --] _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. 2015-01-05 3:03 ` Francisco Jerez @ 2015-01-05 2:36 ` Zhigang Gong 2015-01-05 17:27 ` Francisco Jerez 0 siblings, 1 reply; 7+ messages in thread From: Zhigang Gong @ 2015-01-05 2:36 UTC (permalink / raw) To: Francisco Jerez; +Cc: intel-gfx, Zhigang Gong On Mon, Jan 05, 2015 at 05:03:16AM +0200, Francisco Jerez wrote: > Zhigang Gong <zhigang.gong@intel.com> writes: > > > According to bspec, ROW_CHICKEN3's bit 6 which is to disable > > move of cacheable global atomics to L3 is needed for GT3 D > > stepping. > > > > I enabled it and tested it with HSW GT2 D stepping and GT3 E stepping. > > The atomics works fine in beignet. And it could get more than 10x performance > > improvement with some workload, for an example, the "splat" kernel in darktable, > > without this patch, it consumes 50 seconds in one large raw picture processing. > > But with this patch, the same process only takes less than 1 second. > > > > I tried this already (on HSW GT2 D as well) and I don't think it's > enough to get L3 atomics working reliably. Even though they did seem to > work OK at first glance I observed some corruption issues (e.g. atomic > writes not landing in system memory) when doing atomic writes to > contiguous (as in within the same cache-line) locations in memory. The > "unused" ARB_shader_image_load_store test [1] I sent to the Piglit > mailing list some time ago exposes this IIRC, and probably a couple of > other tests too. Ok, I will find that case and have a try on my systems. I just tested all the atomic related OpenCL conformance test cases without any issues. > > Also this change is going to cause an instant lock-up anytime Mesa uses > atomics because Mesa doesn't change the default L3 way allocation for > the DC, which turns out to be 0 on HSW. This is another issue, IMHO, if the application wants to use atomics, it's better to allocate some L3 space for DC. Otherwise, it could never leverage the "atomics in L3 feature". Based on my test, the performance impact as huge as more than 10x for some workloads. Thanks, Zhigang Gong. > > [1] http://lists.freedesktop.org/archives/piglit/2014-December/013571.html > > > Signed-off-by: Zhigang Gong <zhigang.gong@intel.com> > > --- > > drivers/gpu/drm/i915/intel_pm.c | 10 ++++++---- > > 1 file changed, 6 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c > > index 7d99a9c..8a27802 100644 > > --- a/drivers/gpu/drm/i915/intel_pm.c > > +++ b/drivers/gpu/drm/i915/intel_pm.c > > @@ -5938,10 +5938,12 @@ static void haswell_init_clock_gating(struct drm_device *dev) > > > > ilk_init_lp_watermarks(dev); > > > > - /* L3 caching of data atomics doesn't work -- disable it. */ > > - I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); > > - I915_WRITE(HSW_ROW_CHICKEN3, > > - _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); > > + if (IS_HSW_GT3(dev) && dev->pdev->revision <= 6) { > > + /* L3 caching of data atomics doesn't work -- disable it. */ > > + I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); > > + I915_WRITE(HSW_ROW_CHICKEN3, > > + _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); > > + } > > > > /* This is required by WaCatErrorRejectionIssue:hsw */ > > I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG, > > -- > > 1.8.3.2 > _______________________________________________ > Intel-gfx mailing list > Intel-gfx@lists.freedesktop.org > http://lists.freedesktop.org/mailman/listinfo/intel-gfx _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. 2015-01-05 2:36 ` Zhigang Gong @ 2015-01-05 17:27 ` Francisco Jerez 2015-01-06 8:01 ` Zhigang Gong 0 siblings, 1 reply; 7+ messages in thread From: Francisco Jerez @ 2015-01-05 17:27 UTC (permalink / raw) To: Zhigang Gong; +Cc: intel-gfx, Zhigang Gong [-- Attachment #1.1.1: Type: text/plain, Size: 4272 bytes --] Zhigang Gong <zhigang.gong@linux.intel.com> writes: > On Mon, Jan 05, 2015 at 05:03:16AM +0200, Francisco Jerez wrote: >> Zhigang Gong <zhigang.gong@intel.com> writes: >> >> > According to bspec, ROW_CHICKEN3's bit 6 which is to disable >> > move of cacheable global atomics to L3 is needed for GT3 D >> > stepping. >> > >> > I enabled it and tested it with HSW GT2 D stepping and GT3 E stepping. >> > The atomics works fine in beignet. And it could get more than 10x performance >> > improvement with some workload, for an example, the "splat" kernel in darktable, >> > without this patch, it consumes 50 seconds in one large raw picture processing. >> > But with this patch, the