Summary:
This information is necessary for PPCG to perform correct life range reordering.
With these changes applied we can live-range reorder some of the important
kernels in COSMO.
We also update and rename one test case, which previously could not be optimized
and now is optimized thanks to live-range reordering. To preserve test coverage
we add a new test case scalar-writes-in-scop-requires-abort.ll, which exercises
our automatic abort in case of scalar writes in the kernel.
Reviewers: Meinersbur, bollu, singam-sanjay
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36929
llvm-svn: 311259
Kernel argument sizes now only get appended to the kernel launch parameter list if the OpenCL runtime is selected, not if CUDA runtime is chosen.
Differential revision: D36925
llvm-svn: 311248
When using -polly-ignore-integer-wrapping and -polly-acc-codegen-managed-memory
we add parameter dimensions lazily to the domains, which results in PPCG not
including parameter dimensions that are only used in memory accesses in the
kernel space. To make sure these parameters are still passed to the kernel, we
collect these parameter dimensions and align the kernel's parameter space
before code-generating it.
llvm-svn: 311239
Summary:
When trying to expand memory accesses, the current version of Polly uses statement Level dependences. The actual implementation is not working in case of multiple dependences per statement. For example in the following source code :
```
void mse(double A[Ni], double B[Nj], double C[Nj], double D[Nj]) {
int i,j;
for (j = 0; j < Ni; j++) {
for (int i = 0; i<Nj; i++)
S: B[i] = i;
for (int i = 0; i<Nj; i++)
T: D[i] = i;
U: A[j] = B[j];
C[j] = D[j];
}
}
```
The statement U has two dependences with S and T. The current version of polly fails during expansion.
This patch aims to fix this bug. For that, we use Reference Level dependences to be able to filter dependences according to statement and memory ref. The principle of expansion remains the same as before.
We also noticed that we need to bail out if load come after store (at the same position) in same statement. So a check was added to isExpandable.
Contributed by: Nicholas Bonfante <nicolas.bonfante@insa-lyon.fr>
Reviewers: Meinersbur, simbuerg, bollu
Reviewed By: Meinersbur, simbuerg
Subscribers: pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D36791
llvm-svn: 311165
Summary:
Drop unused parameter dimensions to reduce the size of the sets we are working
with. Especially the computed dependences tend to accumulate a lot of parameters
that are present in the input memory accesses, but often not necessary to
express the actual dependences. As isl represents maps and sets with dense
matrices, reducing the dimensionality of isl sets commonly reduces code
generation performance.
This reduces compile time from 17 to 11 seconds for our test case. While this is
not impressive, this patch helped me to identify the previous two performance
improvements and additionally also increases readability of the isl data
structures we use.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36869
llvm-svn: 311161
Summary:
They are not used and consequently do not even need to be computed. This reduces
the overall compile time for our kernel from 1m33s to 17s.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36868
llvm-svn: 311157
Summary:
This change reduces the overall number of synchronize calls for kernels with
a lot of output data at the cost of additional synchronize calls for kernels
launched in sequence without any device to host transfers in between. As the
latter pattern is a lot less frequent, this seems a better tradeoff.
Even though the above motivation would be motivation enough, this is just
a step towards enabling ppcg to not compute to and from device copy calls
at all, which would be incorrect in case we still relied on these calls to
place our synchronization statements.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: nemanjai, kbarton, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36867
llvm-svn: 311155
This avoid the construction of very large sets and in many cases also keeps the
number of parameters low. As a result, we see a compile time reduction from 5
minutes to only slightly above 1 minute for one of our larger test cases.
llvm-svn: 311127
We add a ScopInliner pass which inlines functions based on a simple heuristic:
Let `g` call `f`.
If we can model all of `f` as a Scop, we inline `f` into `g`.
This requires `-polly-detect-full-function` to be enabled. So, the pass
asserts that `-polly-detect-full-function` is enabled.
Differential Revision: https://reviews.llvm.org/D36832
llvm-svn: 311126
Reuse the machinery built for replacing global arrays to replace malloc/free as
well. Example replacement that was missed earlier:
```
call void \
bitcast (void (i8*)* @free to void (%custom_type*)*) (%custom_type* %13)
```
- Since the `bitcast` is a `ConstantExpr`, `replaceAllUsesWith` would miss
this. We don't miss this anymore.
