The Vulkan spec says:
"pipelineBindPoint is a VkPipelineBindPoint indicating whether the
descriptors will be used by graphics pipelines or compute pipelines.
There is a separate set of bind points for each of graphics and
compute, so binding one does not disturb the other."
Up until now, we've been ignoring the pipeline bind point and had just
one bind point for everything. This commit separates things out into
separate bind points.
Tested-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102897
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
This lets us unify some code between push descriptors and regular
descriptors. It doesn't do much for us yet but it will.
Tested-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
It's now a function which returns the push descriptor set. Since we set
the error on the command buffer, returning the error is a little
redundant. Returning the descriptor set (or NULL on error) is more
convenient.
Tested-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
Initially, these just contain the pipeline in a base struct.
Tested-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
This splits anv_cmd_state_reset into separate init and finish functions.
This lets us share init code with cmd_buffer_create. This potentially
fixes subtle bugs where we may have missed some bit of state that needs
to get initialized on command buffer creation.
Tested-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
Meta has been gone for a long time.
Tested-by: Józef Kucia <joseph.kucia@gmail.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
We're going to want subgroup ID for SPIR-V subgroups eventually anyway.
We really only want to push one and calculate the other from it. It
makes a bit more sense to push the subgroup ID because it's simpler to
calculate and because it's a real API thing. The only advantage to
pushing the base thread ID is to avoid a single SHL in the shader.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
This is a lot more natural than special casing it all over the place.
We still have to do a bit of special-casing in assign_constant_locations
but it's not special-cased quite as bad as it was before.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This moves us away to the array of pointers model and onto a model where
each param is represented by a generic uint32_t handle. We reserve 2^16
of these handles for builtins that get generated by somewhere inside the
compiler and have well-defined meanings. Generic params have handles
whose meanings are defined by the driver.
The primary downside to this new approach is that it moves a little bit
of the work that we would normally do at compile time to draw time. On
my laptop this hurts OglBatch6 by no more than 1% and doesn't seem to
have any measurable affect on OglBatch7. So, while this may come back
to bite us, it doesn't look too bad.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This ensures that everything gets cleaned up properly. In particular,
it fixes a memory leak where we were leaking the push constants
structs.
Valgrind stats on
dEQP-VK.pipeline.push_constant.graphics_pipeline.range_size_128 :
Before:
HEAP SUMMARY:
in use at exit: 2,467,513 bytes in 1,305 blocks
total heap usage: 697,853 allocs, 696,530 frees, 138,466,600 bytes allocated
LEAK SUMMARY:
definitely lost: 1,068 bytes in 11 blocks
indirectly lost: 24,669 bytes in 412 blocks
possibly lost: 0 bytes in 0 blocks
still reachable: 2,441,776 bytes in 882 blocks
suppressed: 0 bytes in 0 blocks
After:
HEAP SUMMARY:
in use at exit: 2,467,381 bytes in 1,304 blocks
total heap usage: 697,853 allocs, 696,531 frees, 138,466,600 bytes allocated
LEAK SUMMARY:
definitely lost: 936 bytes in 10 blocks
indirectly lost: 24,669 bytes in 412 blocks
possibly lost: 0 bytes in 0 blocks
still reachable: 2,441,776 bytes in 882 blocks
suppressed: 0 bytes in 0 blocks
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "17.2 17.1" <mesa-stable@lists.freedesktop.org>
When writing to set > 0, we were just wrongly writing to set 0. This
commit fixes this by lazily allocating each set as we write to them.
We didn't go for having them directly into the command buffer as this
would require an additional ~45Kb per command buffer.
v2: Allocate push descriptors from system memory rather than in BO
streams. (Lionel)
Cc: "17.2 17.1" <mesa-stable@lists.freedesktop.org>
Fixes: 9f60ed98e5 ("anv: add VK_KHR_push_descriptor support")
Reported-by: Daniel Ribeiro Maciel <daniel.maciel@gmail.com>
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The vkCmd*() functions do not report errors, instead, any errors should be
reported by the time we call vkEndCommandBuffer(). This means that we
need to make the driver robust against incosistent and/or imcomplete
command buffer states through the command recording process, particularly,
avoid crashes due to access to memory that we failed to allocate previously.
The strategy used to do this is to track the first error ocurred while
recording a command buffer in the batch associated with it. We use the
batch to track this information because the command buffer may not be
visible to all parts of the driver that can produce errors we need to be
aware of (such as allocation failures during batch emissions).