same process only takes less than 1 second. >> > >> >> I tried this already (on HSW GT2 D as well) and I don't think it's >> enough to get L3 atomics working reliably. Even though they did seem to >> work OK at first glance I observed some corruption issues (e.g. atomic >> writes not landing in system memory) when doing atomic writes to >> contiguous (as in within the same cache-line) locations in memory. The >> "unused" ARB_shader_image_load_store test [1] I sent to the Piglit >> mailing list some time ago exposes this IIRC, and probably a couple of >> other tests too. > Ok, I will find that case and have a try on my systems. I just tested all > the atomic related OpenCL conformance test cases without any issues. > >> >> Also this change is going to cause an instant lock-up anytime Mesa uses >> atomics because Mesa doesn't change the default L3 way allocation for >> the DC, which turns out to be 0 on HSW. > > This is another issue, IMHO, if the application wants to use atomics, > it's better to allocate some L3 space for DC. Otherwise, it could > never leverage the "atomics in L3 feature". Based on my test, > the performance impact as huge as more than 10x for some workloads. > Sure, but this change alone will cause a regression (an irrecoverable system hang) with current releases of Mesa. I agree that Mesa should be fixed eventually to assign L3 space to the DC for some workloads, but it doesn't seem like we have a satisfactory API to do that right now -- The current mechanism used by Beignet (poking the L3 control registers from the batch buffer) is slightly concerning from a security point of view because it allows an arbitrary userspace application to cause misrendering or impact the performance of other clients (since the L3 config registers are not being saved and restored as part of the context) and even crash the whole system. It also doesn't look like it could be made to work with the hardware command checker. > Thanks, > Zhigang Gong. > >> >> [1] http://lists.freedesktop.org/archives/piglit/2014-December/013571.html >> >> > Signed-off-by: Zhigang Gong <zhigang.gong@intel.com> >> > --- >> > drivers/gpu/drm/i915/intel_pm.c | 10 ++++++---- >> > 1 file changed, 6 insertions(+), 4 deletions(-) >> > >> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c >> > index 7d99a9c..8a27802 100644 >> > --- a/drivers/gpu/drm/i915/intel_pm.c >> > +++ b/drivers/gpu/drm/i915/intel_pm.c >> > @@ -5938,10 +5938,12 @@ static void haswell_init_clock_gating(struct drm_device *dev) >> > >> > ilk_init_lp_watermarks(dev); >> > >> > - /* L3 caching of data atomics doesn't work -- disable it. */ >> > - I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); >> > - I915_WRITE(HSW_ROW_CHICKEN3, >> > - _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); >> > + if (IS_HSW_GT3(dev) && dev->pdev->revision <= 6) { >> > + /* L3 caching of data atomics doesn't work -- disable it. */ >> > + I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); >> > + I915_WRITE(HSW_ROW_CHICKEN3, >> > + _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); >> > + } >> > >> > /* This is required by WaCatErrorRejectionIssue:hsw */ >> > I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG, >> > -- >> > 1.8.3.2 > > > > >> _______________________________________________ >> Intel-gfx mailing list >> Intel-gfx@lists.freedesktop.org >> http://lists.freedesktop.org/mailman/listinfo/intel-gfx [-- Attachment #1.2: Type: application/pgp-signature, Size: 212 bytes --] [-- Attachment #2: Type: text/plain, Size: 159 bytes --] _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. 2015-01-05 17:27 ` Francisco Jerez @ 2015-01-06 8:01 ` Zhigang Gong 2015-01-06 14:19 ` Francisco Jerez 0 siblings, 1 reply; 7+ messages in thread From: Zhigang Gong @ 2015-01-06 8:01 UTC (permalink / raw) To: Francisco Jerez; +Cc: intel-gfx, Zhigang Gong On Mon, Jan 05, 2015 at 07:27:26PM +0200, Francisco Jerez wrote: > Zhigang Gong <zhigang.gong@linux.intel.com> writes: > > > On Mon, Jan 05, 2015 at 05:03:16AM +0200, Francisco Jerez wrote: > >> Zhigang Gong <zhigang.gong@intel.com> writes: > >> > >> > According to bspec, ROW_CHICKEN3's bit 6 which is to disable > >> > move of cacheable global atomics to L3 is needed for GT3 D > >> > stepping. > >> > > >> > I enabled it and tested it with HSW GT2 D stepping and GT3 E stepping. > >> > The atomics works fine in beignet. And it could get more than 10x performance > >> > improvement with some workload, for an example, the "splat" kernel in darktable, > >> > without