This allows NIR to CSE more operations. LLVM does this also so the
impact is limited, however doing this in NIR allows other opts to
make progress. For example in radeonsi more loops are unrolled in
Civilization Beyond Earth.
The actual pipeline-db stats are not overwhelming but even in the
negatively affected shaders the NIR is clearly better. It just
happens that the code shuffling and in some cases calls to max
rather than a flt result in the final output from LLVM not
giving as good numbers.
However this is an incremental opt that further passes build off
so the change should be made IMO.
Totals from affected shaders:
SGPRS: 20192 -> 20184 (-0.04 %)
VGPRS: 19516 -> 19524 (0.04 %)
Spilled SGPRs: 437 -> 444 (1.60 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 1527444 -> 1522276 (-0.34 %) bytes
LDS: 6 -> 6 (0.00 %) blocks
Max Waves: 1018 -> 1016 (-0.20 %)
Wait states: 0 -> 0 (0.00 %)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
All CTS pass on Polaris/Vega with LLVM 6, 7 and master, so
I think it's safe to enable the feature.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Not sure if this is all wired up. CTS does pass and the Tangrams
demo works fine on Vega. There are corruption issues on Polaris
but not sure if that related to 16-bit support.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
It was very inconsistently handled; the only things that made use of it
were glsl_to_nir, glspirv, and nir_gather_info. In particular,
nir_lower_io completely ignored it so anyone using nir_lower_io on
64-bit vertex attributes was going to be in for a shock. Also, as of
the previous commit, it's set by every driver that supports 64-bit
vertex attributes. There's no longer any reason to have it be an option
so let's just delete it.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
The code sizes return here get passed to the cache shader insert function,
which then memcpy from the code ptr, and causes all sorts of valgrind
errors like:
==6755== Invalid read of size 8
==6755== at 0x4C32FEE: memcpy@GLIBC_2.2.5 (vg_replace_strmem.c:1021)
==6755== by 0x2305D4C7: radv_pipeline_cache_insert_shaders (radv_pipeline_cache.c:416)
==6755== by 0x2305791D: radv_create_shaders (radv_pipeline.c:2158)
==6755== by 0x2305C523: radv_pipeline_init (radv_pipeline.c:3404)
==6755== by 0x2305C890: radv_graphics_pipeline_create (radv_pipeline.c:3515)
==6755== by 0x230188AB: radv_device_init_meta_blit_color (radv_meta_blit.c:871)
==6755== by 0x2301D50E: radv_device_init_meta_blit_state (radv_meta_blit.c:1278)
==6755== by 0x23011893: radv_device_init_meta (radv_meta.c:352)
==6755== by 0x2300744B: radv_CreateDevice (radv_device.c:1576)
==6755== by 0x5187D0F: ??? (in /usr/lib64/libvulkan.so.1.1.77)
==6755== by 0x518F6A3: ??? (in /usr/lib64/libvulkan.so.1.1.77)
==6755== by 0x5192A42: vkCreateDevice (in /usr/lib64/libvulkan.so.1.1.77)
==6755== Address 0x22a58548 is 4 bytes after a block of size 116 alloc'd
==6755== at 0x4C2EBAB: malloc (vg_replace_malloc.c:299)
==6755== by 0x23089DC4: ac_elf_read (ac_binary.c:144)
==6755== by 0x23090A60: ac_compile_module_to_binary (ac_llvm_helper.cpp:162)
==6755== by 0x23053F06: compile_to_memory_buffer (radv_llvm_helper.cpp:58)
==6755== by 0x23053F06: radv_compile_to_binary (radv_llvm_helper.cpp:98)
==6755== by 0x23052769: ac_llvm_compile (radv_nir_to_llvm.c:3394)
==6755== by 0x23052823: ac_compile_llvm_module (radv_nir_to_llvm.c:3418)
==6755== by 0x23053C05: radv_compile_nir_shader (radv_nir_to_llvm.c:3542)
==6755== by 0x23061B4E: shader_variant_create (radv_shader.c:580)
==6755== by 0x23061CFD: radv_shader_variant_create (radv_shader.c:634)
==6755== by 0x23057765: radv_create_shaders (radv_pipeline.c:2123)
==6755== by 0x2305C523: radv_pipeline_init (radv_pipeline.c:3404)
==6755== by 0x2305C890: radv_graphics_pipeline_create (radv_pipeline.c:3515)
Since we are just inserting the code into the cache, we can avoid these
bad reads and data in the cache by just using the binary code size here.
Fixes: 939e5a382 (radv: add padding for the UMR disassembler)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This uses the common compiler passes abstraction to help radv
avoid fixed cost compiler overheads. This uses a linked list per
thread stored in thread local storage, with an entry in the list
for each target machine.
This should remove all the fixed overheads setup costs of creating
the pass manager each time.
