We have two major requirements: Make sure that only the bottom half of the
physical reg space is used, and make sure that none of our values are live
in an accumulator across a switch.
We have had no reason to separate ability to store in an accumulator from
ability to store in B, but with FS threading, we need to be able to force
values to be stored only in the physical regfiles.
This will eventually be generated at the QIR level, so that
vc4_qir_schedule.c can arrange the separation of tex_strb from tex_result
correctly. It will also be important so that register allocation set the
register classes appropriately for values that are live across the switch.
Fixes fallout from acc23b04 ("ralloc: remove memset from ralloc_size").
We were still depending on zero'd allocations in a couple of places.
Signed-off-by: Rob Clark <robdclark@gmail.com>
This adds some performance counters/metrics for SM50/SM52.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
I'm leaving num_out_sgpr around since it's not in a fast path, and besides
the compiler should be able to optimize it away easily. The alternative
with #if/#endif would be extremely ugly.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The fix in commit 88f791db75 was insufficient
for radeonsi because the vector case was not handled properly. It seems
piglit only covers the scalar case, unfortunately.
Fixes GL45-CTS.shader_bitfield_operation.[iu]mulExtended.*
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This setting is for whether color and alpha have different blend
settings, not for whether blending is enabled on a per-RT basis.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
With ARB_clip_control, GL may also do 0..1 depth clipping, not just
-1..1. This removes clip's reliance on driver type. DX users will need
to be updated to set the new clipHalfZ flag to get proper clipping
functionality.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
Most logic op usage is probably going to end up with normalized
textures. Scale the floating point values and convert to integer before
performing the logic operations.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tim Rowley <timothy.o.rowley@intel.com>
Jonas Pfeil noticed that we were putting passthrough tlb_z writes early in
the shader, despite QIR and QPU scheduling both trying to delay scoreboard
locking for as long as possible.
The problem was that when trying to pair up QPU instructions, at some
point the passthrough tlb_z would be the last one available and it would
get paired, even if the other half would open up other instructions to be
scheduled and we could have paired tlb_z with something later in the
program. Also, since passthrough z is just a mov, it pairs up really
easily.
The proper fix would probably be to flip the order of scheduling
instructions so we went from bottom to top (also relevant for branch delay
slot scheduling).
However, we can do a quick fix here to just not schedule a TLB lock until
there's nothing but TLB left in the program, at a slight instruction cost
(est .61% cycle count in shader-db) but a major fragment shader
parallelism win.
glmark2 results:
texture:texture-filter=linear: +1.24481% +/- 0.626117% (n=15)
bump:bump-render=height: 1.24991% +/- 0.154793% (n=136,133 -- screensaver
outliers removed)
It's much better to just skip the draw call entirely. Getting this
information out of register allocation will also be useful for
implementing threaded fragment shaders, which will need to retry
non-threaded if RA fails.
Cc: <mesa-stable@lists.freedesktop.org>
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>
Trivial. There's some regressions internally, related to overflow
behavior. I'll have to look at it at another time, some interactions
with vsplit/vcache are actually mind-blowing.
This reverts commit 3fa10ffb49.
The first IF statement disabled the second one.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98599
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This patch does two things:
1. It separates the host-CPU code generation from the generic code
generation. This guards against accidently breaking things for
radeonsi in the future.
2. It makes sure we actually use both arguments and don't just compute
a square :-p
Fixes a regression introduced by commit 29279f44b3
Cc: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Instead of doing all the math with scalars, use vectors. This means the
overflow math needs to be done manually, albeit that's only really
problematic for the stride/index mul, the rest has been pretty much
moved outside the shader loop (albeit the mul could actually be optimized
away too), where things are still scalar. Because llvm is complete fail
with the zero-extend widening mul, roll our own even...
To eliminate control flow in the main shader loop fetch, provide fake
buffers (so index 0 is always valid to fetch).
Still uses aos fetch though in the end - mostly because some more code
would be needed to handle unaligned fetches in that path, and because for
most formats it won't make a difference anyway (we generate some truly
horrendous code for things like R16G16_something for instance).
Instanced fetch however stays roughly the same as before, except that
no longer the same element is fetched multiple times (I've seen a reduction
of ~3 times in main shader loop size due to apparently llvm not being able
to deduce it's really all the same with a couple instanced elements).
Also, for elts gathering, use vectorized code as well - provide a fake
elt buffer if there's no valid one bound.
The generated shaders are smaller and faster to compile (not entirely sure
about execution speed, but generally unless there's just single vertices
to handle I would expect it to be faster - there's more opportunities
for future improvements by using soa fetch).
No piglit change.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This is used by shader umul_hi/imul_hi functions (and soon by draw).
It's actually useful separating this out on its own, however the real
reason for doing it is because we're using an optimized sse2 version,
since the code llvm generates is atrocious (since there's no widening
mul in llvm, and it does not recognize the widening mul pattern, so
it generates code for real 64x64->64bit mul, which the cpu can't do
natively, in contrast to 32x32->64bit mul which it could do).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
In the event that multiple threads attempt to install a graph
concurrently, protect the shared list.
Signed-off-by: Steven Toth <stoth@kernellabs.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
We're missing the closedir() to the matching opendir().
Signed-off-by: Steven Toth <stoth@kernellabs.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Instead of trying to maintain a reference counted list of valid HUD
objects, and freeing them accordingly, creating race conditions
between unanticipated multiple threads, simply accept they're
allocated once and never released until the process terminates.
They're a shared resource between multiple threads, so accept
they're always available for use.
Signed-off-by: Steven Toth <stoth@kernellabs.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
v3 [Francisco Jerez]: Loosely based on Serge's v1 of this patch in
order to avoid CL-specific enums in the clover module binary
format. In addition to other changes made in v2: Represent the CL
program binary type as the section type instead of adding a CL
API-specific enum, check that the binary types of the input objects
are valid during clLinkProgram(), pass section type as argument to
build_module_library() instead of using separate function.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Instead, hardcode group sigsel because there are a bunch of unknown
groups, especially on SM50/SM52.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This is similar to NVC0 and GK110 emitters where we emit
reduction operations instead of atomic operations when the
destination is not used.
Found after writing some tests which check if performance counters
return the expected value. In that case, gred_count returned 0
on gm107 while at least gk106 returned the correct value.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