this patch, it consumes 50 seconds in one large raw picture processing. > >> > But with this patch, the same process only takes less than 1 second. > >> > > >> > >> I tried this already (on HSW GT2 D as well) and I don't think it's > >> enough to get L3 atomics working reliably. Even though they did seem to > >> work OK at first glance I observed some corruption issues (e.g. atomic > >> writes not landing in system memory) when doing atomic writes to > >> contiguous (as in within the same cache-line) locations in memory. The > >> "unused" ARB_shader_image_load_store test [1] I sent to the Piglit > >> mailing list some time ago exposes this IIRC, and probably a couple of > >> other tests too. > > Ok, I will find that case and have a try on my systems. I just tested all > > the atomic related OpenCL conformance test cases without any issues. I just found the patchset hasn't been accepted by piglit. I took a look at the source code. And realize that that shader will not work correctly if mesa doesn't allocate some DC in L3 space. Don't know whether you already allocated some DC when you did that test. > > > >> > >> Also this change is going to cause an instant lock-up anytime Mesa uses > >> atomics because Mesa doesn't change the default L3 way allocation for > >> the DC, which turns out to be 0 on HSW. > > > > This is another issue, IMHO, if the application wants to use atomics, > > it's better to allocate some L3 space for DC. Otherwise, it could > > never leverage the "atomics in L3 feature". Based on my test, > > the performance impact as huge as more than 10x for some workloads. > > > Sure, but this change alone will cause a regression (an irrecoverable > system hang) with current releases of Mesa. > > I agree that Mesa should be fixed eventually to assign L3 space to the > DC for some workloads, but it doesn't seem like we have a satisfactory > API to do that right now -- The current mechanism used by Beignet > (poking the L3 control registers from the batch buffer) is slightly > concerning from a security point of view because it allows an arbitrary > userspace application to cause misrendering or impact the performance of I agree with you that the L3 configuration is very dangerous. It's better to do more checking in the kernel space and make sure the user space application will not crash the system via L3 configuration. And, if the kernel provide new API we will use it in beignet ASAP in beignet. > other clients (since the L3 config registers are not being saved and > restored as part of the context) and even crash the whole system. It I found the L3 config registers are part of the IVB context, but not part of HSW's. So if two user applications want to use different L3 config, the user application need to do the config registers' store/restore manually. Is there a way to do this type of thing gracefully in KMD? The new API may need to consider the safely context switching regards to L3 config change. > also doesn't look like it could be made to work with the hardware > command checker. As to the hardware command checker: I just found with current nightly kernel, if we boot the kernel with "i915.enable_ppgtt=2", then we can configure the L3 related registers in user space. But the command checker still don't allow user space to chagne L3 config by default. Thanks, Zhigang Gong. > > > Thanks, > > Zhigang Gong. > > > >> > >> [1] http://lists.freedesktop.org/archives/piglit/2014-December/013571.html > >> > >> > Signed-off-by: Zhigang Gong <zhigang.gong@intel.com> > >> > --- > >> > drivers/gpu/drm/i915/intel_pm.c | 10 ++++++---- > >> > 1 file changed, 6 insertions(+), 4 deletions(-) > >> > > >> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c > >> > index 7d99a9c..8a27802 100644 > >> > --- a/drivers/gpu/drm/i915/intel_pm.c > >> > +++ b/drivers/gpu/drm/i915/intel_pm.c > >> > @@ -5938,10 +5938,12 @@ static void haswell_init_clock_gating(struct drm_device *dev) > >> > > >> > ilk_init_lp_watermarks(dev); > >> > > >> > - /* L3 caching of data atomics doesn't work -- disable it. */ > >> > - I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); > >> > - I915_WRITE(HSW_ROW_CHICKEN3, > >> > - _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); > >> > + if (IS_HSW_GT3(dev) && dev->pdev->revision <= 6) { > >> > + /* L3 caching of data atomics doesn't work -- disable it. */ > >> > + I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); > >> > + I915_WRITE(HSW_ROW_CHICKEN3, > >> > + _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); > >> > + } > >> > > >> > /* This is required by WaCatErrorRejectionIssue:hsw */ > >> > I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG, > >> > -- > >> > 1.8.3.2 > > > > > > > > > >> _______________________________________________ > >> Intel-gfx mailing list > >> Intel-gfx@lists.freedesktop.org > >> http://lists.freedesktop.org/mailman/listinfo/intel-gfx _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. 