[Intel-gfx] [PATCH] drm/i915/uapi: Add DRM_I915_QUERY_GEOMETRY_SUBSLICES
Matt Roper
matthew.d.roper at intel.com
Mon Mar 28 16:01:29 UTC 2022
On Mon, Mar 28, 2022 at 09:44:36AM +0100, Tvrtko Ursulin wrote:
>
> + Joonas
>
> On 25/03/2022 23:03, Francisco Jerez wrote:
> > Matt Atwood <matthew.s.atwood at intel.com> writes:
> >
> > > Newer platforms have DSS that aren't necessarily available for both
> > > geometry and compute, two queries will need to exist. This introduces
> > > the first, when passing a valid engine class and engine instance in the
> > > flags returns a topology describing geometry.
> > >
> > > v2: fix white space errors
> > > v3: change flags from hosting 2 8 bit numbers to holding a
> > > i915_engine_class_instance struct
> > >
> > > Cc: Ashutosh Dixit <ashutosh.dixit at intel.com>
> > > Cc: Matt Roper <matthew.d.roper at intel.com>
> > > Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> > > UMD (mesa): https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14143
> > > Signed-off-by: Matt Atwood <matthew.s.atwood at intel.com>
> > > ---
> > > drivers/gpu/drm/i915/i915_query.c | 68 ++++++++++++++++++++++---------
> > > include/uapi/drm/i915_drm.h | 24 +++++++----
> > > 2 files changed, 65 insertions(+), 27 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/i915_query.c b/drivers/gpu/drm/i915/i915_query.c
> > > index 2dfbc22857a3..fcb374201edb 100644
> > > --- a/drivers/gpu/drm/i915/i915_query.c
> > > +++ b/drivers/gpu/drm/i915/i915_query.c
> > > @@ -9,6 +9,7 @@
> > > #include "i915_drv.h"
> > > #include "i915_perf.h"
> > > #include "i915_query.h"
> > > +#include "gt/intel_engine_user.h"
> > > #include <uapi/drm/i915_drm.h>
> > > static int copy_query_item(void *query_hdr, size_t query_sz,
> > > @@ -28,36 +29,30 @@ static int copy_query_item(void *query_hdr, size_t query_sz,
> > > return 0;
> > > }
> > > -static int query_topology_info(struct drm_i915_private *dev_priv,
> > > - struct drm_i915_query_item *query_item)
> > > +static int fill_topology_info(const struct sseu_dev_info *sseu,
> > > + struct drm_i915_query_item *query_item,
> > > + const u8 *subslice_mask)
> > > {
> > > - const struct sseu_dev_info *sseu = &to_gt(dev_priv)->info.sseu;
> > > struct drm_i915_query_topology_info topo;
> > > u32 slice_length, subslice_length, eu_length, total_length;
> > > int ret;
> > > - if (query_item->flags != 0)
> > > - return -EINVAL;
> > > + BUILD_BUG_ON(sizeof(u8) != sizeof(sseu->slice_mask));
> > > if (sseu->max_slices == 0)
> > > return -ENODEV;
> > > - BUILD_BUG_ON(sizeof(u8) != sizeof(sseu->slice_mask));
> > > -
> > > slice_length = sizeof(sseu->slice_mask);
> > > subslice_length = sseu->max_slices * sseu->ss_stride;
> > > eu_length = sseu->max_slices * sseu->max_subslices * sseu->eu_stride;
> > > total_length = sizeof(topo) + slice_length + subslice_length +
> > > eu_length;
> > > - ret = copy_query_item(&topo, sizeof(topo), total_length,
> > > - query_item);
> > > + ret = copy_query_item(&topo, sizeof(topo), total_length, query_item);
> > > +
> > > if (ret != 0)
> > > return ret;
> > > - if (topo.flags != 0)
> > > - return -EINVAL;
> > > -
> > > memset(&topo, 0, sizeof(topo));
> > > topo.max_slices = sseu->max_slices;
> > > topo.max_subslices = sseu->max_subslices;
> > > @@ -69,27 +64,61 @@ static int query_topology_info(struct drm_i915_private *dev_priv,
> > > topo.eu_stride = sseu->eu_stride;
> > > if (copy_to_user(u64_to_user_ptr(query_item->data_ptr),
> > > - &topo, sizeof(topo)))
> > > + &topo, sizeof(topo)))
> > > return -EFAULT;
> > > if (copy_to_user(u64_to_user_ptr(query_item->data_ptr + sizeof(topo)),
> > > - &sseu->slice_mask, slice_length))
> > > + &sseu->slice_mask, slice_length))
> > > return -EFAULT;
> > > if (copy_to_user(u64_to_user_ptr(query_item->data_ptr +
> > > - sizeof(topo) + slice_length),
> > > - sseu->subslice_mask, subslice_length))
> > > + sizeof(topo) + slice_length),
> > > + subslice_mask, subslice_length))
> > > return -EFAULT;
> > > if (copy_to_user(u64_to_user_ptr(query_item->data_ptr +
> > > - sizeof(topo) +
> > > - slice_length + subslice_length),
> > > - sseu->eu_mask, eu_length))
> > > + sizeof(topo) +
> > > + slice_length + subslice_length),
> > > + sseu->eu_mask, eu_length))
> > > return -EFAULT;
> > > return total_length;
> > > }
> > > +static int query_topology_info(struct drm_i915_private *dev_priv,
> > > + struct drm_i915_query_item *query_item)
> > > +{
> > > + const struct sseu_dev_info *sseu = &to_gt(dev_priv)->info.sseu;
