These are required to get piglit's idiv tests working. The
unsigned<->float conversions are wrong, but are good enough to get
piglit's small ranges of values working.
We're over-allocating our BCL in vc4_draw.c, so this never mattered.
However, new RCL-only blit support might end up here without having set up
any BCL contents.
Just build up arrays for src0/src1, and use create_collect()..
Also add back missing .3d flag for 3d/cube textures.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
I noticed some cases where we where trying to copy-propagate indirect
src's into places they cannot go, like 2nd src for cat3 (mad, etc).
Expand out valid_flags() to be aware of relativ flag, and fix up a few
related spots.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
When we get in a scenario where we cannot schedule any more instructions
due to address register conflict, clone the instruction that writes the
address register, and switch the remaining unscheduled users for the
current address register over to the new clone.
This is simpler and more robust than the previous attempt (which tried
and sometimes failed to ensure all other dependencies of users of the
address register were scheduled first).. hint it would try to schedule
instructions that were not actually needed for any output value.
We probably need to do the same with predicate register, although so far
it isn't so heavily used so we aren't running into problems with it
(yet).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
A bit fugly.. try and make this cleaner.. note if we hoist all the
get_addr() out of the loop we can drop the hashtable and just use
create_addr()..
Signed-off-by: Rob Clark <robclark@freedesktop.org>
It probably *should* be an assert, but for now TGSI f/e isn't very good
about dealing w/ CONST vs ABS/NEG. So for debug builds, print a warning
instead of crashing with an assert for now.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Without this, a3xx breaks.. a4xx would too if it had already implemented
support for passing driver params.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
For a normal MAD (ie. not MADSH), if first source is gpr and second
source is const, we can swap the first two sources to avoid needing a
mov instruction.
This gives back the biggest advantage TGSI f/e had over NIR f/e for
common shaders, since TGSI f/e had this logic in the f/e. Note that
doing this in copy-prop step has the advantage that it will also work
for cases like:
MOV TEMP[b], CONST[x]
MAD TEMP[d], TEMP[a], TEMP[b], TEMP[c]
Signed-off-by: Rob Clark <robclark@freedesktop.org>
So far just the system values that freedreno supports, so we may add
more later.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
With TXD we also have the ddx/ddy sources (before the sampler).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Split out from ttn_tex() since it is kind of a weird instruction that
maps to two NIR opcodes, and it was cleaner this way.
v2: query_levels doesn't take any args
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
We'll need this as well for TXQ. Split this out first to reduce noise
in the next patch.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Since the rest of NIR really would rather have these as variables rather
than registers, create a nir_variable per array. But rather than
completely re-arrange ttn to be variable based rather than register
based, keep the registers. In the cases where there is a matching var
for the reg, ttn_emit_instruction will append the appropriate intrinsic
to get things back from the shadow reg into the variable.
NOTE: this doesn't quite handle TEMP[ADDR[]] when the DCL doesn't give
an array id. But those just kinda suck, and should really go away.
AFAICT we don't get those from glsl. Might be an issue for some other
state tracker.
v2: rework to use load_var/store_var with deref chains
v3: create new "burner" reg for temporarily holding the (potentially
writemask'd) dest after each instruction; add load_var to initialize
temporary dest in case not all components are overwritten
v4: review comments: asserts and use ttn_src_for_indirect() in
ttn_array_deref() so we can drop later patch converting to use vec1 for
addr reg (since ttn_src_for_indirect() handles the imov to vec1 from
tgsi addr component that we want)
v5: rebase: new requirements about parent mem ctx for derefs
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Extract tgsi_dst->Index into a local.. split out from 'gallium/ttn: add
support for temp arrays' for noise reduction..
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Neither the shader nor the key change when doing elts or linear variant, so
this was just annoying (probably mildly useful at some point when we printed
the IR per function too).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
llvm goes crazy when doing that, using way more memory and time, though there's
probably more to it - this points to a very much similar issue as fixed in
8a9f5ecdb1. In any case I've seen a quite
plain looking vertex shader with just ~50 simple tgsi instructions (but with a
dozen or so such indirect constant buffer lookups) go from a terribly high
~440ms compile time (consuming 25MB of memory in the process) down to a still
awful ~230ms and 13MB with this fix (with llvm 3.3), so there's still obvious
improvements possible (but I have no clue why it's so slow...).