This takes a demo app time to compile the radv meta shaders on nocache
and exit from 1.7s to 1s. It also has been reported to take the startup
time of uncached shaders on RoTR from 12m24s to 11m35s (Alex)
v2: fix llvm6 build, inline emit function, handle multiple targets
in one thread
v3: rebase and port onto new structure
v4: rename some vars (Bas)
v5: drag all code into radv for now, we can refactor it out later
for radeonsi if we make it shareable
v6: use a bit more C++ in the wrapper
v7: logic bugs fixed so it actually runs again.
v8: rebase on top of radeonsi changes.
v9: drop some C++ headers, cleanup list entry
v10: use pop_back (didn't have enough caffeine)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This ports radv to the shared code, however due to a bug in LLVM
version prior to 7, radv cannot add target info at this stage,
as it would leak one for every shader compile, however I'd prefer
to keep this llvm damage in the shared code, since it isn't the
driver at fault here. We just add a flag to denote if the driver
can support leaking the target info or not, and the common code
does the right thing depending on the llvm version.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This just splits out the non-shared code and reuses ac_get_llvm_target in radv.
v2: rebase on Marek's patch - fixup brace position/whitespace
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The driver already supports exporting the stencil value.
The following CTS test now pass:
dEQP-VK.pipeline.shader_stencil_export.op_replace
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This commit completely reworks function calls in NIR. Instead of having
a set of variables for the parameters and return value, nir_call_instr
now has simply has a number of sources which get mapped to load_param
intrinsics inside the functions. It's up to the client API to build an
ABI on top of that. In SPIR-V, out parameters are handled by passing
the result of a deref through as an SSA value and storing to it.
This virtue of this approach can be seen by how much it allows us to
delete from core NIR. In particular, nir_inline_functions gets halved
and goes from a fairly difficult pass to understand in detail to almost
trivial. It also simplifies spirv_to_nir somewhat because NIR functions
never were a good fit for SPIR-V.
Unfortunately, there is no good way to do this without a mega-commit.
Core NIR and SPIR-V have to be changed at the same time. This also
requires changes to anv and radv because nir_inline_functions couldn't
handle deref instructions before this change and can't work without them
after this change.
Acked-by: Rob Clark <robdclark@gmail.com>
Acked-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Acked-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Before, we were doing structure splitting in spirv_to_nir.
Unfortunately, this doesn't really work when you think about passing
struct pointers into functions. Doing it later in NIR is a much better
plan.
Acked-by: Rob Clark <robdclark@gmail.com>
Acked-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Acked-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Workaround for bug in llvm that causes the GPU to hang in presence
of nested loops because there is an exec mask issue. The proper
solution is to fix LLVM but this might require a bunch of work.
This fixes a bunch of GPU hangs that happen with DXVK.
Vega10:
Totals from affected shaders:
SGPRS: 110456 -> 110456 (0.00 %)
VGPRS: 122800 -> 122800 (0.00 %)
Spilled SGPRs: 7478 -> 7478 (0.00 %)
Spilled VGPRs: 36 -> 36 (0.00 %)
Code Size: 9901104 -> 9922928 (0.22 %) bytes
Max Waves: 7143 -> 7143 (0.00 %)
Code size slightly increases because it inserts more branch
instructions but that's expected. I don't see any real performance
changes.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105613
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Errors are not that common of a case so we can eat a slight perf
hit in having to call a function and do a runtime check.
In turn this makes debugging random errors happening for end users
easier, because they don't have to have a debug build on hand.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This doesn't nothing special currently because we don't create
any copy_var instructions, but this is needed for the next patch.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Having an entrypoint different than "main" doesn't mean we
have multiple shaders per module.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
When VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT is set we skip NIR
linking optimisations and only run over the NIR optimisation loop
once similar to the GLSLOptimizeConservatively constant used by
some GL drivers.
We need to run over the opts at least once to avoid errors in LLVM
(e.g. dead vars it can't handle) and also to reduce the time spent
compiling the IR in LLVM.
With this change the Blacksmith Unity demos compilation times
go from 329760 ms -> 299881 ms when using Wine and DXVK.
V2: add bit to radv_pipeline_key
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=106246
The driver already supports exporting the Layer and ViewportIndex
built-ins from vertex or tessellation shaders.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Noticed while passing by. Not sure if it impacts anything, but
likely to impact GFX9 more than anything else since we lower
inputs, outputs and locals there.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
If a shader only writes to an output via a constant initializer we
need to lower it before we call nir_remove_dead_variables so that
this pass sees the stores from the initializer and doesn't kill the
output.
Fixes test failures in new work-in-progress CTS tests:
dEQP-VK.spirv_assembly.instruction.graphics.variable_init.output.float
This is ported from anv:
99b57daf4a anv/pipeline: lower constant initializers on output variables earlier
from Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
If AMD_shader_info or RADV_TRACE_FILE is used we might need to
keep trace of LLVM IR.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
