This eliminates the need to walk the list of input variables, recurse
into their types (via logic largely redundant with nir_lower_io), and
interpolate all possible inputs up front. The backend no longer has
to care about variables at all, which eliminates complications from
trying to pack multiple variables into the same location. Instead,
each intrinsic specifies exactly what's needed.
This should unblock Timothy's work on GL_ARB_enhanced_layouts.
Each load_interpolated_input intrinsic corresponds to PLN instructions,
while load_barycentric_at_* intrinsics correspond to pixel interpolator
messages. The pixel/centroid/sample barycentric intrinsics simply refer
to payload fields (delta_xy[]), and don't actually generate any code.
Because we use a single intrinsic for both centroid-qualified variables
and interpolateAtCentroid(), they become indistinguishable. We stop
sending pixel interpolator messages for those, and instead use the
payload provided data, which should be considerably faster.
On Broadwell:
total instructions in shared programs: 9067751 -> 9067570 (-0.00%)
instructions in affected programs: 145902 -> 145721 (-0.12%)
helped: 422
HURT: 209
total spills in shared programs: 2849 -> 2899 (1.76%)
spills in affected programs: 760 -> 810 (6.58%)
helped: 0
HURT: 10
total fills in shared programs: 3910 -> 3950 (1.02%)
fills in affected programs: 617 -> 657 (6.48%)
helped: 0
HURT: 10
LOST: 3
GAINED: 3
The differences mostly appear to be slight changes in MOVs.
v2: Use nir_shader_compiler_options::use_interpolated_input_intrinsics
flag rather than passing it directly to nir_lower_io. Use the
unreachable() macro rather than assert in one place. (Review
feedback from Chris Forbes.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Currently, i965 interpolates all FS inputs at the top of the program.
This has advantages and disadvantages, but I'd like to keep that policy
while reworking this code. We can consider changing it independently.
The next patch will make the compiler generate PLN instructions "on the
fly", when it encounters an input load intrinsic, rather than doing it
for all inputs at the start of the program.
To emulate this behavior, we introduce an ugly pass to move all NIR
load_interpolated_input and payload-based (not interpolator message)
load_barycentric_* intrinsics to the shader's start block.
This helps avoid regressions in shader-db for cases such as:
if (...) {
...load some input...
} else {
...load that same input...
}
which CSE can't handle, because there's no dominance relationship
between the two loads. Because the start block dominates all others,
we can CSE all inputs and emit PLNs exactly once, as we did before.
Ideally, global value numbering would eliminate these redundant loads,
while not forcing them all the way to the start block. When that lands,
we should consider dropping this hacky pass.
Again, this pass currently does nothing, as i965 doesn't generate these
intrinsics yet. But it will shortly, and I figured I'd separate this
code as it's relatively self-contained.
v2: Dramatically simplify pass - instead of creating new instructions,
just remove/re-insert their list nodes (suggested by Jason Ekstrand).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisforbes@google.com> [v1]
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
When working with a non-multisampled render target, asking for "sample"
interpolation locations doesn't make sense. We demote them to centroid.
In a couple of patches, brw_compute_barycentric_modes will begin looking
at these intrinsics to determine the barycentric modes. fs_visitor also
will use them to code-generate pixel interpolator messages or payload
references. Handling the "but what if it's not MSAA?" logic ahead of
time in a NIR pass simplifies things and prevents duplicated logic.
This patch doesn't actually do anything useful yet as we don't generate
these intrinsics. I decided to keep it separate as it's self-contained,
in the hopes of shrinking the "convert everything" patch for reviewers.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Now nir_lower_io can optionally produce load_interpolated_input
and load_barycentric_* intrinsics for fragment shader inputs.
flat inputs continue using regular load_input.
v2: Use a nir_shader_compiler_options flag rather than ad-hoc boolean
passing (in response to review feedback from Chris Forbes).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Backends can normally handle shader inputs solely by looking at
load_input intrinsics, and ignore the nir_variables in nir->inputs.
One exception is fragment shader inputs. load_input doesn't capture
the necessary interpolation information - flat, smooth, noperspective
mode, and centroid, sample, or pixel for the location. This means
that backends have to interpolate based on the nir_variables, then
associate those with the load_input intrinsics (say, by storing a
map of which variables are at which locations).
With GL_ARB_enhanced_layouts, we're going to have multiple varyings
packed into a single vec4 location. The intrinsics make this easy:
simply load N components from location <loc, component>. However,
working with variables and correlating the two is very awkward; we'd
much rather have intrinsics capture all the necessary information.
Fragment shader input interpolation typically works by producing a
set of barycentric coordinates, then using those to do a linear
interpolation between the values at the triangle's corners.