Later patches will use this error information to guard parts of the driver
that may not be safe to execute.
v2: Move the field from the command buffer to the batch so we can track
errors from batch emissions (Jason)
v3: Registering errors in the command buffer's batch during
anv_create_cmd_buffer() is unnecessary, since the command buffer
is freed at the end of the function in that case (Topi)
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
We have a performance problem with dynamic buffer descriptors. Because
we are currently implementing them by pushing an offset into the shader
and adding that offset onto the already existing offset for the UBO/SSBO
operation, all UBO/SSBO operations on dynamic descriptors are indirect.
The back-end compiler implements indirect pull constant loads using what
basically amounts to a texelFetch instruction. For pull constant loads
with constant offsets, however, we use an oword block read message which
goes through the constant cache and reads a whole cache line at a time.
Because of these two things, direct pull constant loads are much faster
than indirect pull constant loads. Because all loads from dynamically
bound buffers are indirect, the user takes a substantial performance
penalty when using this "performance" feature.
There are two potential solutions I have seen for this problem. The
alternate solution is to continue pushing offsets into the shader but
wire things up in the back-end compiler so that we use the oword block
read messages anyway. The only reason we can do this because we know a
priori that the dynamic offsets are uniform and 16-byte aligned.
Unfortunately, thanks to the 16-byte alignment requirement of the oword
messages, we can't do some general "if the indirect offset is uniform,
use an oword message" sort of thing.
This solution, however, is recommended for a few of reasons:
1. Surface states are relatively cheap. We've been using on-the-fly
surface state setup for some time in GL and it works well. Also,
dynamic offsets with on-the-fly surface state should still be
cheaper than allocating new descriptor sets every time you want to
change a buffer offset which is really the only requirement of the
dynamic offsets feature.
2. This requires substantially less compiler plumbing. Not only can we
delete the entire apply_dynamic_offsets pass but we can also avoid
having to add architecture for passing dynamic offsets to the back-
end compiler in such a way that it can continue using oword messages.
3. We get robust buffer access range-checking for free. Because the
offset and range are baked into the surface state, we no longer need
to pass ranges around and do bounds-checking in the shader.
4. Once we finally get UBO pushing implemented, it will be much easier
to handle pushing chunks of dynamic descriptors if the compiler
remains blissfully unaware of dynamic descriptors.
This commit improves performance of The Talos Principle on ULTRA
settings by around 50% and brings it nicely into line with OpenGL
performance.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
We will be using the image layout. Store the full struct directly from
the user.
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This allows the helper to check for llc instead of having to do it
manually at all the call sites.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
It's a bit shorter and easier to work with. Also, we're about to add a
helper called clflush which does the clflush but without any memory
fencing.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This helps Dota 2 on Broadwell by 8-9%. I also hacked up the driver and
used the Sascha "shadowmapping" demo to get some results. Setting
uses_kill to true dropped the framerate on the demo by 25-30%. Enabling
the PMA fix brought it back up to around 90% of the original framerate.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
The spec section 5.2 says:
"vkAllocateCommandBuffers can be used to create multiple command
buffers. If the creation of any of those command buffers fails, the
implementation must destroy all successfully created command buffer
objects from this command, set all entries of the pCommandBuffers
array to VK_NULL_HANDLE and return the error."
Fixes:
dEQP-VK.api.object_management.alloc_callback_fail_multiple.command_buffer_primary
dEQP-VK.api.object_management.alloc_callback_fail_multiple.command_buffer_secondary
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "13.0 17.0" <mesa-stable@lists.freedesktop.org>
Our command buffers already efficiently use a global pool so trimming
doesn't really need to do anything.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This fixes a bunch of new CTS tests which look for exactly this. Even in
the cases where we just call vk_free to free a CPU data structure, we still
handle NULL explicitly. This way we're less likely to forget to handle
NULL later should we actually do something less trivial.
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Dave Airlie <airlied@redhat.com>
This reverts most of commit 52904ba85c.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
This commit moves the allocation and filling out of surface state from
CreateImageView time to BeginRenderPass time. Instead of allocating the
render target surface state as part of the image view, we allocate it in
the command buffer state at the same time that we set up clears. For
secondary command buffers, we allocate memory for the surface states in
BeginCommandBuffer but don't fill them out; instead, we use our new
SOL-based memcpy function to copy the surface states from the primary
command buffer. This allows us to handle secondary command buffers without
the user specifying the framebuffer ahead-of-time.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
While we can simply calculate offsets to get to things such as the
prog_data and the key, it's much more user-friendly if there are just
pointers. Also, it's a bit more fool-proof.
While we're at it, we rework the pipeline cache API to use the
brw_stage_prog_data type directly.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98012
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
This moves all the alloc/free in anv to the generic helpers.
