Summary:
1. Added `torch/csrc/cuda/Event.h` and `torch/csrc/cuda/Event.cpp` to bind Python Event class to C++ implementation.
2. Move all CUDA runtime invocations from `torch/cuda/streams.py` to C++
3. Added tests to cover Stream and Event APIs. ~(event IPC handle tests is introduced in #15974)~
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15937
Differential Revision: D13649001
Pulled By: mrshenli
fbshipit-source-id: 84ca58f35f6ba679a4ba33150ceba678d760d240
Summary:
This PR enables C++ frontend modules to be bound into Python and added as submodules of Python modules. For this, I added lots of pybind11 bindings for the `torch::nn::Module` class, and modified the `torch.nn.Module` class in Python to have a new Metaclass that makes `isinstance(m, torch.nn.Module)` return true when `m` is a C++ frontend module. The methods and fields of C++ modules are bound in such a way that they work seamlessly as submodules of Python modules for most operations (one exception I know of: calling `.to()` ends up calling `.apply()` on each submodule with a Python lambda, which cannot be used in C++ -- this may require small changes on Python side).
I've added quite a bunch of tests to verify the bindings and equality with Python. I think I should also try out adding a C++ module as part of some large PyTorch module, like a WLM or something, and see if everything works smoothly.
The next step for inter-op across our system is ScriptModule <-> C++ Frontend Module inter-op. I think this will then also allow using C++ frontend modules from TorchScript.
apaszke zdevito
CC dzhulgakov
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13481
Differential Revision: D12981996
Pulled By: goldsborough
fbshipit-source-id: 147370d3596ebb0e94c82cec92993a148fee50a7
Summary:
Anywhere we used #include "foo.h", we now say #include <foo.h>
Paths are adjusted to be rooted out of aten/src, torch/lib, or
the root level directory.
I modified CMakeLists.txt by hand to remove TH and THC from
the include paths.
I used the following script to do the canonicalization:
```
import subprocess
import re
import os.path
files = subprocess.check_output(['git', 'ls-files']).decode('utf-8').rstrip().split('\n')
for fn in files:
if not any(fn.endswith(suff) for suff in ['.cu', '.cpp', '.in', '.h', '.hpp', '.cu', '.cuh', '.cc']):
continue
if not any(fn.startswith(pref) for pref in ["aten/", "torch/"]):
continue
with open(fn, 'r') as f:
c = f.read()
def fmt(p):
return "#include <{}>".format(p)
def repl(m):
p = m.group(1)
if p in ["dlfcn.h", "unistd.h", "nvrtc.h", "cuda.h", "cuda_runtime.h", "cstdint", "cudnn.h", "Python.h", "cusparse.h", "cuda_runtime_api.h", "cuda_fp16.h", "cublas_v2.h", "stdint.h", "curand_kernel.h"]:
return fmt(p)
if any(p.startswith(pref) for pref in ["torch/csrc", "c10/", "ATen/", "caffe2/", "TH/", "THC/", "Eigen/", "gtest/", "zdl/", "gloo/", "onnx/", "miopen/"]):
return fmt(p)
for root in ["aten/src", "torch/lib", ""]:
for bad_root in [os.path.dirname(fn), "aten/src/TH", "aten/src/THC", "torch/csrc"]:
new_p = os.path.relpath(os.path.join(bad_root, p), root)
if not new_p.startswith("../") and (os.path.exists(os.path.join(root, new_p)) or os.path.exists(os.path.join(root, new_p + ".in"))):
return fmt(new_p)
print("ERROR: ", fn, p)
return m.group(0)
new_c = re.sub(r'#include "([^"]+)"', repl, c)
if new_c != c:
print(fn)
with open(fn, 'w') as f:
f.write(new_c)
```
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14849
Reviewed By: dzhulgakov
Differential Revision: D13363445
Pulled By: ezyang
fbshipit-source-id: 52361f878a672785f9306c9e9ab2513128092b68
Summary:
If torch.multiprocessing.spawn is used to launch non-daemonic
processes (the default since #14391), the spawned children won't be
automatically terminated when the parent terminates.
On Linux, we can address this by setting PR_SET_PDEATHSIG, which
delivers a configurable signal to child processes when their parent
terminates.
Fixes#14394.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14491
Differential Revision: D13270374
Pulled By: pietern
fbshipit-source-id: 092c9d3c3cea2622c3766b467957bc27a1bd500c
Summary:
Add to the Tensor doc info about `.device`, `.is_cuda`, `.requires_grad`, `.is_leaf` and `.grad`.
Update the `register_backward_hook` doc with a warning stating that it does not work in all cases.
Add support in the `_add_docstr` function to add docstring to attributes.
There is an explicit cast here but I am not sure how to handle it properly. The thing is that the doc field for getsetdescr is written as being a const char * (as all other doc fields in descriptors objects) in cpython online documentation. But in the code, it is the only one that is not const.
I assumed here that it is a bug in the code because it does not follow the doc and the convention of the others descriptors and so I cast out the const.
EDIT: the online doc I was looking at is for 3.7 and in that version both the code and the doc are const. For older versions, both are non const.
