This was the competing idea long ago, but it was only with the rewrite
of the idr as an radixtree and using the radixtree directly ourselves,
along with the realisation that we can store the vma directly in the
radixtree and only need a list for the reverse mapping, that made the
patch performant enough to displace using a hashtable. Though the vma ht
is fast and doesn't require any extra allocation (as we can embed the node
inside the vma), it does require a thread for resizing and serialization
and will have the occasional slow lookup. That is hairy enough to
investigate alternatives and favour them if equivalent in peak performance.
One advantage of allocating an indirection entry is that we can support a
single shared bo between many clients, something that was done on a
first-come first-serve basis for shared GGTT vma previously. To offset
the extra allocations, we create yet another kmem_cache for them.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170816085210.4199-5-chris@chris-wilson.co.uk
When userspace is doing most of the work, avoiding relocs (using
NO_RELOC) and opting out of implicit synchronisation (using ASYNC), we
still spend a lot of time processing the arrays in execbuf, even though
we now should have nothing to do most of the time. One issue that
becomes readily apparent in profiling anv is that iterating over the
large execobj[] is unfriendly to the loop prefetchers of the CPU and it
much prefers iterating over a pair of arrays rather than one big array.
v2: Clear vma[] on construction to handle errors during vma lookup
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170816085210.4199-3-chris@chris-wilson.co.uk
During execbuf, a mandatory step is that we add this request (this
fence) to each object's reservation_object. Inside execbuf, we track the
vma, and to add the fence to the reservation_object then means having to
first chase the obj, incurring another cache miss. We can reduce the
number of cache misses by stashing a pointer to the reservation_object
in the vma itself.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170616140525.6394-1-chris@chris-wilson.co.uk
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
The advent of full-ppgtt lead to an extra indirection between the object
and its binding. That extra indirection has a noticeable impact on how
fast we can convert from the user handles to our internal vma for
execbuffer. In order to bypass the extra indirection, we use a
resizable hashtable to jump from the object to the per-ctx vma.
rhashtable was considered but we don't need the online resizing feature
and the extra complexity proved to undermine its usefulness. Instead, we
simply reallocate the hastable on demand in a background task and
serialize it before iterating.
In non-full-ppgtt modes, multiple files and multiple contexts can share
the same vma. This leads to having multiple possible handle->vma links,
so we only use the first to establish the fast path. The majority of
buffers are not shared and so we should still be able to realise
speedups with multiple clients.
v2: Prettier names, more magic.
v3: Many style tweaks, most notably hiding the misuse of execobj[].rsvd2
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
With the introduce of i915_vma_instance() for obtaining the VMA
singleton for a (obj, vm, view) tuple, we can remove the
i915_vma_create() in favour of a single entry point. We do incur a
lookup onto an empty tree, but the i915_vma_create() were being called
infrequently and during initialisation, so the small overhead is
negligible.
v2: Drop the i915_ prefix from the now static vma_create() function
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170116152131.18089-4-chris@chris-wilson.co.uk
Reading the ggtt_views is much more pleasant without the extra
characters from specifying the union (i.e. ggtt_view.partial rather than
ggtt_view.params.partial). To make this work inside i915_vma_compare()
with only a single memcmp requires us to ensure that there are no
uninitialised bytes within each branch of the union (we make sure the
structs are packed) and we need to store the size of each branch.
v4: Rewrite changelog and add comments explaining the assert.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20170114002827.31315-5-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
In preparation for the next patch to convert to using an anonymous union
and leaving the excess bytes in the union uninitialised, we first need
to make sure we do not compare using those uninitialised bytes. We also
want to preserve the compactness of the code, avoiding a second call to
memcmp or introducing a switch, so we take advantage of using the type
as an encoded size (as well as a unique identifier for each type of view).
v2: Add the rationale for why we encode size into ggtt_view.type as a
comment before the memcmp()
v3: Use a switch to also assert that no two i915_ggtt_view_type have the same
value.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170114002827.31315-3-chris@chris-wilson.co.uk
The fence size/alignment is a combination of the vma size plus object
tiling parameters. Those parameters are rarely changed, making the fence
size/alignemnt roughly constant for the lifetime of the VMA. We can
simplify subsequent calculations by precalculating the size/alignment
required for GGTT vma taking fencing into account (with an update if we
do change the tiling or stride).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170109161613.11881-4-chris@chris-wilson.co.uk