The resulting shader is most likely also faster (certainly seemed so though
I don't have any hard numbers as it may have been influenced by compile times)
since generally fetching constants outside the buffer range is most likely an
app error (that is we expect all indices to be valid).
It is possible this fixes some mysterious vertex shader slowdowns we've seen
ever since we are conforming to newer apis at least partially (the main draw
loop also has similar looking conditionals which we probably could do without -
if not for the fetch at least for the additional elts condition.)
v2: use static vars for the fake bufs, minor code cleanups
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Add SV_GEOMETRY_EMIT special variable type to track the
implicit dependencies between CUT/EMIT_VERTEX/MEM_RING
instructions so GCM/scheduler doesn't reorder them.
Mark emit instructions as unkillable so DCE doesn't eat them.
Enable only for evergreen/cayman as there are a few
unexplained GS piglit regressions on R6xx/R7xx with SB
enabled otherwise.
Signed-off-by: Glenn Kennard <glenn.kennard@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
CF_END could end up emitted in the middle of a shader on cayman
when there was a loop at the very end.
Fixes glsl-1.50-geometry-end-primitive and
ext_transform_feedback-geometry-shaders-basic piglit tests.
Signed-off-by: Glenn Kennard <glenn.kennard@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
arb_stencil_texturing-draw failed under softpipe because we got a float
back from the texturing function, and then tried to U2F it, stencil
texturing returns ints, so we should fix the tiling to retrieve
the stencil values as integers not floats.
Signed-off-by: Dave Airlie <airlied@redhat.com>
We limit y-tiling to 0x20 when depth is involved. However the function is
run for each miplevel, and the hardware expects miplevel 0 to have the
highest tiling settings. Perform the y-tiling limit on all levels of a
3d texture, not just the ones that have depth.
Fixes:
texelFetch fs sampler3D 98x129x1-98x129x9
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Tested-by: Nick Tenney <nick.tenney@gmail.com> # GT216
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
This code to handle absolute values on op3 srcs was a bit too simple,
it really needs a temp reg per src, not one per channel, make it
easier and let sb clean up the mess.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89831
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The NIR compiler frontend is an alternative to the TGSI f/e, producing
the same ir3 IR and using the same backend passes for scheduling, etc.
It is not enabled by default yet, as there are still some regressions.
To enable, use 'FD_MESA_DEBUG=nir'. It is enough to use with, for
example, xonotic or supertuxkart.
With the NIR f/e, scalarizing and a number of other lowering steps
happen in NIR, so we don't have to do them in ir3. Which simplifies the
f/e and allows the lowered instructions to pass through other
optimization stages.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Use the correct sprite replacement depending on the flip of the coord
mode, using either T or 1-T depending on whether we have an upper-left or
lower-left coordinate origin. This fixes all the point sprite piglits.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Copies nouveau_buffer and radeon_buffer. This allows a write to proceed
to an uninitialized part of a buffer even when the GPU is using the
previously-initialized portions.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Waiting on a bo being ready is handled in fd_bo_cpu_prep. No need to
keep separate timestamps around.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
A resource flush is an upload of a hypothetically-staging texture to the
GPU. For a UMA system, this will largely be a no-op or
cache-maintenance. Move the render flush logic into transfer_map where
it belongs, and clear out the transfer_flush function.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
pipe_sampler_view already contains a texture, remove the redundant
tex_resource member which pointed at the same thing.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Since NIR f/e currently encodes immediates in instructions (rather than
passing via const), we need to ensure that when const's are used the get
initialized to the proper values. Otherwise comparing NIR to TGSI
compiler, it will use proper immediate values in one case, and randomly
initialize values in the other. Which confuses ir3test.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Be smarter about propagating copies from const or immed, or with abs/neg
modifiers. Also, realize that absneg.s and absneg.f are really "fancy"
mov instructions.
This opens up the possibility to remove more copies. It helps the TGSI
frontend a bit, but will be really needed for the NIR f/e which builds
everything up in SSA form (ie. will *always* insert a mov from const or
immediate).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Even though in the end, they map to the same bits, the backend will need
to be able to differentiate float abs/neg vs integer abs/neg. Rather
than making the backend figure it out based on instruction opcode (which
when combined with mov/absneg instructions, can be awkward), just split
out different flags for each so the frontend can signal it's intentions
more clearly. Also, since (neg) for bitwise op's is actually a bitwise-
not, split it out into bnot flag.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Add helpers for constructing SSA forms of instructions.
