This is essentially the same problem fixed in an earlier patch for
immediates. Setting the stride to zero will be particularly useful
for my future SIMD lowering pass, because we will be able to just
check whether the stride of a source register is zero and skip
emitting the copies required to unzip it in that case.
Instead of setting stride to zero in every caller of emit_uniformize()
I've changed the function to return the result as its return value
(previously it was being written into a caller-provided destination
register), because this way we can enforce that the result is used with
the correct regioning from the function itself.
The changes to the prototype of its VEC4 counterpart are mainly for
the sake of symmetry, VEC4 registers don't have stride.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
This fixes essentially the same problem as for immediates. Registers
of the UNIFORM file are typically accessed according to the formula:
read_uniform(r, channel_index, array_index) =
read_element(r, channel_index * 0 + array_index * 1)
Which matches the general direct addressing formula for stride=0:
read_direct(r, channel_index, array_index) =
read_element(r, channel_index * stride +
array_index * max{1, stride * width})
In either case if reladdr is present the access will be according to
the composition of two register regions, the first one determining the
per-channel array_index used for the second, like:
read_indirect(r, channel_index, array_index) =
read_direct(r, channel_index,
read(r.reladdr, channel_index, array_index))
where:
read(r, channel_index, array_index) = if r.reladdr == NULL
then read_direct(r, channel_index, array_index)
else read_indirect(r, channel_index, array_index)
In conclusion we can handle uniforms consistently with the other
register files if we set stride to zero. After lowering to a GRF
using VARYING_PULL_CONSTANT_LOAD in demote_pull_constant_loads() the
stride of the source is set to one again because the result of
VARYING_PULL_CONSTANT_LOAD is generally non-uniform.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
When the width field was removed from fs_reg the BROADCAST handling
code in opt_algebraic() started to miss a number of trivial
optimization cases resulting in the ugly indirect-addressing sequence
to be emitted unnecessarily for some variable-indexed texturing and
UBO loads regardless of one of the sources of BROADCAST being
immediate. Apparently the reason was that we were setting the stride
field to one for immediates even though they are typically uniform.
Width used to be set to one too which is why this optimization used to
work previously until the "reg.width == 1" check was removed.
The stride field of vector immediates is intentionally left equal to
one, because they are strictly speaking not uniform. The assertion in
fs_generator makes sure that immediates have the expected stride as
consistency check.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
We only consider a vgrf defined by a given block if the block writes to it
unconditionally. So far we have been checking this by testing that the
instruction is not predicated, however, in the case of BRW_OPCODE_SEL,
the predication is used to select the value to write, not to decide if
the write is actually done. The consequence of this was increased life
spans for affected vgrfs, which could lead to additional register pressure.
Since NIR generates selects for conditional writes this was causing massive
register pressure in a handful of piglit and dEQP tests that had a large
number of select operations with the NIR-vec4 backend.
Fixes the following piglit tests with the NIR-vec4 backend:
spec/glsl-1.50/execution/variable-indexing/vs-output-array-vec4-index-wr-before-gs
spec/glsl-1.50/execution/variable-indexing/gs-input-array-vec4-index-rd
spec/glsl-1.50/execution/variable-indexing/vs-output-array-vec2-index-wr-before-gs
spec/glsl-1.50/execution/variable-indexing/vs-output-array-vec3-index-wr-before-gs
spec/glsl-1.50/execution/variable-indexing/vs-output-array-float-index-wr-before-gs
Fixes 80 dEQP tests with the NIR-vec4 backend in the following category:
dEQP-GLES3.functional.ubo.*
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This fixes crashes in llvmpipe with LLVM 3.8 and also some piglit tests
on radeonsi that use the draw module.
This is just a temporary solution. The correct solution will require
creating a TargetMachine during gallivm initialization and pulling the
DataLayout from there. This will be a somewhat invasive change, and it
will need to be validatated on multiple LLVM versions.
https://llvm.org/bugs/show_bug.cgi?id=24172
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This fixes a compilation warning introduced in commit 05a12c5
(gallium: add interface for writable shader images).
