Summary:
Implementation of the less popular proposal for eliminating overlap between LetStmt and Let: removing both and storing a mapping between Var and value Expr in the Block.
This complicates some tests but simplifies the IR by restricting where variable binding can occur.
I used the unit tests & python integration tests to verify this is correct but I'm unsure of coverage, particularly around the dependency checker in loopnest - ZolotukhinM your review would be useful there.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37606
Differential Revision: D21467483
Pulled By: nickgg
fbshipit-source-id: b402d3fce4cacf35d75f300f0a7dca32a43b6688
Summary:
In the IR Simplifier when doing partial factorization of Round+Mod patterns we divide by the lower number, which could be zero. Add in a quick check against zero avoid the crash.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38055
Differential Revision: D21478486
Pulled By: nickgg
fbshipit-source-id: c5083f672e91662b7d1271d817cade7fa6c39967
Summary:
The IR Simplifier early exits when working with dtypes that are not safe to reorder. There are some cases where we still want to simplify ops in these dtypes: x + 0, x - 0, x * 0 and x * 1. It's safe to eliminate the op here and it reduces clutter in the expr.
Also added a quick simplification of casts which do nothing (their type is the same as the underlying).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37960
Differential Revision: D21457736
Pulled By: nickgg
fbshipit-source-id: 40e20a3b55fc1afb2ec50071812238a08bded2ac
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36291
Move profiler state to be a thread local property,
reuse existing thread local propagation mechanism to ensure
correct profiling of async tasks. This also makes
push/pop callback thread safe and easier to use in e.g.
distributed profilier
Test Plan:
USE_BLAS=MKL USE_MKLDNN=0 USE_CUDA=0 python setup.py develop install
./build/bin/test_jit
./build/bin/test_jit
python test/test_autograd.py
python test/test_jit.py
Differential Revision: D20938501
Pulled By: ilia-cher
fbshipit-source-id: c0c6c3eddcfea8fc7c14229534b7246a0ad25845
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37745
This PR makes it possible to set TLS callbacks and use
them transparently not only in the main thread but also
in any async tasks
Test Plan: Imported from OSS
Differential Revision: D21374873
Pulled By: ilia-cher
fbshipit-source-id: 3be2e121673b32d7694e17e794f3b474826dffe9
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37548
Moving RecordFunction from torch::autograd::profiler into at namespace
Test Plan:
CI
Imported from OSS
Differential Revision: D21315852
fbshipit-source-id: 4a4dbabf116c162f9aef0da8606590ec3f3847aa
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37491
This PR modernizes RecordFunction API and adds thread local callbacks
in addition to the global ones
Changes:
- support for TLS callbacks, this is going to be the foundation of profiler and other tools
- modernize interface around simple set of functions (add|remove|has|clear)(Global|ThreadLocal)(Callback) and adding RecordFunctionCallback to easily construct callbacks to be passed
- we also add `.setShouldRun` into the callback interface to support cases when simple uniform sampling is not enough
- to properly support add/remove introduce the idea of callback handle returned by add
- internal implementation still uses SmallVector to store intermediate state (as before) - in this case these are vector of handles of callbacks that were picked to run
- to speed up runtime we keep these vectors sorted, this way we can quickly enumerate callbacks that need to be run
- added tests for new functionality
Test Plan:
BUILD_BINARY=1 USE_BLAS=MKL USE_MKLDNN=0 USE_CUDA=0 python setup.py
develop install
./build/bin/test_jit
CI
record_function_benchmark: https://gist.github.com/ilia-cher/f1e094dae47fe23e55e7672ac4dcda2f
Imported from OSS
Differential Revision: D21300448
fbshipit-source-id: 6d55c26dbf20b33d35c3f1604dcc07bb063c8c43
Summary:
This PR added more supported operations in CUDA fuser. We are covering major point-wise operations supported in legacy fuser.
