| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Allows us to transform
mul res src0 src1
mov.sat dst -res
into
mul.sat dst src0 -src1
instructions in affected programs: 45246 -> 45054 (-0.42%)
helped: 162
Reviewed-by: Ian Romanick <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It's not correct to CSE these multiplies
mul.sat dst1, -a, b
mul.sat dst2, a, b
by emitting a negated MOV from dst1 to dst2:
mul.sat dst1, -a, b
mov dst2, -dst1
Take 2.0*2.0 for example. The first multiply would produce 0.0 and the
second would produce 1.0.
Fixes bad generated code in 18 to 22 shaders:
instructions in affected programs: 432 -> 464 (7.41%)
helped: 4
HURT: 18
Cc: [email protected]
Reviewed-by: Ian Romanick <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
RGBA8 and BGRA8 unorm formats are compatible with the various
mem_copy functions. Their sRGB counterparts are also compatible
because they're also color-renderable (of importance when the
specified resource is a readbuffer) and they share the same
physical layout.
Signed-off-by: Nanley Chery <[email protected]>
Reviewed-by: Anuj Phogat <[email protected]>
|
|
|
|
|
|
|
|
| |
So the result is of float type if we're implementing the float
overload of imageAtomicExchange. This is the only back-end change
required to support OES_shader_image_atomic AFAICT.
Reviewed-by: Ilia Mirkin <[email protected]>
|
|
|
|
|
|
|
|
|
| |
If the destination is a renderbuffer, dst_tex_image will be NULL. This
fixes the *to_renderbuffer dEQP copy image tests.
Signed-off-by: Ilia Mirkin <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
Cc: [email protected]
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp:244:1: warning:
‘void {anonymous}::fs_copy_prop_dataflow::dump_block_data() const’ defined but not used [-Wunused-function]
fs_copy_prop_dataflow::dump_block_data() const
^~~~~~~~~~~~~~~~~~~~~
From looking at git history, it looks like this is intended to be unused
(ie. just for adding on-demand debug prints)
Signed-off-by: Rob Clark <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
Every stage has a corresponding 3DSTATE_CONSTANT_XS packet, so having
the code to create and emit push constant buffers in genX_vs_state.c
is a little strange. Moving it to a separate file seems more logical.
v2 [Ken]: Rebase on master, explain motivation in the commit message.
Signed-off-by: Ben Widawsky <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
And use it in brw_fs_nir.cpp.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Gen4/5's SEL instruction cannot use conditional modifiers, so min/max
are implemented as CMP + SEL. Handling that after optimization lets us
CSE more.
On Ironlake:
total instructions in shared programs: 6426035 -> 6422753 (-0.05%)
instructions in affected programs: 326604 -> 323322 (-1.00%)
helped: 1411
total cycles in shared programs: 129184700 -> 129101586 (-0.06%)
cycles in affected programs: 18950290 -> 18867176 (-0.44%)
helped: 2419
HURT: 328
Reviewed-by: Francisco Jerez <[email protected]>
|
|
|
|
| |
Reviewed-by: Francisco Jerez <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
On gen7 (Ivy Bridge, Haswell), we will get a GPU hang if an indirect
dispatch is used, but one of the dimensions is 0.
Therefore we use predicated rendering on the GPGPU_WALKER command to
handle this case.
Fixes piglit test: spec/arb_compute_shader/zero-dispatch-size
From the ARB_compute_shader spec, under DispatchCompute:
"If the work group count in any dimension is zero, no work groups are
dispatched."
And then for DispatchComputeIndirect:
... "is equivalent (assuming no errors are generated) to calling
DispatchCompute with <num_groups_x>, <num_groups_y> and
<num_groups_z>" ...
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94100
Signed-off-by: Jordan Justen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
Tested-by: Ilia Mirkin <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
See Ivy Bridge PRM, Volume 2, Part 2, 1.8.4 INTERFACE_DESCRIPTOR_DATA:
DWORD 5, bits 20:16: "This field indicates how much shared local
memory the thread group requires. The amount is specified in 4k
blocks, but only powers of 2 are allowed: 0, 4k, 8k, 16k, 32k and 64k
per half-slice."
For Haswell, see Volume 2d, INTERFACE_DESCRIPTOR_DATA:
DWORD 5, bits 20:16: With text identical to the Ivy Bridge PRM.
For Broadwell, see Volume 2d, INTERFACE_DESCRIPTOR_DATA:
DWORD 6, bits 20:16: With text identical to the Ivy Bridge PRM.