Only partial cat5/cat6 coverage.. but we can add stuff as needed.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We need to pull in libnir.la and it's dependency libglsl_util.la. Also,
_mesa_error_no_memory() must be defined.
Fortunately with libnir.la (vs pulling in all of libglsl.la) we don't
also need libstdc++.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Commit 1a170980a0 started writing to q->data[4]/[5] but kept the
per-query space at 16, which meant that in some cases we would write
past the end of the buffer. Rotate by 32, like nvc0 does. This ensures
that we always have 32 bytes in front of us, and the data writes will go
within the allocated space.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89679
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Tested-by: Nick Tenney <nick.tenney@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Tobias Klausmann <tobias.johannes.klausmann@mni.thm.de>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Commit 09ee907266 added logic to fold immediates into mad operations,
but the emission code is only there for fmad. Only allow it on float
types.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Commit fb63df2215 added 4-byte mad support, but only supported
emission for floats. Disable it for ints for now.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
The hardware only supports 4 MRTs. It should be possible to emulate
support for 8, but doesn't seem worth the trouble.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This complication is unnecessary and makes MRTs more complicated and
likely to generate tons of variants.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Unused as of commit 630ab0d27ba(mesa: remove last of MAX_WIDTH,
MAX_HEIGHT). Update all the remaining references to the defines.
v2: Use the correct variable name in the comments
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
I was previously using temporary disables of VC4 optimization to show the
benefits of improved NIR optimization, but this can get me quick and dirty
numbers for NIR-only improvements without having to add hacks to disable
VC4's code (disabling of which might hide ways that the NIR changes would
hurt actual VC4 codegen).
NIR brings us better optimization than I would have bothered to write
within the driver, developers sharing future optimization work, and the
ability to share device-specific lowering code that we and other
GLES2-level drivers need.
total uniforms in shared programs: 13421 -> 13422 (0.01%)
uniforms in affected programs: 62 -> 63 (1.61%)
total instructions in shared programs: 39961 -> 39707 (-0.64%)
instructions in affected programs: 15494 -> 15240 (-1.64%)
v2: Add missing imov support, and assert that there are no dest saturates.
v3: Rebase on the target-specific algebraic series.
v4: Rebase on gallium-includes-from-NIR changes in mater.
v5: Rebase on variables being in lists instead of hash tables.
v6: Squash in intermediate changes that used the NIR-to-TGSI pass (which
I'm not committing)
This will be used by the VC4 driver for doing device-independent
optimization, and hopefully eventually replacing its whole IR. It also
may be useful to other drivers for the same reason.
v2: Add all of the instructions I was relying on tgsi_lowering to remove,
and more.
v3: Rebase on SSA rework of the builder.
v4: Use the NIR ineg operation instead of doing a src modifier.
v5: Don't use ineg for fnegs. (infer_src_type on MOV doesn't do what I
expect, again).
v6: Fix handling of multi-channel KILL_IF sources.
v7: Make ttn_get_f() return a swizzle of a scalar load_const, rather than
a vector load_const. CSE doesn't recognize that srcs out of those
channels are actually all the same.
v8: Rebase on nir_builder auto-sizing, make the scalar arguments to
non-ALU instructions actually be scalars.
v9: Add support for if/loop instructions, additional texture targets, and
untested support for indirect addressing on temps.
v10: Rebase on master, drop bad comment about control flow and just choose
the X channel, use int comparison opcodes in LIT for now, drop unused
pipe_context argument..
v11: Fix translation of LRP (previously missed because I mis-translated
back out), use nir_builder init helpers.
v12: Rebase on master, adding explicit include of mtypes.h to get
INTERP_QUALIFIER_*
v13: Rebase on variables being in lists instead of hash tables, drop use
of mtypes.h in favor of util/pipeline.h. Use Ken's nir_builder
swizzle and fmov/imov_alu helpers, drop "struct" in front of
nir_builder, use nir_builder directly as the function arg in a lot of
cases, drop redundant members of ttn_compile that are also in
nir_builder, drop some half-baked malloc failure handling.
v14: The indirect uniform src0 should be scalar, not vector (noticed as
odd by robclark, confirmed by cwabbott). Apply Ken's review to
initialize s->num_uniforms and friends, skip ttn_channel for dot
products, and use the simpler discard_if intrinsic.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v13)
Acked-by: Rob Clark <robclark@freedesktop.org>
At the moment to get an EGL image to a dma-buf file descriptor,
you have to use EGL_MESA_drm_image, and then use libdrm to
convert this to a file descriptor.