In an attempt to adapt to legacy executor:
1. added an naive shape propagation pass on pytorch JIT IR;
2. small refactor on graph partitioning;
3. fallback interpreter execution of fusion group;
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37849
Reviewed By: yf225
Differential Revision: D21444320
Pulled By: soumith
fbshipit-source-id: 712e18ab8497f8d58a07e6f8d200cdab52cf0d74
Summary:
In D21209901 TensorPipe added support for a vector of payloads inside each message, instead of a single one, so that users with multiple payloads can send them separately as they are instead of having to copy them into a new block of contiguous memory. The PyTorch agent is using the old API, which is preventing us from deleting it. This change has no effects on over-the-wire format and thus on performance.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37919
ghstack-source-id: 103572164
Test Plan:
On both workers
```
import os
import torch
import torch.distributed.rpc as rpc
os.environ["MASTER_ADDR"] = "127.0.0.1"
os.environ["MASTER_PORT"] = "8765"
```
On worker 0
```
rpc.init_rpc(name="foo", rank=0, backend=rpc.backend_registry.BackendType.TENSORPIPE, world_size=2, rpc_backend_options=rpc.TensorPipeRpcBackendOptions(worker_name_to_id={"foo": 0, "bar": 0}))
```
On worker 1
```
rpc.init_rpc(name="bar", rank=1, backend=rpc.backend_registry.BackendType.TENSORPIPE, world_size=2, rpc_backend_options=rpc.TensorPipeRpcBackendOptions(worker_name_to_id={"foo": 0, "bar": 0}))
```
On worker 0
```
In [15]: rpc.rpc_sync("bar", torch.add, args=(torch.full((2,2), 1), torch.full((2,2), 2)))
Out[15]:
tensor([[3., 3.],
[3., 3.]])
In [16]: rpc.rpc_sync("bar", torch.add, args=(1, 2))
Out[16]: 3
```
Differential Revision: D21425536
fbshipit-source-id: a0ec2be825556b39aff018a2834baf815a6d8fa5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37472
Our convention is for `findX` to return an optional version and `getX`
to assert that the X is there. Fix up `getMethod` to be consistent with
this convention.
Test Plan: Imported from OSS
Differential Revision: D21297543
Pulled By: suo
fbshipit-source-id: b40f56231cc8183e61bbb01fe5c0c113bcb6464d
Summary:
Remove the requirement for the axes provided to reorderAxis to come from a Tensor. We were using that to determine the relevant loops, but we can alternatively determine it by traversing the parents of each provided For.
resistor does this work for you?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37873
Differential Revision: D21428016
Pulled By: nickgg
fbshipit-source-id: b16b2f41cb443dfc2c6548b7980731d1e7d89a35
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36611
Currently Buf represents underlying storage but it didn't have dtype.
That resulted in specifying dtypes in different places and there was no
mechanism to enforce its consistency: e.g. one could've created a kFloat
expression and use a kInt buffer to store its result. Now we're
centralizing where the logic regarding the storage is located and we can
start enforcing semantics rules.
Follow-ups: we can merge Buffer and BufHandle classes as the former is
now a mere wrapper over the latter.
Test Plan: Imported from OSS
Differential Revision: D21027356
Pulled By: ZolotukhinM
fbshipit-source-id: c06aa2c4077fdcde3bb4ca622d324aece79b5a9c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37776
* Remove type-specific size tracking in favor of byte size tracking in Storage and StorageImpl
* Changed numel() and set_numel() to nbytes() and set_nbytes()
* Added enum argument to Storage/StorageImpl constructor to indicate new meaning of the size parameter
* Update all callers of the changed API
Part of issue https://github.com/pytorch/pytorch/issues/33950
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37028
Differential Revision: D21171334
Pulled By: ezyang
fbshipit-source-id: 37329a379de9a3a83cc5e9007e455a3e1c2d10b8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37681
By passing by value, we can std::move, and avoid unnecessarily copying
args that are part of any std::function/lambda state (e.g. in the jit
interpreter, there is a std::vector<> stack passed in the
InterpreterContinuation)
This makes the api also consistent with e.g. folly and best practices.
Added a minor at::launch() benchmark to test/cpp/, the difference is
mostly noticeable when copying the std::function<> internal args is
non-trivial.
Benchmarks pre/post (min over ~5 runs)
NoData: 5.81 us -> 5.63 us (-3.2%)
WithData(0): 6.67 us -> 5.88 us (-11.8%)
WithData(4): 6.98 us -> 6.51 us (-6.7%)
WithData(256): 9.44 us -> 7.89 (-16.5%)
ghstack-source-id: 103322321
Test Plan:
- perf: buck run mode/opt caffe2/test/cpp/api:parallel_benchmark pre/post
- correctness buck test mode/dev-nosan caffe2/test/...
