There was way too much incrementing of things going on. Instead, let's
just start everything off at the right base location, and then increment in
the loop.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
v4:
* Apply similar optimization for shared variable stores as
0cb7d7b4b7. This was causing a
OpenGLES 3.1 CTS failure, but
867c436ca8 fixes that.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
This commit pushes makes uniform offsets be terms of bytes starting with
nir_lower_io. They get converted to be in terms of vec4s or floats when we
cram them in the UNIFORM register file but reladdr remains in terms of
bytes all the way down to the point where we lower it to a pull constant
load.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The one and only place where the FS backend allows reladdr is on uniforms.
For locals, inputs, and outputs, we lower it away before the backend ever
sees it. This commit gets rid of the dead indirect handling code.
Cc: "11.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
While we correctly set output[] for composite varyings, we set completely
bogus values for output_components[], making emit_urb_writes() output
zeros instead of the actual values.
Unfortunately, our simple approach goes out the window, and we need to
recurse into structs to get the proper value of vector_elements for each
field.
Together with the previous patch, this fixes rendering in an upcoming
game from Feral Interactive.
v2: Use pointers instead of pass-by-mutable-reference (Jason, Matt).
Cc: "11.1 11.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Apparently we have literally no support for FS varying struct inputs.
This is somewhat surprising, given that we've had tests for that very
feature that have been passing for a long time.
Normally, varying packing splits up structures for us, so we don't see
them in the backend. However, with SSO, varying packing isn't around
to save us, and we get actual structs that we have to handle.
This patch changes fs_visitor::emit_general_interpolation() to work
recursively, properly handling nested structs/arrays/and so on.
(It's easier to read with diff -b, as indentation changes.)
When using the vec4 VS backend, this fixes rendering in an upcoming
game from Feral Interactive. (The scalar VS backend requires additional
bug fixes in the next patch.)
v2: Use pointers instead of pass-by-mutable-reference (Jason, Matt).
Cc: "11.1 11.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We were including it in headers, which then caused it to be included in
tons of places it wasn't needed.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Previously, we had a rescale_texcoords helper in the FS backend for
handling rescaling of texture coordinates. Now that we can do variants in
NIR, we can use nir_lower_tex to do the rescaling for us. This allows us
to delete the i965-specific code and gives us proper TEXTURE_RECTANGLE and
GL_CLAMP handling in vertex and geometry shaders.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In most cases (when the negate is copy propagated and the MOV removed),
this is two instructions on Gen >= 8 and only two instructions on
earlier platforms -- and it doesn't use the flag register.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
On the vec4 backend, textureSamplesIdentical() will always return
false. There are currently no test cases for the vec4 backend, so we
don't have much confidence in any implementation. We also don't think
anyone is likely to miss it.
v2: Handle immediate value for MCS smarter. Rebase on changes to
nir_texop_sampels_identical (missing second parameter). Suggested by
Jason.
v3: Add Neil's code to handle 16x MSAA in the FS. Also rebase on top of
f9a9ba5e. Stub out the vec4 implementation.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Neil Roberts <neil@linux.intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> [v2]
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz> [v2]
This is the NIR analog to GLSL IR ir_samples_identical.
v2: Don't add the second nir_tex_src_ms_index parameter. Suggested by
Ken and Jason.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
W/UW immediates are 16-bits, but those 16-bits must be replicated
in the high 16-bits of the 32-bit field.
Remove the useless W/UW immediate saturating code, since we'll now be
using the appropriate immediate (and W/UW immediates in the IR can now
no longer be larger than 16-bits).
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Cuts 10k of .text, of which only 776 bytes are the fs_reg constructor
implementations themselves.
text data bss dec hex filename
5204535 214112 27784 5446431 531b1f i965_dri.so before
5193977 214112 27784 5435873 52f1e1 i965_dri.so after
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This allows arbitrary non-constant indices on GS input arrays,
both for the vertex index, and any array offsets beyond that.
All indirects are handled via the pull model. We could potentially
handle indirect addressing of pushed data as well, but it would add
additional code complexity, and we usually have to pull inputs anyway
due to the sheer volume of input data. Plus, marking pushed inputs
as live due to indirect addressing could exacerbate register pressure
problems pretty badly. We'd need to be careful.
v2: Use updated MOV_INDIRECT opcode.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
The 2-bit hardware register file field is ARF, GRF, MRF, IMM.
Rename GRF to VGRF (virtual GRF) so that we can reuse the GRF name to
mean an assigned general purpose register.
