Right now we're using ralloc to GC our NIR instructions, but ralloc has
significant overhead for its recursive nature so it would be nice to use a
simpler mechanism for GCing instructions.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11776>
Driver like radeonsi load varying in a scalar manner, so prefer to pack
varying with different interpolation qualifier into same slot to save
space.
But driver like panfrost/bifrost can load varying in vector manner,
so prefer to pack varying with same interpolation qualifier.
Driver can add interpolation qualifiers which are able to be
packed into same varying slot to pack_varying_options nir option.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12537>
Without this, there's no way to match the UBO nir_variable declarations to
the load_ubo intrinsics referencing their data.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Acked-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12175>
Per-primitive is similar to per-vertex attributes, but applies to all
fragments of the primitive without any interpolation involved.
Because they are regular input and outputs, keep track in shader_info
of which I/O is per-primitive so we can distinguish them after deref
lowering. These fields can be used combined with the regular
`inputs_read`, `outputs_written` and `outputs_read`.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10600>
We can't append instructions following a return/halt instruction
because the control flow helpers will modify the successor of the
block containing the return/halt. And the NIR validator enforces that
the return/halt must have the end of the function as successor.
This tends to happen following lower_shader_calls lowering which
inserts halts. This probably doesn't prevent the optimization, it'll
just happen in one of the return shaders after the halt has been
removed.
v2: Move prev block ending check earlier in the function (Daniel)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12506>
Six opcodes are added: sdot_4x8_iadd, udot_4x8_uadd, sudot_4x8_iadd,
sdot_4x8_iadd_sat, udot_4x8_uadd_sate, and sudot_4x8_iadd_sat. These
represent the combinations of integer dot-product and add that operate
on packed source vectors. That is, the four 8-bit values for each
vector is stored in a single 32-bit integer.
Some hardware may prefer to operate on unpacked byte vectors. When such
hardware comes to Mesa, we'll have to figure out how to name things.
v2: Add nir_op_iudp4a and nir_op_iudp4a_sat instructions. These opcodes
are not 2-source commutative.
v3: Rename all opcodes to be more like some existing 4x8 opcodes.
Suggested by Timur. Change type of packed vector sources to uint32,
change types of constant folding variables to have explicit size, and
delete some extra casts. All suggested by Jason.
v4: Fix typo previously noticed by Alyssa but missed in v2.
v5: Add has_sudot_4x8 flag. Requested by Rhys.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12142>
A lot of CTS tests write a u8vec4 or an i8vec4 to an SSBO. This results
in a lot of shifts and MOVs. When that pattern can be recognized, the
individual 8-bit components can be packed much more efficiently.
v2: Rebase on b4369de27f ("nir/lower_packing: use
shader_instructions_pass")
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9025>
gl_FragDepth default value is gl_FragCoord.z so if a shader does:
gl_FragDepth = gl_FragCoord.z
we can drop this assignment.
v2: use nir_ssa_scalar_resolved and don't do this is gl_FragDepth
is wrote multiple times (Jason)
v3: - move to its own pass (Jason)
- handle var = NULL (Rhys)
v4: refactoring (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10697>
The liveness information will be a superset of real liveness so it's
unlikely something will explode if it tries to use it. However, it is
out-of-date and should be re-run if someone really wants it.
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12186>
This is where it should be rather than having to pass it into the
optimisation pass every time.
It also allows us to call the loop analysis pass without having to
duplicate these options which we will do later in this series.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12064>
For TGSI, we need the coordinate, comparator, bias, and LOD all together
in the first two vec4 args, and by doing it in the backend we were
generating extra MOVs.
softpipe shader-db results:
total instructions in shared programs: 2985416 -> 2953625 (-1.06%)
instructions in affected programs: 499937 -> 468146 (-6.36%)
total temps in shared programs: 544769 -> 565869 (3.87%)
temps in affected programs: 105469 -> 126569 (20.01%)
i915g shader-db:
total instructions in shared programs: 371625 -> 369594 (-0.55%)
instructions in affected programs: 24903 -> 22872 (-8.16%)
total tex_indirect in shared programs: 11381 -> 11365 (-0.14%)
tex_indirect in affected programs: 43 -> 27 (-37.21%)
LOST: 7
GAINED: 16
The temps increase is the pre-existing issue that we never release temps
for NIR regs, which doesn't matter much for softpipe (just memory/cache
footprint) but does for i915g as seen by shaders that no longer compile
(though overall we seem to win).
Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11912>
Only fragment and some compute shaders support implicit derivatives.
They're totally meaningless without helper invocations and some
understanding of the dispatch pattern. We've got code to lower
nir_texop_tex in these shader stages to use an explicit derivative of 0
but it was pretty badly broken:
1. It only handled nir_texop_tex, not nir_texop_txb or nir_texop_lod.
2. It didn't take min_lod into account
3. It was conflated with adding a missing LOD parameter to opcodes
which expect one such as nir_texop_txf. While not really a bug,
this does make it way harder to reason about the code.
4. Unless you set a flag (which most drivers don't), it left the
opcode nir_texop_tex instead of nir_texop_txl which it should have
been.
This reworks it to go through roughly the same path as other LOD
lowering only with a constant lod of 0 instead of calling out to
nir_texop_lod. We also get rid of the lower_tex_without_implicit_lod
flag because most drivers set it and those that don't are probably
subtly broken. If someone really wants to get nir_texop_tex in their
vertex shaders, they can write a new patch to add the flag back in.
