Summary:
Added SPIR Code Generation to the PPCG Code Generator. This can be invoked using
the polly-gpu-arch flag value 'spir32' or 'spir64' for 32 and 64 bit code respectively.
In addition to that, runtime support has been added to execute said SPIR code on Intel
GPU's, where the system is equipped with Intel's open source driver Beignet (development
version). This requires the cmake flag 'USE_INTEL_OCL' to be turned on, and the polly-gpu-runtime
flag value to be 'libopencl'.
The transformation of LLVM IR to SPIR is currently quite a hack, consisting in part of regex
string transformations.
Has been tested (working) with Polybench 3.2 on an Intel i7-5500U (integrated graphics chip).
Reviewers: bollu, grosser, Meinersbur, singam-sanjay
Reviewed By: grosser, singam-sanjay
Subscribers: pollydev, nemanjai, mgorny, Anastasia, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D35185
llvm-svn: 308751
This commit *WILL COMPILE*.
1. `PPCG` now uses `isl_multi_pw_aff` instead of an array of `pw_aff`.
This needs us to adjust how we index array bounds and how we construct
array bounds.
2. `PPCG` introduces two new kinds of nodes: `init_device` and `clear_device`.
We should investigate what the correct way to handle these are.
3. `PPCG` has gotten smarter with its use of live range reordering, so some of
the tests have a qualitative improvement.
4. `PPCG` changed its output style, so many test cases need to be updated to
fit the new style for `polly-acc-dump-code` checks.
Differential Revision: https://reviews.llvm.org/D35677
llvm-svn: 308625
We extended kills in Polly to handle both `phi` nodes and scalars that
are not used within the Scop. Update the comments and choice of
variable names to reflect this.
llvm-svn: 308279
- We should call `preloadInvariantLoads` to make sure that code is
generated for invariant loads in the kernel.
Differential Revision: https://reviews.llvm.org/D35410
llvm-svn: 308187
- There is a conditional branch that is used to switch between the old
and new versions of the code.
- If we detect that the build was unsuccessful, `PPCGCodeGeneration` will
change the runtime check to be always set to false.
- To actually *reach* this runtime check instruction, `PPCGCodeGeneration`
was using assumptions about the layout of the BBs.
- However, invariant load hoisting violates this assumption by inserting
an extra basic block in the middle.
- Fix the assumption on the layout by having `createScopConditionally`
return the conditional branch instruction.
- Use this reference to set to always-false.
llvm-svn: 308010
We need to relax constraints on invariant loads so that they do not
create fake RAW dependences. So, we do not consider invariant loads as
scalar dependences in a region.
During these changes, it turned out that we do not consider `llvm::Value`
replacements correctly within `PPCGCodeGeneration` and `ISLNodeBuilder`.
The replacements dictated by `ValueMap` were not being followed in all
places. This was fixed in this commit. There is no clean way to decouple
this change because this bug only seems to arise when the relaxed
version of invariant load hoisting was enabled.
Differential Revision: https://reviews.llvm.org/D35120
llvm-svn: 307907
Summary:
Add a sequence number that identifies a ptx_kernel's parent Scop within a function to it's name to differentiate it from other kernels produced from the same function, yet different Scops.
Kernels produced from different Scops can end up having the same name. Consider a function with 2 Scops and each Scop being able to produce just one kernel. Both of these kernels have the name "kernel_0". This can lead to the wrong kernel being launched when the runtime picks a kernel from its cache based on the name alone. This patch supplements D33985, by differentiating kernels across Scops as well.
Previously (even before D33985) while profiling kernels generated through JIT e.g. Julia, [[ https://groups.google.com/d/msg/polly-dev/J1j587H3-Qw/mR-jfL16BgAJ | kernels associated with different functions, and even different SCoPs within a function, would be grouped together due to the common name ]]. This patch prevents this grouping and the kernels are reported separately.
Reviewers: grosser, bollu
Reviewed By: grosser
Subscribers: mehdi_amini, nemanjai, pollydev, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D35176
llvm-svn: 307814
- By definition, we can pass something as a `kill` to PPCG if we know
that no data can flow across a kill.
- This is useful for more complex examples where we have scalars that
are local to a scop.
- If the local is only used within a scop, we are free to kill it.
Differential Revision: https://reviews.llvm.org/D35045
llvm-svn: 307260
Summary:
Provide more context to the name of a GPU kernel by prefixing its name with the host function that calls it. E.g. The first kernel called by `gemm` would be `FUNC_gemm_KERNEL_0`.
Kernels currently follow the "kernel_#" (# = 0,1,2,3,...) nomenclature. This patch makes it easier to map host caller and device callee, especially when there are many kernels produced by Polly-ACC.
