This mirrors the vs/gs keys, and will be needed when adding images
support.
The const changes also mirror how the draw code work (as is needed
when we add images)
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This is why we can't have nice things. I'm sure there's someway
to do this with {0} but I really don't have time for that.
Fixes: 2631fd3b0b ("gallivm: rework lp_build_tgsi_soa to take a struct")
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
The parameters were getting messy and I have to add a few more
for compute shaders, so clean it up before proceeding.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This add support for setting shader buffers and passing them
to draw or binding them to the fragment shader jit.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
this isn't used outside this file.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This unifies the explicit rasterization discard as well as the implicit
rasterization disabled logic (which we need for another state tracker),
which really should do the exact same thing.
We'll now toss out the prims early on in setup with (implicit or
explicit) discard, rather than do setup and binning with them, which
was entirely pointless.
(We should eventually get rid of implicit discard, which should also
enable us to discard stuff already in draw, hence draw would be
able to skip the pointless clip and fallback stages in this case.)
We still need separate logic for only null ps - this is not the same
as rasterization discard. But simplify the logic there and don't count
primitives simply when there's an empty fs, regardless of depth/stencil
tests, which seems perfectly acceptable by d3d10.
While here, also fix statistics for primitives if face culling is
enabled.
No piglit changes.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The tgsi_info.num_tokens fix broke llvmpipe's detection of no-op shaders.
Fix the code to check for num_instructions <= 1 instead.
Fixes: 8fde9429c3 ("tgsi: fix incorrect tgsi_shader_info::num_tokens
computation")
Tested-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The new name make the zero-input behavior more obvious. The next
patch adds a new function with different zero-input behavior.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
The comment said it will only represent the lowest 32 regs. This was
not entirely true in practice, since at least on x86 you'll get
masked shifts (unless the compiler could recognize it already and toss
it out). It turns out this actually works out alright (presumably
noone uses it for temp regs) when increasing max sampler views, so
make that behavior explicit.
Albeit it feels a bit hacky (but in any case, explicit behavior there
is better than undefined behavior).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Once a lp_build_sampler_soa or lp_build_sampler_aos object is created,
it should never be modified. Found by inspection.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This probably isn't all that useful for GL, but there are apis where
sample_mask is a valid output even without msaa.
Just discard the pixel if the sample_mask doesn't include the bit for
sample 0.
Reviewed-by: Brian Paul <brianp@vmware.com>
With GALLIVM_DEBUG=perf set, output the relevant stats for shader cache usage
whenever we have to evict shader variants.
Also add some output when shaders are deleted (but not with the perf setting
to keep this one less noisy).
While here, also don't delete that many shaders when we have to evict. For fs,
there's potentially some cost if we have to evict due to the required flush,
however certainly shader recompiles have a high cost too so I don't think
evicting one quarter of the cache size makes sense (and, if we're evicting
based on IR count, we probably typically evict only very few or just one
shader too). For vs, I'm not sure it even makes sense to evict more than
one shader at a time, but keep the logic the same for now.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
This is mostly mechanical search-and-replace, plus touching up the
macros in u_dump_defines.c manually a bit.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Generally we should do tranpose after conversion, if the format has less than
32 bits per channel (if it has 32 bits, conversion is going to be a no-op
anyway...). This is obviously because there's less vectors to deal with.
Though the advantage for 16 bit formats isn't that big, and in fact with AVX
there isn't really any (as the 32bit unpacks can be done with 256bit, but
the smaller ones cannot, although that would change again with proper AVX2
support).
Only makes sense for 2d and not 1d cases. And to keep things easy, only handle
1,2 and 4 channels (rgbx is just fine).
For rgba unorm8 format the backend conversion sums up to these instruction
totals (not counting the movs for SSE2 due to 2-op syntax - generally every 2
unpacks need an additional mov).
SSE2 AVX
transpose: 32 unpack 16 unpack
untwiddle: 0 8 (128bit low/high permutes)
convert: 16 mul + 16 cvt 8 mul + 8 cvt
32->8bit: 12 pack 8 (128bit extract) + 12 pack
When doing transpose/untwiddle afterwards we get:
convert: 16 mul + 16 cvt 8 mul + 8 cvt
32->8bit: 12 pack 8 (128bit extract) + 12 pack
transpose/untwiddle 12 unpack 12 unpack
So for SSE2, this drops 20 unpacks (total instruction count 76->56)
whereas for AVX it replaces the 16 256bit unpacks with 8 128bit ones
and drops the 8 lo/hi permutes (in total 60->48). (Albeit to be fair,
the permutes could be dropped even when doing the transpose first,
they are extremely pointless but we'd need to be able to tell
lp_build_conv to reorder the vectors, for AVX2 we're going to need to
be able to tell lp_build_conv about ordering in any case.)