2015-01-06 8:01 ` Zhigang Gong @ 2015-01-06 14:19 ` Francisco Jerez 0 siblings, 0 replies; 7+ messages in thread From: Francisco Jerez @ 2015-01-06 14:19 UTC (permalink / raw) To: Zhigang Gong; +Cc: intel-gfx, Zhigang Gong [-- Attachment #1.1.1: Type: text/plain, Size: 6426 bytes --] Zhigang Gong <zhigang.gong@linux.intel.com> writes: > On Mon, Jan 05, 2015 at 07:27:26PM +0200, Francisco Jerez wrote: >> Zhigang Gong <zhigang.gong@linux.intel.com> writes: >> >> > On Mon, Jan 05, 2015 at 05:03:16AM +0200, Francisco Jerez wrote: >> >> Zhigang Gong <zhigang.gong@intel.com> writes: >> >> >> >> > According to bspec, ROW_CHICKEN3's bit 6 which is to disable >> >> > move of cacheable global atomics to L3 is needed for GT3 D >> >> > stepping. >> >> > >> >> > I enabled it and tested it with HSW GT2 D stepping and GT3 E stepping. >> >> > The atomics works fine in beignet. And it could get more than 10x performance >> >> > improvement with some workload, for an example, the "splat" kernel in darktable, >> >> > without this patch, it consumes 50 seconds in one large raw picture processing. >> >> > But with this patch, the same process only takes less than 1 second. >> >> > >> >> >> >> I tried this already (on HSW GT2 D as well) and I don't think it's >> >> enough to get L3 atomics working reliably. Even though they did seem to >> >> work OK at first glance I observed some corruption issues (e.g. atomic >> >> writes not landing in system memory) when doing atomic writes to >> >> contiguous (as in within the same cache-line) locations in memory. The >> >> "unused" ARB_shader_image_load_store test [1] I sent to the Piglit >> >> mailing list some time ago exposes this IIRC, and probably a couple of >> >> other tests too. >> > Ok, I will find that case and have a try on my systems. I just tested all >> > the atomic related OpenCL conformance test cases without any issues. > > I just found the patchset hasn't been accepted by piglit. I took a look at > the source code. And realize that that shader will not work correctly if > mesa doesn't allocate some DC in L3 space. Don't know whether you already > allocated some DC when you did that test. > Of course, and it still fails no matter how you allocate the L3. >> > >> >> >> >> Also this change is going to cause an instant lock-up anytime Mesa uses >> >> atomics because Mesa doesn't change the default L3 way allocation for >> >> the DC, which turns out to be 0 on HSW. >> > >> > This is another issue, IMHO, if the application wants to use atomics, >> > it's better to allocate some L3 space for DC. Otherwise, it could >> > never leverage the "atomics in L3 feature". Based on my test, >> > the performance impact as huge as more than 10x for some workloads. >> > >> Sure, but this change alone will cause a regression (an irrecoverable >> system hang) with current releases of Mesa. >> >> I agree that Mesa should be fixed eventually to assign L3 space to the >> DC for some workloads, but it doesn't seem like we have a satisfactory >> API to do that right now -- The current mechanism used by Beignet >> (poking the L3 control registers from the batch buffer) is slightly >> concerning from a security point of view because it allows an arbitrary >> userspace application to cause misrendering or impact the performance of > I agree with you that the L3 configuration is very dangerous. It's better > to do more checking in the kernel space and make sure the user space > application will not crash the system via L3 configuration. > And, if the kernel provide new API we will use it in beignet ASAP in beignet. > >> other clients (since the L3 config registers are not being saved and >> restored as part of the context) and even crash the whole system. It > I found the L3 config registers are part of the IVB context, but not part > of HSW's. So if two user applications want to use different L3 config, > the user application need to do the config registers' store/restore manually. > Is there a way to do this type of thing gracefully in KMD? The new API > may need to consider the safely context switching regards to L3 