Reviewed By: dzhulgakov
Differential Revision: D21355148
fbshipit-source-id: 3567e730845106f1991091e4a892d093e00571c3
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36197
Create APIs to convert between rpc::message and tensorpipe::message
1. tensorpipeSerialize() - converts rpc::message to tensorpipe::message without memory copy (tensors).
2. tensorpipeAllocateMessage - allocates rpc::message based on received tensorpipe descriptor to prepare memory-copy-free receiving.
Test Plan: buck test caffe2/test/cpp/rpc:test_tensorpipe_serialization
Reviewed By: lw
Differential Revision: D20084125
fbshipit-source-id: ffbc310f93443e50261aed752be0fe176610dd2a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37704
If input tensor can not be chunked, run `parallel_apply` on fewer devices
Modfy input tensor dimention in `DataParallelUsesAllAvailableCUDADevices_CUDA` to be chunkable by any number of available CUDA devices
Test Plan: Run `test/cpp/api/parallel` on machine with 6 GPUs
Differential Revision: D21365416
fbshipit-source-id: 60fdfed4a0e6256b2c966c2ea3e8d0bfb298d9a8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37106
Recomputing the aliasdb on every fusion iteration + in every subblock
is hugely expensive. Instead, update it in-place when doing fusion.
The graph fuser pass operates by pushing nodes into a fusion group. So
we start with
```
x, y = f(a, b, c)
```
and end with:
```
x_out, y_out = prim::fusionGroup(a, b, c)
x_in, y_in = f(a_in, b_in, c_in)
-> x_in, y_in
```
We destroy the `x` and `y` `Value*`s in the process. This operation is
easy to express as an update to the aliasDb--`x_out` just takes on all
the aliasing information `x` used to have. In particular, since we know
`f` and `prim::fusionGroup` are purely functional, we don't have to mess
with any write information.
This PR is the bare minimum to get this working, in the interest of
unscrewing the compilation times ASAP.
Followups I want to do:
- We don't have a way of expressing deletion of values in AliasDb. In
`graph_fuser.cpp` we sometimes construct nodes that we end up throwing
away, and we are littering `MemoryDAG` with references to dangling
pointers. Because of the way the pass works, it's fine, but this is
fragile so I want to fix it.
- We should decouple alias analysis from write tracking, to simplify the
job of keeping the write caches consistent as we mutate the aliasing
information.
- the tensorexpr fuser doesn't do this and thus is incorrect today, we
need to update it to work.
Test Plan: Imported from OSS
Differential Revision: D21219179
Pulled By: suo
fbshipit-source-id: 8ae5397b3a0ad90edec2fbc555647091f1ad5284
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36345
During compilation, we spend a huge amount of time in alias analyis.
This PR does a few things to speed it up.
1. Separate the analysis into two phases: one where we build up the
necessary data structures, and the other where we service aliasing
queries. This allows us to defer building indices/maintaining index
consistency until after the "buildup" phase is done.
2. Properly memoize/dynamic program the memory locations lookups.
3. Done naively, setting wildcards invalidates the above memoization,
trigger costly recomputation. So I added a cache-aware `setWildcards`.
Sadly that means you need alias analysis to reach into the guts of
memorydag, but the speedup is worth it.
Sadly, these changes are kind of coupled for correctness reasons, so
they're all here at once.
I used this model (thanks IlyaOvodov) as a provisional benchmark. You
can get it here:
https://www.dropbox.com/s/jlyygn6yygj1jkx/yolov3.zip. Unzip at run
`python test_timing.py`.
Baseline: (752.076s) right before 6bc8ffe824
After optimizing before inlining: (699.593s)
After deferring cache construction: (426.180s)
After cache-aware `setWildcards`: (193.678s)
So a nice 75% speedup to overall compilation. There's a lot more to do
in other places of the compilation pipeline though.
Followup to this PR specifically: Everything that fans out from the
`analyze` call is the "buildup" phase of AliasDB construction. This
should be factored into a separate analysis pass to statically
distinguish the two phases (right now we just null out stuff to
accomplish the same thing dynamically).