Signed-off-by: Jordan Justen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
| |
Alos use the opportunity to mark inputs constant. (Context has to be
given as read-write to intel_miptree_supports_non_msrt_fast_clear()
to support debug output).
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
| |
This will be re-used to initialize auxiliary buffers in lossless
compression case.
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
| |
v2 (Ben): Use combination of msaa_layout and number of samples
instead of introducing explicit type for lossless
compression (intel_miptree_is_lossless_compressed()).
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
| |
v2 (Ben): Use combination of msaa_layout and number of samples
instead of introducing explicit type for lossless
compression (intel_miptree_is_lossless_compressed()).
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
| |
v2 (Ben): Use combination of msaa_layout and number of samples
instead of introducing explicit type for lossless
compression (intel_miptree_is_lossless_compressed()).
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
v2 (Ben): Use combination of msaa_layout and number of samples
instead of introducing explicit type for lossless
compression.
v3 (Ben): Squash with "i965: Resolve color buffer also in
lossless compression case" and clarify simple
non-compressed fast clear case.
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
| |
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This got pushed accidentally in the first place but wasn't reverted
as it didn't regress piglit but instead fixed one newly introduced
test exercising a corner in case in i965 driver. However, saving and
restoring vertex buffer context is complicated and requires more
thought.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94150
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Tapani Palli <[email protected]>
|
|
|
|
|
|
|
|
| |
v2: Forgot to squash in the comment removal
Signed-off-by: Ben Widawsky <[email protected]>
Reviewed-by: Mark Janes <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
brw_draw_upload.c uploads VertexID/InstanceID first, then DrawID.
So we need to assign the attribute mapping in that order as well.
Fixes the following Pigit tests with the vec4 backend:
- arb_shader_draw_parameters-drawid vertexid
- arb_shader_draw_parameters-drawid-indirect basevertex
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Kristian Høgsberg <[email protected]>
|
|
|
|
|
|
|
|
| |
SPIR-V has a concept of a function type that's used fairly heavily. We
could special-case function types in SPIR-V -> NIR but it's easier if we
just add support to glsl_types.
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
| |
It's a bit more descriptive since it is the base type that you get when you
sample from it. Also, the next commit adds a bare "sampler" type and we
need glsl_type::sampler_type available for a public static member.
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
Fixes several of the
"dEQP-GLES31.functional.image_load_store*load_store*single_layer" dEQP
tests that use image formats we implement using untyped surface
messages.
Cc: [email protected]
Tested-by: Ilia Mirkin <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
MAX_COMPUTE_SHARED_SIZE should be set to 32768. This fixes a regression
introduced in be27f77 (mesa: do not use a constant for
MAX_COMPUTE_SHARED_SIZE).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94139
Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
Reviewed-by: Ilia Mirkin <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Until now there has been only one type of color buffer that needs
to resolved - namely single sampled fast clear. As even the
sampler engine in GPU doesn't understand the associated meta data,
the color values need to be always resolved prior to reading them.
From SKL onwards there is new scheme supported called the lossless
compression of single sampled color buffers. This is something that
is understood by the sampling engine and therefore resolving of
these types of buffers is not necessary before sampling.
This patch adds means to make the distinction when considering if
resolve is needed.
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
| |
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In addition to simply calling miptree_create() the higher level
call intel_miptree_create() also considers if the buffer should
be associated with an auxiliary buffer based on the given format.
Here we are allocating an auxiliary buffer which in turn has such
format that would mislead intel_miptree_create_layout() later on
to try to associate the auxiliary buffer with an auxiliary buffer.
To prevent this the actual buffer creation logic was split out
into its own function. Lets invoke that instead.
v2 (Ben): Do not signal msaa layout with explicit argument but
using layout_flags instead.
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
| |
This allows ls, and scripts to get the file names in the correct order of
optimization.
Signed-off-by: Ben Widawsky <[email protected]>
Reviewed-by: Matt Turner <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The L3 partitioning code tries to look at all programs - both render
programs (VS/TCS/TES/GS/FS) and compute (CS).
After calling brw_clear_cache, all prog_data pointers are invalid and
point to freed data. The intention was that flagging the dirty bits for
all programs would cause the next draw call to re-run the atoms for each
program stage, uploading new programs and installing new, valid pointers.
However, this doesn't quite work in our new multi-pipeline world. When
drawing or dispatching a compute workload, we only consider the programs
for the appropriate pipeline: drawing sets up VS/TCS/TES/GS/FS, but not
CS, and vice versa. This leaves pointers dangling a bit longer than
intended.