While we are at it, fix indentation and rename parameters according to
the gallium interface.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Only constbufs must be aligned to 0x100, but since all buffers can be
rebinded as constant buffers they must be also aligned.
This patch prevents this behaviour by aligning everything to 256-byte
increments at buffer creation.
This fixes dmesg fails for the following piglit test:
ext_transform_feedback-immediate-reuse-uniform-buffer -auto -fbo
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
NV50_3D_BIND_TSC only allows to bind 16 samplers, and since we don't
want to do anything with NV50_3D_BIND_TSC2, just limit the maximum
number of samplers to 16 like for nvc0.
This fixes dmesg fails with the following piglit test:
max-samplers
But the test still fails.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Fixes the following piglit test:
occlusion_query_meta_no_fragments
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
As functions are inlined, and nir_lower_global_vars_to_local gets
run, all global variables are lowered to local variables.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It appears that the G80 did not have support for the sampler view
first/last clamping. Put the view's last level in the place of the
texture's so that it doesn't go past what the sampler view allows.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org
There's no need to deal with samplers for texture size queries. That
code also was accidentally setting an invalid sIndirectSrc position, but
it can now just be removed.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org
Switch off hardware-generated binding tables and gather push
constants in the blorp. Blorp requires only a minimal set of
simple constants. There is no need for the extra complexity
to program a gather table entry into the pipeline.
Cc: kenneth@whitecape.org
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
When hardware-generated binding tables are enabled, use the hw-generated
binding table format when uploading binding table state.
Normally, the CS will will just consume the binding table pointer commands
as pipelined state. When the RS is enabled however, the RS flushes whatever
edited surface state entries of our on-chip binding table to the binding
table pool before passing the command on to the CS.
Note that the the binding table pointer offset is relative to the binding table
pool base address when resource streamer instead of the surface state base address.
v2: Fix possible buffer overflow when allocating a chunk out of the
hw-binding table pool (Ken).
v3: Remove extra newline and add missing brace around if-statement (Matt).
v4: Fix broken INTEL_DEBUG=shader_time for hw-generated binding tables.
Document PRM WaStateBindingTableOverfetch workaround.
Cc: kenneth@whitecape.org
Cc: mattst88@gmail.com
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Unlike normal software binding tables where the driver has to manually
generate and fill a binding table array which are then uploaded to the
hardware, the resource streamer instead presents the driver with an option
to fill out slots for individual binding table indices. The hardware
accumulates the state for these combined edits which it then automatically
flushes to a binding table pool when the binding table pointer state
command is invoked.
v2: Clarify binding table edit bit aligment (Topi).
v3: Make comments and function names more clearer (Ken).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
This patch implements the binding table enable command which is also
used to allocate a binding table pool where where hardware-generated
binding table entries are flushed into. Each binding table offset in
the binding table pool is unique per each shader stage that are
enabled within a batch.
Also insert the required brw_tracked_state objects to enable
hw-generated binding tables in normal render path.
v2: - Use MOCS in binding table pool alloc for GEN8
- Fix spurious offset when allocating binding table pool entry
and start from zero instead.
v3: - Include GEN8 fix for spurious offset above.
v4: - Fixup wrong packet length in enable/disable hw-binding table
for GEN8 (Ville).
- Don't invoke HW-binding table disable command when we dont
have resource streamer (Chris).
v5: - Reorder the state cache invalidate flush so it happens in-between
enabling hw-generated binding tables and the previous sw-binding
table GPU state (Chris).
v6: - Do the same fix in v5 for gen7_disable_hw_binding_tables().
- Adhere to coding guidelines and make comments more informative.