Test Plan: Imported from OSS
Differential Revision: D20952727
Pulled By: suo
fbshipit-source-id: 099f797222d7e71e5c04991584adc2c7eab5a70f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32684
Previously we have `clone` and `clone_instance`, where `clone` will clone both type
and value, and `clone_instance` only clone the value, both of them are shallow copies.
We need to re-evaluate whether we should expose them as a user facing API.
I think we should hide `clone`, but `clone_instance` might be useful as well, especially
when we are copying a model with very large weights, people might just want to do shallow copy.
This PR adds a `deepcopy` that might be useful as a user API, which deep copies the values, including
Tensor, but we didn't deepcopy `Blob`, `Capsule`, `Future` or `PyObject`.
For more discussions please see the following issue.
fixes: https://github.com/pytorch/pytorch/issues/32519
Test Plan: Imported from OSS
Differential Revision: D21220756
fbshipit-source-id: 476bf11fe82c08fac36e7457879a09f545ffdc5e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37382
After adding c10::DispatchKey::Profiler the behavior of RecordFunction
observers is also controlled by the dispatch key,
this PR moves the logic outside of the profiler into the record function
Reviewed By: jamesr66a
Differential Revision: D21268320
fbshipit-source-id: 93207e3b55325d20dcc5b1e8f448ab86933321da
Summary:
Today in PyTorch, warnings triggered in C++ are printed to Python users like this:
`../aten/src/ATen/native/BinaryOps.cpp:81: UserWarning: Integer division of tensors using div or / is deprecated, and in a future release div will perform true division as in Python 3. Use true_divide or floor_divide (// in Python) instead.`
This may be unhelpful to Python users, who have complained it's difficult to relate these messages back to their programs. After this PR, warnings that go through the PyWarningHandler and allow it to add context print like this:
```
test/test_torch.py:16463: UserWarning: Integer division of tensors using div or / is deprecated, and in a future release div will perform true division as in Python 3. Use true_divide or floor_divide (// in Python) instead. (Triggered internally at ../aten/src/ATen/native/BinaryOps.cpp:81.)
cpu_result = getattr(cpu_tensor, op_str)(*cpu_args)
```
This relates the warning back to the user's program. The information about the cpp file and line number is preserved in the body of the warning message.
Some warnings, like those generated in the JIT, already account for a user's Python context, and so they specify that they should be printed verbatim and are unaffected by this change. Warnings originating in Python and warnings that go through c10's warning handler, which prints to cerr, are also unaffected.
A test is added to test_torch.py for this behavior. The test relies on uint8 indexing being deprecated and its warning originating from its current header file, which is an unfortunate dependency. We could implement a `torch.warn` function, instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36052
Differential Revision: D20887740
Pulled By: mruberry
fbshipit-source-id: d3515c6658a387acb7fccaf83f23dbb452f02847
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37251
This was broken by recent changes to how we serialize with type tags. We
save a name (like `Dict[str, MyNamedTuple]`) and then relied on the
mobile type parser to resolve that name back into a set of types.
This doesn't work for any NamedTypes as the mobile type parser doesn't
know how to resolve those. The unpickler allows the caller to inject a
type resolver in for this purpose, use that so that when importing in a
non-mobile environment you get the right results.
A second problem also had to be fixed: the SourceImporter type loader
would only load named types directly (e.g. `MyNamedTuple`) and choked if
it was a general type that contained a named tupe (e.g.
`List[MyNamedTuple]`). Fixed that and renamed `loadNamedType` to
`loadType` for clarity.
Test Plan: Imported from OSS
Differential Revision: D21235213
Pulled By: suo
fbshipit-source-id: 16db0f4c5e91a890d67a8687cc8ababa6b94b0f4
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36984
Follow LOG(WARNING) format for c++ side warnings in order to play well with larger services, especially when using glog. I need to hook up into GLOG internals a bit in order to override FILE/LINE without having to change the whole thing to be macros, but it seems to be stable between glog versions.
Note, this also changes caffe2_log_level to warning by default - I think it's a much better default when compiling without glog (or maybe even have info).
With glog output, stderr capture doesn't work any more in tests. That's why we instead use c10-level warnings capture.