Reviewers: grosser, Meinersbur, bollu, philip.pfaffe, kbarton!
Reviewed By: grosser
Subscribers: nemanjai, pollydev
Tags: #polly
Differential Revision: https://reviews.llvm.org/D33985
llvm-svn: 307173
Polly did not use PPCG's live range reordering feature. Teach
PPCGCodeGeneration to use this.
Documentation on this is sparse, so much of the code is conservative.
We currently kill all phi nodes in a Scop by appending them to the
must_kill map we pass to PPCG. I do not have a proof of correctness,
but it seems to be intuitively correct.
We also do not handle `array_order`, which, quoting PPCG, is:
PPCG/gpu.h: "Order dependences on non-scalars."
It seems to consist of RAW dependences between arrays. We need to
pass this information for more complex privatization cases.
Differential Revision: https://reviews.llvm.org/D34941
llvm-svn: 307163
Summary:
Introduce a "hybrid" `-polly-target` option to optimise code for either the GPU or CPU.
When this target is selected, PPCGCodeGeneration will attempt first to optimise a Scop. If the Scop isn't modified, it is then sent to the passes that form the CPU pipeline, i.e. IslScheduleOptimizerPass, IslAstInfoWrapperPass and CodeGeneration.
In case the Scop is modified, it is marked to be skipped by the subsequent CPU optimisation passes.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: grosser
Subscribers: kbarton, nemanjai, pollydev
Tags: #polly
Differential Revision: https://reviews.llvm.org/D34054
llvm-svn: 306863
- In D33414, if any function call was found within a kernel, we would bail out.
- This is an over-approximation. This patch changes this by allowing the
`llvm.sqrt.*` family of intrinsics.
- This introduces an additional step when creating a separate llvm::Module
for a kernel (GPUModule). We now copy function declarations from the
original module to new module.
- We also populate IslNodeBuilder::ValueMap so it replaces the function
references to the old module to the ones in the new module
(GPUModule).
Differential Revision: https://reviews.llvm.org/D34145
llvm-svn: 306284
This commit returns both the start and the exit block that are created
by executeScopConditionally.
In a future commit we will make use of the exit block. Before we would
have to use the implicit property that there won't be any code generated
between polly.start and polly.exiting at the time of use to find the
correct block ('polly.exiting').
All usage location are semantically unchanged.
llvm-svn: 306283
The condition that disallowed code generation in PPCGCodeGeneration with
invariant loads is not required. I haven't been able to construct a
counterexample where this generates invalid code.
Differential Revision: https://reviews.llvm.org/D34604
llvm-svn: 306245
In `PPCGCodeGeneration`, we try to take the references of every `Value`
that is used within a Scop to offload to the kernel. This occurs in
`GPUNodeBuilder::createLaunchParameters`.
This breaks if one of the values is a function pointer, since one of
these cases will trigger:
1. We try to to take the references of an intrinsic function, and this
breaks at `verifyModule`, since it is illegal to take the reference of
an intrinsic.
2. We manage to take the reference to a function, but this fails at
`verifyModule` since the function will not be present in the module that
is created in the kernel.
3. Even if `verifyModule` succeeds (which should not occur), we would
then try to call a *host function* from the *device*, which is
illegal runtime behaviour.
So, we disable this entire range of possibilities by simply not allowing
function references within a `Scop` which corresponds to a kernel.
However, note that this is too conservative. We *can* allow intrinsics
within kernels if the backend can lower the intrinsic correctly. For
example, an intrinsic like `llvm.powi.*` can actually be lowered by the `NVPTX`
backend.
We will now gradually whitelist intrinsics which are known to be safe.
Differential Revision: https://reviews.llvm.org/D33414
llvm-svn: 305185
Summary: This patch ports IslAst to the new PM. The change is mostly straightforward. The only major modification required is making IslAst move-only, to correctly manage the isl resources it owns.
Reviewers: grosser, Meinersbur
Reviewed By: grosser
Subscribers: nemanjai, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D33422
llvm-svn: 303622
- We use the outermost dimension of arrays since we need this
information to generate GPU transfers.
- In general, if we do not know the outermost dimension of the array
(because the indexing expression is non-affine, for example) then we
simply cannot generate transfer code.
- However, for Fortran arrays, we can use the Fortran array
representation which stores the dimensions of all arrays.
- This patch uses the Fortran array representation to generate code that
computes the outermost dimension size.
Differential Revision: https://reviews.llvm.org/D32967
llvm-svn: 303429
Summary: This is a proof of concept of how to port polly-passes to the new PassManager architecture. This approach works ootb for Function-Passes, but might not be directly applicable to Scop/Region-Passes. While we could just run the Analyses/Transforms over functions instead, we'd surrender the nice pipelining behaviour we have now.
