WMMA = "Warp Level Matrix Multiply-Accumulate".
These are the new instructions introduced in PTX6.0 and available
on sm_70 GPUs.
Differential Revision: https://reviews.llvm.org/D38645
llvm-svn: 315601
This patch enables support for .f16x2 operations.
Added new register type Float16x2.
Added support for .f16x2 instructions.
Added handling of vectorized loads/stores of v2f16 values.
Differential Revision: https://reviews.llvm.org/D30057
Differential Revision: https://reviews.llvm.org/D30310
llvm-svn: 296032
Support for barrier synchronization between a subset of threads
in a CTA through one of sixteen explicitly specified barriers.
These intrinsics are not directly exposed in CUDA but are
critical for forthcoming support of OpenMP on NVPTX GPUs.
The intrinsics allow the synchronization of an arbitrary
(multiple of 32) number of threads in a CTA at one of 16
distinct barriers. The two intrinsics added are as follows:
call void @llvm.nvvm.barrier.n(i32 10)
waits for all threads in a CTA to arrive at named barrier #10.
call void @llvm.nvvm.barrier(i32 15, i32 992)
waits for 992 threads in a CTA to arrive at barrier #15.
Detailed description of these intrinsics are available in the PTX manual.
http://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions
Reviewers: hfinkel, jlebar
Differential Revision: https://reviews.llvm.org/D17657
llvm-svn: 293384
Summary:
Specifically, we upgrade llvm.nvvm.:
* brev{32,64}
* clz.{i,ll}
* popc.{i,ll}
* abs.{i,ll}
* {min,max}.{i,ll,u,ull}
* h2f
These either map directly to an existing LLVM target-generic
intrinsic or map to a simple LLVM target-generic idiom.
In all cases, we check that the code we generate is lowered to PTX as we
expect.
These builtins don't need to be backfilled in clang: They're not
accessible to user code from nvcc.
Reviewers: tra
Subscribers: majnemer, cfe-commits, llvm-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D28793
llvm-svn: 292694
Only scalar half-precision operations are supported at the moment.
- Adds general support for 'half' type in NVPTX.
- fp16 math operations are supported on sm_53+ GPUs only
(can be disabled with --nvptx-no-f16-math).
- Type conversions to/from fp16 are supported on all GPU variants.
- On GPU variants that do not have full fp16 support (or if it's disabled),
fp16 operations are promoted to fp32 and results are converted back
to fp16 for storage.
Differential Revision: https://reviews.llvm.org/D28540
llvm-svn: 291956
- Rename the ptx.read.* intrinsics to nvvm.read.ptx.sreg.* - some but
not all of these registers were already accessible via the nvvm
name.
- Rename ptx.bar.sync nvvm.bar.sync, to match nvvm.bar0.
There's a fair amount of code motion here, but it's all very
mechanical.
llvm-svn: 274769
Summary:
Currently clang emits these instructions via inline (volatile) asm in
the CUDA headers. Switching to intrinsics will let the optimizer reason
across calls to these intrinsics.
Reviewers: tra
Subscribers: llvm-commits, jholewinski
Differential Revision: http://reviews.llvm.org/D21160
llvm-svn: 272298
Summary:
Tablegen tries to infer this from the selection DAG patterns defined for
the instructions, but it can't always.
An instructive example is CLZr64. CLZr32 is correctly inferred to have
no side-effects, but the selection DAG pattern for CLZr64 is slightly
more complicated, and in particular the ctlz DAG node is not at the root
of the pattern. Thus tablegen can't infer that CLZr64 has no
side-effects.
Reviewers: jholewinski
Subscribers: jholewinski, tra, llvm-commits
Differential Revision: http://reviews.llvm.org/D17472
llvm-svn: 265089
Summary:
Tablegen was unable to determine that param loads/stores were actually
reading or writing from memory. I think this isn't a problem in
practice for param stores, because those occur in a block right before
we make our call. But param loads don't have to at the very beginning
of a function, so should be annotated as mayLoad so we don't incorrectly
optimize them.
Reviewers: jholewinski
Subscribers: jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D17471
llvm-svn: 262381
Summary:
Previously the machine instructions for bar.sync &co. were not marked as
convergent. This resulted in some MI passes (such as TailDuplication,
fixed in an upcoming patch) doing unsafe things to these instructions.
Reviewers: jingyue
Subscribers: llvm-commits, tra, jholewinski, hfinkel
Differential Revision: http://reviews.llvm.org/D17318
llvm-svn: 261115
Clang may well start emitting these soon, and while it may not be
directly relevant for OpenCL or GLSL, the instructions were just
sitting there waiting to be used.
llvm-svn: 213356
This also uses TSFlags to mark machine instructions that are surface/texture
accesses, as well as the vector width for surface operations. This is used
to simplify some of the switch statements that need to detect surface/texture
instructions
llvm-svn: 213256
This makes the two intrinsics @llvm.convert.from.f16 and
@llvm.convert.to.f16 accept types other than simple "float". This is
only strictly needed for the truncate operation, since otherwise
double rounding occurs and there's no way to represent the strict IEEE
conversion. However, for symmetry we allow larger types in the extend
too.
During legalization, we can expand an "fp16_to_double" operation into
two extends for convenience, but abort when the truncate isn't legal. A new
libcall is probably needed here.
Even after this commit, various target tweaks are needed to actually use the
extended intrinsics. I've put these into separate commits for clarity, so there
are no actual tests of f64 conversion here.
llvm-svn: 213248
The address space of the pointer must be global (1) for these intrinsics. There must also be alignment metadata attached to the intrinsic calls, e.g.
%val = tail call i32 @llvm.nvvm.ldu.i.global.i32.p1i32(i32 addrspace(1)* %ptr), !align !0!0 = metadata !{i32 4}
llvm-svn: 211939
This commit adds intrinsics and codegen support for the surface read/write and texture read instructions that take an explicit sampler parameter. Codegen operates on image handles at the PTX level, but falls back to direct replacement of handles with kernel arguments if image handles are not enabled. Note that image handles are explicitly disabled for all target architectures in this change (to be enabled later).
llvm-svn: 205907
This converter currently only handles global variables in address space 0. For
these variables, they are promoted to address space 1 (global memory), and all
uses are updated to point to the result of a cvta.global instruction on the new
variable.
The motivation for this is address space 0 global variables are illegal since we
cannot declare variables in the generic address space. Instead, we place the
variables in address space 1 and explicitly convert the pointer to address
space 0. This is primarily intended to help new users who expect to be able to
place global variables in the default address space.
llvm-svn: 182254
Vectors were being manually scalarized by the backend. Instead,
let the target-independent code do all of the work. The manual
scalarization was from a time before good target-independent support
for scalarization in LLVM. However, this forces us to specially-handle
vector loads and stores, which we can turn into PTX instructions that
produce/consume multiple operands.
llvm-svn: 174968
The new target machines are:
nvptx (old ptx32) => 32-bit PTX
nvptx64 (old ptx64) => 64-bit PTX
The sources are based on the internal NVIDIA NVPTX back-end, and
contain more functionality than the current PTX back-end currently
provides.
NV_CONTRIB
llvm-svn: 156196
