一個Tensor的一生 - torch.rand篇

置顶 已于 2024-01-01 12:04:39 修改
阅读量613 收藏 9
点赞数 8
分类专栏： PyTorch 文章标签： python pytorch c++
于 2023-12-05 22:43:48 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/keineahnung2345/article/details/134819131
版权

一個Tensor的一生 - torch.rand篇

前言

Life of a Tensor這篇文章中介紹了torch.rand函數從Python API一直到C++底層的調用過程，寫於2019年7月，當時的PyTorch版本為1.1.0。參考該篇文章，本篇的關注點同樣在整個調用流程，不過是基於較新的PyTorch 2.0版。

首先編寫一個torch_rand.py如下：

import torch
torch.rand(3, 4)

如果我們查找torch.rand的定義，會跳到torch/_C/_VariableFunctions.pyi中。

torch/_C/_VariableFunctions.pyi

torch/_C/_VariableFunctions.pyi這個檔案裡面共有8個不同簽名的rand函數：

@overload
def rand(size: Sequence[Union[_int, SymInt]], *, generator: Optional[Generator], names: Optional[Sequence[Union[str, ellipsis, None]]], dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(*size: _int, generator: Optional[Generator], names: Optional[Sequence[Union[str, ellipsis, None]]], dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(size: Sequence[Union[_int, SymInt]], *, generator: Optional[Generator], out: Optional[Tensor] = None, dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(*size: _int, generator: Optional[Generator], out: Optional[Tensor] = None, dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(size: Sequence[Union[_int, SymInt]], *, out: Optional[Tensor] = None, dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(*size: _int, out: Optional[Tensor] = None, dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(size: Sequence[Union[_int, SymInt]], *, names: Optional[Sequence[Union[str, ellipsis, None]]], dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...
@overload
def rand(*size: _int, names: Optional[Sequence[Union[str, ellipsis, None]]], dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...

torch.rand(3, 4)調用的是上面倒數第三個rand函數。

注1：關於此處出現的*size, _int及Sequence等型別，以及, *, 的用法，詳見Python typing函式庫和torch.types。

注2：torch/_C/_VariableFunctions.pyi這個檔案是由gen_pyi.py依據torch/_C/_VariableFunctions.pyi.in這個模板及native_functions.yaml生成的，想知道它具體是怎麼被生成的，詳見PyTorch中的pyi檔案生成機制。

如果把擂同的部份對齊，完全相同的部份省略：

@overload
def rand(size: Sequence[Union[_int, SymInt]], *, generator: Optional[Generator], names: Optional[Sequence[Union[str, ellipsis, None]]], ...
@overload
def rand(*size: _int,                            generator: Optional[Generator], names: Optional[Sequence[Union[str, ellipsis, None]]], ...
@overload
def rand(size: Sequence[Union[_int, SymInt]], *, generator: Optional[Generator], out: Optional[Tensor] = None, ...
@overload
def rand(*size: _int,                            generator: Optional[Generator], out: Optional[Tensor] = None, ...
@overload
def rand(size: Sequence[Union[_int, SymInt]], *,                                 out: Optional[Tensor] = None, ...
@overload
def rand(*size: _int,                                                            out: Optional[Tensor] = None, ...
@overload
def rand(size: Sequence[Union[_int, SymInt]], *,                                 names: Optional[Sequence[Union[str, ellipsis, None]]], ...
@overload
def rand(*size: _int,                                                            names: Optional[Sequence[Union[str, ellipsis, None]]], ...

可以依據參數將它們分為以下四類：

  • size, generator, names

  • size, generator（out可省略）

  • size（out可省略）

  • size, names

這四類中的每一類又各自都有兩個版本，即第一個參數（形狀）接受的是_int或者是Sequence。

如果將Python腳本中的代碼改成：

torch.rand((3,4))

或：

torch.rand([3,4])

就會調用到倒數第四個、Sequence版本的rand函數（跟剛剛看到的倒數第三個、_int版本的是一對）：

@overload
def rand(size: Sequence[Union[_int, SymInt]], *, out: Optional[Tensor] = None, dtype: Optional[_dtype] = None, layout: Optional[_layout] = None, device: Optional[Union[_device, str, None]] = None, pin_memory: Optional[_bool] = False, requires_grad: Optional[_bool] = False) -> Tensor: ...

backtrace

接下來新建一個名為gdbrand的檔案並將以下內容填入其中：

python sys.path.append("/usr/share/gcc/python");
set logging file rand.txt
set logging on
set breakpoint pending on
break at::empty
info breakpoints
run torch_rand.py
bt

