This fixes corruption of push constants on Xe2 due to a mismatch in
the uniform layout implemented by the compiler and assumed by the
driver. To fix it we need to align the push constant ranges computed
by the Vulkan driver to a multiple of the GRF size of the platform.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29926>
Lowering/layout is pretty much the same as direct descriptors. The
caveats is that since the descriptor buffers are not visible from the
binding tables we can't promote anything to the binding table (except
push descriptors).
The reason for this is that there is nothing that prevents an
application to use both types of descriptors and because descriptor
buffers have visible address + capture replay, we can't merge the 2
types in the same virtual address space location (limited to 4Gb max,
limited 2Gb with binding tables).
If we had the guarantee that both are not going to be used at the same
time, we could consider a 2Gb VA for descriptor buffers.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22151>
We make the compiler assume the worst possible case (it's not great
because we have to burn 32 GRFs of potential input data) and then we
push the actual value through push constants.
This enables VK_EXT_gpl usage on zink, which causes two traces to change
their results. Raven is an imperceptible change, blender has missing
original pngs but looks plausible.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22378>
With independent sets, we're not able to compute immediate values for
the index at which to read anv_push_constants::dynamic_offsets to get
the offset of a dynamic buffer. This is because the pipeline layout
may not have all the descriptor set layouts when we compile the
shader.
To solve that issue, we insert a layer of indirection.
This reworks the dynamic buffer offset storage with a 2D array in
anv_cmd_pipeline_state :
dynamic_offsets[MAX_SETS][MAX_DYN_BUFFERS]
When the pipeline or the dynamic buffer offsets are updated, we
flatten that array into the
anv_push_constants::dynamic_offsets[MAX_DYN_BUFFERS] array.
For shaders compiled with independent sets, the bottom 6 bits of
element X in anv_push_constants::desc_sets[] is used to specify the
base offsets into the anv_push_constants::dynamic_offsets[] for the
set X.
The computation in the shader is now something like :
base_dyn_buffer_set_idx = anv_push_constants::desc_sets[set_idx] & 0x3f
dyn_buffer_offset = anv_push_constants::dynamic_offsets[base_dyn_buffer_set_idx + dynamic_buffer_idx]
It was suggested by Faith to use a different push constant buffer with
dynamic_offsets prepared for each stage when using independent sets
instead, but it feels easier to understand this way. And there is some
room for optimization if you are set X and that you know all the sets in
the range [0, X], then you can still avoid the indirection. Separate
push constant allocations per stage do have a CPU cost.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15637>
This makes UBO loads in the variable pointers or bindless case work just
like SSBO loads in the sense that they use A64 messages and 64-bit
global addresses. The primary difference is that we have an
optimization in anv_nir_lower_ubo_loads which uses a (possibly
predicated) block load message when the offset is constant so we get
roughly the same performance as we would from plumbing load_ubo all the
way to the back-end.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8635>
Instead of load_global_constant_offset/bounded, we want to use the
Intel-specific block load intrinsic whenever we can. This way we get
the same wide block loads that we usually use for constant offset UBO
pulls with a binding table.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8635>
This has the advantage of giving us cheaper address calculations because
we can calculate in 32 bits first and then do a single 64x32 add. It
also lets us delete a bunch of code for dealing with descriptor
dereferences (vulkan_resource_reindex, and friends) because our bindless
SSBO pointers are now vec4s regardless of whether or not we're doing
bounds checking. This also unifies UBOs and SSBOs. The one down-side
is that, in certain variable pointers cases, it may end up burning more
memory and/or increasing register pressure. This seems like a worth-
while trade-off.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8635>
Identify if view_index is used only for position calculation, and use
Primitive Replication to implement Multiview in Gen12. This feature
allows storing per-view position information in a single execution of
the shader, treating position as an array.
The shader is transformed by adding a for-loop around it, that have an
iteration per active view (in the view_mask). Stores to the position
now store into the position array for the current index in the loop,
and load_view_index() will return the view index corresponding to the
current index in the loop.