Please let me know if this should not be done. And if it should be done if there is a cleaner way to do it !
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14339
Differential Revision: D13243266
Pulled By: ezyang
fbshipit-source-id: 75b7838f7cd6c8dc72b0c61950e7a971baefaeeb
Summary:
This is the next minimal step towards moving _C into cmake. For now,
leave _C in setup.py, but reduce it to an empty stub file. All of its
sources are now part of the new torch-python cmake target.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12742
Reviewed By: soumith
Differential Revision: D13089691
Pulled By: anderspapitto
fbshipit-source-id: 1c746fda33cfebb26e02a7f0781fefa8b0d86385
Summary:
There are still a few work to be done:
- Move logging and unify AT_WARN with LOG(ERROR).
- A few header files are still being plumbed through, need cleaning.
- caffe2::EnforceNotMet aliasing is not done yet.
- need to unify the macros. See c10/util/Exception.h
This is mainly a codemod and not causing functional changes. If you find your job failing and trace back to this diff, usually it can be fixed by the following approaches:
(1) add //caffe2/c10:c10 to your dependency (or transitive dependency).
(2) change objects such as at::Error, at::Optional to the c10 namespace.
(3) change functions to the c10 namespace. Especially, caffe2::MakeString is not overridden by the unified c10::str function. Nothing else changes.
Please kindly consider not reverting this diff - it involves multiple rounds of rebasing and the fix is usually simple. Contact jiayq@ or AI Platform Dev for details.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12354
Reviewed By: orionr
Differential Revision: D10238910
Pulled By: Yangqing
fbshipit-source-id: 7794d5bf2797ab0ca6ebaccaa2f7ebbd50ff8f32
Summary:
To illustrate the benefits of this commit, I'll use the time/iter I got from one of the JIT benchmarks on my machine.
| Run | Time |
|----------------------------------------------|-------------------------|
| No profiler | 45ms |
| With profiler | 56ms |
| Use `clock_gettime` instead of `std::chrono` | 48ms |
| Touch all pages on block allocation | 48ms (less jitter) |
| Use `const char*` instead of `std::string` | 47ms (even less jitter) |
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11773
Differential Revision: D9886858
Pulled By: apaszke
fbshipit-source-id: 58f926f09e95df0b11ec687763a72b06b66991d0
Summary:
Will use USE_DISTRIBUTED for both c10d and THD
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11237
Differential Revision: D9647825
Pulled By: teng-li
fbshipit-source-id: 06e0ec9b5e2f8f38780fc88718f8499463e9e969
Summary:
This was lingering after #10731.
cc orionr
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11240
Differential Revision: D9645437
Pulled By: pietern
fbshipit-source-id: d02c33354b094be3bb0872cf54a45721e20c4e7d
Summary:
How did we get so many uses of `NULL` again?
ezyang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11047
Differential Revision: D9566799
Pulled By: goldsborough
fbshipit-source-id: 83469f352ac69aa65bdaf1a1a21f922d892e0db3
Summary:
Currently our `skipIfLapack` has uses a try-catch block and regex match the error message. It is highly unreliable. This PR adds `hasLAPACK` and `hasMAGMA` on ATen context, and expose the flags to python.
Also fixes refcounting bug with `PyModule_AddObject`. The method steals reference, but we didn't `Py_INCREF` in some places before calling it with `Py_True` or `Py_False`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11024
Differential Revision: D9564898
Pulled By: SsnL
fbshipit-source-id: f46862ec3558d7e0058ef48991cd9c720cb317e2
Summary:
These could use some autograd tests, which are coming in a later PR, but using them in autograd is probably pretty rare.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9947
Reviewed By: ezyang
Differential Revision: D9032778
Pulled By: gchanan
fbshipit-source-id: fa5a6509d3bac31ea4fae25143e82de62daabfbd
Summary:
Usually DLPack consumer is expected to call DLManagedTensor's
deleter to signal that it doesn't need the contents.
This patch calls the deleter when freeing unconsumed
DLPack capsules created by PyTorch.
Test script:
```
import torch
import torch.utils.dlpack
import gc
for i in range(10000):
a = torch.randn(1000,1000, dtype=torch.float32, device='cuda')
b = torch.utils.dlpack.to_dlpack(a)
gc.collect()
```
Before patch: consume all GPU ram.
After patch: constant GPU ram consumption.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9297
Differential Revision: D8781571
Pulled By: soumith
fbshipit-source-id: 2ebadec6c857646220d632ca64110af430dbd52f
* Bag of fixes
* Rename tensor_range.h to tensor_list_view.h
* Post rebase fixes
* Rename torch::tensor namespace to torch::tensors due to name conflict
* Avoid recursion in Module::to
* Some 0-sized dimension support, port catArray away from resizeLegacy.
The goal of this PR is to port catArray away from resizeLegacy (so we can delete the legacy resize calls), but since catArray has some weird behavior because
we don't have arbitrary 0-sized dimension support, I made some effort to fix these both in one pass.
The major changes here are:
1) catArray uses the new resize API, no longer the old resizeLegacy API.
