We currently use CL_INVOCATION_COUNT for the GL_PRIMITIVES_GENERATED
query, which involves passing all primitives to the clipper. When
rasterizer discard is enabled, we program the clipper in REJECT_ALL
mode, rather than using the SOL stage's "Rendering Disable" feature.
See commit f09b91f782 for an explanation
of why we implement GL_PRIMITIVES_GENERATED this way.
Apparently the SOL stage's "Rendering Disable" feature is a lot faster
than having the clipper reject all primitives. It's safe to use when
no GL_PRIMITIVES_GENERATED query is active, as we don't care about
CL_INVOCATION_COUNT incrementing.
This patch makes us use SO_RENDERING_DISABLE when no query is active,
but continues falling back to the clipper in REJECT_ALL mode when the
queries are enabled. It brings back the perf_debug for the clipper
case (which I removed in commit 1f9445ff57, thinking it wasn't useful).
Improves performance in Gl32GSCloth by 84.8303% +/- 2.07132% (n = 10)
on my Broadwell GT2 laptop.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
They're basically the same. Let's avoid the code duplication.
v2: Fix SO_BUFFER_ENABLE stuff to only happen on Gen < 8 (caught
by Jason Ekstrand).
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Constant propagation on arrays doesn't make a lot of sense. If the
array is only accessed with constant indexes, then opt_array_splitting
would split it up. Otherwise, we have variable indexing. If there's
multiple accesses, then constant propagation would end up replicating
the data.
The lower_const_arrays_to_uniforms pass creates uniforms for each
ir_constant with array type that it encounters. This means that it
creates redundant uniforms for each copy of the constant, which means
uploading too much data. It can even mean exceeding the maximum number
of uniform components, causing link failures.
We could try and teach the pass to de-duplicate the data by hashing
constants, but it makes more sense to avoid duplicating it in the first
place. We should promote constant arrays to uniforms, then propagate
the uniform access.
Fixes the TressFX shaders from Tomb Raider, which exceeded the maximum
number of uniform components by a huge margin and failed to link.
On Broadwell:
total instructions in shared programs: 9067702 -> 9068202 (0.01%)
instructions in affected programs: 10335 -> 10835 (4.84%)
helped: 10 (Hoard, Shadow of Mordor, Amnesia: The Dark Descent)
HURT: 20 (Natural Selection 2)
loops in affected programs: 4 -> 0
The hurt programs appear to no longer have a constarray uniform, as
all constants were successfully propagated. Apparently before this
patch, we successfully unrolled a loop containing array access, but
only after promoting constant arrays to uniforms. With this patch,
we unroll it first, so all array access is direct, and the array
is split up, and individual constants are propagated. This seems
better.
Cc: mesa-stable@lists.freedesktop.org
Reported-by: Karol Herbst <nouveau@karolherbst.de>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
There's really no point in looking at ir_dereference_array of a
constant. It also misses cases like:
(assign () (var_ref tmp) (constant (array ...) ...))
No changes in shader-db, but keeps it working after the next commit.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
The scalar backend currently doesn't support variable indexing on
temporary arrays, but it does support it on uniform arrays, and
some stages support it for input arrays. Make sure these are
propagated through before exploding indirects into piles of
if-ladders unnecessarily.
On Broadwell, no instruction count change in shader-db.
total cycles in shared programs: 80675652 -> 80674928 (-0.00%)
cycles in affected programs: 649972 -> 649248 (-0.11%)
helped: 386
HURT: 165
This will help avoid code quality regressions in a future commit.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
The new uniform may need precise as well.
Fixes copy propagation of constant array uniforms in Tomb Raider shaders.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Previously, we failed to split constant arrays. Code such as
int[2] numbers = int[](1, 2);
would generates a whole-array assignment:
(assign () (var_ref numbers)
(constant (array int 4) (constant int 1) (constant int 2)))
opt_array_splitting generally tried to visit ir_dereference_array nodes,
and avoid recursing into the inner ir_dereference_variable. So if it
ever saw a ir_dereference_variable, it assumed this was a whole-array
read and bailed. However, in the above case, there's no array deref,
and we can totally handle it - we just have to "unroll" the assignment,
creating assignments for each element.
This was mitigated by the fact that we constant propagate whole arrays,
so a dereference of a single component would usually get the desired
single value anyway. However, I plan to stop doing that shortly;
early experiments with disabling constant propagation of arrays
revealed this shortcoming.
This patch causes some arrays in Gl32GSCloth's geometry shaders to be
split, which allows other optimizations to eliminate unused GS inputs.
The VS then doesn't have to write them, which eliminates the entire VS
(5 -> 2 instructions). It still renders correctly.
