| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Jason suggested adding an assert(function->impl) here. All callers
of this function actually want ->impl, so I decided just to change
the API.
We also change the nir_lower_io_to_temporaries API here. All but one
caller passed nir_shader_get_entrypoint(), and with the previous commit,
it now uses a nir_function_impl internally. Folding this change in
avoids the need to change it and change it back.
v2: Fix one call I missed in ir3_compiler (caught by Eric).
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
Reviewed-by: Connor Abbott <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
This changes the pass internals to work with a nir_function_impl
directly rather than a nir_function. The next patch will change
the API.
v2: Rebase after framebuffer fetch landed.
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
Reviewed-by: Connor Abbott <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
The reason why it was safe for the scheduler to ignore the side
effects of framebuffer write instructions was that its side effects
couldn't have had any influence on any other instruction in the
program, because we weren't doing framebuffer reads, and framebuffer
writes were always non-overlapping. We need actual memory dependency
analysis in order to determine whether a side-effectful instruction
can be reordered with respect to other instructions in the program.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
We weren't checking the fs_inst::target field when comparing whether
two instructions are equal. For FB writes it doesn't matter because
they aren't CSE-able anyway, but this would have become a problem with
FB reads which are expression-like instructions.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
brw_set_dp_read_message() was setting the data cache as send message
SFID on Gen7+ hardware, ignoring the target cache specified by the
caller. Some of the callers were passing a bogus target cache value
as argument relying on brw_set_dp_read_message not to take it into
account. Fix them too.
Reviewed-by: Iago Toral Quiroga <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
hardware.
This is not enabled on the original Gen4 part because it lacks surface
state tile offsets so it may not be possible to sample from arbitrary
non-zero layers of the framebuffer depending on the miptree layout (it
should be possible to work around this by allocating a scratch surface
and doing the same hack currently used for render targets, but meh...).
On Gen9+ even though it should mostly work (feel free to force-enable
it in order to compare the coherent and non-coherent paths in terms of
performance), there are some corner cases like 1D array layered
framebuffers that cannot be handled easily by the non-coherent path
because of the incompatible layout in memory of 1D and 2D miptrees (it
should be possible to work around this too by doing state-dependent
recompiles, but it's hard to care enough since Gen9 has native support
for coherent render target reads...)
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This is a no-op if the platform supports coherent framebuffer fetch,
-- If it doesn't we just need to flush the render cache and invalidate
the texture cache in order for previous rendering to be visible to
framebuffer fetch.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
| |
This iterates over the list of attached render buffers and binds
appropriate surface state structures to the binding table block
allocated for shader framebuffer read.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
brw_emit_surface_state.
This allows the caller to bind a miptree using a texture target other
than the one it it was created with. The code should work even if the
memory layouts of the specified and original targets don't match, as
long as the caller only intends to access a single slice of the
miptree structure.
This will be exploited by the next commit in order to support
non-coherent framebuffer fetch of a single layer of a 3D texture
(since some generations lack the minimum array element control for 3D
textures bound to the sampler unit), and multiple layers of a 1D array
texture (since binding it as an actual 1D array texture would require
state-dependent recompiles because the same shader couldn't
simultaneously work for 1D and 2D array textures due to the different
texel fetch coordinate ordering).
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit does three different things in a single pass in order to
keep the amount of churn low: Remove the for_gather boolean argument
which was unused, pass the isl_view argument by value rather than by
reference since I'll have to modify it from within the function, and
add a target argument to allow callers to bind textures using a target
other than the original. The prototype of the function now looks
like:
void brw_emit_surface_state(struct brw_context *brw,
struct intel_mipmap_tree *mt,
GLenum target, struct isl_view view,
uint32_t mocs, uint32_t *surf_offset, int surf_index,
unsigned read_domains, unsigned write_domains);
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
structures.
This surface state control has been supported by all hardware
generations since G45.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
| |
textures.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
| |
The logic to calculate the right layout and dimensionality for a given
GL texture target is going to be useful elsewhere, factor it out from
intel_miptree_get_isl_surf().
