Scalarizing preambles in NIR isn't really necessary, we can do it more
efficiently in the backend. This makes the final NIR a lot less annoying to
read; the backend IR was already nice to read thanks to all the scalarized moves
being copypropped. Plus, this is a lot simpler.
No shader-db changes.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21122>
Move the decision of "can I copyprop this uniform?" from copyprop to a
standalone lowering pass. This is more straightforward and will enable the next
patch. This has the side effect of sinking load_preamble instructions, for a
nice reduction in register pressure. Instruction count increase is from
rematerializing some moves, which should be more than balanced out by the
reduced register pressure.
total instructions in shared programs: 1523285 -> 1523317 (<.01%)
instructions in affected programs: 1148 -> 1180 (2.79%)
helped: 0
HURT: 13
HURT stats (abs) min: 1.0 max: 4.0 x̄: 2.46 x̃: 2
HURT stats (rel) min: 0.69% max: 7.69% x̄: 3.65% x̃: 2.61%
95% mean confidence interval for instructions value: 1.78 3.14
95% mean confidence interval for instructions %-change: 2.16% 5.15%
Instructions are HURT.
total bytes in shared programs: 10444532 -> 10444724 (<.01%)
bytes in affected programs: 7386 -> 7578 (2.60%)
helped: 0
HURT: 13
HURT stats (abs) min: 6.0 max: 24.0 x̄: 14.77 x̃: 12
HURT stats (rel) min: 0.63% max: 7.14% x̄: 3.40% x̃: 2.48%
95% mean confidence interval for bytes value: 10.68 18.85
95% mean confidence interval for bytes %-change: 2.02% 4.78%
Bytes are HURT.
total halfregs in shared programs: 419444 -> 416434 (-0.72%)
halfregs in affected programs: 27080 -> 24070 (-11.12%)
helped: 634
HURT: 0
helped stats (abs) min: 1.0 max: 30.0 x̄: 4.75 x̃: 2
helped stats (rel) min: 2.90% max: 54.55% x̄: 13.13% x̃: 8.51%
95% mean confidence interval for halfregs value: -5.08 -4.41
95% mean confidence interval for halfregs %-change: -14.03% -12.23%
Halfregs are helped.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21122>
Instead of printing the raw location number, which is pretty hard to interpret,
let's print the name of the location. Example output:
vec4 16 ssa_2 = intrinsic load_interpolated_input (ssa_0, ssa_1) (base=0,
component=0, dest_type=float16 /*144*/, io location=VARYING_SLOT_VAR0 slots=1
mediump /*8388768*/)
One of the "regressions" from moving to purely lowered I/O with all variables
removed is a lack of debuggability, since otherwise these location strings don't
show up anywhere in the printed shader! By contrast this should make the lowered
I/O nice to read like the early I/O.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Acked-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21091>
create_shader_state passes ownership of the NIR to the driver, so we need to
free it when we destroy the shader CSO later. Use ralloc to manage this in a
uniform way between graphics and compute. Strategy from Panfrost.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21062>
This adds the basic scaffolding for compute kernels. There's a bit of churn to
make sure we don't need to hang onto the kernel NIR, since it's never used for
anything else except looking up the shader stage.
The compute kernels aren't actually wired up here, but they do get compiled.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21062>
Implement custom border colours, as required by OpenGL's CLAMP_TO_BORDER and
Vulkan with customBorderColor. This uses an extended sampler descriptor, which
has space for the custom border values. The trouble is that the border must be
packed into an internal interchange format that depends on the original format
in a complex way. That said, we're not solving NP-complete problems here, and it
passes the tests (dEQP-GLES31.functional.texture.border_clamp.* and piglit
texwrap).
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20570>
Vulkan spec 18.8. Primitives Generated Queries:
When a generated primitive query for a vertex stream is active,
the primitives-generated count is incremented every time a
primitive emitted to that stream reaches the transform feedback
stage, whether or not transform feedback is active.
We can see the order of stages in chapter 27 Fixed-Function
Vertex Post-Processing, which shows that the transform feedback
stage is before rasterization (and therefore culling).
Conclusion is that culled primitives should be included
in the primitives generated query.
This commit makes sure to emit the primitives generated query
code before culling and uses the input primitive count passed
to the current wave instead of the exec mask after culling.
Cc: mesa-stable
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21037>
Match iadd(x, #y). The format shift will get constant-folded away and, if y
is sufficiently small, the constant will be inlined by the AGX backend
optimizer. This gets rid of piles of 64-bit arithmetic from lowering UBOs. It
probably doesn't matter for perf since that's happening in preamble shaders but
it *is* noisy.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21108>
The offset is in vec4s, not words (unlike the component). This doesn't matter
right now since we get everything lowered (offset -> 0) but it will come up if
we implement clip distances natively (instead of lowering in FS).
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21097>
This works around bugs in a LOT of applications, since fp16 texture coordinates
are almost never appropriate even though it's a valid implementation of the GLES
spec. It also doesn't seem to matter for perf.
Code from the Bifrost compiler which implements the same workaround for slightly
different reasons.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21082>
This speeds up glReadPixels. Instead of reading from the write-combined
framebuffer and converting colours on the CPU, this blits on the GPU to a
writeback staging resource with the colour conversion for free, and memcpies
from the writeback staging resource on the CPU.
In general, due to textures being write combined and tiled/compressed by default
by staging resources being linear writeback, blit-based texture transfer should
win out (you were going to blit anyway), particularly when format conversion is
involved
33% reduction in wall clock time for grim at 4K. No change in deqp-gles2
runtime.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21063>
We only need 4 byte alignment, not 8 bytes. This isn't a big difference in
practice, but it probably reduces padding in some cases. More importantly, it
corrects our XML to match what the hardware actually does, which is great.
(There is exactly enough room for a 40-bit address with 4 byte alignment.)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21118>
