When binding a layered texture, the layer is already 0. There's no need
to special case this.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
This ports over Chris Forbes' equivalent fixes in gen7_misc_state.c
from commit 77d55ef481.
No Piglit changes on Sandybridge.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Color clears can be performed via two separate shaders - one is the
generic "meta clear" shader (in meta.c); the other is the i965 specific
"repclear" shader (in brw_meta_fast_clear.c).
Giving them separate names makes them distinguishable when reading
INTEL_DEBUG=shader_time output.
v2: Call it "meta repclear", as suggested by Jason.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Paul's original code had emit_control_data_bits() skip the URB write if
vertex_count was 0. This meant wrapping every control data write in a
conditional write.
We accumulate control data bits in a single UD (32-bit) register. For
simple shaders that don't emit many vertices, the control data header
will be <= 32-bits long, so we only need to write it once at the end of
the shader.
For shaders with larger headers, we write out batches of control data
bits at EmitVertex(), when (vertex_count * bits_per_vertex) % 32 == 0.
On the first EmitVertex() call, the above expression will evaluate to
true simply because vertex_count == 0. But we want to avoid emitting
the control data bits, because we haven't accumulated 32-bits worth yet.
In other words, the vertex_count != 0 check is really only necessary in
the EmitVertex() batching case, not the end-of-thread case.
This saves a CMP/IF/ENDIF in every shader that uses EndPrimitive() or
multiple streams. The only downside is that a shader which emits no
vertices at all will execute an additional URB write---but such shaders
are pointless and not worth optimizing.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
When the new hash table implementation was added to Mesa it claimed to be much
faster, see commits 35fd61bd99 and 72e55bb688.
The set implementation follows the same implementation strategy so this should
be faster and there was no need to store a data field.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Don't assume that $(top_srcdir)/.git is a directory. It may be a
gitlink file [1] if $(top_srcdir) is a submodule checkout or a linked
worktree [2].
[1] A "gitlink" is a text file that specifies the real location of
the gitdir.
[2] Linked worktrees are a new feature in Git 2.5.
Cc: "10.6, 10.5" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
If we split addr/pred, the original instruction could have originated
from a different block. If we don't fixup the block ptr we hit asserts
later (in debug builds).
NOTE: perhaps we don't want to try to preserve addr/pred reg's across
block boundaries.. this at least needs some thought in case addr/pred
writes end up inside a conditional block..
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The address and predicate register are special, they don't get assigned
in RA. So do a better job of ignoring them rather than hitting later
asserts.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Some, but not all, state trackers will explicitly unref (and set to
NULL) the previous *fence before calling pipe->flush(). So driver
should use fence_ref() which will unref the old fence if not NULL.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Acked-by: Eric Anholt <eric@anholt.net>
Some, but not all, state trackers will explicitly unref (and set to
NULL) the previous *fence before calling pipe->flush(). So driver
should use fence_ref() which will unref the old fence if not NULL.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Acked-by: Chia-I Wu <olvaffe@gmail.com>
Some, but not all, state trackers will explicitly unref (and set to
NULL) the previous *fence before calling pipe->flush(). So driver
should use fence_ref() which will unref the old fence if not NULL.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
XA was never unref'ing last_fence in the various call paths to
pipe->flush(). Add this to xa_context_flush() and update the other
open-coded calls to pipe->flush() to use xa_context_flush() instead.
This fixes a memory leak reported with xf86-video-freedreno.
Reported-by: Nicolas Dechesne <nicolas.dechesne@linaro.org>
Cc: "10.5 10.6" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Rob Clark <robclark@freedesktop.org>
After tearing it out another level or two, and just passing the key and
vp directly, we can finally remove this struct. It also eliminates a
pointless memcpy() of the key.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
At this point, the brw_vs_compile structure only contains the key and
gl_vertex_program pointer. We may as well pass and store them directly;
it's simpler and more convenient (key-> instead of vs_compile->key...).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Nothing outside of vec4_visitor uses it, so we may as well keep it
internal.
Commit db9c915abc for the vec4 backend.
(The empty class will be going away soon.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
This is more consistent with how we do it in the FS backend, and reduces
a tiny bit of duplication. It'll also allow for a bit more tidying.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
This patch makes us only issue the performance warning about register
spilling if we actually spilled registers. We also use scratch space
for indirect addressing and the like.
This is basically commit c51163b0cf for
the vec4 backend.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Jason plumbed this through a while back in the FS backend, but
apparently we were just passing NULL in the vec4 backend.