This extension just provides an API modelled on EGL_MESA_drm_image,
to return a dma-buf file descriptor.
v2: update spec for new API proposal
add internal queries to get the fourcc back from intel driver.
v2.1: add gallium pieces.
v2.2: add offsets to spec and API, rename fd->fds, stride->strides
in API. rewrite spec a bit more, add some q/a
v2.3:
add modifiers to query interface and 64-bit type for that (Daniel Stone)
specifiy what happens to num fds vs num planes differences. (Chad Versace)
v2.4:
fix grammar (Daniel Stone)
Signed-off-by: Dave Airlie <airlied@redhat.com>
We've seen some cases where performance can hurt quite a bit.
Technically, the more simple the function the more overhead there is
for using a function for this (and the less benefits this provides).
Hence don't do this if we expect the generated code to be simple.
There's an even more important reason why this hurts performance,
which is shaders reusing the same unit with some of the same inputs,
as llvm cannot figure out the calculations are the same if they
are performned in the function (even just reusing the same unit without
any input being the same provides such optimization opportunities though
not very much). This is something which would need to be handled by IPO
passes however.
When nvc0_push_vbo calls nouveau_scratch_done it does not mean
scratch buffers can be freed immediately. It means "when hardware
advances to this place in the command stream the scratch buffers
can be freed".
To fix it, just postpone scratch runout destruction after current
fence is signalled.
The bug existed for a very long time. Nobody noticed, because
"scratch runout" code path is rarely executed.
Fixes hang at the very beginning of first mission in "Serious Sam 3"
on nve7/gk107. It manifested as:
nouveau E[ PFIFO][0000:01:00.0] read fault at 0x000a9e0000 [PTE] from GR/GPC0/PE_2 on channel 0x007f853000 [Sam3[17056]]
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
This will help encoding VUI into the bitstream
v2: make backward compatible
Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
The framerate will be used for video usability info support by VCE driver
Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Just announce support for 4 components.
While here also increase the max/min texel offsets (the limit is completely
artificial, was chosen because that's what other hardware did, however there's
other drivers using larger limits).
Over a thousand little piglits skip->pass.
v2: update docs/GL3.txt
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This is quite trivial, essentially just follow all the same code you'd
use with linear min/mag (and no mip) filter, then just skip the filtering
after looking up the texels in favor of direct assignment of the right channel
to the result. (This is though not true for the multi-offset version if we'd
want to support it - for this would probably need to do something along the
lines of 4x nearest sampling due to the necessity of doing coord wrapping
individually per texel.)
Supports multi-channel formats.
From the SM5 gather cap bit, should support non-constant offsets, plus shadow
comparisons (the former untested), but not component selection (should be
easy to implement but all this stuff is not really exposable anyway for now).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This has got a bit out of control with more and more parameters added.
Worse, whenever something in there changes all callees have to be updated
for that, even though they don't really do much with any parameter in there
except pass it on to the actual sampling function.
Hence simply put almost everything into a struct. Also instead of relying
on some arguments being NULL, be explicit and set this in a key (which is
just reused for function generation for simplicity). (The code still relies
on them being NULL in the end for now.)
Technically there is a minimal functional change here for shadow sampling:
if shadow sampling is done is now determined explicitly by the texture
function (either sample_c or the gl-style tex func inherit this from target)
instead of the static texture state. These two should always match, however.
Otherwise, it should generate all the same code.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This cleans up more instructions generated by uniform array indexing
multiplies.
total instructions in shared programs: 39989 -> 39961 (-0.07%)
instructions in affected programs: 896 -> 868 (-3.12%)
This cleans up some pointless operations generated by the in-driver mul24
lowering (commonly generated by making a vec4 index for a matrix in a
uniform array).
I could fill in other operations, but pretty much anything else ought to
be getting handled at the NIR level, I think.
total uniforms in shared programs: 13423 -> 13421 (-0.01%)
uniforms in affected programs: 346 -> 344 (-0.58%)
The hardware just uses the low 24 lines, saving us an AND to drop the high
bits.
total uniforms in shared programs: 13433 -> 13423 (-0.07%)
uniforms in affected programs: 356 -> 346 (-2.81%)
total instructions in shared programs: 40003 -> 39989 (-0.03%)
instructions in affected programs: 910 -> 896 (-1.54%)
Fixes a crash in genymotion with several threads compiling shaders
concurrently.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89746
Cc: 10.5 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This does not (yet) support different coordinate origins, so the tests
still fail due to fbo flipping.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This appears to need the A2XX version of the point list, so select it at
draw time if necessary.