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is required because the sampler unit used to fetch from the
framebuffer is unable to interpret non-color-compressed fast-cleared
single-sample texture data. Roughly the same limitation applies for
surfaces bound to texture or image units, but unlike texture sampling,
non-coherent framebuffer fetch is by definition non-coherent with
previous rendering, so the brw_render_cache_set_check_flush() call can
be omitted except after resolve.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
intel_miptree_resolve_color().
This will allow optimizing out the cache flush in some cases when
resolving wasn't necessary.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
| |
This gets the non-coherent framebuffer fetch path hooked up to the NIR
front-end.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This gets rid of the duplication of logic between nir_setup_outputs()
and get_frag_output() by allocating fragment output temporaries lazily
whenever get_frag_output() is called. This makes nir_setup_outputs()
a no-op for the fragment shader stage.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The problem with the current approach is that driver output locations
are represented as a linear offset within the nir_outputs array, which
makes it rather difficult for the back-end to figure out what color
output and index some nir_intrinsic_load/store_output was meant for,
because the offset of a given output within the nir_output array is
dependent on the type and size of all previously allocated outputs.
Instead this defines the driver location of an output to be the pair
formed by its GLSL-assigned location and index (I've borrowed the
bitfield macros from brw_defines.h in order to represent the pair of
integers as a single scalar value that can be assigned to
nir_variable_data::driver_location). nir_assign_var_locations is no
longer useful for fragment outputs.
Because fragment outputs are now allocated independently rather than
within the nir_outputs array, the get_frag_output() helper becomes
necessary in order to obtain the right temporary register for a given
location-index pair.
The type_size helper passed to nir_lower_io is now type_size_dvec4
rather than type_size_vec4_times_4 so that output array offsets are
provided in terms of whole array elements rather than in terms of
scalar components (dvec4 is the largest vector type supported by the
GLSL so this will cause all individual fragment outputs to have a size
of one regardless of the type).
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
| |
Most likely we had only ever used this macro on bitfields of less than
31 bits -- That's going to change shortly.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
I'm about to change how fragment shader output locations are
represented, so the generic nir_intrinsic_store_output implementation
that assumes that outputs are just contiguous elements in the big
nir_outputs array won't work anymore. This somewhat simplified
implementation of nir_intrinsic_store_output for fragment shaders
should be functionally equivalent to the current fall-back one.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
| |
v2: Memoize sample ID, misc codestyle changes. (Ken)
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
| |
This will be required for the next commit since the non-coherent path
makes use of the fragment coordinates implicitly, so they need to be
calculated.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The result of a framebuffer fetch from a multisample FBO is inherently
per-sample, so the spec requires at least those sections of the shader
that depend on the framebuffer fetch result to be executed once per
sample.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Unfortunately due to the inconsistent meaning of some surface state
structure fields, we cannot re-use the same binding table entries for
sampling from and rendering into the same set of render buffers, so we
need to allocate a separate binding table block specifically for
render target reads if the non-coherent path is in use.
The slight noise is due to the change of
brw_assign_common_binding_table_offsets to return the next available
binding table index rather than void.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Some of the following changes in this series are specific to the
non-coherent path, so I need some way to tell whether the coherent or
non-coherent path is in use. The flag defaults to the value of the
gl_extensions::MESA_shader_framebuffer_fetch enable so that it can be
overridden easily on hardware that supports both framebuffer fetch
extensions in order to test the non-coherent path, like:
MESA_EXTENSION_OVERRIDE=-GL_EXT_shader_framebuffer_fetch
(Of course trying to force-enable the coherent framebuffer fetch
extension on hardware without native support won't work and lead to
assertion failures).
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This boolean flag was being used for two different things:
- To set the brw_wm_prog_data::dual_src_blend flag. Instead we can
just set it based on whether the dual_src_output register is valid,
which will be the case if the shader writes the secondary blending
color.
