It appears we never had a test in piglit or deqp sampling from a null
surface...
It turns out this triggers a hang on IVB only. Updating the null
surface format to R32_UINT fixes the hang on ivb and doesn't affect
other platforms, so set it by default for all platforms.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Gitlab: https://gitlab.freedesktop.org/mesa/mesa/issues/1872
Cc: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This should improve texture sampling performance on GC3000.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
We're missing the offset of the slice in the subslice mask...
This worked for most platforms that don't have first slice fused off
because we would reread the same mask from slice0 again and again...
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: c1900f5b0f ("intel: devinfo: add helper functions to fill fusing masks values")
Gitlab: https://gitlab.freedesktop.org/mesa/mesa/issues/1869
Reviewed-by: Mark Janes <mark.a.janes@intel.com>
We were supplying __DRI2_THROTTLE_SWAPBUFFER, rather than the obvious
choice of __DRI2_THROTTLE_COPYSUBBUFFER. This meant that we hit the
swap-based frame throttling. glXCopySubBuffer doesn't seem like it's
intended to be a frame boundary, so we'd like to avoid this throttling.
Tested-by: Michel Dänzer <mdaenzer@redhat.com> # DRI3 only
Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
glXCopySubBufferMESA copies data from the back buffer to the front,
so it needs to perform a MSAA downsampling operation just like
glXSwapBuffers would.
Currently, the CopySubBuffer implementations supply a throttle reason
of __DRI2_THROTTLE_SWAPBUFFERS, so they hit this path and work today.
But we'd like to avoid swapbuffer throttling in this case, so the next
patch will change that reason.
Tested-by: Michel Dänzer <mdaenzer@redhat.com> # DRI3 only
Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Not much to do to enable this, just make sure to always write to the
GGTT :)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I thought I fixed this, but I guess I must have broken it again.
Fixes various dEQP-VK.draw.* tests
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Working on the algebraic implementation, I was being driven nuts by my
editor not highlighting and handling indentation for the C code. It turns
out that it's basically not pass-specific code, and we can move it over to
the relevant .c file. Replaces 30KB of code with 34KB of data on my i965
build. No perf diff on shader-db (n=3)
Reviewed-by: Ian Romanick <ian.d.romainck@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This lets us memoize range analysis work across instructions. Reduces
runtime of shader-db on Intel by -30.0288% +/- 2.1693% (n=3).
Fixes: 405de7ccb6 ("nir/range-analysis: Rudimentary value range analysis pass")
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Having passes generate these is just making more work for copy
propagation (and thus probably calling more optimization passes)
later. Noticed while trying to debug nir_opt_algebraic()
top-to-bottom having O(n^2) behavior due to not finding new matches in
replacement code.
Reviewed-by: Ian Romanick <ian.d.romainck@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This matches what we do for uses_sample_qualifier, and what we
do in ir_set_program_inouts.cpp as well.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This simplifies ACO and allows the lowered code to be optimized (in
particular, constant folded).
Totals from affected shaders:
SGPRS: 1776 -> 1776 (0.00 %)
VGPRS: 1436 -> 1436 (0.00 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 203452 -> 203564 (0.06 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 103 -> 103 (0.00 %)
At least some of the code size increase seems to be from literals being
applied to instructions as a result of constant folding.
v2: remove fmod/frem handling in init_context()
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
0 is __DRI2_THROTTLE_SWAPBUFFER, which doesn't really make sense here.
Avoids dri_flush() throttling twice for the same glFlush call with front
buffer rendering, as described in
https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2057 .
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
If the instruction interpolateAtCentroid is used the extra interpolator
must also be enabled in the state.
Fixes: fs-interpolateatcentroid-block
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Now that we have live_out calculated per block as metadata, calculating
liveness of an instruction at a given point in the program becomes O(n)
to the size of the block worst-case, rather than O(n) the program.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Callers should have liveness info ready. Ideally we'd have a nice
metadata tracking framework like NIR to handle this automatically, but
for now this will allow us to make forward progress... when we're about
to do something with liveness, invalidate everything ahead to force a
clean calculation.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This will allow us to explicitly invalidate liveness analysis results so
we can cache liveness results.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
By definition, once liveness analysis has occurred:
live_out = OR {succ} succ->live_in
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
There are unfortunately two distinct liveness analysis passes in the
compiler right now -- one good (but complex) pass used by RA based on
solving data flow equations, and one awful (but simple) pass used for
dead code elimination and bundling based on an abstract walk of the AST.