Differential Revision: https://reviews.llvm.org/D36825
llvm-svn: 311121
- If we have global arrays, we would like to rewrite them to global
pointers which are allocated using `cudaMallocManaged`.
- If we have allocas in a function, we would like to rewrite them to
heap-allocations with `cudaMallocManaged` and `cudaFree`.
- With these rewrite mechanisms, we can offload _any_ function to the
GPU with no code rewrite whatsover.
Differential Revision: https://reviews.llvm.org/D36516
llvm-svn: 311080
Summary:
This pass detangles induction variables from functions, which take variables by
reference. Most fortran functions compiled with gfortran pass variables by
reference. Unfortunately a common pattern, printf calls of induction variables,
prevent in this situation the promotion of the induction variable to a register,
which again inhibits any kind of loop analysis. To work around this issue
we developed a specialized pass which introduces separate alloca slots for
known-read-only references, which indicate the mem2reg pass that the induction
variables can be promoted to registers and consquently enable SCEV to work.
We currently hardcode the information that a function
_gfortran_transfer_integer_write does not read its second parameter, as
dragonegg does not add the right annotations and we cannot change old dragonegg
releases. Hopefully flang will produce the right annotations.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: mgorny, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36800
llvm-svn: 311066
Summary:
Before, if we fail to parse a jscop file, this will be reported as an
error and importing is aborted. However, this isn't actually strong
enough, since although the import is aborted, the scop has already been
modified and is very likely broken. Instead, make this a hard failure
and throw an LLVM error. This new behaviour requires small changes to
the tests for the legacy pass, namely using `not` to verify the error.
Further, fixed the jscop file for the
base_pointer_load_is_inst_inside_invariant_1 testcase.
Reviewed By: Meinersbur
Split out of D36578.
llvm-svn: 310599
Summary:
I pulled out all functionality into static functions, and use those both
in the legacy passes and in the new ones.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: Meinersbur
Subscribers: llvm-commits, pollydev
Differential Revision: https://reviews.llvm.org/D36578
llvm-svn: 310597
Summary:
During code generation for a Scop we modify the IR of a function.
While this shouldn't affect a Scop in the formal sense, the implementation
caches various information about the IR such as SCEV expressions for bounds or
parameters. This cached information needs to be updated or invalidated. To this
end, SPMUpdater allows passes to report when they've invalidated a Scop to the
PassManager, which will then flush and recompute all Scops. This in turn
invalidates all iterators, so references to Scops shouldn't be held.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: grosser
Subscribers: llvm-commits, pollydev
Differential Revision: https://reviews.llvm.org/D36524
llvm-svn: 310551
We are working towards removing uses of Scop::getStmtFor(BB). In this
patch, we remove dependency of Scop::getStmtFor(Inst) on getStmtFor(BB).
To do so, we introduce a map of instructions to their corresponding scop
statements and use it to get the instructions' statement.
Contributed-by: Nandini Singhal <cs15mtech01004@iith.ac.in>
Differential Revision: https://reviews.llvm.org/D35663
llvm-svn: 310494
This pass is useful to automatically convert a codebase that uses malloc/free
to use their managed memory counterparts.
Currently, rewrite malloc and free to the `polly_{malloc,free}Managed` variants.
A future patch will teach ManagedMemoryRewrite to rewrite global arrays
as pointers to globally allocated managed memory.
Differential Revision: https://reviews.llvm.org/D36513
llvm-svn: 310471
Codegen with -polly-parallel queried the unmapped MemoryAccess, but only
the MemoryKind after mapping is relevant for codegen.
This should fix various fails of the
perf-x86_64-penryn-O3-polly-parallel-fast buildbot.
llvm-svn: 310466
The previous value of "polly-delicm" was forgotten to to be changed when
ForwardOpTree was split from DeLICM.
Thanks to Tobias for noticing!
llvm-svn: 310465
distributeDomain() and filterKnownValInst() are used in a scop
of ForwardOpTree that limits the number of isl operations.
Therefore some isl functions may return null after any operation.
Remove assertion that assume non-null results and handle
isl_*_foreach returning isl::stat::error.
I hope this fixes the crash of the asop buildbot at ihevc_recon.c.
llvm-svn: 310461
Previously, we used to compute this with `elementSizeInBits / 8`. This
would yield an element size of 0 when the array had element size < 8 in
bits.
To fix this, ask data layout what the size in bytes should be.
