I recently made us try two different things that tried to reduce register
pressure so that we would be more likely to allocate successfully. But
now that we have the logic for trying two, we can make the first thing we
try be the normal, not-prioritizing-register-pressure heuristic.
This means one less scheduling pass in the common case of that heuristic
not producing spills, plus the best schedule we know how to produce, if
that one happens to succeed. This is important, because our register
allocation produces a lot of possibly avoidable dependencies for the
post-register-allocation schedule, despite ra_set_allocate_round_robin().
GLB2.7: 1.04127% +/- 0.732461% fps improvement (n=31)
nexuiz: No difference (n=5)
lightsmark: 0.838512% +/- 0.300147% fps improvement (n=86)
minecraft apitrace: No difference (n=15)
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
The canary is basically just to give a better debugging message when you
ralloc_free() something that wasn't rallocated. Reduces maximum memory
usage of apitrace replay of the dota2 demo by 60MB on my 64-bit system (so
half that on a real 32-bit dota2 environment).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I think I was thinking of the batch command packet cache when I pasted
this in, but this counter is only used for dumping out streamed state for
INTEL_DEBUG=batch and for putting annotations in our aub files.
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Commit 2f89662 added the driconf option 'clamp_max_samples'. In that
commit, the option did not alter the context version. The neglect to
alter the context version is a fatal issue for some apps.
For example, consider running Chromium with clamp_max_samples=0.
Pre-patch, Mesa creates a GL 3.0 context but clamps GL_MAX_SAMPLES to
0. This violates the GL 3.0 spec, which requires GL_MAX_SAMPLES >= 4.
The spec violation causes WebGL context creation to fail in many
scenarios because Chromium correctly assumes that a GL 3.0 context
supports at least 4 samples.
Since the driconf option was introduced largely for Chromium, the issue
really needs fixing.
This patch fixes calculation of the context version to respect the
post-clamped value of GL_MAX_SAMPLES. This in turn fixes WebGL on
Chromium when clamp_max_samples=0.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>
clamp_max_samples() and intel_quantize_num_samples() each maintained
their own list of which MSAA modes the hardware supports. This patch
removes the duplication by making intel_quantize_num_samples() use the
same list as clamp_max_samples(), the list maintained in
brw_supported_msaa_modes().
By removing the duplication, we prevent the scenario where someone
updates one list but forgets to update the other.
Move function `brw_context.c:static brw_supported_msaa_modes()` to
`intel_screen.c:(non-static) intel_supported_msaa_modes()` and patch
intel_quantize_num_samples() to use the list returned by that function.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>
This simplifies the loop logic in a subsqequent patch that refactors
intel_quantize_num_samples() to use brw_supported_msaa_modes().
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Chad Versace <chad.versace@linux.intel.com>
These degenerate instructions can often be emitted by state trackers
when the semantics of instructions don't match precisely.
Reviewed-by: Brian Paul <brianp@vmware.com>
Several of links the were contributed by Keith Whitwell and Roland Scheidegger.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
From section 6.1.18 (Renderbuffer Object Queries) of the GL 3.2 spec,
under the heading "If the value of FRAMEBUFFER_ATTACHMENT_OBJECT_TYPE
is TEXTURE, then":
If pname is FRAMEBUFFER_ATTACHMENT_LAYERED, then params will
contain TRUE if an entire level of a three-dimesional texture,
cube map texture, or one-or two-dimensional array texture is
attached. Otherwise, params will contain FALSE.
Fixes piglit tests:
- spec/!OpenGL 3.2/layered-rendering/framebuffer-layered-attachments
- spec/!OpenGL 3.2/layered-rendering/framebuffertexture-defaults
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
v2: Don't include "EXT" in the error message, since this query only
makes sensen in context versions that have adopted
glGetFramebufferAttachmentParameteriv().
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Previously we were using the code path for validating
glFramebufferTextureLayer(). But glFramebufferTexture() allows
additional texture types.
Fixes piglit tests:
- spec/!OpenGL 3.2/layered-rendering/gl-layer-cube-map
- spec/!OpenGL 3.2/layered-rendering/framebuffertexture
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
v2: Clarify comment above framebuffer_texture().
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
From section 4.4.7 (Layered Framebuffers) of the GLSL 3.2 spec:
When the Clear or ClearBuffer* commands are used to clear a
layered framebuffer attachment, all layers of the attachment are
cleared.
This patch fixes the fast depth clear path.
Fixes piglit test "spec/!OpenGL 3.2/layered-rendering/clear-depth".
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
From section 4.4.7 (Layered Framebuffers) of the GLSL 3.2 spec:
When the Clear or ClearBuffer* commands are used to clear a
layered framebuffer attachment, all layers of the attachment are
cleared.