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The test (file == BAD_FILE) works on registers for which the constructor
has not run because BAD_FILE is zero. The next commit will move
BAD_FILE in the enum so that it's no longer zero.
In the case of this->outputs, the constructor was being run implicitly,
and we were unnecessarily memsetting is to zero.
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
FS_OPCODE_GET_BUFFER_SIZE is calculated with a resinfo's sampler message.
This patch adjusts the number of registers written by the opcode
following what the PRM spec says about the number of registers written
by the SIMD8 and SIMD16's writeback messages for sampler messages.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Many intrinsics only apply to a particular stage (such as discard).
In other cases, we may want to interpret them differently based on
the stage (such as load_primitive_id or load_input).
The current method isn't that pretty - we handle all intrinsics in
one giant function. Sometimes we assert on stage, sometimes we forget.
Different behaviors are handled via if-ladders based on stage.
This commit introduces new nir_emit_<stage>_intrinsic() functions,
and makes nir_emit_instr() call those. In turn, those fall back to
the generic nir_emit_intrinsic() function for cases they don't want
to handle specially.
This makes it clear which intrinsics only exist in one stage, and makes
it easy to handle inputs/outputs differently for various stages.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
For these nir intrinsics, we emit the same code as
nir_intrinsic_memory_barrier:
* nir_intrinsic_memory_barrier_atomic_counter
* nir_intrinsic_memory_barrier_buffer
* nir_intrinsic_memory_barrier_image
We treat these nir intrinsics as no-ops:
* nir_intrinsic_group_memory_barrier
* nir_intrinsic_memory_barrier_shared
v3:
* Add comment for no-op cases (curro)
v4:
* Moving comment to a separate patch authored by curro
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The scalar VS backend has never handled float[] and vec2[] outputs
correctly (my original code was broken). Outputs need to be padded
out to vec4 slots.
In fs_visitor::nir_setup_outputs(), we tried to process each vec4 slot
by looping from 0 to ALIGN(type_size_scalar(type), 4) / 4. However,
this is wrong: type_size_scalar() for a float[2] would return 2, or
for vec2[2] it would return 4. This looked like a single slot, even
though in reality each array element would be stored in separate vec4
slots.
Because of this bug, outputs[] and output_components[] would not get
initialized for the second element's VARYING_SLOT, which meant
emit_urb_writes() would skip writing them. Nothing used those values,
and dead code elimination threw a party.
To fix this, we introduce a new type_size_vec4_times_4() function which
pads array elements correctly, but still counts in scalar components,
generating correct indices in store_output intrinsics.
Normally, varying packing avoids this problem by turning varyings into
vec4s. So this doesn't actually fix any Piglit or dEQP tests today.
However, if varying packing is disabled, things would be broken.
Tessellation shaders can't use varying packing, so this fixes various
tcs-input Piglit tests on a branch of mine.
v2: Shorten the implementation of type_size_4x to a single line (caught
by Connor Abbott), and rename it to type_size_vec4_times_4()
(renaming suggested by Jason Ekstrand). Use type_size_vec4
rather than using type_size_vec4_times_4 and then dividing by 4.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Do it in the visitor, like we do for other opcodes.
v2: use const, get rid of useless surf_index temporary (Curro)
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Right now the generator marks direct surfaces as used but leaves marking of
indirect surfaces to the caller. Just make the callers handle marking in both
cases for consistency.
v2: Use const and remove useless surf_index temporary (Curro)
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
V3: clamp array index to the correct size (the size of the current array
rather than the inner array) Francisco Jerez.
V2: avoid useless zero-initialization and addition for the first AoA level,
avoid redundant temporary, make use of type_size_scalar(), rename aoa_size
to element_size, assign the indirect indexing temporary directly to
image.reladdr, and replace while loop with a for loop. All suggested
by Francisco Jerez.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
This is hidden behind INTEL_SCALAR_GS=1 for now, as we don't yet support
instanced geometry shaders, and Orbital Explorer's shader spills like
crazy. But the infrastructure is in place, and it's largely working.
v2: Lots of rebasing.
v3: (feedback from Kristian Høgsberg)
- Handle stride and subreg_offset correctly for ATTRs; use a helper.
- Fix missing emit_shader_time_end() call.
- Delete dead code after early EOT in static vertex case to avoid
tripping asserts in emit_shader_time_end().
- Use proper D/UD type in intexp2().
- Fix "EndPrimitve" and "to that" typos.
