Acked-by: Dave Airlie <airlied@redhat.com>
We already use GFX9 and I don't want us to have confusing naming
in the driver. GFXn naming is better from the driver perspective,
because it's the real version of the gfx portion of the hw. Also,
CIK means Bonaire-Kaveri-Kabini, it doesn't mean CI.
It shouldn't confuse our SDMA, UVD, VCE etc. code much. Those have
nothing to do with GFXn and they have their own version numbers.
The overall goal is to support unaligned loads from vertex buffers
natively on SI.
In the unaligned case, we fall back to the general case implementation in
ac_build_opencoded_load_format. Since this function is fully general,
we will also use it going forward for cases requiring fully manual format
conversions of dwords anyway.
This requires a different encoding of the fix_fetch array, which will now
contain the entire format information if a fixup is required.
Having to check the alignment of vertex buffers is awkward. To keep the
impact on the fast path minimal, the si_context will keep track of which
vertex buffers are (not) at least dword-aligned, while the
si_vertex_elements will note which vertex buffers have some (at most dword)
alignment requirement. Vertex buffers should be dword-aligned most of the
time, which allows a fast early-out in almost all cases.
Add the radeonsi_vs_fetch_always_opencode configuration variable for
testing purposes. Note that it can only be used reliably on LLVM >= 9,
because support for byte and short load is required.
v2:
- add a missing check to si_bind_vertex_elements
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Needed to track context rolls caused by streamout and ACQUIRE_MEM.
ACQUIRE_MEM can occur outside of draw calls.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110355
v2: squashed patches and done more rework
Cc: 19.0 <mesa-stable@lists.freedesktop.org>
Move the definition of radeonsi_clear_db_cache_before_clear there,
as well as radeonsi_enable_nir.
This removes the AMD_DEBUG=nir option.
We currently still have two places for options: the driconf machinery
and AMD_DEBUG/R600_DEBUG. If we are to have a single place for options,
then the driconf machinery should be preferred since it's more flexible.
The only downside of the driconf machinery was that adding new options
was quite inconvenient. With this change, a simple boolean option can
be added with a single line of code, same as for AMD_DEBUG.
One technical limitation of this particular implementation is that while
almost all driconf features are available, the translation machinery doesn't
pick up the description strings for options added in si_debvug_options. In
practice, translations haven't been provided anyway, and this is intended
for developer options, so I'm not too worried. It could always be added
later if anybody really cares.
v2:
- use bool instead of uint8_t for options
- si_debug_options.inc -> si_debug_options.h
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
so that bound compute shader resources won't be added when they are not
needed and same for graphics.
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The OpenMAX state tracker will use this.
RadeonSI is adapted to use pipe_grid_info::last_block instead of its
internal state.
Acked-by: Leo Liu <leo.liu@amd.com>
v2: use tc.stream_uploader in si buffer_transfer_map if not called from
the driver thread
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com> (v1)
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
We should get fewer context rolls with the SET_CONTEXT_REG optimization,
but it would have been for nothing if the scissor state rolled the context
anyway. Don't emit the scissor state if there is no context roll.
R600_DEBUG=gisel will tell LLVM to use GlobalISel rather than
SelectionDAG for instruction selection.
v2: mareko: move the helper to src/amd/common
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Tom Stellard <tstellar@redhat.com>
As precursor to moving init to common code, just rename the struct
and move it.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This allows to avoid having to see garbage in Dying Light loading screen
at least, which probably expects Windows/NV behavior of all allocations
being zeroed by default.
Analogous to radv flag with the same name.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Remembering latest states of registers to eliminate redunant SET_CONTEXT_REG packets
Signed-off-by: Sonny Jiang <sonny.jiang@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Fixes truncation warning in gcc 8.1
Fixes: 8539c9bf31 ("gallium/radeon: add the kernel version into the renderer string")
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
The compiler queue was limited to 3 threads, so shader-db running
on a 16-thread CPU would have a bottleneck on the 3-thread queue.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Tested-by: Benedikt Schemmer <ben at besd.de>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
It will contain more variables.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Tested-by: Benedikt Schemmer <ben at besd.de>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reduce swizzle constraints to the ALPHA_IS_ON_MSB constraint and the clear
value of 1.
This significantly changes the DCC fast clear code, and fixes fast clear
for RGB formats without alpha.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
(This patch doesn't enable the behavior. It will be enabled in a later
commit.)
Draw calls from multiple IBs can be executed in parallel.
v2: do emit partial flushes on SI
v3: invalidate all shader caches at the beginning of IBs
v4: don't call si_emit_cache_flush in si_flush_gfx_cs if not needed,
only do this for flushes invoked internally
v5: empty IBs should wait for idle if the flush requires it
v6: split the commit
If we artificially limit the number of draw calls per IB to 5, we'll get
a lot more IBs, leading to a lot more partial flushes. Let's see how
the removal of partial flushes changes GPU utilization in that scenario:
With partial flushes (time busy):
CP: 99%
SPI: 86%
CB: 73:
Without partial flushes (time busy):
CP: 99%
SPI: 93%
CB: 81%
Tested-by: Benedikt Schemmer <ben@besd.de>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
so that the draw is started as soon as possible.
v2: only prefetch the API VS and VBO descriptors
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Using mesa OpenCL failed on a big endian PowerPC machine because
si_vgt_param_key is using bitfields and a 32 bit int for an
index into an array.
Fix si_vgt_param_key to work correctly on both little endian
and big endian machines.
