Reworks:
* Don't pack params, just memcpy param struct (s-b Jason)
* Old subject: "intel/blorp: Emit compute program if
params.cs_prog_data is set"
* Various cleanups of push-const size/alignment (s-b Jason)
* Fix subslice count by moving to devinfo (s-b Ken)
* Simplify cw.InterfaceDescriptor code (s-b Ken)
* Drop some comments from i965 (s-b Ken)
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11564>
Found this nugget while looking at the ACO driver. It seems sensible to
avoid SLM fences if there is no SLM. This also makes the check depend
on SLM usage rather than just shader stage which will be useful if we
ever implement task/mesh because task shaders also have SLM.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13092>
Start off making everything look like LSC where we have three types of
fences: TGM, UGM, and SLM. Then, emit the actual code in a generation-
aware way. There are three HW generation cases we care about:
XeHP+ (LSC), ICL-TGL, and IVB-SKL. Even though it looks like there's a
lot to deduplicate, it only increases the number of ubld.emit() calls
from 5 to 7 and entirely gets rid of the SFID juggling and other
weirdness we've introduced along the way to make those cases "general".
While we're here, also clean up the code for stalling after fences and
clearly document every case where we insert a stall.
There are only three known functional changes from this commit:
1. We now avoid the render cache fence on IVB if we don't need image
barriers.
2. On ICL+, we no longer unconditionally stall on barriers. We still
stall if we have more than one to help tie them together but
independent barriers are independent. Barrier instructions will
still operate in write-commit mode and still be scheduling barriers
but won't necessarily stall.
3. We now assert-fail for URB fences on LSC platforms. We'll be adding
in the new URB fence message for those platforms in a follow-on
commit.
It is a big enough refactor, however, that other minor changes may be
present.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13092>
The populate_base_prog_key will set VARYING depending if the pipeline
flag is used. Later, when full subgroups flag is set, it will flip to
UNIFORM -- which for compute shaders is effectively the same, so don't
bother setting it again.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12946>
The lowering passes will soon be moved to another function, so there
won't be any choice.
As a side benefit, this allows eliminating the uses_atomic_load_store
**pointer** parameter from brw_nir_lower_storage_image. For some reason
crocus was passing false instead of NULL.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12858>
This consolidates several duplicated pieces of code into devinfo.
max_scratch_ids is an array that provides the max number of threads
for the rendering and compute stages.
This fixes some exceptions missed by crocus for scratch ids on haswell
and cherryview.
It also fills out devinfo->max_scratch_ids properly for stages VS
through CS on Gfx12.5. But, functionally this should not make a
difference as Gfx12.5 already uses COMPUTE for all stages.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12799>
By default, Mesa's X11 Vulkan WSI will wait for buffers to be ready
before submitting them to Xwayland when the swapchain is created
with the IMMEDIATE mode.
This is undesirable when the Wayland compositor already monitors
fences. A Wayland compositor may want to know the delay between
the buffer submition and the end of the GPU work, this is impossible
to measure if the WSI waits for the buffer to be ready before
submission.
Since most compositors don't monitor fences, let's introduce a driconf
option for this for now. We can reconsider once more compositors
have better support for fences.
Signed-off-by: Simon Ser <contact@emersion.fr>
Acked-by: Michel Dänzer <mdaenzer@redhat.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11290>
For now, only mark the 4x8BitPacked variants as accelerated.
Applications are unlikely to use the "add with saturate" opcodes from
VK_INTEL_shader_integer_functions2, so, technically, all of the
AccumulatingSaturating variants "[provide] a performance advantage over
user-provided code composed from elementary instructions..." on all
Intel platforms. If we encounter an application that cares, we can do
things differently then. Ditto for the non-packed 8Bit, 4-element
vector variants.
v2: Don't memset props as this also zeros sType and pNext. Noticed by
Georg Lehmann in !12617.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12624>
Jordan Justen