| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
| |
Instead of changing the format on the existing template
which makes error handling not nice and confuses coverity.
CoverityID: 1337953
Signed-off-by: Julien Isorce <[email protected]>
Reviewed-by: Christian König <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
pipe_resource_reference(&res, NULL) will decrement reference counting,
i.e. p_atomic_dec(res->count). But the va surface still has the initial
reference since it has created the resource. So calling vaDestroyImage
on a derived image calls VaDestroyBuffer but the decrementation won't
reach 0. It is just wrong for vlVaDestroyBuffer to rely on the
export_refcount flag. Finally the vaapi intel driver has the same logic.
Signed-off-by: Julien Isorce <[email protected]>
Reviewed-by: Christian König <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
This change makes sure to remove arrays when checking if type
is a double.
The check for the end of the first slot of a multi-slot double
is also fixed by bumping the check to 4 rather than 3.
Previously we were we not reserving the last component.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
Since this extension allows more than one varying to share a single
location we can't just count the number of slots a varying takes and
add it to the total.
Instead we now reuse the reserved varyings bitfield to determine how
many slots are reserved for explicit locations instead.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We were programming the number of threads per subslice, when we should
have been programming the total number of threads on the GPU as a whole.
Thanks to Curro and Jordan for helping track this down!
On Skylake GT3e:
- Improves performance in Unreal's Elemental Demo by roughly 1.5-1.7x.
- Improves performance in Synmark's Gl43CSDof by roughly 3.7x.
- Improves performance in Synmark's Gl43GSCloth by roughly 1.18x.
On Broadwell GT2:
- Improves performance in Unreal's Elemental Demo by roughly 1.2-1.5x.
- Improves performance in Synmark's Gl43CSDof by roughly 2.0x.
- Improves performance in Synmark's Gl43GSCloth by 1.47035% +/-
0.255654% (n=25).
On Haswell GT3e:
- Improves performance in Unreal's Elemental Demo (in GL 4.3 mode)
by roughly 1.10x.
- Improves performance in Synmark's Gl43CSDof by roughly 1.18x.
- Decreases performance in Synmark's Gl43CSCloth by -1.99484% +/-
0.432771% (n=64).
On Ivybridge GT2:
- Improves performance in Unreal's Elemental Demo (in GL 4.2 mode)
by roughly 1.03x.
- Improves performance in Synmark's G/43CSDof by roughly 1.25x.
- No change in Synmark's Gl43CSCloth (n=28).
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I don't know that anything actually guarantees this, but if we exceed
the limits, we may end up overflowing and trashing random buffers that
happen to be nearby in the VMA space, leading to rendering corruption,
hangs, or worse.
We should really fix this properly. However, the pitfall has existed
for ages, so for now we should at least detect it.
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
| |
These are linear, not powers of two, and much more limited.
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Most scratch stages use power of two sizes, in kilobytes, where
0 means 1kB. But compute shaders on Haswell have a minimum of 2kB,
and use a representation where 0 = 2kB.
This meant that we were effectively telling the hardware to allocate
each thread twice as much space as we meant to, while simultaneously
not allocating that much space in the buffer, leading to overflows.
Note that the existing code is completely wrong for Ivybridge,
but that will take additional work to sort out, so I've left it
as is for now. A subsequent commit will take care of that.
Together with the previous patches, this fixes rendering corruption
on Synmark's Gl43CSDof on Haswell.
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Curro figured this out by investigating the simulator. Apparently
there's also a workaround in the Windows driver. I'm not sure it's
actually documented anywhere.
We were underallocating the scratch buffer by a factor of 128/70.
v2: Rename threads_per_subslice to scratch_ids_per_subslice
(suggested by Jordan Justen).
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We were allocating enough space for the number of threads per subslice,
when we should have been allocating space for the number of threads in
the entire GPU.
Even though we currently run with a reduced thread count (due to a bug),
we might still overflow the scratch buffer because the address
calculation is based on the FFTID, which can depend on exactly which
threads, EUs, and threads are executing. We need to allocate enough
for every possible thread that could run.
