It's a very odd case to hit in the real world. However, there are some
CTS tests which switch back and forth between dispatch and clear without
changing the pipeline.
Fixes: bc612536eb "anv: Emit a dummy MEDIA_VFE_STATE before switching..."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
When we moved from allocating BOs directly to using the BO cache, we
lost the EXEC_OBJECT_CAPTURE flag on all our state buffers.
Fixes: 3119b96bdf "anv: Allocate block pool BOs from the cache"
Fixes: ee77938733 "anv: Allocate batch and fence buffers from..."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Instead of doing a dummy submit on the command buffer for the fence or a
dummy semaphore and trusting in implicit sync, this commit moves us to
take advantage of implicit sync and just use the WSI image BO as the
fence. Both semaphores and fences require a tiny bit of extra plumbing
to do this but the result is that we can get rid of a bunch of the extra
synchronization we're doing today.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
In 83b943cc2f, we started making all VkDeviceMemory BOs resident all
the time. One unfortunate side-effect of this is that every
vkQueueSubmit sets EXEC_OBJECT_WRITE on every WSI memory object which
means that X server or Wayland compositor, instead of waiting on the
last vkQueueSubmit to actually write the buffer, now waits on the last
vkQueueSubmit to from that driver instance relative to whenever the
compositor's GL driver instance calls execbuf. This potentially leads
to a lot of extra synchronization that we didn't intend to have.
Instead, this commit makes it so that we leave WSI memory objects with
EXEC_OBJECT_ASYNC most of the time and only unset EXEC_OBJECT_ASYNC and
set EXEC_OBJECT_WRITE in the dummy execbuf that we do as part of
vkQueuePresent. This should hopefully result in tighter integration
with the compositor, lower latency, and better performance.
Testing with DOOM 2016, this seems to reduce latency by at least a frame
if not two and makes the game much more responsive. Testing was,
however, subjective, so we don't have any hard data on that.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Otherwise, we're trusting in the execbuf_add_bo which sets
EXEC_OBJECT_WRITE to to always be the first one that gets called. This
is likely true for fences but it seems somewhat fragile.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
The primary difference between the KHR and EXT versions of the extension
is that the KHR provides the address at AllocateMemory time for replay
so we can replay it safely without moving to a sparse address model.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This function has a lot of possible extensions and some of them we can
easily handle on-the-fly so it's easier to just have a loop than to find
each structure manually.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
When a BO is flagged as having a client visible address, we put it in
its own heap. We also support the client explicitly specifying an
address in said heap. If an address collision happens, we return false
from anv_vma_alloc which turns into a VK_ERROR_OUT_OF_DEVICE_MEMORY.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
We already have a mechanism for specifying that we want a fixed address
provided by the driver internals. We're about to let the client start
specifying addresses in some very special scenarios as well so we want
to pass this through to the allocation function.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Our VMA allocations are really independent from the memory heaps we
expose via the API. The only thing that really matters is the GTT size
so we can make the high heap the right size.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
util_vma_heap_alloc will already return 0 if it doesn't have enough
space. The only thing the vma_*_available tracking was doing was
preventing us from allocating too much on any given heap. Now that
we're tracking that in the heap itself, we can drop these.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
We're already tracking the amount of memory used in each heap. This
commit just makes us start rejecting memory allocations if the heap
would grow too large.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
I think the reason why we only do this for primaries is that we didn't
expect to have blorp calls in secondaries. However, you are allowed to
have a full render pass in a secondary command buffer so resolves and
clears can end up in there. We should just always invalidate.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This separates "has" from "use" which will make the next commit a bit
cleaner.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
In ee77938733, we started using the BO cache for anv_bo_pool and
stopped using the bo_flags parameter. However, we never dropped it from
the struct or the init function.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
See also 246261f0ad
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
CID: 1455892
Fixes: 246261f0ad ("anv: prepare the driver for delayed submissions")
Use this new instruction introduced in Gen12. The instruction itself is
smaller, and it also allows us to emit a single instruction to all
stages that have the same push constant buffers (e.g. when they don't
have constant buffers).
There's one restriction to use this instruction, though: the length
field is only 5 bits long, so we need to check whether we can use it,
and fallback to the old 3DSTATE_CONSTANT_XS if that field is >= 32.
v2:
- Rebased on top of the lasted changes from Jason.
- Added review suggestions by Caio.
- Removed struct push_bos and merged some code into
anv_nir_compute_push_layout().
v3:
- Remove code churn due to gen8+ workaround in
anv_nir_compute_push_layout(). This code has been removed in an earlier
commit, and implemented in cmd_buffer_emit_push_constant().
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Add a helper function to get the push range address. Once we have a
separate function for emitting gen12 push constants, we can use this
helper and avoid duplicating code.
v3: Do not add range->start to the address in gen7 (Caio).
v4: Do not drop range->start from gen7 (Caio, Jason).
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Store push_ranges in ascending order, and only "shift" them to the end
of the array during state packet emission.