The feature is controlled by setting the environment variable
ANV_PRIMITIVE_REPLICATION_MAX_VIEWS, which defaults to 2 if unset.
For pipelines with view counts larger than that, the regular
instancing will be used instead of Primitive Replication. To disable
it completely set the variable to 0.
v2: Don't assume position is set in vertex shader; remove only stores
for position; don't apply optimizations since other passes will
do; clone shader body without extract/reinsert; don't use
last_block (potentially stale). (Jason)
Fix view_index immediate to contain the view index, not its order.
Check for maximum number of views supported.
Add guard for gen12.
v3: Clone the entire shader function and change it before reinsert;
disable optimization when shader has memory writes. (Jason)
Use a single environment variable with _DEBUG on the name.
v4: Change to use new nir_deref_instr.
When removing stores, look for mode nir_var_shader_out instead
of the walking the list of outputs.
Ensure unused derefs are removed in the non-position part of the
shader.
Remove dead control flow when identifying if can use or not
primitive replication.
v5: Consider all the active shaders (including fragment) when deciding
that Primitive Replication can be used.
Change environment variable to ANV_PRIMITIVE_REPLICATION.
Squash the emission of 3DSTATE_PRIMITIVE_REPLICATION into this patch.
Disable Prim Rep in blorp_exec_3d.
v6: Use a loop around the shader, instead of manually unrolling, since
the regular unroll pass will kick in.
Document that we don't expect to see copy_deref or load_deref
involving the position variable.
Recover use_primitive_replication value when loading pipeline from
the cache.
Set VARYING_SLOT_LAYER to 0 in the shader. Earlier versions were
relying on ForceZeroRTAIndexEnable but that might not be
sufficient.
Disable Prim Rep in cmd_buffer_so_memcpy.
v7: Don't use Primitive Replication if position is not set, fallback
to instancing; change environment variable to be
ANV_PRIMITVE_REPLICATION_MAX_VIEWS and default it to 2 based on
experiments.
Reviewed-by: Rafael Antognolli <rafael.antognolli@intel.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/2313>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/2313>
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>
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>
Instead of taking a whole pipeline (which could be anything!), just take
a physical device and robust_buffer_access boolean. This makes it
easier to verify that only the things in the hash actually affect
pipeline compilation.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
This is part of the device groups extension/feature but it's a decent
chunk of work in its own right so it's worth breaking into its own
patch. The mechanism we use is fairly straightforward: we just push the
base work group id into the shader and add it to the work group id we
get from dispatch.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
The Vulkan spec states that VkPipelineLayout objects must not be
destroyed while any command buffer that uses them is in the recording
state, but it permits them to be destroyed otherwise. This means that
applications are allowed to free pipeline layouts after command recording
is finished even if there are pipeline objects that still exist and were
created with these layouts.
There are two solutions to this, one is to use reference counting on
pipeline layout objects. The other is to avoid holding references to
pipeline layouts where they are not really needed.
This patch takes a step towards the second option by making the
pipeline shader compile code take pipeline layout from the
VkGraphicsPipelineCreateInfo provided rather than the pipeline
object.
A follow-up patch will remove any remaining uses of the layout field
so we can remove it from the pipeline object and avoid the need
for reference counting.
v2: Use ANV_FROM_HANDLE, remove unnecessary braces (Jason)
Suggested-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This pass implements all the implicit conversions required by the
VK_KHR_sampler_ycbcr_conversion specification.
It also inserts plane sources onto sampling instructions that we then
let the pipeline layout pass deal with, when mapping things correctly
to descriptors.
v2: Add new file to meson build (Lionel)
Use nir_frcp() rather than (1.0f / x) (Jason)
Reuse nir_tex_instr_dest_size() rather than handwritten one (Jason)
Return progress (Jason)
Account for array of samplers (Jason)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
v2 (Jason Ekstrand):
- Take a view_mask rather than a whole subpass
- Build the view mask into the VS shader key
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
All of the code that did something special based on vec4 vs. scalar is
bogus. In the backend, everything is now in units of bytes and the vec4
backend can handle full std140 packing so we don't need to do anything
special anymore.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94998