Fixes rendering corruption in Synmark's Gl43CSDof on Gen8+.
Earlier platforms need additional bug fixes.
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
We'll use this for compute shader thread counts and scratch space
calculations shortly.
Note that subslices are referred to as "half slices" on Ivybridge.
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Skylake changes the representation of shared local memory size:
Size | 0 kB | 1 kB | 2 kB | 4 kB | 8 kB | 16 kB | 32 kB | 64 kB |
-------------------------------------------------------------------
Gen7-8 | 0 | none | none | 1 | 2 | 4 | 8 | 16 |
-------------------------------------------------------------------
Gen9+ | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
The old formula would substantially underallocate the amount of space.
This fixes GPU hangs on Skylake when running with full thread counts.
v2: Fix the Vulkan driver too, use a helper function, and fix the table
in the comments and commit message.
Cc: "12.0" <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This was disabled due to occasionally incorrect behavior when trying to
upload data. It later became apparent that nvc0 also had a similar but
slightly different issue, which was resolved in commit e50c01d5. This
takes the same logic as nvc0 and applies it to nv50 (which has somewhat
different interfaces).
Unfortunately I did not note down precisely what was broken with UBOs
when removing the support from nv50, but I've tested a bunch of local
traces, and none of them appear to regress. This should hopefully
improve performance when UBOs are used, but this was not directly
verified.
Signed-off-by: Ilia Mirkin <[email protected]>
Reviewed-by: Samuel Pitoiset <[email protected]>
|
|
|
|
| |
Signed-off-by: Ilia Mirkin <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
This was fixed in revision 47 of the ARB_dsa spec in Oct 22, 2015. Since
it's horrible to have differing APIs across library versions, we should
attempt to minimize the impact by backporting it as far as possible and
hope no one notices.
Signed-off-by: Ilia Mirkin <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
Cc: "11.2 12.0" <[email protected]>
|
|
|
|
|
|
|
|
| |
This brings in the fixed glClearNamedFramebufferfi definition, as well
as a lot of GLsizei -> GLsizeiptr changes.
Signed-off-by: Ilia Mirkin <[email protected]>
Cc: "11.2 12.0" <[email protected]>
|
|
|
|
|
|
|
|
| |
This brings in defines from GL_EXT_window_rectangles and fixes the
glClearNamedFramebufferfi definition.
Signed-off-by: Ilia Mirkin <[email protected]>
Cc: "11.2 12.0" <[email protected]>
|
|
|
|
| |
Signed-off-by: Brian Paul <[email protected]>
|
|
|
|
| |
Reviewed-by: Charmaine Lee <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
This basically disallows all 8-bit x 3 and 16-bit x 3 formats for
textures and render targets. Some 3-component formats were already
disallowed before. This avoids problems with GL_ARB_copy_image.
v2: the previous version of this patch disallowed all 3-component formats
Reviewed-by: Charmaine Lee <[email protected]>
Reviewed-by: Roland Scheidegger <[email protected]>
|
|
|
|
| |
Reviewed-by: Charmaine Lee <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Mesa and gallium don't have a complete set of matching 3-component
texture formats. For example, 8-bit sRGB unorm. To fully support
the GL_ARB_copy_image extension we need to have support for all of
these formats: RGB8_UNORM, RGB8_SNORM, RGB8_SRGB, RGB8_UINT, and
RGB8_SINT using the same component order. Since we don't have that,
disable the 3-component formats for now.
v2: Simplify 3-component format check, per Marek.
Also check that target != PIPE_BUFFER.
Reviewed-by: Marek Olšák <[email protected]>
Reviewed-by: Charmaine Lee <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
1. Try to choose R8G8B8A8 unorm/srgb formats before others in an
effort to try to match component ordering for UINT/SINT/etc.
2. If we can't get a format such as PIPE_FORMAT_A16_UNORM, try
PIPE_FORMAT_R16G16B16A16_UNORM before shallower formats.