Cc: kenneth@whitecape.org
Cc: syrjala@sci.fi
Cc: chris@chris-wilson.co.uk
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Check first if the hardware and kernel supports resource streamer. If this
is allowed, tell the kernel to enable the resource streamer enable bit on
MI_BATCHBUFFER_START by specifying I915_EXEC_RESOURCE_STREAMER
execbuffer flags.
v2: - Use new I915_PARAM_HAS_RESOURCE_STREAMER ioctl to check if kernel
supports RS (Ken).
- Add brw_device_info::has_resource_streamer and toggle it for
Haswell, Broadwell, Cherryview, Skylake, and Broxton (Ken).
v3: - Update I915_PARAM_HAS_RESOURCE_STREAMER to match updated kernel.
v4: - Always inspect the getparam.value (Chris Wilson).
v5: - Fold redundant devinfo->has_resource_streamer check in context create
into init screen.
Cc: kenneth@whitecape.org
Cc: chris@chris-wilson.co.uk
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
This gives the kernel a chance to validate and lock down the data,
without having to deal with mmap zapping.
With this, GLBenchmark stops on a texture relocations, because we'd
recycled a shader BO as another shader and failed to revalidate, since we
weren't clearing the cached validation state on mmap faults.
In particular, we were incorrectly accepting s3tc (and lots of others)
for CompressedTexSubImage3D (but not CompressedTexImage3D) calls with 3d
targets. At this time, the only allowed formats for these calls are the
bptc ones, since none of the specific extensions allow it (astc hdr would).
Also, fix up a bug in _mesa_target_can_be_compressed - 3d target needs to
be allowed for bptc formats.
Reviewed-by: Brian Paul <brianp@vmware.com>
On a release build, this makes the rest of vc4_qpu_validate.c go away
(the compiler didn't know that our qpu helper function calls had no
side effects).
These are really useful hints to the compiler in the absence of link-time
optimization, and I'm going to use them in VC4.
I've made the const attribute be ATTRIBUTE_CONST unlike other function
attributes, because we have other things in the tree #defining CONST for
their own unrelated purposes.
v2: Alphabetize.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
Before, we were setting payload_last_use_ip for unused payload
registers to 0, which made them interfere with whatever the first
instruction wrote to due to the workaround for SIMD16 uniform arguments.
Just use -1 to mean "unused" instead, and then skip setting any
interferences for unused payload registers.
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
GAINED: 1
LOST: 0
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
regs_read() will handle LINTERP for us since the previous commit. In
addition, we were being too conservative, since it will only read 2
registers on SIMD8.
instructions in affected programs: 9061 -> 8893 (-1.85%)
helped: 10
HURT: 0
GAINED: 0
LOST: 0
All of the changes were due to spills being eliminated, mostly in KSP
shaders.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
If a source register in the push constant registers uses more than one
register, then we wouldn't update payload_last_use_ip for subsequent
registers.
Unlike most uniform data pushed into registers, the CS gl_LocalInvocationID
data varies per execution channel. Therefore for SIMD16 mode, we have vec16
data in the payload. In this case we then need to mark 2 registers in
payload_last_use_ip as last used by the instruction. There's a similar
situation for the z and w coordinates of gl_FragCoord for fragment shaders,
where it had only happened to work before because of some bogus interferences
which the next commit removes.
(Connor: added bit about gl_FragCoord to commit message)
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Connor Abbott <connor.w.abbott@intel.com>
The second source always stays within the same SIMD8 register.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
This prevents an assertion failure in brw_fs_live_variables.cpp,
fs_live_variables::setup_one_write: Assertion `var < num_vars' failed.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This prevents an assertion failure in brw_fs_live_variables.cpp,
fs_live_variables::setup_one_read: Assertion `var < num_vars' failed.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
A fragment program from "Pixel Piracy" contains redundant OPTION
directives:
!!ARBfp1.0
OPTION ARB_precision_hint_fastest;
OPTION ARB_fog_exp2;
OPTION ARB_precision_hint_fastest;
OPTION ARB_fog_exp2;
...