This patch fixes the blorp clear path for color buffers.
Fixes piglit test "spec/!OpenGL 3.2/layered-rendering/clear-color".
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
From section 4.4.7 (Layered Framebuffers) of the GLSL 3.2 spec:
When the Clear or ClearBuffer* commands are used to clear a
layered framebuffer attachment, all layers of the attachment are
cleared.
This patch fixes meta clears to properly clear all layers of a layered
framebuffer attachment. We accomplish this by adding a geometry
shader to the meta clear program which sets gl_Layer to a uniform
value. When clearing a layered framebuffer, we execute in a loop,
setting the uniform to point to each layer in turn.
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
In order to properly clear layered framebuffers, we need to know how
many layers they have. The easiest way to do this is to record it in
the gl_framebuffer struct when we check framebuffer completeness.
This patch replaces the gl_framebuffer::Layered boolean with a
gl_framebuffer::NumLayers integer, which is 0 if the framebuffer is
not layered, and equal to the number of layers otherwise.
v2: Remove gl_framebuffer::Layered and make gl_framebuffer::NumLayers
always have a defined value. Fix factor of 6 error in the number of
layers in a cube map array.
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Prevents a GPU page fault if somehow the uniform bo gets evicted
before the screen_create pushbuf has been submitted.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Previously, we checked for interstage uniform interface block link
errors in validate_interstage_interface_blocks(), which is only called
on pairs of adjacent shader stages. Therefore, we failed to detect
uniform interface block mismatches between non-adjacent shader stages.
Before the introduction of geometry shaders, this wasn't a problem,
because the only supported shader stages were vertex and fragment
shaders, therefore they were always adjacent. However, now that we
allow a program to contain vertex, geometry, and fragment shaders,
that is no longer the case.
Fixes piglit test "skip-stage-uniform-block-array-size-mismatch".
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
v2: Rename validate_interstage_interface_blocks() to
validate_interstage_inout_blocks() to reflect the fact that it no
longer validates uniform blocks.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
v3: Make validate_interstage_inout_blocks() skip uniform blocks.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Previously, when attempting to link a vertex shader and a geometry
shader that use different GLSL versions, we would sometimes generate a
link error due to the implicit declaration of gl_PerVertex being
different between the two GLSL versions.
This patch fixes that problem by only requiring interface block
definitions to match when they are explicitly declared.
Fixes piglit test "shaders/version-mixing vs-gs".
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
v2: In the interface_block_definition constructor, move the assignment
to explicitly_declared after the existing if block.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
From section 7.1 (Built-In Language Variables) of the GLSL 4.10
spec:
Also, if a built-in interface block is redeclared, no member of
the built-in declaration can be redeclared outside the block
redeclaration.
We have been regarding this text as a clarification to the behaviour
established for gl_PerVertex by GLSL 1.50, so we apply it regardless
of GLSL version.
This patch enforces the rule by adding an enum to ir_variable to track
how the variable was declared: implicitly, normally, or in an
interface block.
Fixes piglit tests:
- gs-redeclares-pervertex-out-after-global-redeclaration.geom
- vs-redeclares-pervertex-out-after-global-redeclaration.vert
- gs-redeclares-pervertex-out-after-other-global-redeclaration.geom
- vs-redeclares-pervertex-out-after-other-global-redeclaration.vert
- gs-redeclares-pervertex-out-before-global-redeclaration
- vs-redeclares-pervertex-out-before-global-redeclaration
Cc: "10.0" <mesa-stable@lists.freedesktop.org>
v2: Don't set "how_declared" redundantly in builtin_variables.cpp.
Properly clone "how_declared".
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Unfortunately, our hardware only has one set of aggregating performance
counters shared between all 3D programs, and their values are not saved
or restored by hardware contexts. Also, at least on Sandybridge and
Ivybridge, the counters lose their values if the GPU goes to sleep.
To work around both of these problems, we have to snapshot the
performance counters at the beginning and end of each batch, similar to
how we handle query objects on platforms that don't support hardware
contexts. I call these "bookend" snapshots.
Since there can be multiple performance monitors active at a time, we
store the bookend snapshots in a global BO, shared by all monitors.
For monitors that span multiple batches, acquiring results involves
adding up three segments:
BeginPerfMonitor --> End of Batch 1 ("head")
Start of Batch 2 --> End of Batch 2
... ("middle")
Start of Batch N-1 --> End of Batch N-1
Start of Batch N --> EndPerfMonitor ("tail")
Monitors that refer to bookend BO snapshots are considered "unresolved".