config change. > Yeah, relying on userspace saving and restoring the L3 config values sounds rather fragile to me, and may not work if e.g. a context hangs and gets kicked out. IMHO it would be preferrable to keep track of the L3 config and save/restore it from the kernel on context switch. >> also doesn't look like it could be made to work with the hardware >> command checker. > > As to the hardware command checker: > I just found with current nightly kernel, if we boot the kernel with > "i915.enable_ppgtt=2", then we can configure the L3 related registers > in user space. But the command checker still don't allow user space to > chagne L3 config by default. > Ugh, apparently that's because the hardware command checker is not enabled at all in that case... > Thanks, > Zhigang Gong. > >> >> > Thanks, >> > Zhigang Gong. >> > >> >> >> >> [1] http://lists.freedesktop.org/archives/piglit/2014-December/013571.html >> >> >> >> > Signed-off-by: Zhigang Gong <zhigang.gong@intel.com> >> >> > --- >> >> > drivers/gpu/drm/i915/intel_pm.c | 10 ++++++---- >> >> > 1 file changed, 6 insertions(+), 4 deletions(-) >> >> > >> >> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c >> >> > index 7d99a9c..8a27802 100644 >> >> > --- a/drivers/gpu/drm/i915/intel_pm.c >> >> > +++ b/drivers/gpu/drm/i915/intel_pm.c >> >> > @@ -5938,10 +5938,12 @@ static void haswell_init_clock_gating(struct drm_device *dev) >> >> > >> >> > ilk_init_lp_watermarks(dev); >> >> > >> >> > - /* L3 caching of data atomics doesn't work -- disable it. */ >> >> > - I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); >> >> > - I915_WRITE(HSW_ROW_CHICKEN3, >> >> > - _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); >> >> > + if (IS_HSW_GT3(dev) && dev->pdev->revision <= 6) { >> >> > + /* L3 caching of data atomics doesn't work -- disable it. */ >> >> > + I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE); >> >> > + I915_WRITE(HSW_ROW_CHICKEN3, >> >> > + _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE)); >> >> > + } >> >> > >> >> > /* This is required by WaCatErrorRejectionIssue:hsw */ >> >> > I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG, >> >> > -- >> >> > 1.8.3.2 >> > >> > >> > >> > >> >> _______________________________________________ >> >> Intel-gfx mailing list >> >> Intel-gfx@lists.freedesktop.org >> >> http://lists.freedesktop.org/mailman/listinfo/intel-gfx [-- Attachment #1.2: Type: application/pgp-signature, Size: 212 bytes --] [-- Attachment #2: Type: text/plain, Size: 159 bytes --] _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings. 2015-01-05 1:05 [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings Zhigang Gong 2015-01-05 3:03 ` Francisco Jerez @ 2015-01-05 7:05 ` shuang.he 1 sibling, 0 replies; 7+ messages in thread From: shuang.he @ 2015-01-05 7:05 UTC (permalink / raw) To: shuang.he, intel-gfx, zhigang.gong Tested-By: PRC QA PRTS (Patch Regression Test System Contact: shuang.he@intel.com) -------------------------------------Summary------------------------------------- Platform Delta drm-intel-nightly Series Applied PNV 363/364 363/364 ILK 364/366 364/366 SNB +2 443/450 445/450 IVB 496/498 496/498 BYT 288/289 288/289 HSW +3-1 542/564 544/564 BDW 415/417 415/417 -------------------------------------Detailed------------------------------------- Platform Test drm-intel-nightly Series Applied *SNB igt_kms_flip_modeset-vs-vblank-race DMESG_WARN(3, M35) PASS(1, M35) SNB igt_kms_plane_plane-position-hole-pipe-B-plane-1 DMESG_WARN(1, M35)PASS(3, M35M22) PASS(1, M35) *HSW igt_kms_flip_dpms-vs-vblank-race-interruptible DMESG_WARN(2, M40) PASS(1, M40) HSW igt_kms_flip_single-buffer-flip-vs-dpms-off-vs-modeset DMESG_WARN(1, M40)PASS(3, M40M19) DMESG_WARN(1, M40) HSW igt_kms_plane_plane-panning-bottom-right-pipe-C-plane-1 TIMEOUT(2, M40)PASS(2, M19M40) PASS(1, M40) HSW igt_pm_rpm_modeset-non-lpsp-stress-no-wait NSPT(1, M19)DMESG_WARN(1, M40)PASS(2, M40) PASS(1, M40) Note: You need to pay more attention to line start with '*' _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx ^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2015-01-06 14:22 UTC | newest] Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2015-01-05 1:05 [PATCH] drm/i915/hsw: enable atomic in L3 for some steppings Zhigang Gong 2015-01-05 3:03 ` Francisco Jerez 2015-01-05 2:36 ` Zhigang Gong 2015-01-05 17:27 ` Francisco Jerez 2015-01-06 8:01 ` Zhigang Gong 2015-01-06 14:19 ` Francisco Jerez 2015-01-05 7:05 ` shuang.he
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.