(With different ordering going into conversion, it would be possible
to do 4 unpacks + 4 pshufbs instead of 12 unpacks, but that might not
be better, and not all cpus can do it. Proper AVX2 support should eliminate
the 8 128bit extracts, reduce these 12 packs to 6 and the 12 unpacks to 2
pshufb + 2 permq ideally (+ 2 final 128bit extracts).)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
For rgbx formats, there is no point in doing alpha conversion again (and
with different tranpose even, so llvm can't eliminate it).
Albeit it looks like there's some minimal changes needed in the blend code
(found by code inspection, no test seemed to complain) if we do this -
the blend factors are already sanitized if we have no destination alpha,
however for src_alpha_saturate it looks like it still might make a
difference (note that we forced has_alpha to true before for some formats
and nothing complained, but this seems safer).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
llvm has _huge_ problems trying to load things like <4 x i8> vectors and
stitching such loads together to form 128bit vectors. My understanding
of the problem is that the type legalizer tries to extend that to
really a <4 x i32> vector and not a <16 x i8> vector with the 4 elements
first then followed by padding, so the shuffles for then combining things
together are more or less impossible - you can in fact see the pmovzxd
llvm generates. Pre-4.0 llvm just gives up on it completely and does a 30+
pextrb/pinsrb sequence instead.
It looks like current llvm has fixed this behavior (my guess would be
due to better shuffle combination and load/shuffle folds), but we can
avoid this by just loading as <1 x i32> values, combine that and only
cast at the end. (I suspect it might also work if we'd pad the loaded
vectors immediately before shuffling them together, instead of directly
stitching 2 such vectors together pairwise before combining the pair.
But this _might_ lose the ability to load the values directly into
their right place in the vector with pinsrd.). But using 32bit values
is probably easier for llvm as it will never give it funny ideas how
the vector should look like.
(This is possibly only a problem for 1x8bit formats, since 2x8bit will
end up fetching 64bit hence only two vectors are stitched together,
not 4, but we use the same strategy anyway.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Using bit replication. This path now resembles something which might make
sense. (The logic was mostly copied from llvmpipe fs backend.)
I am not convinced though it is actually faster than SoA sampling (actually
I'm quite certain it's always a loss with AVX).
With SoA it's just shift/mask/cvt/mul for getting the colors, whereas
there's still roughly 3 shifts, 3 or/and per channel for AoS
(i.e. for SoA it's exactly the same as it would be for a rgba8 format,
whereas the extra effort for AoS is significant). The filtering
might still be faster (albeit with FMA the instruction count gets down
quite a bit there on the SoA float filtering path on new cpus). And those
small unorm formats often don't have an alpha channel (which makes things
worse relatively for AoS path).
(This also fixes a trivial bug in the llvmpipe fs code this was derived
from, albeit it was only relevant for 4-bit channels.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
simd instruction sets usually have comparisons for equal, not unequal.
So use a different comparison against the mask itself - which also means
we don't need a all-zero as well as a all-one (for the pxor) reg.
Also add code to avoid scalar expansion of i1 values which we definitely
shouldn't do. There's problems with this though with llvm select
interaction, so it's disabled (basically using llvm select instead of
intrinsics may still produce atrocious code, even in cases where we
figured it should not, albeit I think this could probably be fixed
with some better selection of optimization passes, but I have zero
idea there really).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Applies on top of v3 of Tom's gallivm change.
v2:
- Tom Stellard: Use enums instread of strings.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Aaron Watry <awatry@gmail.com>
CC: Tom Stellard <thomas.stellard@amd.com>
CC: Jan Vesely <jan.vesely@rutgers.edu>
We only did depth clamp when the value was written from the fs.
This is very wrong both for d3d10 and GL, and only passed the
corresponding piglit test due to pure luck (it no longer does
with the enhanced test).
Also, interpolation clamped values to 1.0 always, which can legitimately
happen if depth clip is disabled, so fix that as well (untested).
There is one unresolved issue left, d3d10 always does depth clamping,
whereas GL does not (but does [0,1] clamp instead for fs depth outputs)
- this information isn't in any gallium state object, leave it as-is
for now (though it looks like llvmpipe misses the [0,1] clamp as well).
This (with the previous patch) fixes piglit depth-clamp-range test.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The gallium contract would be that bind flags must indicate all possible
bindings a resource might get used, but fact is the mesa state tracker does
not set bind flags correctly, and this is more or less unfixable due to GL.
This caused a bug with piglit arb_uniform_buffer_object-rendering-dsa
since 6e6fd911da - the commit is correct,
but it caused us to miss updates to fs UBOs completely, since the
corresponding buffer didn't have the appropriate bind flag set (thus we
wouldn't check if it is indeed currently bound).
See the discussion about this starting here:
https://lists.freedesktop.org/archives/mesa-dev/2016-June/119829.html
So, update the bind flags when we detect such usage.