2) As 1) is the last usage of resizeLegacy, it is deleted.
3) If compiled with USE_TH_SIZE_ZERO_DIM, catArray will work and properly check shapes for n-dimensional empty tensors.
4) However, we retain the old behavior of "ignoring" size [0] tensors in catArray. We previously allowed this because we didn't have n-dimensional empty tensors.
5) To get the above to work, we also add support for n-dimensional empty tensors for narrow and slice (ifdef USE_TH_SIZE_ZERO_DIM).
6) We change the stride formula for empty tensors to match NumPy; basically, we never multiply by 0 as the size, always at least 1, so the
strides are monotonically increasing in the empty tensor case.
7) We print the size of empty tensors if size != [0]; this matches NumPy behavior (even in cases where the size could be inferred from the brackets.
8) For test purposes, we add torch._C._use_zero_size_dim() to add tests for the above.
* Fix flake8.
* Address review comments.
* Build and install c10d from tools/build_pytorch_libs.sh
* Create initial Python bindings for c10d
* clang-format
* Switch link order to include more symbols
* Add bindings and tests for ProcessGroupGloo
* Add broadcast test
* Separate build flag for c10d
* Explicit PIC property
* Skip c10d tests if not available
* Remove c10d from Windows blacklist
Let it skip by itself because it won't be available anyway.
* Make lint happy
* Comments
* Move c10d module into torch.distributed
* Close tempfile such that it is deleted
* Test if ASAN is actually working as part of ASAN tests.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Drop explicit use of libstdc++, we should not care.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Build with DEBUG=1
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Increase main thread stack size when using ASAN.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
This makes the JIT tracer much more robust, by allowing it to record
dependencies on tensor sizes. For example, if you were to trace this
function
def fn(x):
return x.view(x.size(1), -1)
before this patch, then it would embed the actual value of x.size(1)
in the trace as a constant, making it very hard to have e.g. batch size
independent traces. Now, this will correctly record the dependency, and
will retrieve the size of x at every run.
* Split set_default_tensor_type(dtype) into set_default_dtype(dtype).
* Fix flake8.
The difference between this one and set_default_tensor_type is that it only sets scalar type what determines the type + device of a tensor returned from a factory function with defaults is the default tensor type + the current device (if the default tensor type is cuda). This just changes the scalar type of the default tensor type.
We do eventually want to deprecate set_default_tensor_type; it is not clear how to do that in a sensible and backwards compatible way.
* Separate cuda-ness from dtype.
There are no longer torch.cuda.int64, etc; only torch.int64 that correspond to at::ScalarType.
At the python arg parser level, the corresponding ATen type is selected from the combination of (ScalarType, Layout, Device).
There is also currently unused code in here for support ScalarType in native_functions; this will be used for specifying aggregate types
on reduction functions.
* Fix test_autograd.
* Add defaults to randint_like.
* Track is_cuda in py tensor types.
* Fix test_sparse.
* Fix multiprocessing.
* Fix rnn.
* Fix test_nn.
* Fix flake8.
* Add string-style devices to all tensors.
Previously, tensors only had a 'get_device' method which would throw an exception on a CPU tensor. This made it necessary to if/else code that
was meant to be device agnostic.
This PR implements the following:
1) Adds a 'device' property to all tensors that returns a string representation of the device for all tensors.
For cpu tensors this is 'cpu'. For cuda tensors this is 'cuda:X', where X is the cuda device ordinal.
2) Adds a DeviceSpec class. This is just a helper class for separating device_type and device_index specification and to allow partial specification.
For example, you can call DeviceSpec('cuda'), DeviceSpec('cuda:0'), DeviceSpec('cuda', 1).
Also has backwards compatibility support for specifying integers, which are treated as cuda devices.
DeviceSpecs have the following properties:
a) device_type: string representation of the device type (i.e. 'cpu' or 'cuda')
b) device_index: integer for the device index (None if not specified)
c) cuda_device_index: for backwards compatibility; behaves roughly like `get_device` did previously. I.e. if a function previously took integers for cuda devices,
it can now take DeviceSpecs (or strings), and can maintain the old functionality by calling `old_index = DeviceSpec(old).cuda_device_index`.
3) tensor methods and torch. functions that took integer devices can now take integers, strings, or DeviceSpecs. For example:
torch.randn((2,3), dtype=torch.cuda.float32, device='cuda:1')
TODO in future PRs:
A) Split out cuda from dtype so you don't need to overspecify cuda-ness
B) We currently only support strings/DeviceSpecs in tensor methods and torch. functions. We should have equivalents torch.cuda.device(...), torch.cuda.device_of, etc.
at the torch. level that work on strings/DeviceSpecs
* Add deviceInt64 to python arg parser.
* device_str.
* Remove device_str.
* remove device prefix from attributes.
* Use const char * instead of string.
* Move autogpu index out of Device.
* comment on is_default.
* Rename torch.DeviceSpec to torch.device.
* comment.
* Fix tests.
* Fix flake8.
* Fix sparse_coo_tensor parameter name.
* Improve error message.