- To decide whether to call emit_single_fb_write() once, or in a loop
that would iterate only once, which seems pretty useless.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
This requires emitting a series of copies at the top of the program
from each output variable to the corresponding temporary. The initial
copy can be skipped for non-framebuffer fetch outputs whose initial
value is undefined, and the final copy needs to be skipped for
read-only outputs (i.e. gl_LastFragData), since it would be illegal to
emit a store output intrinsic for it.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The NIR representation of framebuffer fetch is the same as the GLSL
IR's until interface variables are lowered away, at which point it
will be translated to load output intrinsics. The GLSL-to-NIR pass
just needs to copy the bits over to the NIR program.
Reviewed-by: Kenneth Graunke <[email protected]>
|
|
|
|
| |
Based vaguely on a patch by jonasarrow on github.
|
|
|
|
|
|
|
| |
We need the source to be in r0-r3, so make a new register class for it.
It will be up to the surrounding passes to make sure that the r0-r3
allocation of its source won't conflict with anything other class
requirements on that temp.
|
|
|
|
| |
Extracted from a patch by jonasarrow on github.
|
|
|
|
|
| |
Extracted and fixed up from a patch by jonasarrow on github. This ended
up not getting used for ddx/ddy, but seems like it might still be useful.
|
|
|
|
| |
We need MUL rotates to do ddx/ddy support.
|
| |
|
|
|
|
| |
Caught problems in the upcoming DDX/DDY implementation.
|
|
|
|
|
| |
This will be used in the ddx/ddy support for "Am I the top half?" or "Am I
the left half?" checks.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Respect intel_miptree_slice::x_offset,y_offset and
intel_mipmap_tree::offset. All three may be non-zero when glReadPixels
is called on an EGLImage created from the non-base slice of a miptree.
Patch 2/2 that fixes test
'dEQP-EGL.functional.image.create.gles2_cubemap_*'.
Reported-by: Haixia Shi <[email protected]>
Diagnosed-by: Haixia Shi <[email protected]>
Cc: [email protected]
Reviewed-by: Kenneth Graunke <[email protected]>
Change-Id: I4b397b27e55a743a7094d29fb0a6a4b6b34352b0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When glEGLImageTargetRenderbufferStorageOES() was given an EGLImage
created from the non-base slice of a miptree,
intel_image_target_renderbuffer_storage() forgot to apply the intra-tile
offsets __DRIimage::tile_x,tile_y to the miptree layout.
This patch fixes the problem with a quick hack suitable for
cherry-picking. A proper fix requires more thorough plumbing in
intel_miptree_create_layout() and brw_tex_layout().
Patch 1/2 that fixes test
'dEQP-EGL.functional.image.create.gles2_cubemap_*'.
Reported-by: Haixia Shi <[email protected]>
Diagnosed-by: Haixia Shi <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
Cc: [email protected]
Change-Id: I8a64b0048a1ee9e714ebb3f33fffd8334036450b
|
|
|
|
|
|
|
|
|
|
| |
This pulls isl and genxml into a single make file so that they can properly
build in parallel. This isn't terribly important now as genxml just
generates sources which happens serially first anyway but it will be more
important as we add more stuff to src/intel.
Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Matt Turner <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The tests assumed that isl would be in the include path but that usually
isn't the case. Instead, we usually have src/intel and you need to add an
"isl/" prefix.
Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Matt Turner <[email protected]>
|
|
|
|
|
|
|
|
|
| |
If the surface has a layout of GEN4_2D then we need to compute a normal 2D
alignment and not use the magic linewar 1D alignment.
Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Chad Versace <[email protected]>
|
|
|
|
|
|
| |
Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Chad Versace <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The Sky Lake 1D layout is only used if the surface is linear. For tiled
surfaces such as depth and stencil the old gen4 2D layout is used.
Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Topi Pohjolainen <[email protected]>
Reviewed-by: Chad Versace <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In some programs, we can have very deep dominance trees and the recursion
can cause us to risk stack overflows. Instead, we replace the recursion
with a pair of loops, one at the start and one at the end. This is
functionally equivalent to what we had before and it's actually a bit
easier to read in the new form without the recursion.
Signed-off-by: Jason Ekstrand <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97225
Reviewed-by: Connor Abbott <[email protected]>
Reviewed-by: Matt Turner <[email protected]>
|