Reviewed-by: Marek Olšák <[email protected]>
Reviewed-by: Charmaine Lee <[email protected]>
|
|
|
|
|
|
|
| |
This primarily means added support for copying between compressed
and uncompressed formats.
Reviewed-by: Charmaine Lee <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
>From OpenGL 4.0 spec, section 4.3.2 "Copying Pixels":
"The pixels corresponding to these buffers are copied from the source
rectangle bounded by the locations (srcX0, srcY 0) and (srcX1, srcY 1)
to the destination rectangle bounded by the locations (dstX0, dstY 0)
and (dstX1, dstY 1). The lower bounds of the rectangle are inclusive,
while the upper bounds are exclusive."
So, the rectangles sharing just an edge shouldn't overlap.
-----------
| |
------- ---
| | |
| | |
------- ---
Cc: "12.0" <[email protected]>
Signed-off-by: Anuj Phogat <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
>From OpenGL 4.0 spec, section 4.3.2 "Copying Pixels":
"The pixels corresponding to these buffers are copied from the source
rectangle bounded by the locations (srcX0, srcY 0) and (srcX1, srcY 1)
to the destination rectangle bounded by the locations (dstX0, dstY 0)
and (dstX1, dstY 1). The lower bounds of the rectangle are inclusive,
while the upper bounds are exclusive."
So, the rectangles sharing just an edge shouldn't overlap.
-----------
| |
------- ---
| | |
| | |
------- ---
Cc: "12.0" <[email protected]>
Signed-off-by: Anuj Phogat <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
|
|
|
|
|
|
|
| |
This converts one other place to using the new helper.
Reviewed-by: Nicolai Hähnle <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
|
|
|
|
|
|
|
|
| |
This converts to testing for 64-bit types and renames some things
in anticipation of 64-bit integer support.
Reviewed-by: Nicolai Hähnle <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
|
|
|
|
|
|
|
|
| |
This just uses the same form across the fetches.
Reviewed-by: Roland Scheidegger <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This just makes some generic code that currently emits double
suitable for emitting 64-bit values.
Reviewed-by: Roland Scheidegger <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
|
|
|
|
|
|
|
|
|
| |
Currently this just doubles, but we'll convert users to this
so making adding 64-bit integers easier.
Reviewed-by: Roland Scheidegger <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reworks the #if guards a bit. When Emil originally wrote them, he
just guarded everything. However, part of what anv_entrypoints_gen.py
generates is a hash table for looking up entrypoints based on their name.
This table *cannot* get out of sync between C and python regardless of
preprocessor flags. In order to prevent this, this commit makes us use
void pointers in the dispatch table for those entrypoints which aren't
available. This means that the dispatch table size and entry order is
constant and it should never get out-of-sync with the python.
Signed-off-by: Jason Ekstrand <[email protected]>
Acked-by: Emil Velikov <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This is a bit cleaner than generating the types ourselves when making the
table.
Signed-off-by: Jason Ekstrand <[email protected]>
Acked-by: Emil Velikov <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When an applications specifies mip levels _before_ setting a mipmap texture
filter, we will initially guess a single texture level. When the second level
image is created, we try to allocate the full texture -- however, we get the
base level size guess wrong if that size is odd. This leads to yet another
re-allocation of the texture later during st_finalize_texture.
Even worse, this re-allocation breaks a (reasonable) assumption made by
st_generate_mipmaps, because the re-allocation in the finalization call will
again allocate a single-level pipe texture (based on the non-mipmap texture
filter!). As a result, mipmap generation fails in interesting ways.
All of this can be avoided by just using the fact that we already know the
size of the base level.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95529
Cc: 12.0 <[email protected]>
Reviewed-by: Brian Paul <[email protected]>
|
|
|
|
|
|
|
|
| |
At this point, the limits are probably more-or-less correct. If there is
an invalid limit, that's a bug not a FINSHME.
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
| |
This gets ANV_ENABLE_PIPELINE_CACHE=false working again.