Experimentally, always using the A2XX version causes hangs when PSIZE
isn't actually emitted.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
The division is probably a holdover from the days when the fixed point
inline functions generated by headergen were broken.
Also reduce the maximum point size to 4092 (vs 4096), which is what the
blob does.
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
The SZ2 field contains the layer size of a lower miplevel. It only
contains 4 bits, which limits the maximum layer size it can describe. In
situations where the next miplevel would be too big, the hardware
appears to keep minifying the size until it hits one of that size.
Unfortunately the hardware's ideas about sizes can differ from
freedreno's which can still lead to issues. Minimize those by stopping
to minify as soon as possible.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
These functions looked quite complicated, even though what they actually did
was trivial (ever since we dropped swizzled rendering). Also drop lookup of
format block per bytes done for each block, and do it once per scene instead.
This improves everybody's favorite "benchmark" by 3% or so, though
lp_rast_shade_quads_all() which calls this shows up still quite high for a
function which does little more than call the jit function.
(This would most likely be much better handled by the jit function itself,
the strides are passed through anyway already, though for being able to
handle layers it would definitely add some complexity.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
When using the texel fetch functions rather than ordinary texturing,
the arguments are all int vecs instead of float vecs, not to mention
the actual function would look completely different. Hence this must
be included in the texture function name (which serves as the key)
otherwise things crash badly when a shader accesses the same texture
and sampler unit with both txf/ld and ordinary texturing instructions
with otherwise matching keys.
Multiply operations can have a post-factor on them, which other ops
don't support. Only perform the peephole optimizations when there is no
post-factor involved.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89758
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
There are issues with inlining everything, most notably llvm will use much
more memory (and be slower) when compiling. Ideally we'd probably use
functions for shader functions too but texture sampling usually is responsible
for quite some IR (it can easily reach 80% of total IR instructions) so this
seems like a good start.
This still generates a different function for all different combinations just
like before, however it is possible llvm is missing some optimization
opportunities - it is believed though such opportunities should be somewhat
rare, but at least for now it can still be switched off (at compile time only).
It should probably make compiled code also smaller because the same function
should be used for different variants in the same module (so for the
opaque/partial or linear/elts variants).
No piglit change (though it does indeed speed up unrealistic tests like
fp-indirections2 by a factor of 30 or so).
Has a small negative performance impact in openarena - I suspect this could
be fixed by running some IPO passes (despite the private linkage, llvm right
now does NO optimization at all wrt anything going past the call, even if
there's just one caller - so things like values stored before the call and then
always written by the function etc. will not be optimized away, nor will dead
arguments (which we mostly shouldn't have) be eliminated, always constant
arguments promoted etc.).
v2: use proper return values instead of pointer function arguments.
llvm supports aggregate return values, which do wonders here eliminating
unnecessary stack variables - everything in fact will be returned in registers
even without any IPO optimizations. It makes the code simpler too.
With this I could not measure a peformance impact in openarena any longer
(though since there's still no constant value propagation etc. into the tex
functions this does not mean it couldn't have a negative impact elsewhere).
v3: fix some minor issues suggested by Jose, and do disassembly (and the
profiling) without hacks.
Reviewed-by: Jose Fonseca <jfonseca@vmware.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>
The data in memory is in big endian format and needs to be converted
into CPU byte order. So the patch actually reversed what needs to be done.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The CUBE_ARRAY case uses r[4]. Make sure that the stack variable is
there.
Noticed by Coverity.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Does not appear to be used in tree. Coverity spotted some errors in the
bitmask stuff, but the whole thing appears to be unused.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Brian Paul <brianp@vmware.com>
Make use of the builtin ffs macros and split out ffsll
to a seperate block. Needed for at least OpenBSD which
does not have ffsll in libc.
Signed-off-by: Jonathan Gray <jsg@jsg.id.au>
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
16 / cpp happens to be the same as utile_w on the only raster format
supported (4 bytes per pixel), but simulator/hw source code generally
talks in terms of utiles.