- Assert that invocations == 1 so we know this is missing.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
v2:
- Add a few const qualifiers for good measure.
- Drop unneeded retype()s (Matt)
- Convert timestamp to SIMD8/16, as fs_visitor::get_timestamp() returns
SIMD4 (Connor)
v3:
- Remove unneeded temporary + MOV (Connor)
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Normally, we could read gl_Layer from bits 26:16 of R0.0. However, the
specification requires that bogus out-of-range 32-bit values written by
previous stages need to appear in the fragment shader as-written.
Instead, we pass in the full 32-bit value from the VUE header as an
extra flat-shaded varying. We have the SF override the value to 0
when the previous stage didn't actually write a value (it's actually
defined to return 0).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Write groups of enabled components together.
Signed-off-by: Kristian Høgsberg Kristensen <krh@bitplanet.net>
Now that we don't read each component one-by-one, we don't need the
temoprary vgrf for the offset. More importantly, this register was type
UD while the nir source was type D. This broke copy propagation and left
a redundant MOV in the generated code.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Kristian Høgsberg Kristensen <krh@bitplanet.net>
The emit_untyped_read and emit_untyped_write helpers already uniformize
the surface index argument. No need to do it before calling them.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Kristian Høgsberg Kristensen <krh@bitplanet.net>
Instead of looping through single-component reads, read all components
in one go.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Signed-off-by: Kristian Høgsberg Kristensen <krh@bitplanet.net>
v2: remove useless source_stencil_to_render_target (Ken)
Squash in the actual packing function, which also got to
v2:
Move the definition of the OPCODE outside of FB_WRITE opcodes (Matt)
Reorder the regioning to be in VWH order (Matt)
Don't retype src in the backend, just assert instead (Matt)
Rename the debug prints to something better (Matt)
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It's stored in bits 31:27 of g1 (along with the URB handles).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
NIR is typeless so this is the only way to keep track of the
type to select the proper atomic to use.
v2:
- Use imin,imax,umin,umax for the intrinsic names (Connor Abbott)
- Change message for unreachable paths (Michael Schellenberger)
Tested-by: Markus Wick <markus@selfnet.de>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, we used nir_lower_io with the scalar type_size function,
which mapped VERT_ATTRIB_* locations to...some numbers. Then, in
fs_visitor::nir_setup_inputs(), we created temporaries indexed by
those numbers, and emitted MOVs from the actual ATTR registers to
those temporaries. Virtually all of these were copy propagated away,
but it's still ugly.
This patch reworks our input lowering to produce NIR lower_input
intrinsics that properly index into the ATTR file, so we can access
it directly.
No changes in shader-db.
v2: Fix unreachable() message (Ken), update commit message (Matt).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
If a non-const sample number is given to interpolateAtSample it will
now generate an indirect send message with the sample ID similar to
how non-const sampler array indexing works. Previously non-const
values were ignored and instead it ended up using a constant 0 value.
The generator will try to determine if the sample ID is dynamically
uniform via nir_src_is_dynamically_uniform. If not it will query the
pixel interpolator in a loop, once for each different live sample
number. The next live sample number is found using emit_uniformize. If
multiple live channels have the same sample number then they will be
handled in a single iteration of the loop. The loop is necessary
because the indirect send message doesn't seem to have a way to
specify a different value for each fragment.
This fixes the following two Piglit tests:
arb_gpu_shader5-interpolateAtSample-nonconst
arb_gpu_shader5-interpolateAtSample-dynamically-nonuniform
v2: Handle dynamically non-uniform sample ids.
v3: Remove the BREAK instruction and predicate the WHILE directly.
Make the tokens arrays const. (Matt Turner)
v4: Iterate over the live channels instead of each possible sample
number.
v5: Don't special case immediate values in
brw_pixel_interpolator_query. Make a better wrapper for the
function to set up the PI send instruction. Ensure that the SHL
instructions are scalar. (Francisco Jerez).
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Unfortunately, we can't get rid of them entirely. The FS backend still
needs gl_program for handling TEXTURE_RECTANGLE. The GS vec4 backend still
needs gl_shader_program for handling transfom feedback. However, the VS
needs neither and we can substantially reduce the amount they are used.
One day we will be free from their tyranny.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The texunit variable we create and assign in nir_emit_texture gets passed
through two more layers of function calls before it gets to its sole use in
rescale_texcoord. The best part is that we already pass the sampler into
rescale_texcoord so we can just look it up there.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