* Remove device_ prefix from C++ device object.
* Allocate static strings.
* Return not implemented from rich compare.
* Move torch::Device to THPDevice.
* Remove cuda index.
* Py_RETURN_NOTIMPLEMENTED doesn't exist in python2.
This changes type(tensor) to return `torch.Tensor` instead of
`torch.autograd.Variable`.
This requires a few implementation changes:
- torch.Tensor is now a regular Python class instead of a
pseudo-factory like torch.FloatTensor/torch.DoubleTensor
- torch.autograd.Variable is just a shell with a __new__ function.
Since no instanes are constructed it doesn't have any methods.
- Adds torch.get_default_dtype() since torch.Tensor.dtype returns
<attribute 'dtype' of 'torch._C._TensorBase' objects>
We had a bug in the Buck build of PyTorch due to symbols from _C
being present in two shared libraries that were both loaded at
runtime. This caused global variables to be initialized twice and
destructed twice on exit. The second destruction often caused
segfaults on exit.
This attempts to detect that sort of situation early on. If
Module.cpp is compiled twice, the symbol
pytorch_duplicate_guard()::initialized will be shared. The second
initialization will print an error message and abort.
* Introduce torch.layout and split layout from dtypes.
Tensors (and tensor types) now have a 'layout' attribute that returns either 'torch.strided' or 'torch.sparse_coo'.
Previously, dtypes were 1-to-1 with ATen types/PyTensorTypes; the impetus behind this decision was to make things easy in the common case
(i.e. specifying a type in a factory function). But this doesn't really follow for sparity, which isn't a common case.
It also doesn't properly represent the concept or a dtype, which in numpy are proper scalar types (i.e. roughly the type returned from indexing the
last dimension of an n-d array). But this should be the same whether or not the tensor is represented via strides, sparsity, etc.
This is accomplished by:
1) having the dtype of tensor return the (device-type, scalar-type) combination, i.e. torch.cuda.float32, so both
torch.cuda.FloatTensor and torch.cuda.sparse.FloatTensor have the same dtype
2) Adding a layout parameter to python functions, where the combination of (dtype, layout) maps to an ATen type that is used for dispatch.
* Formatting, make init throw python_error.
* Fix cuda not enabled error message.
* Fix test.
This is the first of three PRs that #5537 will be split into.
This PR adds mkl headers to included files, and provides helper functions for MKL fft and cuFFT.
In particular, on POSIX, headers are using mkl-include from conda, and on Windows, it is from a new file @yf225 and I made and uploaded to s3.
* add mkl-include to required packages
* include MKL headers; add AT_MKL_ENABLED flag; add a method to query MKL availability
* Add MKL and CUFFT helpers
- Remove some uses of mega-header THP.h
- Use HANDLE_TH_ERRORS in functions that may throw
- Move NumPy includes to common header
- Delete unused allocator
* Add dtype to torch.Tensor, torch.FloatTensor, etc.
* Support passing dtypes to set_default_tensor_type.
* Check dtype exception.
* Correctly handle new type initialization order.
* Move handling of torch.Storage alias to C++.
* Delete function that erroneously reappeared.
This deletes most of the dead Tensor code paths, including the TensorMethods cwrap and generic/Tensor.cpp.
This also moves the THNN.cwrap/.cpp generation to generate_code which can use ninja if installed.
This replaces the torch.Tensor constructors with factories that produce
Variables. Similarly, functions on the torch module (e.g. torch.randn)
now return Variables.
To keep the PR to a reasonable size, I've left most of the unused tensor
code. Subsequent PRs will remove the dead code, clean-up calls to
torch.autograd.Variable, and rename Variable to Tensor everywhere.
There are some breaking changes because Variable and Tensors had
slightly different semantics. There's a list of those changes here:
https://github.com/pytorch/pytorch/wiki/Breaking-Changes-from-Variable-and-Tensor-merge
* Add numpy-style dtypes to Variable factories.
1) Add numpy-style dtypes corresponding to torch tensor types. These are:
torch.float16, torch.float32, torch.float64, torch.uint8, torch.int8, torch.int16, torch.int32, torch.int64
as well as torch.cuda, torch.sparse, and torch.cuda.sparse equivalents.
2) Adds "legacy" names for the above dtypes that correspond more closely to existing tensor names. These are:
torch.half, torch.float, torch.double, torch.short, torch.int, torch.long.
torch.byte and torch.char don't exist because they either don't match numpy semantics or differ on different architectures.
3) Adds a "dtype" parameter to Variable factories (e.g. zeros, ones) that allows the user to specify the type without changing the default tensor type.
4) Adds a "dtype" getter to Variables that return the canonical dtype from 1)
This PR is missing the following useful features that should be added in the future:
A) We only add the "dtype" parameter to auto-generated factories; hand-written factories like in tensor_new.cpp don't support this yet.
B) We don't allow type conversions to use dtypes; that should be added to type(param) or a new function.
C) We don't yet have a "device" parameter for these factories; right now, they will only create Variables on the default device.
* backend_to_string can be private.
* Define python binding argument indexes in a more simple way.