Let's move RA's pass into shared code so we can work on unifying.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This allows us to fill in ctx->temp_count explicitly, even if we haven't
squished down the MIR.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We already enforce this with the SSA/register distinction in the
backend. There is no need to duplicate this logic merely for an assert.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Now that we have track inter-batch dependencies, the flush done in
panfrost_set_framebuffer_state() is no longer needed. Let's get rid of
it.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Now that we have all the pieces in place to support pipelining batches
we can get rid of the drmSyncobjWait() at the end of
panfrost_batch_submit().
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We don't have to flush all batches when we're only interested in
reading/writing a specific BO. Thanks to the
panfrost_flush_batches_accessing_bo() and panfrost_bo_wait() helpers
we can now flush only the batches touching the BO we want to access
from the CPU.
This fixes the dEQP-GLES2.functional.fbo.render.texsubimage.* tests.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This is needed if we want to free the panfrost_batch object at submit
time in order to not have to GC the batch on the next job submission.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Will be useful to make the ioctl(WAIT_BO) call conditional on BOs that
are not exported/imported (meaning that all GPU accesses are known
by the context).
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This will allow us to only flush batches touching a specific resource,
which is particularly useful when the CPU needs to access a BO.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
And use it in panfrost_flush() to flush all batches, and not only the
one currently bound to the context.
We also replace all internal calls to panfrost_flush() by
panfrost_flush_all_batches() ones.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
The panfrost_fence logic currently waits on the last submitted batch,
but the batch serialization that was enforced in
panfrost_batch_submit() is about to go away, allowing for several
batches to be pipelined, and the last submitted one is not necessarily
the one that will finish last.
We need to make sure the fence logic waits on all flushed batches, not
only the last one.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
The idea is to track which BO are being accessed and the type of access
to determine when a dependency exists. Thanks to that we can build a
dependency graph that will allow us to flush batches in the correct
order.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We'll soon need to freeze a batch not only when it's flushed, but also
when another batch depends on us, so let's add a helper to avoid
duplicating the logic.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We just replace the per-context out_sync object by a pointer to the
the fence of the last last submitted batch. Pipelining of batches will
come later.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
So we can implement fine-grained dependency tracking between batches.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
So we can store the flags as data and keep the BO as a key. This way
we keep track of the type of access done on BOs.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
The type of access being done on a BO has impacts on job scheduling
(shared resources being written enforce serialization while those
being read only allow for job parallelization) and BO lifetime (the
fragment job might last longer than the vertex/tiler ones, if we can,
it's good to release BOs earlier so that others can re-use them
through the BO re-use cache).
Let's pass extra access flags to panfrost_batch_add_bo() and
panfrost_batch_create_bo() so the batch submission logic can take the
appropriate when submitting batches. Note that this information is not
used yet, we're just patching callers to pass the correct flags here.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We know a shader will be used by a batch when
panfrost_patch_shader_state() is called, so let's add the shader BO at
that time.
Suggested-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
The CTS finally has agreed to drop the requirement for a
565-no-depth-no-stencil config for ES 3.0. Hence we can now remove the
code to satisfy this requirement using a pbuffer-only visual with
whatever other buffers the driver happens to have given us.
This reverts commit 82607f8a90,
commit 6ad31c4ff3 and
commit dacb11a585.
v2:
- Reference the VK-GL-CTS issue (Eric E.).
v3:
- Don't revert
fc21394bc4 ("egl: Quiet warning about front buffer rendering for pixmaps/pbuffers")
(Kenneth).
References: VK-GL-CTS issue 1601.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Andres Gomez <agomez@igalia.com>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
On 64 bits platforms, some atomic operations like __sync_fetch_and_add()
have constant time, but on 32 bits platforms they are implemented with a
loop and might take much longer.
Additionally, it seems like if their operands are not aligned to 64
bits, they also require extra memory accesses. From the Intel
Architecture's Developer Manual Vol. 1, 4.1.1:
"A word or doubleword operand that crosses a 4-byte boundary or a
quadword operand that crosses an 8-byte boundary is considered
unaligned and requires two separate memory bus cycles for access."
Forcing the u64 field to be aligned to 64 bits seems to make the unit
tests that are stressing this finish much faster.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
v1 by Topi Pohjolainen
v2,v3 by Anuj Phogat:
- Apply for gen >= 11
- Remove wa_bug_xxx function
- Use helper functions
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
RADV and RadeonSI both lower these two NIR instructions.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This lowers fmod and frem at NIR level like RadeonSI. fmod is
already lowered directly in NIR->LLVM, and frem will be lowered by
LLVM anyways.
This fixes a LLVM crash with:
dEQP-VK.glsl.builtin.precision_fp16_storage32b.frem.compute.scalar.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