We represent this by introducing five new load_barycentric_* intrinsics:
- load_barycentric_pixel (ordinary variable)
- load_barycentric_centroid (centroid qualified variable)
- load_barycentric_sample (sample qualified variable)
- load_barycentric_at_sample (ARB_gpu_shader5's interpolateAtSample())
- load_barycentric_at_offset (ARB_gpu_shader5's interpolateAtOffset())
Each of these take the interpolation mode (smooth or noperspective only)
as a const_index, and produce a vec2. The last two also take a sample
or offset source.
We then introduce a new load_interpolated_input intrinsic, which
is like a normal load_input intrinsic, but with an additional
barycentric coordinate source.
The intention is that flat inputs will still use regular load_input
intrinsics. This makes them distinguishable from normal inputs that
need fancy interpolation, while also providing all the necessary data.
This nicely unifies regular inputs and interpolateAt functions.
Qualifiers and variables become irrelevant; there are just
load_barycentric intrinsics that determine the interpolation.
v2: Document the interp_mode const_index value, define a new
BARYCENTRIC() helper rather than using SYSTEM_VALUE() for
some of them (requested by Jason Ekstrand).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Gen7/7.5 call it "Rendering Disable" while Gen8/9 prefix it with "API".
Pick one for consistency, and so we can share code between generations.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The bulk of this is the same. There are just a couple fields that only
exist on one generation or another, and we can easily handle those with
an #ifdef.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
We set the cull mode, but forgot the enable bit. Gen8 uses this.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The Gen7/7.5 clip code used APIMODE_OGL, while the Gen8+ clip code used
APIMODE_D3D. The meaning hasn't changed, so one of these must be wrong.
It appears that the hardware documentation is completely wrong. It
claims that the "API Mode" bit means:
0h APIMODE_OGL NEAR_VP boundary == 0.0 (NDC)
1h APIMODE_D3D NEAR_VP boundary == -1.0 (NDC)
However, DirectX typically uses 0.0 for the near plane, while unextended
OpenGL uses -1.0. i965's gen6_clip_state.c uses APIMODE_D3D for the
GL_ZERO_TO_ONE case, so I believe the meanings are backwards from what
the documentation says.
Section 23.2 ("Primitive Clipping") of the Vulkan 1.0.21 specification
contains the following equations:
-w_c <= x_c <= w_c
-w_c <= y_c <= w_c
0 <= z_c <= w_c
This means that Vulkan follows D3D semantics.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Gen6-7.5 use CLIPMODE_REJECT_ALL, while Gen8+ just used REJECT_ALL.
Being consistent will let me unify code, and I prefer having the prefix.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Refactoring to leave existing simd_* intrinsics in "simdintrin.h" unchanged,
adding corresponding simd16_* intrinsics in "simd16intrin.h" on the side,
with emulation, that we can use piecemeal, rather than the all-or-nothing
approach to bring up avx512.
Signed-off-by: Tim Rowley <timothy.o.rowley@intel.com>
Add support for enhanced attribute swizzling. Currently supports constant
source overrides to handle PrimitiveID support. No support yet for input
select swizzling or wrap shortest. Removes obsoleted linkageMask and
associated code.
Signed-off-by: Tim Rowley <timothy.o.rowley@intel.com>
Moved the setting into the existing component control code. Fixes bad
interaction between attribute/component setting for vertex/instance ID
and component packing.
Signed-off-by: Tim Rowley <timothy.o.rowley@intel.com>
From the Sky Lake PRM:
"For SURFTYPE_CUBE: For Sampling Engine Surfaces and Typed Data Port
Surfaces, the range of this field is [0,340], indicating the number of
cube array elements (equal to the number of underlying 2D array elements
divided by 6). For other surfaces, this field must be zero."
In other words, the depth field for cube maps is in number of cubes not
number of 2-D slices so we need to divide by 6. ISL will do this correctly
for us assuming that we provide it with the correct array bounds which it
expects to be in 2-D slices. It appears as if we've been doing this wrong
ever since we first added cube map arrays for Sandy Bridge and the change
to ISL made things slightly worse. While we're at it, we now need to remoe
the shader hacks we've always done since they were only needed because we
were setting the depth field six times too large.
v2: Fix the vec4 backend as well (not sure how I missed this).
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
This matches what we do for cube maps where logical_depth0 is in number of
face-layers rather than number of cubes. This does mean that we will
temporarily be setting the surface bounds too loose for cube map textures
but we are already setting them too loose for cube arrays and we will be
fixing that in the next commit anyway.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Cc: "12.0 11.2 11.1" <mesa-stable@lists.freedesktop.org>
The GL API and mesa internals do this differently than we do. In GL, there
is no depth parameter for 1-D arrays and height is used. In the i965
miptree code we do the sane thing and make height == 1 and use depth for
number of slices. This makes for a mismatch every time we create a 1-D
array texture from GL. Instead of actually solving this problem, we just
said "1-D is hard, let's make sure it works no matter which way we pass the
parameters" and called it a day.