Signed-off-by: Bas Vermeulen <bas@daedalean.ai>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Most files in gallium/radeon now include si_pipe.h.
chip_class and family are now here:
sscreen->info.family
sscreen->info.chip_class
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This requires out-of-band creation of fences, and will be signaled to
the pipe_context::flush implementation by a special TC_FLUSH_ASYNC flag.
v2:
- remove an incorrect assertion
- handle fence_server_sync for unsubmitted fences by
relying on the improved cs_add_fence_dependency
- only implement asynchronous flushes on amdgpu
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
It's inaccurate. Instead, see the copyright and use "git log" and
"git blame" to know the authorship.
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
It appears that flushing the DB metadata is actually not sufficient
since the driver uses the new VS blit shaders. This looks quite
strange though, but it seems like we need to flush DB for fixing
the corruption.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102955
Fixes: 69ccb9dae7 (radeonsi: use new VS blit shaders (VS inputs in SGPRs)
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
We'll use it in the scissors / clip / guardband state.
v2: avoid a performance regression on r600 when applied to
(pre-fork) stable branches
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This fixes the extremely unlikely case that an application uses
0x80000000 or 0x3f800000 as border color for an integer texture and
helps in the also, but perhaps slightly less, unlikely case that 1 is
used as a border color.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
The hardware usually does this automatically. However, we upgrade
depth to Z32_FLOAT to enable TC-compatible HTILE, which means the
hardware no longer clamps the comparison value for us.
The only way to tell in the shader whether a clamp is required
seems to be to communicate an additional bit in the descriptor
table. While VI has some unused bits in the resource descriptor,
those bits have unfortunately all been used in gfx9. So we use
an unused bit in the sampler state instead.
Fixes dEQP-GLES3.functional.texture.shadow.2d.linear.equal_depth_component32f
and many other tests in dEQP-GLES3.functional.texture.shadow.*
Fixes: d4d9ec55c5 ("radeonsi: implement TC-compatible HTILE")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
We do not enable this by default for additive blending, since it slightly
breaks OpenGL invariance guarantees due to non-determinism.
Still, there may be some applications can benefit from white-listing
via the radeonsi_commutative_blend_add drirc setting without any real
visible artifacts.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
This option enables a performance optimization where typical non-blending
draws with depth buffer may be rasterized out-of-order (on VI+, multi-SE
chips).
This optimization can lead to incorrect results when an applications
renders multiple objects with the same Z value at the same pixel, so we
will never enable it by default. But there may be applications that could
benefit from white-listing.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
This does not take commutative blending into account yet.
R600_DEBUG=nooutoforder disables it.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
When the HS wave is empty, the hardware writes the LS VGPRs starting at
v0 instead of v2. Workaround by shifting them back into place when
necessary. For simplicity, this is always done in the LS prolog.
According to the hardware team, this will be fixed in future chips,
so take that into account already.
Note that this is not a bug fix, as the bug was already worked
around by commit 166823bfd2 ("radeonsi/gfx9: add a temporary workaround
for a tessellation driver bug"). This change merely replaces the
workaround by one that should be better.
v2: add workaround code to shader only when necessary
v3: clarify the prefer_mono comment
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This fixes GL45-CTS.shader_image_load_store.basic-glsl-earlyFragTests.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This increases performance, but it was tuned for Raven, not Vega.
We don't know yet how Vega will perform, hopefully not worse.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
We have come to the conclusion that it doesn't improve performance.
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Currently, when the array is full it is resized but it can grow
over and over because we don't try to re-use descriptor slots.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Using VRAM address as bindless handles is not a good idea because
we have to use LLVMIntToPTr and the LLVM CSE pass can't optimize
because it has no information about the pointer.
Instead, use slots indexes like the existing descriptors. Note
that we use fixed 16-dword slots for both samplers and images.
This doesn't really matter because no real apps use image handles.
This improves performance with DOW3 by +7%.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Use the auto logger facility, so that CS chunks will be interleaved
with other log info.
v2:
- fix some crashes when not using CE
- fix skipping "previous" chunks of current (unflushed) IB
- fix error handling in si_begin_cs_debug
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
so that we don't rely on si_pm4_state_enabled_and_changed, allowing us
to move prefetches after draw calls.
v2: ckear the dirty mask after unbinding shaders
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de> (v1)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com> (v1)
I'd like to be able to move the prefetch call site around.
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This fixes corruption with bindless textures in Dawn Of War 3.
The do_update_surf_dirtiness mechanism was complicated and dirty_level_mask
was only updated after the first draw call. The problem is bindless textures
are checked for decompression every draw call and we would only decompress
after the first draw call. The solution is to set dirtiness after the last
draw call to the framebuffer, so the (unconditional) decompression of
bindless textures happens at the right time.
Cc: 17.2 <mesa-stable@lists.freedesktop.org>
Tested-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
For inputs and outputs, indirect indexing is lowered by the GLSL compiler.
For temporaries, use alloca and disable the "promote-alloca" pass.
In the future, we could switch all codepaths to alloca permanently and
just rely on the "promote-alloca" pass.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Use the mechanism of si_decompress_textures, but instead of doing
the actual decompression, just flag the DB cache flush there.
This removes a lot of unnecessary DB cache flushes.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Needed for updating all resident texture descriptors when
dirty_tex_counter changes.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This is done by introducing a separate list.
si_decompress_textures() is now 5x faster.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
When texture buffers are invalidated the addr in the resident
descriptor has to be updated but we can't create a new descriptor
because the resident handle has to be the same.
Instead, use the WRITE_DATA packet which allows to update memory
directly but graphics/compute have to be idle in case the GPU is
reading the descriptor.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>