We don't need this workaround with the new 3DSTATE_CONSTANT_ALL packet.
So instead of applying the workaround here just for GEN < 12 (which
requires and extra loop through all the ranges to figure out if we
should shift them or not), we simply move the whole logic to the state
emission code. At that point, in a later commit, we are already looping
through all of the ranges anyway to check which packet we will be using,
so we might as well implement the workaround there, where it is going to
be used.
v3: Move gen8+ workaround to the state emission code (Caio).
v4: Add explanation of why we moved the workaroudn (Caio).
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
gl_Viewport is also in the VUE header so we need to whack the read
offset to 0 and emit a default (no overrides) SBE_SWIZ entry in that
case as well.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
In the case of promoted extensions we can end up with an entrypoint that
we support being an alias of an entrypoint we do not support. For
instance, if an extension gets promoted from EXT to KHR, the EXT entry-
points may be aliases of the KHR ones. We want to leave everything as
EXT until we get around to advertising the KHR so that we don't break
things when we update the XML and headers.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
If both are zero (the common case), we can emit a null vertex buffer
rather than emitting a vertex buffer with zeros in it. The packing of
the VERTEX_BUFFER_STATE is faster because no relocation is emitted and
we can avoid creating the vertex buffer which means one less
anv_state_stream_alloc.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This is a bit more natural because we're already getting an anv_state
most places in the pipeline. The important part here, however, is that
we're no longer calling anv_block_pool_map on every alloc_binding_table
call. While it's probably pretty cheap, it is potentially a linear walk
over the list of BOs and it was showing up in profiles.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Instead of blindly dirtying descriptors and push constants the moment we
see a pipeline change, check to see if it actually changes the bind
layout or push constant layout. This doubles the runtime performance of
one CPU-limited example running with the Dawn WebGPU implementation when
running on my laptop.
NOTE: This effectively reverts beca63c6c0. While it was a nice
optimization, it was based on prog_data and we can't do that anymore
once we start allowing the same binding table to be used with multiple
different pipelines.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Instead of dirtying all graphics or all compute based on binding point,
we're now much more careful. We first check to see if the actual
descriptor set changed and then only dirty the stages used by that
descriptor set. For dynamic offsets, we keep a bitfield per-stage of
which offsets are actually used in that stage and we only dirty push
constants and descriptors if that stage has dynamic offsets AND those
offsets actually change.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
It theoretically could be more efficient but the real point here is that
it's no longer really a matter of dealing with special cases and then
the "real" thing. The way we're handling binding tables, it's more of a
multi-step process and a switch is more natural.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This substantially reworks both the state setup side of push constant
handling and the pipeline compile side. The fundamental change here is
that we're no longer respecting the prog_data::param array and instead
are just instructing the back-end compiler to leave the array alone.
This makes the state setup side substantially simpler because we can now
just memcpy the whole block of push constants and don't have to
upload one DWORD at a time.
This also means that we can compute the full push constant layout
up-front and just trust the back-end compiler to not mess with it.
Maybe one day we'll decide that the back-end compiler can do useful
things there again but for now, this is functionally no different from
what we had before this commit and makes the NIR handling cleaner.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This moves the compute stuff into a anv_push_constants::cs sub-struct.
It also moves dynamic offsets into the push constants. This means we
have to duplicate the data per-stage but that doesn't seem like the end
of the world and one day we may wish to make dynamic offsets per-stage
anyway.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
It turns off that emitting push constants is one of the hottest paths in
the driver and ANY work we do there costs us. By pre-computing things a
bit ahead of time, we shave 5% off the runtime of a CPU-limited example
running with the Dawn WebGPU implementation.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
The bounds checking is actually less safe than just pushing the data.
If the bounds checking actually ever kicks in and it's not on the last
UBO push range, then the shrinking will cause all subsequent ranges to
be pushed to the wrong place in the GRF. One of the behaviors we
definitely don't want is for OOB UBO access to result in completely
unrelated UBOs returning garbage values. It's safer to just push the
UBOs as-requested. If we're really concerned about robustness, we can
emit shader code to do bounds checking which should be stupid cheap (a
CMP followed by SEL).
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
As of 2d78e55a8c, nir_intrinsic_load_constant with a constant offset
is constant-folded so we should never end up with any that trigger
brw_nir_analyze_ubo_ranges.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This lets us stop tracking the pipeline layout. It also means less
indirection on a very hot path. As an extra bonus, we can make some of
our data structures smaller. No measurable CPU overhead improvement.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
In the early days of the driver we allowed layout to be VK_NULL_HANDLE
and used that for some internal pipelines when we wanted to be lazy.
Vulkan doesn't actually allow NULL layouts, however, so there's no
reason to have this check.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This was causing uninitialized value to end up propagated to the
3DSTATE_DEPTH_BOUNDS packet, leading to asserts on packet
building due to the value being greater than 1.
Fixes: 939ddccb7a ("anv: Add support for depth bounds testing.")
Reviewed-by: Plamena Manolova <plamena.manolova@intel.com>