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
This way the the bind map (which we're caching) is mostly independent of
the pipeline layout. The only coupling remaining is that we pull the array
size of a binding out of the layout. However, that size is also specified
in the shader and should always match so it's not really coupled. This
rendering issues in Dota 2.
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
Since applications are allowed to specify some set of bindings which need
not be dense they also need not be in order. For most things, this doesn't
matter, but it could result getting the wrong dynamic offsets. This adds a
quick-and-dirty sort to ensure that everything is always in increasing
order of binding index.
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
| |
This allows for some extra validation and makes it easier to see what's
going on when poking around in gdb.
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
| |
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
With glx of gstreamer-vaapi, the temporary pixmap for front buffer gets
renewed in each frame, so when we receive a new pixmap, should get a new
front buffer for it.
This also fixes Totem player playback corruption.
Signed-off-by: Leo Liu <[email protected]>
Reviewed-by: Michel Dänzer <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
From original commit, the macro "if HAVE_DRI3" was in Makefile.sources,
this file is shared with SCons, SCons is not able to parse this marco,
the SCons build failed. Jose quickly gave two approaches and quick fix
with his second approach, thanks Jose for the solutions and fixes.
This patch is Jose's first approach, and it's more proper, because the
dri3 c file should not be included to build when DRI3 is not enabled.
Signed-off-by: Leo Liu <[email protected]>
Acked-by: Emil Velikov <[email protected]>
Cc: "12.0" <[email protected]>
|
|
|
|
|
|
|
|
| |
Besides the old JIT bug, it seems the X86 backend on LLVM 3.3 doesn't
handle llvm.fmuladd and instead it fall backs to a C function. Which in
turn causes a segfault on Windows.
Reviewed-by: Roland Scheidegger <[email protected]>
|
|
|
|
| |
Reviewed-by: Roland Scheidegger <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
As suggested by Roland Scheidegger.
Use the same logic as f16c, since fma requires VEX encoding.
But disable FMA on LLVM 3.3 without MCJIT.
Reviewed-by: Roland Scheidegger <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This fixes a problem with the CE preamble and restoring only stuff in the
preamble when needed.
To illustrate suppose we have two graphics IB's 1 and 2, which are submitted in
that order. Furthermore suppose IB 1 does not use CE ram, but IB 2 does, and we
have a context switch at the start of IB 1, but not between IB 1 and IB 2.
The old code put the CE RAM loads in the preamble of IB 2. As the preamble of
IB 1 does not have the loads and the preamble of IB 2 does not get executed, the
old values are not load into CE RAM.
Fix this by always restoring the entire CE RAM.
v2: - Just load all descriptor set buffers instead of load and store the entire
CE RAM.
- Leave the ce_ram_dirty tracking in place for the non-preamble case.
v3: - Fixed parameter alignment.
- Rebased to master (Nicolai's descriptor series).
Signed-off-by: Bas Nieuwenhuizen <[email protected]>
Reviewed-by: Nicolai Hähnle <[email protected]>
|
|
|
|
|
|
|
| |
This should fix the MSVC linker failures that arose with commit
5e2d25894b962aae9158261897e13843377e3b95.
Trivial.
|
|
|
|
|
|
|
|
|
|
|
| |
The reality is that this doesn't matter, because we manually emit the
ARL to the sampler reladdr, and those arguments don't get an extra load
later, so it's effectively just a boolean. However having the types be
wrong is confusing and could trigger very odd bugs should usage change
down the line.
Signed-off-by: Ilia Mirkin <[email protected]>
Reviewed-by: Brian Paul <[email protected]>
|
|
|
|
|
|
|
|
| |
Adding 64-bit integers support was going to make this file worse,
just remove the tabs from it now.
Acked-by: Timothy Arceri <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
|
|
|
|
|
|
|
|
|
| |
An update in graphics specs has deleted the halign and valign fields
from XY_FAST_COPY_BLT command. See mesa commit 97f0f91.
Cc: Ben Widawsky <[email protected]>
Signed-off-by: Anuj Phogat <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|