We delay resolving them (and adding up deltas to obtain the results) as
long as possible to avoid blocking on mapping monitor->oa_bo.
We can also run out of space in the bookend BO, at which point we have
to resolve all unresolved monitors. Then we can throw away the
snapshots and begin writing at the beginning of the buffer.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
In order to use the Observability Architecture effectively, we'll need
to take snapshots of the OA counters via MI_REPORT_PERF_COUNT at the
start and end of each batch.
Experimentation reveals that we need to flush before and after each
MI_REPORT_PERF_COUNT to get working values. For simplicitly, I chose to
use intel_batchbuffer_emit_mi_flush(), which unfortunately expands to
triple pipe controls on Sandybridge.
We may want to start computing per-generation reserved batch space to
avoid the insanity of Sandybridge's PIPE_CONTROL cost. That said, much
of this cost existed before I rewrote the query object support to use
hardware contexts, so it's at least not entirely new.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Currently, this only considers the monitor start and end snapshots.
This is woefully insufficient, but allows me to add a bunch of the
infrastructure now and flesh it out later.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
We need to start OA at the beginning of each batch where monitors are
active. OACONTROL isn't part of the hardware context, so to avoid
leaving counters enabled for other applications, we turn them off at the
end of the batch too.
We also need to start them at BeginPerfMonitor time (unless they've
already been started). We stop them when the monitor last ends as well.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
We'll need to write this register to start/stop performance counters.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
MI_REPORT_PERF_COUNT writes a snapshot of the Observability Architecture
counters to a buffer. Exactly how it works varies between generations:
Ironlake requires two packets, Sandybridge has to use GGTT, and Ivybridge
and later use PPGTT.
v2: Assert that we didn't use more space than we reserved (suggested
by Eric Anholt).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Using the OA counters requires some per-batch work. When starting and
ending a batch, it's useful to know whether any monitors are actually
interested in OA data.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
In addition to listing the counter names, we include several "remap"
tables. Confusingly, counters are documented with names like "A23",
are written to some buffer offset other than 23, and exposed by core
Mesa under a counter ID that is different still.
The first is inevitable; MI_REPORT_PERF_COUNT writes certain counters to
fixed locations in the buffer. The latter could be avoided, but core
Mesa uses the "Counters" array index as the ID for a counter. We could
do remapping there, but it would just complicate the core Mesa code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
This is fairly simple:
- At BeginPerfMonitor time, take an opening snapshot.
- At EndPerfMonitor time, take a closing snapshot.
- The first time the application asks for results, subtract the two and
store that value. Then free the BO containing the snapshots.
- On subsequent requests for the results, just return the saved value.
- On reset, throw away the results.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
For now, we only support these on Gen6+, since that's what currently
uses hardware contexts. When we add Ironlake hardware context support,
we can add pipeline statistics register support for that as well.
In theory, we could support pipeline statistics counters even without
hardware contexts, but it would be annoyingly painful.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Since we don't support any counters, there are zero groups.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
The Observability Architecture counters are 32-bit unsigned values, and
the Pipeline Statistics Register counters are 64-bit unsigned values.
These convenience macros make it easy to create those types of counters.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
It's useful to see the state of all outstanding monitors; the start
of a new batch seems like a reasonable time to print them out.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
These stub functions will be filled out in later patches.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
This will enable debugging printfs for the AMD_performance_monitor code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Without hardware contexts, the pipeline statistics registers are
free-running and include data from every 3D application running.
In order to find out the contributions of one particular context, we
need to take a snapshot at the start and end of each batch.
Previously, we emitted the PIPE_CONTROL necessary to capture
PS_DEPTH_COUNT when drawing primitives. Special tracking ensured it
happened only on the first draw of the batch, rather than on every draw.
Moving this to brw_new_batch increases symmetry, since the final
snapshot has always been in brw_finish_batch, which is just a few lines
below. It should be basically equivalent.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
The new intel_batchbuffer_emit_render_ring_prelude() hook will be called
when switching from BLT or UNKNOWN_RING to RENDER_RING. This provides a
place to emit state that should go at the start of each render ring
batch, with minimal overhead.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
When we first create a batch buffer, it's empty. We don't actually
know what ring it will be targeted at until the first BEGIN_BATCH or
BEGIN_BATCH_BLT macro.
Previously, one could determine the state of the batch by checking
brw->batch.ring (blit vs. render) and brw->batch.used != 0 (known vs.
unknown).
This should be functionally equivalent, but the tri-state enum is a bit
clearer.
v2: Catch three explicit require_space callers (thanks to Carl and Eric).
v3: Split the boolean -> enum change from the UNKNOWN_RING change.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