I'd like to compile as much of the device-specific code as possible
when building for simulator, and using if (using_simulator) instead of
ifdefs helps.
This reverts commit 20346808cf.
The conversion is actually done since these are the *B macro variants
and no vtx format is supplied, which makes them go through the translate
module.
This restores the following piglit tests to passing:
draw-vertices user
gl-2.0-vertexattribpointer
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
The storage size for local kernel args can be queried before the
arguments are set by using the CL_KERNEL_LOCAL_MEM_SIZE param
of clGetKernelWorkGroupInfo().
The spec says that if local kernel arguments have not been specified,
then we should assume their size is 0.
v2:
- Implement using c++11 member initialization.
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Cc: 10.5 10.4 <mesa-stable@lists.freedesktop.org>
The pipe's get_vendor method returns something more akin to a driver
vendor string in most cases, instead of the actual device vendor. Use
get_device_vendor instead, which was introduced specifically for this
purpose.
Signed-off-by: Giuseppe Bilotta <giuseppe.bilotta@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The only hackish ones are llvmpipe and softpipe, which currently return
the same string as for get_vendor(), while ideally they should return
the CPU vendor.
Signed-off-by: Giuseppe Bilotta <giuseppe.bilotta@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This will be needed by Clover to return the correct information
to CL_DEVICE_VENDOR info queries.
Signed-off-by: Giuseppe Bilotta <giuseppe.bilotta@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Should have been part of 429a4355259(galahad: remove driver). Seems like
I've erroneously committed the trimmed patch.
Reported-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
There's 2 reasons why we'd want to use the global context:
1) There still seems to be one memory "leak" left when using multiple llvm
contexts (it is not a true leak as the memory disappears into some still
addressable pool but nevertheless the memory consumption grows). See
http://cgit.freedesktop.org/~jrfonseca/llvm-jitstress/
2) These contexts get kinda big - even when disposing modules etc. after
compiling a shader the LLVMContext can easily be over 100kB. So when there's
lots of llvm contexts arounds it adds up.
The downside is that at least right now this is absolutely not thread safe,
so this only works safely in environments where multiple pipe contexts are not
used concurrently.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This add primitive restart support to the prim conversion.
This involves changing the API for the translate functions
as we need to pass the prim restart index and the original
number of indices into the translate functions.
primitive restart is support for quads, quad strips
and polygons.
This deal with the case where the actual number of output
primitives is less than the initially calculated number,
by filling the rest of the output primitives with the restart
index, the other option is to reduce the output prim number,
but that will make the generator code a bit messier.
Reviewed-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
One more case we need to handle. One of the src instructions for the
indirect could also end up being ourself.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
radeon_llvm_emit_prepare_cube_coords uses coords[4] in some cases (TXB2 etc.)
Discovered by Coverity. Reported by Ilia Mirkin.
Cc: 10.5 10.4 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Massive list of constant data. Annotate it as such.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Both of which are no longer used. Use designated initializer to make
things obvious as people add/remove TGSI_OPCODEs.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
... rather than the local one in inst_info->tgsi_opcode.
This will allow us to simplify struct r600_shader_tgsi_instruction.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Before this actually ran into an infinite loop printing out "invalid"...
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Squash this silly typo introduced with commit c63eb5dd5ec(auxiliary/os: get
the mmap/munmap wrappers working with android)
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Remove the forward declaration and make use of the DEBUG_PRINT macro for
debug builds.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
v2: Don't use the intrinsics, the shader backend can recognize these
patterns and generates optimal code automatically.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This requires enabling the optional GL provoking vertex behavior for quads.
+ some cosmetic changes, so that the register is set exactly the same as
on r600.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
The fragment shader multiplies the alpha channel with gl_SampleMaskIn.
If blending is enabled, it looks like MSAA.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
This will be used for line and polygon smoothing.
This is GCN-only even though it's in shared code.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
I have to use the BFE instrinsics, because BFE is one of the most complex
instructions that can't be matched easily. BFE has 3 conditional branches
and one of them is quite big.
In the isel DAG, lowered BFE has 27 nodes (including leafs).
