isl has a giant comment that explains the hardware's padding
requirements. (Hint: Cache lines and page faults). But the comment is in
the wrong place, in isl_calc_linear_row_pitch(), which is unrelated to
padding.
The important parts of that comment were copied to
isl_apply_surface_padding() long ago. So drop the misplaced comment.
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
All the plumbing is in place so the extension just needs to be
advertised.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This doesn't really "do" anything because the default tiling for the
winsys buffer is X tiled. We do however want the X tiled modifier to
work correctly from the API perspective, which would imply that if you
set this modifier, and later do a get_modifier, you get back at least X
tiled.
Running with a modified kmscube, here are the bandwidth measurements.
Linear:
Read bandwidth: 1039.31 MiB/s
Write bandwidth: 1453.56 MiB/s
Y-tiled:
Read bandwidth: 458.29 MiB/s
Write bandwidth: 542.12 MiB/s
X-tiled:
Read bandwidth: 575.01 MiB/s
Write bandwidth: 606.25 MiB/s
Cc: Kristian Høgsberg <krh@bitplanet.net>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Acked-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This patch begins introducing how we'll actually handle the potentially
many modifiers coming in from the API, how we'll store them, and the
structure in the code to support it.
Prior to this patch, the Y-tiled modifier would be entirely ignored. It
shouldn't actually be used until this point because we've not bumped the
DRIimage extension version (which is a requirement to use modifiers).
Measuring later in the series with kmscube:
Linear:
Read bandwidth: 1048.44 MiB/s
Write bandwidth: 1483.17 MiB/s
Y-tiled:
Read bandwidth: 471.13 MiB/s
Write bandwidth: 589.10 MiB/s
Similar functionality was introduced and then reverted here:
commit 6a0d036483
Author: Ben Widawsky <ben@bwidawsk.net>
Date: Thu Apr 21 20:14:58 2016 -0700
i965: Always use Y-tiled buffers on SKL+
v2: Use last set bit instead of first set bit in modifiers to address
bug found by Daniel Stone.
v3: Use the new priority modifier selection thing. This nullifies the
bug fixed by v2 also.
v4: Get rid of modifier compaction which originally served another
purpose and now serves none (Jason)
Signed-off-by: Ben Widawsky <benjamin.widawsky@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Acked-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
At image creation create a path for dealing with the linear modifier.
This works exactly like the old usage flags where __DRI_IMAGE_USE_LINEAR
was specified.
During development of this patch series, it was decided that a lack of
modifier was an insufficient way to express the required modifiers. As a
result, 0 was repurposed to mean a modifier for a LINEAR layout.
NOTE: This patch was added for v3 of the patch series.
v2: Rework the algorithm for modifier selection to go from a bitmask
based selection to this priority value.
v3: Make DRM_FORMAT_MOD_INVALID allowed at selection as a way of
identifying no modifiers found (because 0 is LINEAR) (Jason)
v4: Remove the logic to prune unknown modifiers (like those from other
vendors) and simply handle is in select_best_modifier (Jason)
Requested-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
New to the patch series after reordering things for landing smaller
chunks.
This will essentially enable modifiers from clients that were just
enabled in previous patches. A client could use the modifiers by
setting all of them at create, but had no way to actually query them
after creating the surface (ie. stupid clients could be broken before
this patch, but in more ways than this).
Obviously, there are no modifiers being actually stored yet - so this
patch shouldn't do anything other than allow the API to get back 0 (or
the LINEAR modifier).
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
I intend to need to get to the devinfo structure, and storing the screen
is an easy way to do that.
It seems to be the consensus that you cannot share an image between
multiple screens.
Scape-goat: Rob Clark <robdclark@gmail.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Acked-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Recent glibc generates this warning:
brw_performance_query.c:1648:13: warning: In the GNU C Library, "minor" is defined
by <sys/sysmacros.h>. For historical compatibility, it is
currently defined by <sys/types.h> as well, but we plan to
remove this soon. To use "minor", include <sys/sysmacros.h>
directly. If you did not intend to use a system-defined macro
"minor", you should undefine it after including <sys/types.h>.
min = minor(sb.st_rdev);
So, include sys/sysmacros.h to shut up the warning.
v2: Use the AC_HEADER_MAJOR defines to figure out the right header
(thanks to Jonathan Gray for helping me not break non-glibc systems)
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Emil Velikov <emli.velikov@collabora.com>
This was used for aubdumping (deleted a while ago) and INTEL_DEBUG=bat
decoding (deleted recently).
While we're changing parameters, delete the wrapper macro and make the
actual function brw_state_batch instead of __brw_state_batch.
This subsumes a patch by Emil Velikov to drop this from BLORP.
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
This deletes all of our handwritten code in favor of autogenerated
genxml-based decoding. This should be much more usable, as the old
code isn't entirely accurate - we updated some things for new
generations, but not everything.
Aubinator has one annoying limitation: it has no idea how many entries
to print when encountering e.g. 3DSTATE_BINDING_TABLE_POINTERS_VS. It
picks an arbitrary number, which may skip decoding valid data, and may
print extra garbage entries.
We do a better job here by making brw_state_batch track the size of the
data stored at a particular batchbuffer offset. Then, we can divide by
the structure size to obtain the exact number of entries.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
This should give substantially better decoding, as the public libdrm
decoder hasn't been properly maintained in years.
For now, we reuse the existing state dumping mechanism. We'll improve
that in the next patch.
To avoid increasing the size of the driver, we restrict this feature
to debug builds of Mesa. There's probably very little use for it in
release builds anyway.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
This will be used for color output in debug messages.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Fix build with Python < 2.7.
File "src/compiler/nir/nir_builder_opcodes_h.py", line 46, in <module>
from nir_opcodes import opcodes
File "src/compiler/nir/nir_opcodes.py", line 178, in <module>
unop_convert("{}2{}{}".format(src_t[0], dst_t[0], bit_size),
ValueError: zero length field name in format
Fixes: 762a6333f2 ("nir: Rework conversion opcodes")
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Like done in another place in that same file.
CID 1250588
Signed-off-by: Julien Isorce <jisorce@oblong.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Simplifies the write code a bit and handles EINTR.
V2: (Timothy Arceri) Drop EINTR handling. To do it
properly we would need a retry limit but it's
probably best to just avoid trying to write if
we hit EINTR and try again next time we see
the program.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
There is no need to hardcode it, we can just use blob_key[0].
This is needed because the next patches are going to change how cache
keys are computed.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
This will allow to hash additional data into the cache keys or even
change the hashing algorithm easily, should we decide to do so.
v2: don't try to compute key (and crash) if cache is disabled
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Per pixel stats are cached but were not always being flushed as threads
moved from one draw context to the next. Added an explicit flush to allow
all archrast objects to flush any cached events.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Performance is now 50x faster with archrast now that we're properly
filtering out all of the rdtsc begin/end.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