This commit fixes the one GL -> i965 transition point where we weren't
already handling 1-D array textures to do the right thing and then replaces
the magic fixup code with an assert that you're doing the right thing.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Cc: "12.0 11.2 11.1" <mesa-stable@lists.freedesktop.org>
This 'last' variable used in FindGLXFunction(...) may become negative,
but has been defined as unsigned int resulting in an overflow,
finally resulting in a segfault when accessing _glXDispatchTableStrings[...].
Fixed this by definining it as signed int. 'first' variable also needs to be
defined as signed int. Otherwise condition for while loop fails due to C
implicitly converting signed to unsigned values before comparison.
Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Stefan Dirsch <sndirsch@suse.de>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Current implementation of the DRI image loader does not free the images
created in get_back_bo() and so leaks memory. Moreover, it creates a new
image every time the DRI driver queries for buffers, even if the backing
native buffer has not changed. leaking memory again.
This patch adds missing call to destroyImage() in droid_enqueue_buffer()
and a check if image is already created to get_back_bo() to fix the
above.
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
It is much easier to debug issues when the application gives some
meaningful error messages. This patch adds few to the EGL Android
platform backend.
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
It might return NULL if specific config variant is unsupported.
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
As the spec allows for {server,client}_wait_sync to be called without
currently bound context, while our implementation requires context
pointer.
v2: Add a mutex and acquire it for the duration of
brw_fence_client_wait() and brw_fence_is_completed() as suggested
by Chad.
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Without this, if a configuration is, say, available only on GLES2/3, but
not on GLES1, and is rejected by the dri module's bindContext call,
eglMakeCurrent fails with error "EGL_SUCCESS".
In this patch, we set error to EGL_BAD_MATCH, which is what CTS/dEQP
dEQP-EGL.functional.surfaceless_context expect.
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Nicolas Boichat <drinkcat@chromium.org>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
There are some unused variables left after previous clean-ups triggering
compiler warnings. Let's remove them.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
An earlier patch fixed the problem for classic drivers, however Gallium
was still left broken. This patch applies the same workaround to
Gallium, when compiled for Android. Following is a quote from the
original patch:
0cbc90c57c mesa: dri: Add shared glapi to LIBADD on Android
/system/vendor/lib/dri/*_dri.so actually depend on libglapi: without
this, loading the so file fails with:
cannot locate symbol "__emutls_v._glapi_tls_Context"
On non-Android (non-bionic) platform, EGL uses the following
workflow, which works fine:
dlopen("libglapi.so", RTLD_LAZY | RTLD_GLOBAL);
dlopen("dri/<driver>_dri.so", RTLD_NOW | RTLD_GLOBAL);
However, bionic does not respect the RTLD_GLOBAL flag, and the dri
library cannot find symbols in libglapi.so, so we need to link
to libglapi.so explicitly. Android.mk already does this.
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Signed-off-by: Nicolas Boichat <drinkcat@chromium.org>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Analogous to previous commit.
Note: scons always uses OOT builds, while the in-tree generated files
could be created either manually or by the autoconf build.
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>
Cc: Alexander von Gluck IV <kallisti5@unixzen.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
In the case of building in out-of-tree fashion, while having generated
in-tree sources, the latter [likely stale] files will be used.
Flip the order to prevent that.
Cc: "11.2 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
This allows Gallium drivers to advertise the subpixel precision
for floating point viewports bounds.
v2:
- Set ViewportSubpixelBits in st_init_limits.
Signed-off-by: Józef Kucia <joseph.kucia@gmail.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
MS images have to be handled explicitly and I don't plan to implement
them for now.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
On Maxwell, images binding is slightly different (and much better)
regarding Fermi and Kepler because a texture view needs to be uploaded
for each image and this is going to simplify the thing a lot.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Currently, we can store 32 tex handles of 32-bits integer each and
that fits perfectly with the underlying hardware except on GM107+
which requires to upload a texture view for each images.
This patch increases the number of storable texture handles in the
driver constant buffer from 32 to 40 because we expose 8 images.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
For intrinsics we don't care about, just skip to the next loop iteration
and process the next instruction. We don't want to execute the rest of
the code.
This was a bug in commit cdfc05ea6e.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Fixes a 10-20% performance regression in OglCSDof caused by commit
5a8c89038a, which made images (in the
image load/store sense) use BDW_MOCS_PTE instead of BDW_MOCS_WB.
This seems sketchy, as the default PTE value is supposed to be
WB LLC eLLC, which is the same as our MOCS WB setting. It's only
supposed to change when using a surface for display, which won't
ever happen for images. Something may be wrong in the kernel...
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