No other change in shader-db.
v2: Drop !AOA check and improve a comment (feedback from Tim Arceri).
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
opt_constant_propagation.cpp contains constant folding code which can
actually do constant propagation in some cases. It was happily
propagating constants into the left-hand-side of assignments.
For example,
(assign () (var_ref temp) (constant ...))
would brilliantly be turned into:
(assign () (constant ...) (constant ....))
This is a bigger hammer than necessary - it prevents propagation
into the left-hand-side altogether. We could certainly do better
someday. Notably, the constant propagation pass itself already
takes this approach - it's just the constant propagation pass's
built-in constant folding code (which actually propagates, too)
that was broken.
No change in shader-db, but prevents regressions after future commits.
It seems plausible that this could be hit today, but I haven't seen it
happen.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Without vertex elements originating directly from vertex fetcher
are not passed to wm-state correctly.
Signed-off-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Never be dependent on "draw 0", instead have a bool that makes the draw
dependent on the previous draw or not dependent at all.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Handle SGV stores separate from the stream fetch code.
Because of this change, there is a potential to jit an extra unused store.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Was trying to store an extra uninitialized component.
Only affects component packing, which isn't enabled (yet).
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Earlier MSVC 2013 releases have troubles compiling some of our C99 code,
so make sure we have Update 4 to avoid confusion.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Brian Paul <brianp@vmware.com>
With commit fb9fe35, we start using transfer_inline_write
for memcpy TexSubImage path, but that triggers a regression with
texture array in the svga driver.
With this patch, the direct map code will update the texture array
correctly.
Fixes VMware bug 1679293.
Tested with MTT piglit, glretrace, conform.
Reviewed-by: Brian Paul <brianp@vmware.com>
Currently with the SetVertexBuffers optimization, we avoid emitting
redundant DXSetVertexBuffers commands. However, these buffers surfaces
will still need to be referenced, otherwise, in the case of linux,
the subsequent surface discard map will map to the existing mob instead
of a new one, causing rendering artifacts.
With this patch, we'll call resource_rebind() to reference the resources
even if we are avoiding the actual set command. This fixes the
rendering artifacts in the window title area running with unity in
Ubuntu 14.04
Tested with piglit, glretrace.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>
This patch fixes three issues with vertex buffer references:
(1) Instead of copy the vertex buffer resource handles to the hw state
in the context structure, use pipe_resource_reference to properly
reference the vertex buffer resources in the context.
(2) Make sure to unbind those unused vertex buffer resources.
(3) Force to rebind the vertex buffer resources at the first draw of each
command buffer to make sure the vertex buffer resources are paged in.
Reviewed-by: Brian Paul <brianp@vmware.com>
Instead of copy the index buffer resource handle to the hw state in
the context structure, use pipe_resource_reference to properly reference
the index buffer resource in the context.
Reviewed-by: Brian Paul <brianp@vmware.com>
We already store these in gl_shader and gl_program here we
remove it from gl_shader_program and just use the values
from gl_shader.
This will allow us to keep the shader cache restore code as
simple as it can be while making it somewhat clearer where these
values originate from.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
We already store this in gl_shader and gl_program here we
remove it from gl_shader_program and just use the values
from gl_shader.
This will allow us to keep the shader cache restore code as
simple as it can be while making it somewhat clearer where these
values originate from.
V2: remove unnecessary NULL check
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Iago Toral <itoral@igalia.com>
This solves a race condition where we can end up having different stages
stomp on each other because they're all trying to scratch in the same BO
but they have different views of its layout.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
While we're here, we also fixup MEDIA_VFE_STATE and rename the field in
3DSTATE_VS on gen6-7.5 to be consistent with the others.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
The pack header generation scripts can't handle the case where you have
two addresses in the same dword; they just take whatever is the last one.
This meant that the MCS address wasn't properly getting handled. Since we
don't care about append counters, we can just re-arrange the XML for now.
Reviewed-by: Chad Versace <chad.versace@intel.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
ISL was being a bit too clever for its own good and lowering the format for
us. This is all well and good *if* we always want to lower it. However,
the GL driver selectively lowers the format depending on whether the
surface is write-only or not.
Reviewed-by: Chad Versace <chad.versace@intel.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Ivy Bridge and above can handle up to 2^31 elements for RAW buffer
surfaces.
Reviewed-by: Chad Versace <chad.versace@intel.com>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Brian Paul
Kenneth Graunke
Topi Pohjolainen
Ardinartsev Nikita
Tim Rowley
Jose Fonseca
Charmaine Lee
Timothy Arceri
Edward O'Callaghan
Christian Gmeiner
Jason Ekstrand