The L3 configuration code tries to inspect the prog_data for all shader
stages, so that we avoid having to reconfigure it when swapping back and
forth between render and compute workloads. So we can't have dangling
pointers.
The fix is simple: have brw_clear_cache NULL out stale prog_data
pointers, making it safe to inspect. The next L3 configuration pass
will see either the render shaders or compute shader as missing for
one go around, but will pick them up when both pipelines have run.
In other words, we'll simply reconfigure L3 twice, which is safe,
if a tiny bit wasteful - but then again, we just threw every compiled
shader we had on the floor and started recompiling the from scratch,
which is massively more wasteful, so it's not much of a concern.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93790
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Francisco Jerez <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
| |
To get _mesa_num_tex_faces() prototype.
|
|
|
|
| |
To get _mesa_num_tex_faces() prototype.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Auxiliary buffers are always created with sample number of zero
which effectively prevents intel_miptree_create_layout() from trying
to associate auxiliary buffers with auxiliary buffers.
Now that there is more direct path available lets start using it
instead and stop even checking for such (im)possibility.
v2 (Ben): Do not signal msaa layout with explicit argument but
using layout_flags instead.
Signed-off-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently the logic allocating and setting up miptrees is closely
combined with decision making when to re-allocate buffers in
X-tiled layout and when to associate colors with auxiliary buffers.
These auxiliary buffers are in turn also represented as miptrees
and are created by the same miptree creation logic calling itself
recursively. This means considering in vain if the auxiliary buffers
should be represented in X-tiled layout or if they should be
associated with auxiliary buffers again.
While this is somewhat unnecessary, this doesn't impose any problems
currently. Miptrees for auxiliary buffers are created as simgle-sampled
fusing the consideration for multi-sampled compression auxiliary
buffers. The format in turn is such that is not applicable for
single-sampled fast clears (that would require accompaning auxiliary
buffer).
But once the driver starts to support lossless compression of color
buffers the auxiliary buffer will have a format that would itself
be applicable for lossless compression. This would be rather
difficult and ugly to detect in the current miptree creation logic,
and therefore this patch seeks to separate the association logic
from the general allocation and setup steps.
v2 (Ben):
- Do not reconsider for X-tiling in intel_miptree_create()
as it was just forced to Y-tiling in miptree_create().
- Do not drop checks for allocation failures.
Signed-off-by: Topi Pohjolainen <[email protected]>
|
|
|
|
|
|
|
|
| |
This makes the logic a little more explicit and helps to keep
subsequent patches easier to read.
Suggested-by: Ben Widawsky <[email protected]>
Signed-off-by: Topi Pohjolainen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Part of brw_try_draw_prims() is a check to validate textures
(brw_validate_textures()). In case of textures that currently have
only level zero but are marked for mipmap generation, i965 driver
will decide to replace the underlying buffer with a larger one
capable of holding also the additional levels. This results into
blit from the original buffer to the newly allocated (see
intel_miptree_copy_teximage()). This blit is currently handled with
blitter engine and hence it won't effect the ongoing draw operation.
However, this blit in turn may trigger color resolve on the source
buffer. In principle, this should be possible with fast cleared
buffers but I only started hitting it when I enabled lossless
compression (that reguires similar resolve to fast cleared buffers).
Now, the color resolve is a meta operation and uses the same drawing
path we are already in middle of. After quite a bit of debugging I
realized that the resolve will modify the current vbo setup but it
won't restore it afterwards resulting in the original draw call
using wrong vertex data.
When brw_try_draw_prims() gets called, the vbo logic in the Mesa
core (see vbo_draw_arrays()) has just bound the vbo (see
vbo_bind_arrays() and recalculate_input_bindings()). Color resolve
operation will overwrite the vbo setup by calling vbo_bind_arrays()
against the resolve rectangle (see brw_draw_rectlist()). Once the
color resolve is done the vbo setup is left to the resolve rectangle
state and the original drawing call yields bogus results.
This patch aims to restore the original state after the color
resolve by calling vbo_bind_arrays() yet again after the vertex
array state in the core context have been restored.
Now having said all this, I'd also like to state that I'm quite
uncomfortable with the nested meta operations. Ths original draw
call in this case is in fact a meta operation itself. It is a blit
from level zero to level one when generating the additional mipmap
levels (see _mesa_meta_GenerateMipmap()). Imagine the complexity
if the blit in the middle from buffer to another would go to meta
path also instead of blitter.
I would very tempted to try to move all the resolves to happen
before a meta operation is started.