We already allow redundant ARB_precision_hint_fastest directives, but
disallow the redundant (yet consistent) ARB_fog_exp2 directives, failing
to compile the program.
The specification seems to contradict itself - the main text says that
only one fog application option may be specified, but then backpedals,
indicating the intent is to disallow /contradictory/ flags. One of the
issues suggests that specifying contradictory ones is stupid, but
allowed, and only the last one should take effect.
Accepting multiple redundant (but consistent) directives seems harmless,
and like a reasonable interpretation of the specification. It also
fixes a fragment program found in the wild.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
With the last few patches a way was provided to influence lower layer miptree
layout and allocation decisions via flags (replacing bools). For simplicity, I
chose not to touch the tiling requests because the change was slightly less
mechanical than replacing the bools.
The goal is to organize the code so we can continue to add new parameters and
tiling types while minimizing risk to the existing code, and not having to
constantly add new function parameters.
v2: Rebased on Anuj's recent Yf/Ys changes
Fix non-msrt MCS allocation (was only happening in gen8 case before)
v3: small fix in assertion requested by Chad
v4: Use parens to get the order right from v3.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
With the last few patches a way was provided to influence lower layer miptree
layout and allocation decisions via flags (replacing bools). For simplicity, I
chose not to touch the tiling requests because the change was slightly less
mechanical than replacing the bools.
The goal is to organize the code so we can continue to add new parameters and
tiling types while minimizing risk to the existing code, and not having to
constantly add new function parameters.
v2: Rebased on Anuj's recent Yf/Ys changes
Fix non-msrt MCS allocation (was only happening in gen8 case before)
v3: small fix in assertion requested by Chad
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com> (v2)
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com> (v2)
Reviewed-by: Chad Versace <chad.versace@intel.com> (v2)
This in principle simple calculation was being open-coded in a number
of places (in a series I haven't yet sent for review there will be a
couple more), all of them were subtly broken in one way or another:
None of them were handling the HW_REG case correctly as pointed out by
Connor, and fs_inst::regs_read() was handling the stride=0 case rather
naively. This patch solves both problems and factors out the
calculation as a new fs_reg method.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This gets rid of two no16() fall-backs and should allow better
scheduling of the generated IR. There are no uses of usubBorrow() or
uaddCarry() in shader-db so no changes are expected. However the
"arb_gpu_shader5/execution/built-in-functions/fs-usubBorrow" and
"arb_gpu_shader5/execution/built-in-functions/fs-uaddCarry" piglit
tests go from 40 to 28 instructions. The reason is that the plain ADD
instruction can easily be CSE'ed with the original addition, and the
b2i negation can easily be propagated into the source modifier of
another instruction, so effectively both operations are performed with
just one instruction.
v2: Rely on carry_to_arith() and borrow_to_arith() to lower these
(Ilia Mirkin).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Booleans are represented as 0/-1 on modern hardware which means we can
just negate them to convert them into a numeric type. Negation has
the benefit that it can be implemented using a source modifier which
can easily be propagated into some other instruction. shader-db
results on HSW:
total instructions in shared programs: 6349082 -> 6346693 (-0.04%)
instructions in affected programs: 40948 -> 38559 (-5.83%)
helped: 123
HURT: 1
GAINED: 1
LOST: 0
Reviewed-by: Matt Turner <mattst88@gmail.com>
PIPE_CAPs and TGSI support will be added later. The TGSI support should be
straightforward. We only need to split TGSI_FILE_RESOURCE into TGSI_FILE_IMAGE
and TGSI_FILE_BUFFER, though duplicating all opcodes shouldn't be necessary.
The idea is:
* ARB_shader_image_load_store should use set_shader_images.
* ARB_shader_storage_buffer_object should use set_shader_buffers(slots 0..M-1)
if M shader storage buffers are supported.
* ARB_shader_atomic_counters should use set_shader_buffers(slots M..N)
if N-M+1 atomic counter buffers are supported.
PIPE_CAPs can describe various constraints for early DX11 hardware.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