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
All code that would have once called this can now call the gen-specific
version. The switching version is no longer needed.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
It really should have gone here all along. We were trying a bit too hard
to make it gen-agnostic just because it didn't have any #if's.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
vkBeginCommandBuffer and vkCmdExecuteCommands both call into the
gen-specific emit_state_base_address function and vkEndCommandBuffer
belongs with begin.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
This has two primary advantages. First, it means that the batch_chain code
knows less about the actual command buffer contents which is good because
improves separation. Second, it means that it only gets re-emitted once
after all of the secondaries instead of once after each secondary which is
just wasteful. It also has the advantage of cleaning the code up a bit.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Initially, we had intended set_subpass to be an interesting function that
did whatever (presumably a lot) setup we needed for a subpass. In reality,
it just sets a pointer and a dirty bit and then emits depth and stencil
state. When we call BeginCommandBuffer on a secondary, there's no point in
setting depth and stencil state since it will already be set by the
primary. Instead, the only thing we need to do at the start of a secondary
is set the subpass pointer and the dirty bit.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
The original pipeline cache the Kristian wrote was based on a now-false
premise that the shaders can be stored in the pipeline cache. The Vulkan
1.0 spec explicitly states that the pipeline cache object is transiant and
you are allowed to delete it after using it to create a pipeline with no
ill effects. As nice as Kristian's design was, it doesn't jive with the
expectation provided by the Vulkan spec.
The new pipeline cache uses reference-counted anv_shader_bin objects that
are backed by a large state pool. The cache itself is just a hash table
mapping keys hashes to anv_shader_bin objects. This has the added
advantage of removing one more hand-rolled hash table from mesa.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97476
Acked-by: Kristian Høgsberg Kristensen <krh@bitplanet.net>
All of these worked before because they were depending on prog_data to be
null. Soon, we won't be able to depend on a nice prog_data pointer and
it's nice to be more explicit anyway.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
A ternary is clearer because the range member is assigned one of two values
dependant on one condition.
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This became unused due to commit 612e35b2c6 .
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This way the the bind map (which we're caching) is mostly independent of
the pipeline layout. The only coupling remaining is that we pull the array
size of a binding out of the layout. However, that size is also specified
in the shader and should always match so it's not really coupled. This
rendering issues in Dota 2.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
The old method pushed data for each channels uvec3 data of
gl_LocalInvocationID.
The new method pushes 1 dword of data that is a 'thread local ID'
value. Based on that value, we can generate gl_LocalInvocationIndex
and gl_LocalInvocationID with some calculations.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The cross thread constant support appears on Haswell. It allows us to
upload a set of uniform data for all threads without duplicating it
per thread.
We also support per-thread data which allows us to store a per-thread
ID in one of the uniforms that can be used to calculate the
gl_LocalInvocationIndex and gl_LocalInvocationID variables.
v4:
* Support the old local ID push constant layout as well (Jason)
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This is in contrast to emitting it directly in vkCmdPipelineBarrier. This
has a couple of advantages. First, it means that no matter how many
vkCmdPipelineBarrier calls the application strings together it gets one or
two PIPE_CONTROLs. Second, it allow us to better track when we need to do
stalls because we can flag when a flush has happened and we need a stall.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Instead of blasting it out as part of the pipeline, we put it in the
command buffer and only blast it out when it's really needed. Since the
PUSH_CONSTANT_ALLOC commands aren't pipelined, they immediately cause a
stall which we would like to avoid.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Avoid excessive state emission. Relevant state for an action command
will get set by the user:
From Chapter 5. Command Buffers,
When a command buffer begins recording, all state in that command
buffer is undefined.
[...]
Whenever the state of a command buffer is undefined, the application
must set all relevant state on the command buffer before any state
dependent commands such as draws and dispatches are recorded, otherwise
the behavior of executing that command buffer is undefined.
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Kristian Høgsberg Kristensen <kristian.h.kristensen@intel.com>
Previously, we would always emit all of the render targets in the subpass.
This commit changes it so that we compact render targets just like we do
with other resources. Render targets are represented in the surface map by
using a descriptor set index of UINT16_MAX.
Both logic and indentation suggests that the ; were not intended here.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Jason Ekstrand
Lionel Landwerlin
Iago Toral Quiroga
Anuj Phogat
Craig Stout
Nanley Chery
Dave Airlie
Kenneth Graunke
Jordan Justen
Ronie Salgado
Kristian Høgsberg Kristensen
Thomas Hindoe Paaboel Andersen