Reviewers: Meinersbur, grosser
Reviewed By: grosser
Subscribers: pollydev, sanjoy, nemanjai, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D31459
llvm-svn: 302902
Summary: PPCGCodeGeneration now attaches the size of the kernel launch parameters at the end of the parameter list. For the existing CUDA Runtime, this gets ignored, but the OpenCL Runtime knows to check for kernel-argument size at the end of the parameter list. (The resulting parameters list is twice as long. This has been accounted for in the corresponding test cases).
Reviewers: grosser, Meinersbur, bollu
Reviewed By: bollu
Subscribers: nemanjai, yaxunl, Anastasia, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32961
llvm-svn: 302515
Summary:
When compiling for GPU, one can now choose to compile for OpenCL or CUDA,
with the corresponding polly-gpu-runtime flag (libopencl / libcudart). The
GPURuntime library (GPUJIT) has been extended with the OpenCL Runtime library
for that purpose, correctly choosing the corresponding library calls to the
option chosen when compiling (via different initialization calls).
Additionally, a specific GPU Target architecture can now be chosen with -polly-gpu-arch (only nvptx64 implemented thus far).
Reviewers: grosser, bollu, Meinersbur, etherzhhb, singam-sanjay
Reviewed By: grosser, Meinersbur
Subscribers: singam-sanjay, llvm-commits, pollydev, nemanjai, mgorny, yaxunl, Anastasia
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32431
llvm-svn: 302379
This reverts commit 17a84e414adb51ee375d14836d4c2a817b191933.
Patches should have been submitted in the order of:
1. D32852
2. D32854
3. D32431
I mistakenly pushed D32431(3) first. Reverting to push in the correct
order.
llvm-svn: 302217
Summary:
When compiling for GPU, one can now choose to compile for OpenCL or CUDA,
with the corresponding polly-gpu-runtime flag (libopencl / libcudart). The
GPURuntime library (GPUJIT) has been extended with the OpenCL Runtime library
for that purpose, correctly choosing the corresponding library calls to the
option chosen when compiling (via different initialization calls).
Additionally, a specific GPU Target architecture can now be chosen with -polly-gpu-arch (only nvptx64 implemented thus far).
Reviewers: grosser, bollu, Meinersbur, etherzhhb, singam-sanjay
Reviewed By: grosser, Meinersbur
Subscribers: singam-sanjay, llvm-commits, pollydev, nemanjai, mgorny, yaxunl, Anastasia
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32431
llvm-svn: 302215
generation.
This needs changes to GPURuntime to expose synchronization between host
and device.
1. Needs better function naming, I want a better name than
"getOrCreateManagedDeviceArray"
2. DeviceAllocations is used by both the managed memory and the
non-managed memory path. This exploits the fact that the two code paths
are never run together. I'm not sure if this is the best design decision
Reviewed by: PhilippSchaad
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32215
llvm-svn: 301640
Added a small change to the way pointer arguments are set in the kernel
code generation. The way the pointer is retrieved now, specifically requests
global address space to be annotated. This is necessary, if the IR should be
run through NVPTX to generate OpenCL compatible PTX.
The changes do not affect the PTX Strings generated for the CUDA target
(nvptx64-nvidia-cuda), but are necessary for OpenCL (nvptx64-nvidia-nvcl).
Additionally, the data layout has been updated to what the NVPTX Backend requests/recommends.
Contributed-by: Philipp Schaad
Reviewers: Meinersbur, grosser, bollu
Reviewed By: grosser, bollu
Subscribers: jlebar, pollydev, llvm-commits, nemanjai, yaxunl, Anastasia
Tags: #polly
Differential Revision: https://reviews.llvm.org/D32215
llvm-svn: 301299
Summary:
A couple of the utilities used to analyze or build IR make explicit use of the legacy PM on their interface, to access analysis results. This patch removes the legacy PM from the interface, and just passes the required results directly.
This shouldn't introduce any function changes, although the API technically allowed to obtain two different analysis results before, one passed by reference and one through the PM. I don't believe that was ever intended, however.
Reviewers: grosser, Meinersbur
Reviewed By: grosser
Subscribers: nemanjai, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D31653
llvm-svn: 299423
If a SCoP is most probably sequential, then it's better to run it on a CPU.
Hence, there's no point in running it on a GPU.