* add all_declared_types, still need to hook it up to THPDType.
* Fix all_declared_types for missing types (it's Sparse + Half).
* Ensure cuda dtypes are created even if compiled with NO_CUDA=1.
* Fix case where dtype is provided but dispatch is via namespace.
This happens in ones_like, empty_like, randn_like.
There is some question if we should do:
1) at::ones_like(tensor).toType(dtype)
2) at::ones_like(tensor.toType(dtype))
I did the former because this matches with the numpy documentation, i.e.:
"Overrides the data type of the result." and it's easier to implement.
Note that the above causes an extra copy, either of the input or output.
Here's a better implementation:
1) Make zeros_like, ones_like native functions that take an optional type (named dtype?).
2) Match the type argument with the dtype, so we don't have two different parameters.
3) Call at::zeros_like(input, type) -> at::native::zeros_like(input, type) -> type.zeros(input.sizes())
* Don't return from maybe_initialize_cuda.
* Don't leak DType name.
* Address cpp review comments.
* Share code between sparse and non-sparse test_dtypes.
* Rewrite _like functions as native function with explicit type parameter.
* Use type 'Type' instead of 'dtype' for consistency.
* Address review comments.
* Handle arg_idx when there is requires_grad but no dtype in python_binding_arguments.
* Enable scalars if compiled with WITH_SCALAR environment variable.
We are pretty close to enabling scalars (0-dimensional arrays); this allows turning them on
for development purposes and to be able to write code that works both with and without scalars enabled.
WITH_SCALARS is currently broken with distributions, but should work for test_torch, test_autograd, test_nn.
* Fix unsqueeze.
* Fix wrap dim, wrapping with Scalar.
* Use ATen infer_size implementation rather than TH.
The only substantitive difference between the two implementations is in how empty sizes are handled;
in ATen these are treated as scalars (i.e., can be expanded to anything), whereas in TH they are treated
as a special case of empty tensors (i.e., can't be expanded to anything). Therefore, this change is
necessary to support scalars (0-dimensional tensors). We could also take a bool parameter for determining
how we treat empty tensors but this seems unnecessary: if one tries to expand an empty tensors (as a result
of an infer_size calculation), the expansion will fail.
* Make changes for review.
* Attempt to fix windows build.
* long -> int.
This adds overrides in VariableType for the xxx_out ATen functions and
implements Python bindings. There is no support for automatic
differentiation. If any of the inputs (or outputs) requires grad, then the
function will throw an exception unless it's running in "no-grad" mode.
The bindings for calling torch.xxx functions on Variables are moved to a
different object. Previously, they were static method on VariableBase.
This change prevents users from accidentally calling static methods as if
they were instance methods.
* Fix the inconsistency of `polygamma` on Tensor and Variable.
Signed-off-by: HE, Tao <sighingnow@gmail.com>
* Regression test for #4466, polygamma works on variables.
Signed-off-by: HE, Tao <sighingnow@gmail.com>
* Add macro IMPLEMENT_STATELESS_SWAP to dispatch stateless methods on Variables correctly.
When call stateless methods with more than one arguments and the `self` comes second,
the `self` argument needs to be swapped to the first position before dispatching.
The macro `IMPLEMENT_STATELESS_ADDXX` is still reserved for deprecated `add**`
methods.
Signed-off-by: HE, Tao <sighingnow@gmail.com>
- Rename THNN convolution to have thnn_ prefix.
- Propagate CuDNN benchmark and deterministic to at::Context
- Add 'convolution', 'convNd' and 'conv_transposeNd' native wrappers, with defaults
The conv_transposeNd wrappers are updated to have the same argument
order as Python.
- torch.nn.functional directly dispatches to the native wrappers
- Make it possible to turn off tracing for some native wrappers, so I don't
have to write symbolics for all the functions above
- Spectral ops can now make use of CuDNN convolution if possible
- Better commentary on cudnn_batch_norm
- Turn on DCE for all JIT tests.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Refactor cudnn code layout / make build more robust.
When I previously moved cuDNN into ATen, I wasn't too familiar with the
ATen native function directory layout, and so I did a number of
suboptimal things. This commit fixes those problems.
- If NO_CUDA was set but cuDNN is installed on your system, we'd incorrectly
assume that CUDNN was enabled, to hilarious effect.
- We now distinguish between cudnn implementation files and cudnn
native function files. The native files now live in ATen/native/cudnn,
and are *unconditionally compiled*, even when we are not building with cuDNN.
This means that we can unconditionally declare cudnn functions in yaml
and they are always available, even if they are broken. The cuDNN specific
files live in 'cudnn', they are *never* installed, and they are used
purely for implementation purposes. I had to add stub implementations of
all ATen functions to achieve this.
- I had written headers for at::native functions manually, but codegen
will generate them for me automatically. So I deleted the headers.
That lets me get rid of some header install logic as well.
- There's a new note about ATen preprocessor philosophy.
This removes volatile from Variable. The functionality is mostly
replaced by a global (thread-local) flag, which is controlled by
torch.set_grad_enabled() and the context manager torch.no_grad().
