This is basically a re-write of the slab allocator into a design where
multiple child pools are linked to a parent pool. The intention is that
every (GL, pipe) context has its own child pool, while the corresponding
parent pool is held by the winsys or screen, or possibly the GL share group.
The fast path is still used when objects are freed by the same child pool
that allocated them. However, it is now also possible to free an object in a
different pool, as long as they belong to the same parent. Objects also
survive the destruction of the (child) pool from which they were allocated.
The slow path will return freed objects to the child pool from which they
were originally allocated. If that child pool was destroyed, the corresponding
page is considered an orphan and will be freed once all objects in it have
been freed.
This allocation pattern is required for pipe_transfers that correspond to
(GL) buffer object mappings when the mapping is created in one context
which is later destroyed while other contexts of the same share group live
on -- see the bug report referenced below.
Note that individual drivers do need to migrate to the new interface in
order to benefit and fix the bug.
v2: use singly-linked lists everywhere
v3: use p_atomic_set for page->u.num_remaining
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97894
This is motivated by the fact that p_atomic_read and p_atomic_set may
somewhat surprisingly not do the right thing in the old version: while
stores and loads are de facto atomic at least on x86, the compiler may
apply re-ordering and speculation quite liberally. Basically, the old
version uses the "relaxed" memory ordering.
The new ordering always uses acquire/release ordering. This is the
strongest possible memory ordering that doesn't require additional
fence instructions on x86. (And the only stronger ordering is
"sequentially consistent", which is usually more than you need anyway.)
I would feel more comfortable if p_atomic_set/read in the old
implementation were at least using volatile loads and stores, but I
don't see a way to get there without typeof (which we cannot use here
since the code is compiled with -std=c99).
Eventually, we should really just move to something that is based on
the atomics in C11 / C++11.
Acked-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
This clears the last bits of the usecases of the hash table
located in mesa/program, allowing us to remove it.
V2: Rebase on top of changes to Makefile.sources
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
It is included through the util/hash_table include in
the program hash_table, so this should be safe.
This will be needed when we start converting each use of
the program_hash_table, as some places need this function.
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
This just ports the simpler endian detection bits, addrlib
sharing wants this outside gallium.
Acked-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
GL_EXT_packed_float, 2.1.B Unsigned 10-Bit Floating-Point Numbers:
0.0, if E == 0 and M == 0,
2^-14 * (M / 32), if E == 0 and M != 0,
2^(E-15) * (1 + M/32), if 0 < E < 31,
INF, if E == 31 and M == 0, or
NaN, if E == 31 and M != 0,
In the second case (E == 0 and M != 0), we were multiplying the mantissa
by 2^-20, when we should have been multiplying by 2^-19 (which is
2^(-14 + -5), or 2^-14 * 2^-5, or 2^-14 / 32).
The previous section defines the formula for 11-bit numbers, which is:
2^-14 * (M / 64), if E == 0 and M != 0,
In other words, we had accidentally copy and pasted the 11-bit code
to the 10-bit case, and neglected to change the exponent.
Fixes dEQP-GLES3.functional.pbo.renderbuffer.r11f_g11f_b10f_triangles
when run with surface dimensions of 1536x1152 or 1920x1080.
Cc: mesa-stable@lists.freedesktop.org
References: https://code.google.com/p/chrome-os-partner/issues/detail?id=56244
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Stephane Marchesin <stephane.marchesin@gmail.com>
Reviewed-by: Antia Puentes <apuentes@igalia.com>
The lrint() and lrintf() functions are pretty slow and make some
texture transfers very inefficient. This patch makes a better effort
at using those intrisics for 32-bit gcc and MSVC.
Note, this patch doesn't address the use of SSE4.1 with MSVC.
v2: get rid of the ROUND_WITH_SSE symbol, per Matt.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
v2: Split into two patches.
v3: Fix off by one problem.
Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de>
Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
As requested with the initial creation of util/bitscan.h
now move other bitscan related functions into util.
v2: Split into two patches.
Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de>
Tested-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Both the rgb9e5 and r11g11b10 formats are defined based on how they are
packed into a 32-bit integer. It makes sense that the functions that
manipulate them take an explicitly sized type.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
There are a number of reasons for this refactor. First, format_rgb9e5.h is
not something that a user would expect to define such a generic union.
Second, defining it requires checking for endianness which is ugly. Third,
90% of what we were doing with the union was float <-> uint32_t bitcasts
and the remaining 10% can be done with a sinmple left-shift by 23.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The rgb9e5 format is a packed format defined in terms of slicing up a
single 32-bit value. The bitfields are far more confusing than simple
shifts and require that we check the endianness.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Before, it would happily copy list_head next/prev (ie. pointer to the
*from* list_head), leaving things in a confused state and causing much
mayhem.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The functions are also useful for mesa.
Introduce src/util/bitscan.{h,c}. Move ffs function
implementations from src/mesa/main/imports.{h,c}.