Now that piglit is no longer falling back to old compiler for any tests,
we can remove it. Hurray \o/
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Deadlock can occur if we schedule an address register write, yet some
instructions which depend on that address register value also depend on
other unscheduled instructions that depend on a different address
register value. To solve this, before scheduling an address register
write, ensure that all the other dependencies of the instructions which
consume this address register are already scheduled.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Add an array_insert() macro to simplify inserting into dynamically sized
arrays, add a comment, and remove unused prototype inherited from the
original freedreno.git/fdre-a3xx test code, etc.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
For example if width were 65, the first slice would get 96 while the
second would get 32. However the hardware appears to expect the second
pitch to be 64, based on halving the 96 (and aligning up to 32).
This fixes texelFetch piglit tests on a3xx below a certain size. Going
higher they break again, but most likely due to unrelated reasons.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
We only program in one layer size per texture, so that means that all
levels must share one size. This makes the piglit test
bin/texelFetch fs sampler2DArray
have the same breakage as its non-array version instead of being
completely off, and makes
bin/ext_texture_array-gen-mipmap
start passing.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Fix typo in comment introduced by 70dc8a
Signed-off-by: Alexandre Demers <alexandre.f.demers@gmail.com>
Signed-off-by: Jose Fonseca <jfonseca@vmware.com>
std::unique_ptr takes ownership of MM, and a double delete could ensure
in case of an error, as pointed out by Chris Vine in
https://bugs.freedesktop.org/show_bug.cgi?id=89387
Reviewed-by: Chris Vine <chris@cvine.freeserve.co.uk>
The maximum value of a Gallium HUD's panel is automatically adjusted
when the current value is greater than the max. If we set the
pipe_query_driver_info::max_value to UINT64_MAX, the maximum value is
never adjusted and this results in a flat line instead of a pretty curve
which is correctly scaled.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Lets define R600_MAX_VIEWPORTS instead of using 16 here and there
in the code when looping through viewports and scissors. It is
easier to understand what this number represents.
v2: Missed a case where R600_MAX_VIEWPORTS should have been used.
Signed-off-by: Alexandre Demers <alexandre.f.demers@gmail.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
This fixes the GL_COMPRESSED_RED_RGTC1 part of piglit's rgtc-teximage-01
test as well as the precision part of Wine's 3dc format test (fd.o bug
89156).
The Z component seems to contain a lower precision version of the
result, probably a temporary value from the decompression computation.
The Y and W component contain different data that depends on the input
values as well, but I could not make sense of them (Not that I tried
very hard).
GL_COMPRESSED_SIGNED_RED_RGTC1 still seems to have precision problems in
piglit, and both formats are affected by a compiler bug if they're
sampled by the shader with a swizzle other than .xyzw. Wine uses .xxxx,
which returns random garbage.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89156
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Cc: 10.5 10.4 <mesa-stable@lists.freedesktop.org>
This means dropping CL_FP_DENORM from the current return value.
v2:
- Add comments about minimum values for OpenCL 1.2.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
This fixes ARB_texture_query_levels to actually return the desired
value.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Fixes: 1f3ca56b ("freedreno: use util_copy_framebuffer_state()")
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Cc: "10.4 10.5" <mesa-stable@lists.freedesktop.org>
Now that relative-dst works, we should never fall back to the old
compiler. (Which is almost true, other than a couple edge case sched
fails in piglit).
So replace glsl130 flag to force GLSL 130 and integers on a3xx/a4xx with
a glsl120 flag to force GLSL 120 and !integers.
If this commit breaks any game/app/etc use FD_MESA_DEBUG=glsl120 as a
workaround and please let me know.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Enable the 'sphinx.ext.graphviz' extension, and add in a section for
driver specific docs, with freedreno compiler docs beneath. The
goal is for more complete compiler docs, and hopefully some docs about
other parts of the driver (such as how tiling works, etc).
Note that there is also a Distribution -> Drivers section. Although
that appears to be simply just a list of drivers. Not sure if that
should move under the 'Drivers' section or left alone. I did add a
one-line section for freedreno in the existing Distribution -> Drivers
section.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Eric Anholt
Rob Clark
Nick Sarnie
Vinson Lee
Roland Scheidegger
Dave Airlie
Glenn Kennard
Ilia Mirkin
Stéphane Marchesin
Emil Velikov
Jose Fonseca
Marcin Ślusarz
Tom Stellard
Leo Liu
Michel Dänzer
Jan Vesely
Christian König
Jonathan Gray
Giuseppe Bilotta
Dave Airlie
Brian Paul
Marek Olšák
Alexandre Demers
Samuel Pitoiset
Stefan Dösinger