In C++, the flag is exposed through GradMode::is_enabled() and GradMode::set_enabled()
Fixes#3627
* Support CPU Apply directly in ATen and implement standard_gamma using it.
Main changes in this PR:
1) Added a TH_APPLY-style templatized function for CPU apply calls (currently only 2 and 3 tensor argument
versions are supported, but more are easy to add). In fact, this is basically identical to TH_APPLY, except
it uses ATen functions and the API is a template instead of a macro. The template takes an operation that
is performed on the data (and an indicator to signal early termination); i.e. you don't need to know that
x_data is a pointer to the current data location of x.
2) Refactors the ATen dispatch code to easily generate dispatch code for different subsets of the scalar types.
This is in preference to the template_scalar path, which requires valid specialization of each scalar type. Valid
specializations are particularly annoying with CUDA because you most likely can't put the specializations
in a header so need to write some sort of for-all-scalar-type macro to get the correct specializations.
Currently, we only generate dispatch_all (all scalar types, the equivalent existed already), and
dispatch_cpu_floating_types (which is used by standard_gamma).
3) Implements standard_gamma using the above changes (this is an arbitrary choice, it was the latest
apply macro to be committed). The forward is bound via Declarations.yaml,
the backward via the Apply template, and then they are hooked together in derivatives.yaml. This eliminates
needing to change TH at all going forward, which means one can write idiomatic C++ instead of the TH-style macros
(e.g. TH_MATH_NAME).
* Generate Dispatch code with nicer spacing.
* Small cleanups.
* Fix typo.
* Add TODOs for changing macros, remove dead code.
* Use a lambda function.
* Get rid of early exit.
* Rename Scalar,ScalarType template parameters to CScalar.
* Reorder _standard_gamma_grad parameters.
* Add comments explaining calling convention.
* Don't generate Dispatch.h anymore.
* Get rid of backend specific checks in dispatch.
* Fix empty/scalar check.
This is not currently used by anything, but eventually ATen
will need to make decisions about whether or not to use
CuDNN functions or not, which means we need to propagate
this variable to ATen.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Bind cauchy_, exponential_, normal_, uniform_ functions to THPVariable.
Also changes the error messages around Generator parser; previously, you'd get an error
like: torch._C.Generator is not a torch.Generator; now the check is proper but returns
that only None is supported.
* Support passing Generators to ATen Variable-bound methods.
This involves changing THPGenerator to have an at::Generator rather than a THGenerator.
TH getRNGState, setRNGState are still called directly because they are not bound from ATen yet;
they should probably be on the Generators and return (opaque) GenerateState objects.
* Fix default values.
* Properly use THRandom_initialSeed.
* update standard gamma to use new default generator.
Implements from_numpy using ATen tensors. Variable.from_numpy is a
convenient placeholder for the variant that returns Variables until we
merge Tensor and Variable.
The behavior is slightly changed:
- from_numpy() on an empty array now returns an empty tensor instead of
throwing an exception. The shape may not be preserved.
- CharTensor(ndarray) used to throw an exception. It now copies the
ndarray. Copying is implemented via ATen toType.
* Comprehensive rewrite of Torch CuDNN bindings / a bit of ATen infra
The executive summary is that this moves the torch/csrc/cudnn
library into ATen, adding a number of new cudnn_ methods to ATen
for batchnorm, convolution, affine grid generator and grid sampler.
ATen infra changes:
- TensorGeometry was moved to ATen
- TensorGeometry was modified to make its interface resemble that of
Tensor; in particular, sizes is no longer a field, it's a method.
- AT_CUDA_ENABLED macro is set via ATen/Config.h header which is
generated at cmake configure time.
Fixes https://github.com/zdevito/ATen/issues/168
- Change AT_CUDA_ENABLED macro to be a function macro, so that we
error if it is not defined
- Introduce a new TensorArg class, which is a Tensor plus a little
metadata. This helps us give good error messages when checking
dimensions/shapes of tensors.
Fixes https://github.com/zdevito/ATen/issues/169
- Also introduce a TensorGeometryArg class, for when you don't
need the actual tensor data (which is most of the time.)
- Add ATen/Check.h, which contains a number of utility functions
for testing shapes, types and devices of input tensors. This
will be particulary useful for native methods, which don't get
code generated input testing code. These functions take a
'CheckedFrom' argument, at the moment just a string, which
specifies some extra information about what function was
doing the actual checking; this greatly improves error messages.
- Many check functions take initializer lists, which let you
test that all tensors have some property. This API is
peculiar, in that we IGNORE undefined tensors in this case.
This is handled by filterDefined.
- Add AT_CUDNN_ENABLED macro
- CuDNN linking from ATen was improved; for example, we now actually
add the CuDNN headers to our include path.
- Add some missing override specifiers to some methods
- We now actually build tests with CUDA functionality accessible
(previously, AT_CUDA_ENABLED was not defined, meaning that
the headers were missing all CUDA-only functionality.)
- Native functions now support giving explicit names to return
outputs in yaml. This makes it possible to hook into the NN
autogenerated derivatives codepath using native functions.