Differential Revision: https://reviews.llvm.org/D36459
llvm-svn: 310448
We introduce another level of alias metadata to distinguish the individual
non-aliasing accesses that have inter iteration alias-free base pointers
marked with "Inter iteration alias-free" mark nodes. To distinguish two
accesses, the comparison of raw pointers representing base pointers is used.
In case of, for example, ublas's prod function that implements GEMM, and
DeLiCM we can get accesses to same location represented by different raw
pointers. Consequently, we create different alias sets that can prevent
accesses from, for example, being sinked or hoisted.
To avoid the issue, we compare the corresponding SCEV information instead
of the corresponding raw pointers.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D35761
llvm-svn: 310380
Currently, only convex isolation sets can be efficiently processed by isl.
Consequently, as a temporary solution, we use a different algorithm for partial
tile isolation that helps to build convex isolation sets in some cases.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D36278
llvm-svn: 310374
This allows us to get rid of stores that are overwritten within the very same
basic block, without ever being read beforehand. This simplification is
necessary for delicm to run on pb4's correlation.
llvm-svn: 310369
To do this, we replicate what `CodeGeneration` does. We expose
`markNodeUnreachable` from `CodeGeneration` to `PPCGCodeGeneration`.
Differential Revision: https://reviews.llvm.org/D36457
llvm-svn: 310350
It is possible that partial writes are empty (write is never executed).
In this case, when in PHINode's incoming edge is never taken such that
the incoming write becomes an empty partial write, if enabled. The
issue is that when converting the union_map to an map, it's space
cannot be derived from the union_map itself. Rather, we need to
determine its space independently.
This fixes test-suite's MultiSource/Benchmarks/ASC_Sequoia/CrystalMk.
llvm-svn: 310348
In certain cases delicm might decide to not leave the original array write in
the loop body, but to remove it and instead leave a transformed phi node as
write access. This commit teached the matmul pattern detection to order the
memory accesses according to when the access actually happens and use this
information to detect the new pattern. This makes pattern based matmul
optimization work for 2mm and 3mm in polybench 4 after
polly-position=before-vectorizer has been enabled.
llvm-svn: 310338
Polly has traditionally always been executed at the beginning of the pass
pipeline as LLVM's inliner and DeLICM passes introduced plenty of scalar
dependences which prevented any kind of useful high-level loop optimizations
later in the pass pipeline. With DeLICM now being available, Polly can also
run optimizations when folded into the pass pipeline. This has the benefit
that Polly should now be more effective on C++ code and as an additional bonus,
no additional early canonicalization phase must be run. As a result, Polly
touches the code only if it applies a transformation. Code that does not
benefit from Polly is not touched and consequently will have the very same
execution time as without Polly enabled. Random performance changes, as could
sometimes be observed with polly-position=early are consequently not possible
any more. If performance is changed, this is due to Polly is choosing to
perform a transformation. If this choice is wrong, it can be fixed directly
in Polly.
http://polly.llvm.org/docs/Architecture.html#polly-in-the-llvm-pass-pipeline
llvm-svn: 310319
This allows us to remove more scalar dependences. While this feature is still
rather experimental, we want to give it sufficient test coverage.
llvm-svn: 310314
Two write statements which write into the very same array slot generally are
conflicting. However, in case the value that is written is identical, this
does not cause any problem. Hence, allow such write pairs in this specific
situation.
llvm-svn: 310311
This commit implements the initial version of fully-indexed static
expansion.
```
for(int i = 0; i<Ni; i++)
for(int j = 0; j<Ni; j++)
S: B[j] = j;
T: A[i] = B[i]
```
After the pass, we want this :
```
for(int i = 0; i<Ni; i++)
for(int j = 0; j<Ni; j++)
S: B[i][j] = j;
T: A[i] = B[i][i]
```
For now we bail (fail) in the following cases:
- Scalar access
- Multiple writes per SAI
- MayWrite Access
- Expansion that leads to an access to the original array
Furthermore: We still miss checks for escaping references to the array
base pointers. A future commit will add the missing escape-checks to
stay correct in those cases. The expansion is still locked behind a
CLI-Option and should not yet be used.
Patch contributed by: Nicholas Bonfante <bonfante.nicolas@gmail.com>
Reviewers: simbuerg, Meinersbur, bollu
Reviewed By: Meinersbur
Subscribers: mgorny, llvm-commits, pollydev
Differential Revision: https://reviews.llvm.org/D34982
llvm-svn: 310304