Additionally I still feel that work I did earlier in the spring/
summer time moving meta operations to use direct state upload
bypassing the core context would make sense.
v2: Force input recalculation by setting the flag explicitly
v3: Do not attempt to restore vbo for opengles1 which doesn't
support vertex buffer objects.
Signed-off-by: Topi Pohjolainen <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
Validation may kick off copies and subsequently color resolves.
Color resolves (and the copies themselves if ending up in meta path)
will overwrite the internal driver state but are not prepared to
restore it. Instead of adding that capability the validation can be
simply performed before the state is updated.
Signed-off-by: Topi Pohjolainen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Setting brw->ctx.NewDriverState and brw->ctx.NewGLState affects
the dirty bits for the current pipeline. But, we need to flag
everything dirty on *both* pipelines, so that when we switch
back, we'll realize our programs are stale and re-upload them.
To accomplish this, flag the saved state for both pipelines.
Only one of them should matter, but this way we don't have to
check which we need to set. It's harmless to set the other.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93790
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
Tested-by: Ilia Mirkin <[email protected]>
|
|
|
|
|
|
|
|
| |
I think this was just missed; Curro and I were probably writing
code simultaneously and forgot to combine them at the end.
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
When uploading state for the compute pipeline, we don't want to
look at VS/TCS/TES/GS/FS programs, as they might be stale, and
aren't relevant anyway. Likewise, the render pipeline shouldn't
look at CS.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93790
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
We really need to stop pulling information directly out of shaders for
state setup. For one thing, if we want any sort of an on-disk shader
cache, having all of this metadata in one place is going to be crucial.
Also, passing it all through prog_data cleans up the compiler <-> state
setup API substantially.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
It's extremely FS specific so the fact that we have a stage check in the
middle of it is rather bogus. While were here, we rename
setup_payload_gen4 and setup_payload_gen6 to make it obvious that they are
both FS specific.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
| |
Signed-off-by: Chris Forbes <[email protected]>
Reviewed-by: Ben Widawsky <[email protected]>
|
|
|
|
|
|
|
|
| |
The indirect dispatch registers were whitelisted in command parser
version 5. (Version 5 is available as of Linux 4.4)
Signed-off-by: Jordan Justen <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
| |
to allow LinkShader to free the GLSL IR.
Reviewed-by: Ian Romanick <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
When NIR was originally drafted, there was no easy way to determine if
something was constant or not. The result was that we had lots of
special-casing for constant values such as this. Now that load_const
instructions are SSA-only, it's really easy to find constants and this
isn't really needed anymore.
Reviewed-by: Connor Abbott <[email protected]>
Reviewed-by: Rob Clark <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This fixes an assertion failure in [at least] one of the Unreal Engine Linux
demo/games that uses DXT1 compression. Specifically, the "Vehicle Game".
At some point, the game ends up trying to blit mip level whose size is 2x2,
which is smaller than a DXT1 block. As a result, the assertion in the blit path
is triggered. It should be safe to simply make sure we align the width and
height, which is sadly an example of compression being less efficient.
NOTE: The demo seems to work fine without the assert, and therefore release
builds of mesa wouldn't stumble over this. Perhaps there is some unnoticeable
corruption, but I had trouble spotting it.
Thanks to Jason for looking at my backtrace and figuring out what was going on.
v2: Use NPOT alignment to make sure ASTC is handled properly (Ilia)
Remove comment about how this doesn't fix other bugs, because it does.
Cc: "11.0 11.1" <[email protected]
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93358
Signed-off-by: Ben Widawsky <[email protected]>
Tested-by: Matt Turner <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
tl;dr: For many types of GL object, we can *NEVER* use the Gen function.
In OpenGL ES (all versions!) and OpenGL compatibility profile,
applications don't have to call Gen functions. The GL spec is very
clear about how you can mix-and-match generated names and non-generated
names: you can use any name you want for a particular object type until
you call the Gen function for that object type.
Here's the problem scenario:
- Application calls a meta function that generates a name. The first
Gen will probably return 1.
- Application decides to use the same name for an object of the same
type without calling Gen. Many demo programs use names 1, 2, 3,
etc. without calling Gen.
- Application calls the meta function again, and the meta function
replaces the data. The application's data is lost, and the app
fails. Have fun debugging that.
Fixes piglit 'object-namespace-pollution glGetTexImage-compressed
renderbuffer' test.
Signed-off-by: Ian Romanick <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92363
Reviewed-by: Anuj Phogat <[email protected]>
|