CuDNN rewrite changes:
- torch/csrc/cudnn now uses ATen (rather than passing around
THVoidTensor) and lives in ATen. This lets us remove tensorPointer
shenanigans. The functions are exposed to ATen as native functions
described in aten/src/ATen/cudnn/cuDNN.yaml
- ATen now builds and links against CuDNN when enabled. The cmake
package script was taken from Caffe2.
- Some header reorganization was done to help reduce dependencies
on headers (this reorg is no longer used but I've kept it)
- Rename CHECK to CUDNN_CHECK
- Rip out old shape/type testing code in favor of modern ATen/Check.h
interface using TensorArg. In many cases, increase the robustness of
the checking code.
- Change the inputs of the public facing functions, so that they can
be bound by ATen
- Delete THCState*; this is retrieved from the global ATen context
- Delete cudnnHandle_t, this is retrieved from the global Handles.h
- Delete cudnnDataType_t, this is retrieved from the Tensor type
- Delete Convolution class, instead its constituent arguments are
passed individually
- Change functions to return tensors, rather than take an appropriately
sized output tensor as an input.
- Redo how transposed convolution / backward convolution is implemented
(knock on effect of returning tensors). Previously it was assumed
that you would always pass an appropriately sized output tensor, but
we don't want to do this anymore. For backwards, we instead give
the desired output tensor (input, really) size, because that is
readily available. For *transposed* convolution, however, we take
output_padding, and otherwise do the shape calculation.
- Redo how legacy group convolution is implemented (knock on effect from
porting cudnn to ATen.) Previously, group convolution was implemented
by manually constructing sizes and strides and then outputting
appropriate, with macros switching between individual groups and
all-at-once based on CuDNN version. Now, the code looks exactly what
you'd expect: there's a top-level wrapping function that supports
group convolution no matter the version of CuDNN, and a low-level
wrapper which supports only what CuDNN supports. The top-level
function conditions on CuDNN version, and invokes the low-level
interface 1 or n times.
- There is now a debugging printer for tensor descriptors.
- Convolution struct is replaced with ConvolutionArgs, which is not
part of the public API but is used internally to conveniently
pass around all of the arguments needed for Convolution.
- Add some constexprs for well-known dimensions, reduce amount of
magic numbers in code.
- Put 'deterministic' in to ConvParams. Fixes#3659
- Lots more comments.
- Some pessimizations, in the name of code clarity:
- The descriptors are initialized on every invocation of convolution
forward/backward. Previously, the descriptors were cached, so that
you didn't have to initialize them again on backwards. This is
difficult to support in the ATen interface so I didn't support it.
- Legacy group convolution initializes its workspace for *every* group
it performs. I did not feel motivated to fix this because the
legacy codepath is already quite slow.
- Affine grid generator and grid sampler automatically call contiguous
on their arguments as necessary.
- Batchnorm input checking is greatly beefed up, it now checks for
the following input characteristics:
- Definedness
- GPU location
- Type
- Contiguity
- Size
PyTorch binding code changes
- batchnorm now uses consistent var/data naming
- batchnorm and convolution make use of new ATen bindings
- Affine grid generator and grid sampler make use of ATen CuDNN
bindings via derivatives.yaml. This means I had to restructure
the code a little, since the THNN bindings still go through
a legacy Python class.
- I fixed some warnings:
- s/friend class/friend struct/ on InterpreterStateImpl
- Removed pessimizing move 'detached' in torch/csrc/autograd/variable.cpp
- Removed unused pack_list on Scalar
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
GCC 4.8 buildfix
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Add TensorGeometry to ATen.h
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
CUDNN_CHECK
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Update TODO comment
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Delete return in cudnn_grid_sampler
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
s/cudnnSetStreamToCurrent/setCuDNNStreamToCurrent/g
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Don't allocate a new vector when filtering defined.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Remove Check overloads, convert to pass references.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Some more microbenchmarking.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Add torch.take and Tensor.put_
These are similar to numpy.take and numpy.put. The take function allows
you to linearly index into a tensor without viewing it as a 1D tensor
first. The output has the same shape as the indices. The put function
copies value into a tensor also using linear indices.
This includes some changes to the dispatch code for torch.xxx functions:
- Since Variable.addmm is an instance-method, the self argument has to
come first. The dispatch code swaps the first two arguments if
necessary to suppor the deprecated signatures where 'alpha' or 'beta'
comes before the 'self' tensor.
- Delete IMPLEMENT_STATELESS_REVERSED. These functions require output
arguments to be passed in using the keyword 'out'. They were meant to
handle torch.gt(out, a, b), but we haven't allowed that for a while.
ATen has it's own default CPU RNG. Use this as the default in PyTorch so
that random functions called through ATen have the same behavior as
random functions called through TensorMethods
Along the way I added converters for Variable and TracingInput. Variable should
probably be moved to a more widely known spot.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Previously, our AST was a DAG, where shared Nodes indicated a computation
should be reused. This commit rewrites the IR into a new functional
representation which represents sharing explicitly using variable
bindings.