This patch passes brw in as intended.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Adding new shader stages to a switch statement is less confusing than an
if-else-if ladder where all but the first case are fragment shader
specific (but don't claim to be).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
It's only used inside #ifdef DEBUG. Cuts ~1.7k of .text, and more
importantly prevents a larger code size regression in the next commit
when the .used field is replaced and calculated on demand.
text data bss dec hex filename
4945468 195152 26192 5166812 4ed6dc i965_dri.so before
4943740 195152 26192 5165084 4ed01c i965_dri.so after
And surround the emit and total fields with #ifdef DEBUG to prevent
such mistakes from happening again.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
This patch can cause an infinite recursion if the previous patch titled, "i965:
Track finished batch state" isn't present (backporters take notice).
v2: Sent out the wrong patch originally. This patches switches the order of
flushes, doing the generic flush before the CC_STATE, and the required
workaround flush afterwards
v3: Only perform workaround for render ring
Add text to the BATCH_RESERVE comments
v4 (By Ken): Rebase; update citation to mention PRM and Wa name; combine two
blocks.
http://otc-mesa-ci.jf.intel.com/job/bwidawsk/171/
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We need to check what the 3D pipe is able to handle for the mixer, not what
the decoder is able to decode. This fixes output of resolutions like 720x1280.
Signed-off-by: Christian König <christian.koenig@amd.com>
CC: mesa-stable@lists.freedesktop.org
Before validating vertex arrays we need to check if a VBO is present.
Checking if vb->buffer is not NULL fixes the issue.
Fixes the following piglit test:
gl-3.1-vao-broken-attrib
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
According to nv50, this should be src->ms_y instead of src->ms_x. This
code is here since 2012, so it's probably a typo error which has never
been detected since a long time. I didn't do a full piglit run to check
if it fixes some other weird issues.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
With the exception of gen8, the sole user of the workaround bo are for
emitting pipe controls. Move it out of the purview of the batchbuffer
and into the pipecontrol.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Martin Peres <martin.peres@linux.intel.com>
Since there was an ABI break and linking twice against libudev.so.0 and
libudev.so.1 causes the application to quickly crash, we first check if
the application is currently linked against libudev before dlopening a
local handle. However for backwards/forwards compatability, we need to
inspect the application for current linkage against all known versions
first. Not doing so causes a crash when both libraries are present and
so mesa chooses libudev.so.1 but the application was linked against
libudev.so.0.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Emil Velikov:
I'm ever so slightly conserned that RTLD_NOLOAD is not part of the POSIX
standard, thus it's missing on some platforms (*BSD seems ok, while
Solaris, MacOS are not).
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
Move the query for the TIMESTAMP register from context init to the
screen, so that it is only queried once for all contexts.
On 32bit systems, some old kernels trigger a hw bug resulting in the
TIMESTAMP register being shifted and the low 32bits always zero. Detect
this by repeating the read a few times and check the register is
incrementing every 80ns as expected and not stuck on zero (as would be
the case with the buggy kernel/hw.).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Martin Peres <martin.peres@linux.intel.com>
Reviewed-by: Martin Peres <martin.peres@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Matrix vertex attributes have their columns padded out to vec4s, which
I was failing to account for. Scalar NIR expects them to be packed,
however.
Fixes 1256 dEQP tests on Broadwell.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Mark Janes <mark.a.janes@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
This allows drivers to report queries in units of microseconds and
have the HUD display "us" (microseconds), "ms" (milliseconds) or "s"
(seconds) on the graph.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Instead of using a boolean 'is bytes' value, use the pipe_driver_query_type
enum type. This will let is add support for time values in the next patch.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This was probably disabled due to a combination of several bugs in the
generator code (fixed earlier in this series) and a misunderstanding
of the hardware spec. The documentation for most control flow
instructions mentions among other restrictions:
"Instruction compression is not allowed."
This however doesn't have any implications on 16 wide not being
supported, because none of the control flow instructions have
multi-register operands (control flow instructions are not compressed
on more recent hardware either, except maybe SNB's IF with inline
compare). In fact Gen4-5 had 16-wide control flow masks and stacks,
and the spec mentions in several places that control flow instructions
push and pop 16 channels worth of data -- Otherwise there doesn't seem
to be any indication that it shouldn't work.
Causes no piglit regressions, and gives the following shader-db
results on ILK:
total instructions in shared programs: 4711384 -> 4711384 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
GAINED: 1215
LOST: 0
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