Move bit scan related functions from
src/gallium/auxiliary/util/u_math.h. Merge platform
handling with what is available from within mesa.
v2: Try to fix MSVC compile.
Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de>
Juha-Pekka found this back in May 2015:
<1430915727-28677-1-git-send-email-juhapekka.heikkila@gmail.com>
From the discussion, obviously it would be preferable to make
ralloc_size no longer return zeroed memory, but Juha-Pekka found that
it would break Mesa.
In <56AF1C57.2030904@gmail.com>, Juha-Pekka mentioned that patches
exist to fix i965 when ralloc_size is fixed to not zero memory, but
the patches have not made their way to mesa-dev yet.
For now, let's stop doing the double zeroing of rzalloc buffers.
v2:
* Move ralloc_size code to rzalloc_size, and add a comment as
suggested by Ken.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In optimized builds, visit(ir_expression *) experiences inlining with gcc that
leads the function to have a roughly 32KB stack frame. This is a problem given
that the function is called recursively. In non-optimized builds, the stack
frame is much smaller, hence one gets crashes that happen only in optimized
builds.
Arguably there is a compiler bug or at least severe misfeature here. In any
case, the easy thing to do for now seems to be moving the bulk of the
non-recursive code into a separate function. This is sufficient to convince my
version of gcc not to blow up the stack frame of the recursive part. Just to be
sure, add the gcc-specific noinline attribute to prevent this bug from
reoccuring if inliner heuristics change.
v2: put ATTRIBUTE_NOINLINE into macros.h
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95133
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95026
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92850
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robdclark@gmail.com>
These were being defined in SCons, but it's not practical:
- we actually need to include Gallium headers from external source trees, with
completely disjoint build infrastructure, and it's unsustainable to
replicate the HAVE_xxx checks or even hard-coded defines across
everywhere.
- checking compiler version via command line doesn't really work due to
Clang essentially being like a cameleon which can fake either GCC or
MSVC
There's no change for autoconf.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
This is mostly for variables that are only used in asserts and cause
unused-but-set-variable warnings in release builds. Could just use
UNUSED directly, but MAYBE_UNUSED should be less confusing and is
similar to what the Linux kernel has.
And yes __attribute__((unused)) can be used on variables on both GCC 4.2
(oldest supported by mesa) and clang 3.0 (just some random old version,
not sure what's the minimum for mesa).
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Chad Versace <chad.versace@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
This adds functions for splicing one list into another. These have
more-or-less the same API as the kernel list splicing functions. The
implementation, however, was stolen from the Wayland list implementation.
Reviewed-by: Mark Janes <mark.a.janes@intel.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Fixes intermittent Vulkan CTS failures within the test groups:
dEQP-VK.api.object_management.multithreaded_per_thread_device
dEQP-VK.api.object_management.multithreaded_per_thread_resources
dEQP-VK.api.object_management.multithreaded_shared_resources
Signed-off-by: Mark Janes <mark.a.janes@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94904
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Except:
- u_cache_test -- too long
- translate_test -- unreliable (it's probably testing corner cases that
translate module doesn't care about.)
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
src/util/hash_table.h:111:23: warning: ‘_mesa_fnv32_1a_offset_bias’ defined but not used [-Wunused-const-variable]
static const uint32_t _mesa_fnv32_1a_offset_bias = 2166136261u;
^~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
We no longer need to build any part of Mesa with Windows SDK 7.0.7600 or
MSVC 2008. MSVC 2013 will be the oldest we support.
In practice this means people are now free to declare variables in the
middle of blocks, on the whole Mesa tree.
Care should still be taken with variable length arrays and void pointer
arithmetic.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Hella-acked-by: Ian Romanick <ian.d.romanick@intel.com>
The equivalent of the last patch for the hash table. I'm not aware of
any issues this fixes.
v2:
- use entry_is_deleted (Timothy)
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
When we delete entries in the hash set, we mark them "deleted" by
setting their key to the deleted_key, which points to a dummy
deleted_key_value. When searching for an entry, we normally skip over
those, but set_add() had some code for searching for duplicate entries
which forgot to skip over deleted entries. This led to a segfault inside
the NIR vectorization pass, since its key comparison function
interpreted the memory where deleted_key_value resides as a pointer and
tried to dereference it.
v2:
- add better commit message (Timothy)
- use entry_is_deleted (Timothy)
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
I was cleverly using one iteration to obtain a pointer to the last item
in ralloc's singly list child list, while also setting parents.
Unfortunately, I forgot to set the parent on that last item.
Cc: "11.1 11.0 10.6" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Undefining the NDEBUG is relevant for release build, as they are the
ones that set it.
[Emil Velikov: split from previous patch]
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Kind of a handy function. And I'll want it available outside of i965
for common nir-pass helpers.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Nicolai Hähnle <nhaehnle@gmail.com>