> > > +
> > > + if (query_item->flags != 0)
> > > + return -EINVAL;
> > > +
> > > + return fill_topology_info(sseu, query_item, sseu->subslice_mask);
> > > +}
> > > +
> > > +static int query_geometry_subslices(struct drm_i915_private *i915,
> > > + struct drm_i915_query_item *query_item)
> > > +{
> > > + const struct sseu_dev_info *sseu;
> > > + struct intel_engine_cs *engine;
> > > + struct i915_engine_class_instance classinstance;
> > > +
> > > + if (GRAPHICS_VER_FULL(i915) < IP_VER(12, 50))
> > > + return -ENODEV;
> > > +
> > > + classinstance = *((struct i915_engine_class_instance *)&query_item->flags);
> > > +
> > > + engine = intel_engine_lookup_user(i915, (u8) classinstance.engine_class,
> > > + (u8) classinstance.engine_instance);
> > > +
> > > + if (!engine)
> > > + return -EINVAL;
> > > +
> > > + sseu = &engine->gt->info.sseu;
> > > +
> > > + return fill_topology_info(sseu, query_item, sseu->geometry_subslice_mask);
> > > +}
> > > +
> > > static int
> > > query_engine_info(struct drm_i915_private *i915,
> > > struct drm_i915_query_item *query_item)
> > > @@ -485,6 +514,7 @@ static int (* const i915_query_funcs[])(struct drm_i915_private *dev_priv,
> > > query_engine_info,
> > > query_perf_config,
> > > query_memregion_info,
> > > + query_geometry_subslices,
> > > };
> > > int i915_query_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> > > diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> > > index 05c3642aaece..b539c83a4034 100644
> > > --- a/include/uapi/drm/i915_drm.h
> > > +++ b/include/uapi/drm/i915_drm.h
> > > @@ -2687,10 +2687,11 @@ struct drm_i915_perf_oa_config {
> > > struct drm_i915_query_item {
> > > /** @query_id: The id for this query */
> > > __u64 query_id;
> > > -#define DRM_I915_QUERY_TOPOLOGY_INFO 1
> > > -#define DRM_I915_QUERY_ENGINE_INFO 2
> > > -#define DRM_I915_QUERY_PERF_CONFIG 3
> > > -#define DRM_I915_QUERY_MEMORY_REGIONS 4
> > > +#define DRM_I915_QUERY_TOPOLOGY_INFO 1
> > > +#define DRM_I915_QUERY_ENGINE_INFO 2
> > > +#define DRM_I915_QUERY_PERF_CONFIG 3
> > > +#define DRM_I915_QUERY_MEMORY_REGIONS 4
> > > +#define DRM_I915_QUERY_GEOMETRY_SUBSLICES 5
> > > /* Must be kept compact -- no holes and well documented */
> > > /**
> > > @@ -2714,6 +2715,9 @@ struct drm_i915_query_item {
> > > * - DRM_I915_QUERY_PERF_CONFIG_LIST
> > > * - DRM_I915_QUERY_PERF_CONFIG_DATA_FOR_UUID
> > > * - DRM_I915_QUERY_PERF_CONFIG_FOR_UUID
> > > + *
> > > + * When query_id == DRM_I915_QUERY_GEOMETRY_SUBSLICES must have a valid
> > > + * i915_engine_class_instance struct.
> >
> > To get back to our previous discussion off-list, I find this interface
> > kind of confusing, since it's expecting an engine ID as argument, but it
> > returns the set of subslices available to the *render* engine regardless
> > of the engine class specified. I think it would make sense to rename
> > this to DRM_I915_QUERY_ENGINE_SUBSLICES or similar and have the mask
> > returned be the set of subslices actually available to the engine that
> > was specified (e.g. the compute subslice mask if a compute engine is
> > specified, or an error if the engine specified doesn't have any
> > connection to subslices). Alternatively, if this is really only meant
> > to work for the render engine, maybe the engine class should be dropped
> > from "flags", only the engine instance is necessary -- I think that
> > would prevent programming errors and would leave additional room in
> > "flags" for future expansion.
>
> I have asked a similar question and AFAIR Matt explained it was an arch
> level direction to have it like it is. Not sure of the reasoning.
>
> I wouldn't take the option of implying render and only having instance in
> flags, but returning error for engines not applicable sounds good to me. If
> there isn't a good reason not to do it.
GEOMETRY_SUBSLICES doesn't really have any meaning for non-RCS engines,
so returning an error if the query is performed against a media,
blitter, or compute engine is probably fine.
MattA also has a COMPUTE_SUBSLICES query coming as well, and I believe
the compute subslices query would be relevant to both render and compute
engines since gpgpu workloads can be submitted to either. So keeping
the engine type (and maintaining a consistent signature for the two
queries) is probably the way to go.
Matt
>
> Regards,
>
> Tvrtko
--
Matt Roper
Graphics Software Engineer
VTT-OSGC Platform Enablement
Intel Corporation
(916) 356-2795
More information about the Intel-gfx
mailing list