This lowering isn't needed for RADV because AMDGCN has two
instructions. It will be disabled for RADV in an upcoming series.
While we are at it, factorize a little bit.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Fixes: 58bcebd987 ("spirv: Allow [i/u]mulExtended to use new nir opcode")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
the naming is a bit confusing no matter how you look at it. Within SPIR-V
"global" memory is memory accessible from all threads. glsl "global" memory
normally refers to shader thread private memory declared at global scope. As
we already use "shared" for memory shared across all thrads of a work group
the solution where everybody could be happy with is to rename "global" to
"private" and use "global" later for memory usually stored within system
accessible memory (be it VRAM or system RAM if keeping SVM in mind).
glsl "local" memory is memory only accessible within a function, while SPIR-V
"local" memory is memory accessible within the same workgroup.
v2: rename local to function as well
v3: rename vtn_variable_mode_local as well
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
For now, it's hidden behind a cap. Hopefully, we can eventually drop
that along with all the manual offset code in spirv_to_nir.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Instead of baking in uvec2 for UBO and SSBO pointers and uint for push
constant and shared memory pointers, make it configurable.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
We're going to want to do more deref optimizations going forward and
this gives us a central place to do them. Also, cast propagation will
get a bit more complicated with the addition of ptr_as_array derefs.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
This workaround has been introduced by 135e4d434f for fixing
DXVK GPU hangs with many games. It is no longer needed since
LLVM r345718.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
On some GPUs, especially older Intel GPUs, some math instructions are
very expensive. On those architectures, don't reduce flow control to a
csel if one of the branches contains one of these expensive math
instructions.
This prevents a bunch of cycle count regressions on pre-Gen6 platforms
with a later patch (intel/compiler: More peephole select for pre-Gen6).
v2: Remove stray #if block. Noticed by Thomas.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
That flow control may be trying to avoid invalid loads. On at least
some platforms, those loads can also be expensive.
No shader-db changes on any Intel platform (even with the later patch
"intel/compiler: More peephole select").
v2: Add a 'indirect_load_ok' flag to nir_opt_peephole_select. Suggested
by Rob. See also the big comment in src/intel/compiler/brw_nir.c.
v3: Use nir_deref_instr_has_indirect instead of deref_has_indirect (from
nir_lower_io_arrays_to_elements.c).
v4: Fix inverted condition in brw_nir.c. Noticed by Lionel.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
User are encouraged to switch to LLVM 7.0 released in September 2018.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Helpful for debugging compiler backend problems: this allows us to
easily retrieve the LLVM IR from RenderDoc.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This implementation should work and potential bugs can be
fixed during the release candidates window anyway.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Totals from affected shaders:
SGPRS: 1112 -> 1112 (0.00 %)
VGPRS: 1492 -> 1196 (-19.84 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 112172 -> 101316 (-9.68 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 93 -> 98 (5.38 %)
Wait states: 0 -> 0 (0.00 %)
All affected shaders are from "Batman: Arkham City" over DXVK.
The pass detects that the temporary array created by DXVK for
storing TCS inputs is a copy of the input arrays and allows
us to avoid copying all of the input data and then indirecting
on it with if-ladders, instead we just do indirect indexing.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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>
Otherwise, nir_lower_clip_cull_distance_arrays might report
wrong number of output clips/culls because it relies on
shader output variables and some of them might be dead.
This fixes a rendering issue with Dolphin and Super Mario
Sunshine.
Fixes: b0c643d8f5 ("spirv: Use NIR per-member splitting")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107610
CC: 18.2 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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 is prep work for moving this to a per-thread struct
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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>
Signed-off-by: Rob Clark <robdclark@gmail.com>
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>
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>