Note we update this value for now only in places which matter for us - that
is creating sampler/surface view, or binding constant buffer. There's plenty
more places (setting streamout buffers, vertex/index buffers, ...) where
things can be set with the wrong bind flags, but the bind flags there never
matter.
While here also make sure we only set dirty constant bit when it's a fs
constant buffer - totally doesn't matter if it's vs/gs.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Use PIPE_SWIZZLE_* everywhere.
Use X/Y/Z/W/0/1 instead of RED, GREEN, BLUE, ALPHA, ZERO, ONE.
The new enum is called pipe_swizzle.
Acked-by: Jose Fonseca <jfonseca@vmware.com>
The blend code would do a conditional assignment based on it, causing valgrind
to complain. Since that variable was actually unused in this case, this
doesn't fix anything but the warning.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94955
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
Only provide a fallback for LLVM 3.3.
One less dependency on LLVM C++ interface.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
They can't actually be 0 (as position is there) but should avoid confusion.
This was supposed to have been done by af7ba989fb
but I accidentally pushed an older version of the patch in the end...
Also prettify slightly. And make some notes about the confusing and useless
fs input "map".
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
compressed textures are very slow because decoding is rather complex
(and because there's no jit code code to decode them too for non-technical
reasons).
Thus, add some texture cache which holds a couple of decoded blocks.
Right now this handles only s3tc format albeit it could be extended to work
with other formats rather trivially as long as the result of decode fits into
32bit per texel (ideally, rgtc actually would decode to more than 8 bits
per channel, but even then making it work for it shouldn't be too difficult).
This can improve performance noticeably but don't expect wonders (uncompressed
is unsurprisingly still faster). It's also possible it might be slower in
some cases (using nearest filtering for example or if there's otherwise not
many cache hits, the cache is only direct mapped which isn't great).
Also, actual decode of a block relies on util code, thus even though always
full blocks are decoded it is done texel by texel - this could obviously
benefit greatly from simd-optimized code decoding full blocks at once...
Note the cache is per (raster) thread, and currently only used for fragment
shaders.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Generated by running:
git grep -l INLINE src/gallium/ | xargs sed -i 's/\bINLINE\b/inline/g'
git grep -l INLINE src/mesa/state_tracker/ | xargs sed -i 's/\bINLINE\b/inline/g'
git checkout src/gallium/state_trackers/clover/Doxyfile
and manual edits to
src/gallium/include/pipe/p_compiler.h
src/gallium/README.portability
to remove mentions of the inline define.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Acked-by: Marek Olšák <marek.olsak@amd.com>
Pretty trivial, fixes the issue that we're expected to be able to blit
stencil surfaces (as the blit just relies on util blitter code which needs
stencil export to do it).
2 piglits skip->pass, 11 fail->pass
v2: prettify, keep different stencil ref value handling out of depth/stencil
test itself.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
The callbacks used for getting the dynamic texture/sampler state were using
the jit_context from the generated jit function. This works just fine, however
that way it's impossible to generate separate functions for texture sampling,
as will be done in the next commit. Hence, pass this pointer through all
interfaces so it can be passed to a separate function (technically, it would
probably be possible to extract this pointer from the current function instead,
but this feels hacky and would probably require some more hacks if we'd use
real functions instead of inlining all shader functions at some point).
There should be no difference in the generated code for now.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Addresses MSVC warnings "result of 32-bit shift implicitly converted to
64 bits (was 64-bit shift intended?)", which can often be symptom of
bugs, but in these cases were all benign.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
We'll need to update gallivm for the interface changes in LLVM 3.6, and
the fewer the number of older LLVM versions we support the less hairy that
will be.
As consequence HAVE_AVX define can disappear. (Note HAVE_AVX meant
whether LLVM version supports AVX or not. Runtime support for AVX is
always checked and enforced independently.)
Verified llvmpipe builds and runs with with LLVM 3.3, 3.4, and 3.5.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This is one step to make llvmpipe thread safe as mandated by the OpenGL
standard. Using the global LLVMContext is obviously a problem for
that kind of use pattern. The patch introduces two LLVMContext
instances that are private to an OpenGL context and used for all
compiles. One is put into struct draw_llvm and the other
one into struct llvmpipe_context.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Signed-off-by: Mathias Froehlich <Mathias.Froehlich@web.de>
2ea923cf57 had the side effect of IR counting
now being done after IR optimization instead of before. Some quick analysis
shows that there's roughly 1.5 times more IR instructions before optimization
than after, hence the effective shader cache size got quite a bit smaller.
Could counter this with an increase of the instruction limit but it probably
makes more sense to count them after optimizations, so move that code.
Reviewed-by: Brian Paul <brianp@vmware.com>
Dave Airlie
Eric Engestrom
Marek Olšák
Roland Scheidegger
Brian Paul
Ian Romanick
Nicolai Hähnle
Aaron Watry
Rob Clark
Jose Fonseca
Edward O'Callaghan
Ilia Mirkin
Eric Anholt
Mathias Fröhlich