This is needed to allow lowering blend operations in fragment shaders
when the blend operation uses dynamic blend constants.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13060>
In NIR image instructions, the array index always follows the last
image coordinate, meaning that array index is in coord.y for 1D
arrays. But the current panfrost ABI wants it in coord.z regardless of
the image dimension.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13078>
The Vulkan copy shaders sometimes need to partially write a texel and
issue a load on the FRAGOUT variable in that case, but they do know
the format of the tile buffer in advance in that case. Let's not add an
RT_CONVERSION sysval if we can avoid it.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Acked-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12095>
Allow extracting components Y, Z or W from an 8bit vector.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12095>
Apparently, CLPER_V7 is missing from Mali G31, but CLPER_V6 works. Fixes
INSTR_INVALID_ENC faults and failures in
dEQP-GLES3.functional.shaders.derivate.* on Dvalin.
Technically not an errata but an implementation difference. I suspect
Mali G51 will need this as well, should we ever allowlist it.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12478>
ST_TILE lets us specify an explicit sample, whereas BLEND replicates to
all samples. This fully fixes the interaction between blend shaders and
multisampling on Bifrost, manifesting as
dEQP-GLES3.functional.fragment_ops.random.* failures with the
configuration rgba8888d24s8ms4.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12478>
Use the explicit sample mode and set the sample ID in the pixel indices
structure to the current sample ID. This fixes tilebuffer loads in blend
shaders on multisampled framebuffers.
Make sure the new routine is broken out to a helper for use with ST_TILE
in the next commit.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12478>
Map nir_op_fsat/etc to FCLAMP pseudo ops, instead of FADD. There are
significantly fewer knobs on FCLAMP, meaning significantly fewer things
to get wrong.
This fixes two(!) classes of bugs:
* Swizzles (failing to lower/compose swizzles on clamps)
* Numerical bugs (incorrectly treating +0.0 as an additive identity)
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12205>
Simplifies pattern matching. This commit by itself fixes multiple
numerical issues -- the previous fabsneg check failed to check the round
mode or the sign of the zero. That will break Vulkan/OpenCL.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12205>
This ensures we can constant fold the ALU ops used to lower:
* explicit LOD calculations
* array textures
* texture offsets
* multisample indices
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12205>
It's an optimization pass -- omitting it should not cause MMU faults
(!). Make sure the UBO push mask is set regardless of whether the pass
is called, and just call the pass when required.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12328>
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>
Otherwise there is an uninitialized read, and the register allocation
will fail. (In the sense of failing a precondition. This manifests as
synthetic interference leading to higher register pressure and useless
moves. The allocation itself is ok, but it indicates a real bug.)
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12130>
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>
The enums alias. Fixes the following warning under clang:
../src/panfrost/bifrost/bifrost_compile.c:2515:25: warning: implicit conversion from enumeration type 'enum bifrost_texture_fetch' to different enumeration type 'enum bifrost_lod_mode' [-Wenum-conversion]
BIFROST_TEXTURE_FETCH_TEXEL;
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12003>
Gets rid of the silly inheritance everywhere, which has caused _far_
more problems in practice than it has fixed. It was an idea I tried
before the pandemic. It didn't work. I'm finally cleaning it up.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11936>
Unlike the dependency analysis for the skip bits which is a function of
the data flow graph, the thread termination analysis is dependent on the
actual sequence of instructions. As such, it must be done after
scheduling to be correct in the presence of out-of-order scheduling.
Furthermore it's specified in terms of clauses, not instructions.
Reflecting this in our code gets a nice simplification. As a side effect
this puts extra td flags on subsequent clauses, which matches the DDK's
behaviour. (Maybe td is just a hint?)
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10961>
Set the .skip bit on texture instructions and the terminate discarded
threads bit on the clause header based on data flow analysis of helper
invocations. This code is adapted from Midgard, which requires the same
analysis with a few details changed.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11733>
Redundant with the texture operation descriptor, but we don't want to
parse that in the rest of the compiler. Handling it as a pseudo-modifier
lets us share a code path with TEXS.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11733>
Based on analysis of results from the Mali Offline Compiler. I am
uncertain how well these translate to real life, and they are
normalized counts only...
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11733>
Bifrost v7 trades off register pressure and occupancy. If we restrict to
[R0, R15] U [R48, R63], we get full occupancy, but if we use the full
register file, we only get half occupancy. Try to allocate just 32
registers, and only use the full 64 registers if that would spill.
Clever heuristics could make this both more effective (live range
splitting, shuffling, spilling if deemed acceptable) and cheaper at
compile-time (tracking maximum liveness to determine if it's possible to
hit at all). For now, this should suffice.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11733>
Since the GLSL compilers will pack together flat varyings with no regard
to type, under the assumption the backend can deal with it. I guess we
can deal with it then... Fixes fails in
dEQP-GLES31.functional.separate_shader.random.*
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>
If the shader packs multiple varyings into the same location with
different location_frac, we'll need to lower to a single varying store
that collects all of the channels together. This is not trivial during
code gen, but it is trivial to do in NIR right before codegen by relying
on nir_lower_io_to_temporaries. Since we're guaranteed all varyings will
be written exactly once, in the exit block, we can scan the shader
linearly and collect stores together in a single pass.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>
We have no native way to swizzle out a nonzero component in a load, but
we can simply load extra components and do the swizzle in shader
instructions. This is inefficient, since it loads data to discard
immediately, but it's required for conformance in some edge cases.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>