Test Plan:
Run unittest in both glog and non-glog build mode:
glog:
```
W0416 12:06:49.778215 3311666 exception_test.cpp:23] Warning: I'm a warning (function TestBody)
```
no-glog:
```
[W exception_test.cpp:23] Warning: I'm a warning (function TestBody)
```
Reviewed By: ilia-cher
Differential Revision: D21151351
fbshipit-source-id: fa926d9e480db5ff696990dad3d80f79ef79f24a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37050
With this change curly braces are printed as a part of Block rather than
a part of the enclosing statement. It allows us, for instance, to more
easily see nested blocks: now they will be printed each in its own
curly-braced scope.
As a side effect, I had to change how we print loop options. Previously
we did it like this:
```
for (...) { // <loop options>
<loop body (Block)>
}
```
Now, since everything in between { and } is a part of the block, we have
to do it the following way:
```
for (...) /* <loop options> */ {
<loop body (Block)>
}
```
Note the change from '//' to '/* .. */' for the loop option comments.
Test Plan: Imported from OSS
Differential Revision: D21171851
Pulled By: ZolotukhinM
fbshipit-source-id: 39f51a9e15aec03b6527b0634fd4b9e01a912cda
Summary:
Some IR optimizations were leaving superfluous Blocks in the IR, this PR adds simplification and merging of enclosing Block statements to the IR Simplifier, e.g.
```
Block {
Stmt 1
Block {
Stmt 2
}
Block {}
}
```
becomes
```
Block {
Stmt 1
Stmt 2
}
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37013
Differential Revision: D21166208
Pulled By: nickgg
fbshipit-source-id: 6dcdf863980d94731a8ddf184882c4a5b7259381
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36976
The bounds check and the read were swapped in two places - I noticed
ASAN complaining in an unrelated change on an erroneous buffer.
Adding a couple simple test cases.
ghstack-source-id: 102606986
Test Plan: buck test mode/dev caffe2/test/cpp/rpc:
Differential Revision: D21148936
fbshipit-source-id: 7ec5007535f7310437ac1b9a72852a223b9dd29a
Summary:
When doing elimination of For loops which execute once, e.g. `for i = 0; i < 1; ++i { thing; } => thing;` we do var substitution while the temporary simplifier ExprNodes still exist, which could put them in an invalid state and leave unsimplified terms in the expression. The fix is to apply substitution before simplifying the body of the for loop.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36965
Differential Revision: D21145248
Pulled By: nickgg
fbshipit-source-id: d874600c7a098fc05b8ef3109e516e2eaa2c24e0
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36742
Now, you can define a custom class inside a TORCH_LIBRARY block.
It looks very similar to what you did before. Instead of
```
static auto m = torch::class_<Class>("Namespace", "Class").def("foo", foo);
```
you write
```
TORCH_LIBRARY(Namespace, m) {
m.class_<Class>("Class")
.def("foo", foo);
}
```
All the old usages still work, but at some point we should start
updating the tutorials when we're ready to go 100% live with the
new pybind11 style API.
custom class API previously lived in torch/ folder and in torch
namespace, so for consistency, the new TORCH_LIBRARY also got
moved to torch/library.h The definition of Library::class_ is in the
bottom of that header because I need all of the class_ constructors
available, but there is a circular dependency between the two headers.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D21089648
Test Plan: Imported from OSS
Pulled By: ezyang
fbshipit-source-id: 8d54329c125242605336c22fa1642aae6940b507
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36215
Make it possible to disable observers, e.g. to avoid
infinite recursion if an observer uses an operator
Test Plan:
USE_BLAS=MKL USE_MKLDNN=0 USE_CUDA=0 python setup.py develop install
./build/bin/test_jit
Differential Revision: D20912676
Pulled By: ilia-cher
fbshipit-source-id: 29760cdfe488a02f943f755967b78779d6dbcef3
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36745
As we hold a mutex for our custom C++ Node, when calling reentrant
backward from custom C++ function, we will cocurrently holding many
mutexes up to MAX_DEPTH. TSAN only allow 65 mutexes at once, otherwise
it will complain. This PR lower the limit according to TSAN.
TSAN Reference: https://github.com/google/sanitizers/issues/950
Test Plan: Imported from OSS
Differential Revision: D21072604
Pulled By: wanchaol
fbshipit-source-id: 99cd1acab41a203d834fa4947f4e6f0ffd2e70f2