Fixes: e382890e25 "nir: set default lod to texture opcodes that..."
Fixes: d5ac5d6e83 "nir: Add option to lower tex to txl when..."
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11775>
It's intel-specific, used to get at MSAA compression information.
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11775>
This patch also adds has_iadd3 bit to give more control if backend
supports ternary add instruction or not.
v2:
- Add patterns in late optimization (Connor Abbott)
Suggested-by: Alyssa/Jason
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11596>
We say that they're for debug only but we don't really have a good
policy around when to set them and when not to. In particular,
nir_lower_system_values and nir_lower_vars_to_ssa which are the chief
producers of SSA values which might reasonably have a name do not bother
to set one. We have some names set from things like BLORP and RADV's
meta shaders but AFAICT, they're setting a name more because it's there
than because they actually care.
Also, most things other than nir_clone and nir_serialize don't bother to
try and preserve them. You can see in the diffstat of this commit
exactly what passes attempt to preserve names. Notably missing from the
list is opt_algebraic which is the single largest source of SSA def
churn and it happily throws names away.
These observations lead me to question whether or not names are actually
useful at all or if they're just taking up space (8B per instruction)
and wasting CPU cycles (to ralloc_strdup on the off chance we do have
one). I don't think I can think of a single time in recent history
where I've been debugging a shader issue and a SSA value name has been
there and been useful. If anything, the few times they are there, they
just throw me off because they mess up the indentation in nir_print.
iris shader-db on my system gets runtime -2.07734% +/- 1.26933% (n=5)
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5439>
On qualcomm, we have shared registers similar to SGPR's on AMD. However,
there is no readlane or readfirstlane primitive. shared registers can
only be written to when just one lane is active. This means that we have
to lower readInvocation(val, id) to something like:
if (gl_SubgroupInvocation == id) {
scalar_reg = val;
}
return scalar_reg;
However it's a bit difficult to actually get the value of
gl_SubgroupInvocation in the backend, because for compute it requires
some calculations and we don't have any CSE support in the backend. This
intrinsic lets us turn it into
"readInvocationCond(val, id == gl_SubgroupInvocation)" in NIR at which
point the backend code generation is a lot easier.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6752>
Qualcomm has a mode with a subgroup size of 128, so just emitting larger
integer operations and then lowering them later isn't an option. This
makes the pass able to handle the lowering itself, so that we don't have
to go down to 64-thread wavefronts when ballots are used.
(The GLSL and legacy SPIR-V extensions only support a maximum of 64
threads, but I guess we'll cross that bridge when we come to it...)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6752>
Lower it to a vote instead of a ballot. This was only used for AMD, and
in that case they're pretty much the same. However Qualcomm has a vote
builtin, which we want to use instead of ballots.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6752>
Reduces the work that other shader passes have to do to look at dead code,
and possibly extra rounds around the optimization loop if dce wasn't the
last pass in it.
shader-db runtime -1.12919% +/- 0.264337% (n=49) on SKL.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11628>
In addition to register pressure benefits from getting more fp16/int16,
this avoids i2imp's from standing in the way of loop unrolling.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11545>
Sometimes you might want to find a constant source without going through
all the copy prop and constant folding to make your source be a
load_const.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11613>
Really copying Jason's pass.
Changes:
- Instead of all the intel lowering introduce rt_{execute_callable,trace_ray,resume}
- Add the ability to use scratch intrinsics directly.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10339>
About half or more of the text here is actually from Connor Abbot. I've
edited it a bit to bring it up-to-date and make a few things more clear.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11438>
For debug on Android, it's useful to be able to print shaders to the
android log interface, since you don't usually have stdout/stderr.
Reviewed-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9262>
Be consistent with other usages in Vulkan and SPIR-V, and the recently
added workgroup_size field.
Acked-by: Emma Anholt <emma@anholt.net>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11190>
These are similar to AYUV, but the channel ordering is different... in
such a way that there's no RGBA format that will make the channels line
up right.
v2: Rebase on bc438c91d9 ("nir/lower_tex: ignore texture_index if
tex_instr has deref src")
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9610>
Many fragment shaders do a discard using relatively little information
but still put the discard fairly far down in the shader for no good
reason. If the discard is moved higher up, we can possibly avoid doing
some or almost all of the work in the shader. When this lets us skip
texturing operations, it's an especially high win.
One of the biggest offenders here is DXVK. The D3D APIs have different
rules for discards than OpenGL and Vulkan. One effective way (which is
what DXVK uses) to implement DX behavior on top of GL or Vulkan is to
wait until the very end of the shader to discard. This ends up in the
pessimal case where we always do all of the work before discarding.
This pass helps some DXVK shaders significantly.
v2 (Jason Ekstrand):
- Fix a couple of typos (Grazvydas, Ian)
- Use the new nir_instr_move helper
- Find all movable discards before moving anything so we don't
accidentally re-order anything and break dependencies
v3 (Pierre-Eric): remove the call to nir_opt_conditional_discard based
on Daniel Schürmann comment.
v4 (Pierre-Eric):
- handle demote intrinsics and drop derivatives_safe_after_discard
- add early return if discards/demotes aren't used
v5 (Pierre-Eric):
- use pass_flags instead of instr set (Daniel Schürmann)
v6 (Daniel Schürmann):
- cleanup and fix pass_flags handling
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10522>