Reviewers: grosser
Subscribers: nemanjai
Tags: #polly
Contributed-by: Singapuram Sanjay <singapuram.sanjay@gmail.com>
Differential Revision: https://reviews.llvm.org/D30864
llvm-svn: 297578
Marking a pass as preserved is necessary if any Polly pass uses it, even
if it is not preserved within the generated code. Not marking it would
cause the the Polly pass chain to be interrupted. It is not used by any
Polly pass anymore, hence we can remove all references to it.
llvm-svn: 295983
Instead of keeping two separate maps from Value to Allocas, one for
MemoryType::Value and the other for MemoryType::PHI, we introduce a single map
from ScopArrayInfo to the corresponding Alloca. This change is intended, both as
a general simplification and cleanup, but also to reduce our use of
MemoryAccess::getBaseAddr(). Moving away from using getBaseAddr() makes sure
we have only a single place where the array (and its base pointer) for which we
generate code for is specified, which means we can more easily introduce new
access functions that use a different ScopArrayInfo as base. We already today
experiment with modifiable access functions, so this change does not address
a specific bug, but it just reduces the scope one needs to reason about.
Another motivation for this patch is https://reviews.llvm.org/D28518, where
memory accesses with different base pointers could possibly be mapped to a
single ScopArrayInfo object. Such a mapping is currently not possible, as we
currently generate alloca instructions according to the base addresses of the
memory accesses, not according to the ScopArrayInfo object they belong to. By
making allocas ScopArrayInfo specific, a mapping to a single ScopArrayInfo
object will automatically mean that the same stack slot is used for these
arrays. For D28518 this is not a problem, as only MemoryType::Array objects are
mapping, but resolving this inconsistency will hopefully avoid confusion.
llvm-svn: 293374
To benefit of the type safety guarantees of C++11 typed enums, which would have
caught the type mismatch fixed in r291960, we make MemoryKind a typed enum.
This change also allows us to drop the 'MK_' prefix and to instead use the more
descriptive full name of the enum as prefix. To reduce the amount of typing
needed, we use this opportunity to move MemoryKind from ScopArrayInfo to a
global scope, which means the ScopArrayInfo:: prefix is not needed. This move
also makes historically sense. In the beginning of Polly we had different
MemoryKind enums in both MemoryAccess and ScopArrayInfo, which were later
canonicalized to one. During this canonicalization we just choose the enum in
ScopArrayInfo, but did not consider to move this shared enum to global scope.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D28090
llvm-svn: 292030
This makes polly generate a CFG which is closer to what we want
in LLVM IR, with a loop preheader for the original loop. This is
just a cleanup, but it exposes some fragile assumptions.
I'm not completely happy with the changes related to expandCodeFor;
RTCBB->getTerminator() is basically a random insertion point which
happens to work due to the way we generate runtime checks. I'm not
sure what the right answer looks like, though.
Differential Revision: https://reviews.llvm.org/D26053
llvm-svn: 285864
In case sequential kernels are found deeper in the loop tree than any parallel
kernel, the overall scop is probably mostly sequential. Hence, run it on the
CPU.
llvm-svn: 281849
Offloading to a GPU is only beneficial if there is a sufficient amount of
compute that can be accelerated. Many kernels just have a very small number
of dynamic compute, which means GPU acceleration is not beneficial. We
compute at run-time an approximation of how many dynamic instructions will be
executed and fall back to CPU code in case this number is not sufficiently
large. To keep the run-time checking code simple, we over-approximate the
number of instructions executed in each statement by computing the volume of
the rectangular hull of its iteration space.
llvm-svn: 281848
We may generate GPU kernels that store into scalars in case we run some
sequential code on the GPU because the remaining data is expected to already be
on the GPU. For these kernels it is important to not keep the scalar values
in thread-local registers, but to store them back to the corresponding device
memory objects that backs them up.
We currently only store scalars back at the end of a kernel. This is only
correct if precisely one thread is executed. In case more than one thread may
be run, we currently invalidate the scop. To support such cases correctly,
we would need to always load and store back from a corresponding global
memory slot instead of a thread-local alloca slot.
llvm-svn: 281838
Our alias checks precisely check that the minimal and maximal accessed elements
do not overlap in a kernel. Hence, we must ensure that our host <-> device
transfers do not touch additional memory locations that are not covered in
the alias check. To ensure this, we make sure that the data we copy for a
given array is only the data from the smallest element accessed to the largest
element accessed.
We also adjust the size of the array according to the offset at which the array
is actually accessed.
An interesting result of this is: In case array are accessed with negative
subscripts ,e.g., A[-100], we automatically allocate and transfer _more_ data to
cover the full array. This is important as such code indeed exists in the wild.
llvm-svn: 281611
We do not need the size of the outermost dimension in most cases, but if we
allocate memory for newly created arrays, that size is needed.
Reviewed-by: Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D23991
llvm-svn: 281234
Philipp Schaad