We offer a few justifications for this new style:
1. The new representation is not all that different from the
old one; it is about as easy to construct, and the lack of an
explicit graph doesn't negatively impact our ability to interpret
the graph, since we've chosen, as a matter of design, to NOT have
the IR participate in the actual execution of a graph.
2. The new let-binding representation has an implicit ordering,
which we can use to conveniently keep track of the original order
the trace showed up as. This automatically gives us a topsort,
and gives us an easier to read textual representation of our
IR:
%14 = Embedding %11, %0, -1, None, 2, False, False
%15 = Dropout %14, 0.2, True, False
%16 = Index %12, 0
%17 = Index %12, 1
%18 = Index %13, 0
%19 = Index %13, 1
%20 = Index %15, 0
%21 = Linear %20, %1, %3
%22 = Linear %16, %2, %4
3. It moves us closer to a Futhark style language
(http://futhark-lang.org/publications/pldi17.pdf).
Major aspects of the diff
- Node is replaced with Expr and Arg, a pair of mutually recursive
structures which represent our new language. In BNF, the language
looks like this:
a ::= c | %i
e ::= %i, ... = e
| PyOp e, ...
| Ret %i, ...
Technically, Ret is not actually a return (no control flow is involved),
it just tuples up a series of tensors (identified by variables).
One important invariant is that locals are always tensors; they
are never constants (this is asymmetric with Args.)
- Arguments support Python constants. This is an important piece because
many operators take extra Python literals like integers and tuples in
order to specify extra parameters about how an operator operates. Adding
this was essential to getting word_language_model to work.
- As both Expr and Arg have multiple variants, there is new infrastructure
for doing case on the variants using ExprVisitor and ArgVisitor. The
strategy here is adapted from WebAssembly's visitors, although we have
generalized to permit arbitrary argument forwarding, which is necessary
to support tail-recursive visitor calls. TCO is important because our
interpreter may recurse arbitrarily deep into a stack of nested lets.
If users wish, they can also manually case on the type tag.
- Tracing is now turned on and off using _tracer_enter/_tracer_exit in
torch._C. _tracer_enter accepts a list of variables which are to be
treated as arguments; _tracer_exit accepts the list of traced variables
which should be returned when you reexecute the trace, and returns
the trace expression which can be reexecuted. GlobalTracingState
is a global variable which tracks whether or not we are tracing or not.
- You use run_forward to execute a trace on some set of parameters.
- When under tracing, variables keep track, via trace_local, what the
name of their variables in the IR are.
Here is a simple runner which leaks memory but can be used to JIT models:
import torch.autograd.function as F
import torch._C
def jit(model):
import types
real_forward = model.forward
def forward(self, *args):
def flatten(x):
return tuple(F._iter_variables(x))
if not hasattr(self, "saved_trace"):
torch._C._tracer_enter(tuple(self.parameters()) + flatten(args))
out = real_forward(*args)
self.saved_trace = torch._C._tracer_exit(flatten(out))
self.saved_outs = out
return out
else:
flat_out = Variable._execution_engine.run_forward(self.saved_trace, tuple(self.parameters()) + flatten(args))
return F._unflatten(flat_out, self.saved_outs)
Major problems:
- Sanity checking is spotty at best, especially when users pass in variables.
- The interpreter leaks tensor memory from the store. When we add back def-use
we should be able to deallocate tensors as soon as we know they are no longer
necessary.
- The interpreter needs to reach feature parity with the old execution engine.
From there, we need to see if backwards can be subsumed as well.
- I still have no confidence in having memory managed everything correctly.
This requires a close look.
- Rather than return an *open* expression as a trace, we should return a
*lambda* instead, which knows about how many formal parameters it
requires.
- The IR is not introspectable from Python at the moment, but this is simply a
matter of implementing all the binding code.
- The tracer is NOT reentrant (you can't trace while you're inside a trace.)
Furthermore, no sanity checking is done if you try to incorrectly reuse
things from one trace in another.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Setting torch.utils.backcompat.broadcast.warning.enabled=True
will cause Python warnings in the case where broadcast occurs
but previously 1-d view style pointwise ops occured.
Otherwise, on many machines, the size of the OpenMP thread pool will
change between MKL and our OpenMP enabled functions. The constant thread
creation and destruction results in worse performance and leaks memory
on GCC 5.4
The core autograd Variable, Function, and Engine no longer depend on the
Python API. This let's us implement functions in C++. In the future, we
can also multithread engine and release the GIL for most of the
non-Python backwards.
This hooks into the (internal) ForkingPickler class in multiprocessing
to reduce tensors, storages, and CUDA events instead of our queue from
joblib. This makes it easier to use the standard multiprocessing classes
in later versions of Python.
This also exposes:
- Tensor/Storage.share_memory_()
- Module.share_memory()
These methods move the CPU tensors and storages to shared memory. If
you're using the "fork" method of multiprocessing, these objects can be
directly inherited instead of serialized through a queue.
See issue #20
The torch.Size class is a tuple subclass which distinguishes sizes from
other tuples so that torch.Tensor(size) is interpreted as size instead
of data.