開啟gdb python，輸入source gdbrand指令執行上述腳本，就會得到以下的backtrace：

#0  at::empty (size=..., options=..., memory_format=memory_format@entry=...) at /root/Documents/pytorch/build/aten/src/ATen/Functions.h:2652
#1  0x00007f04620d60b1 in at::native::rand (size=..., generator=..., dtype=..., layout=..., device=..., pin_memory=...)
    at /root/Documents/pytorch/c10/util/Optional.h:204
#2  0x00007f04620d61fc in at::native::rand (size=..., dtype=..., layout=..., layout@entry=..., device=..., pin_memory=...)
    at /root/Documents/pytorch/aten/src/ATen/native/TensorFactories.cpp:781
#3  0x00007f0462dd2c28 in at::(anonymous namespace)::(anonymous namespace)::wrapper_CompositeExplicitAutograd__rand (pin_memory=..., device=..., layout=..., 
    dtype=..., size=...) at /root/Documents/pytorch/build/aten/src/ATen/RegisterCompositeExplicitAutograd.cpp:2214
#4  c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<at::Tensor(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>), at::(anonymous namespace)::(anonymous namespace)::wrapper_CompositeExplicitAutograd__rand>, at::Tensor, c10::guts::typelist::typelist<c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > >::operator() (args#4=..., args#3=..., args#2=..., args#1=..., args#0=..., this=<optimized out>)
    at /root/Documents/pytorch/aten/src/ATen/core/boxing/impl/WrapFunctionIntoFunctor.h:13
#5  c10::impl::wrap_kernel_functor_unboxed_<c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<at::Tensor(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>), at::(anonymous namespace)::(anonymous namespace)::wrapper_CompositeExplicitAutograd__rand>, at::Tensor, c10::guts::typelist::typelist<c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > >, at::Tensor(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)>::call(c10::OperatorKernel *, c10::DispatchKeySet, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>) (functor=<optimized out>, args#0=..., args#1=..., args#2=..., 
    args#3=..., args#4=...) at /root/Documents/pytorch/aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor.h:463
#6  0x00007f0462889309 in c10::callUnboxedKernelFunction<at::Tensor, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > (dispatchKeySet=..., functor=<optimized out>, unboxed_kernel_func=<optimized out>)
    at /root/Documents/pytorch/aten/src/ATen/core/boxing/KernelFunction_impl.h:50
#7  c10::KernelFunction::call<at::Tensor, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > (dispatchKeySet=..., opHandle=..., this=0x2075db8) at /root/Documents/pytorch/aten/src/ATen/core/boxing/KernelFunction_impl.h:91
#8  c10::Dispatcher::redispatch<at::Tensor, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> >(c10::TypedOperatorHandle<at::Tensor (c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)> const&, c10::DispatchKeySet, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>) const (this=<optimized out>, currentDispatchKeySet=..., op=...)
    at /root/Documents/pytorch/aten/src/ATen/core/dispatch/Dispatcher.h:656
#9  c10::TypedOperatorHandle<at::Tensor (c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)>::redispatch(c10::DispatchKeySet, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>) const (args#4=..., args#3=..., args#2=..., args#1=..., args#0=..., currentDispatchKeySet=..., this=0x7f046a780150 <at::_ops::rand::redispatch(c10::DispatchKeySet, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)::op>) at /root/Documents/pytorch/aten/src/ATen/core/dispatch/Dispatcher.h:492
#10 at::_ops::rand::redispatch (dispatchKeySet=..., size=..., dtype=..., layout=..., layout@entry=..., device=..., pin_memory=...) at /root/Documents/pytorch/build/aten/src/ATen/Operators_2.cpp:5220
#11 0x00007f0462c05329 in at::(anonymous namespace)::rand (pin_memory=..., device=..., layout=..., dtype=..., size=...) at /root/Documents/pytorch/build/aten/src/ATen/RegisterBackendSelect.cpp:365
#12 c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<at::Tensor(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>), at::(anonymous namespace)::rand>, at::Tensor, c10::guts::typelist::typelist<c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > >::operator() (args#4=..., args#3=..., args#2=..., args#1=..., args#0=..., this=<optimized out>) at /root/Documents/pytorch/aten/src/ATen/core/boxing/impl/WrapFunctionIntoFunctor.h:13
#13 c10::impl::wrap_kernel_functor_unboxed_<c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<at::Tensor(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>), at::(anonymous namespace)::rand>, at::Tensor, c10::guts::typelist::typelist<c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > >, at::Tensor(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)>::call(c10::OperatorKernel *, c10::DispatchKeySet, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>) (functor=<optimized out>, args#0=..., args#1=..., args#2=..., args#3=..., args#4=...) at /root/Documents/pytorch/aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor.h:463
#14 0x00007f04628e3422 in c10::callUnboxedKernelFunction<at::Tensor, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > (dispatchKeySet=..., functor=<optimized out>, unboxed_kernel_func=<optimized out>) at /root/Documents/pytorch/aten/src/ATen/core/boxing/KernelFunction_impl.h:50
#15 c10::KernelFunction::call<at::Tensor, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> > (dispatchKeySet=..., opHandle=..., this=0x2076638) at /root/Documents/pytorch/aten/src/ATen/core/boxing/KernelFunction_impl.h:91
#16 c10::Dispatcher::call<at::Tensor, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool> >(c10::TypedOperatorHandle<at::Tensor (c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)> const&, c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>) const (op=..., this=<optimized out>) at /root/Documents/pytorch/aten/src/ATen/core/dispatch/Dispatcher.h:639
#17 c10::TypedOperatorHandle<at::Tensor (c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)>::call(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>) const (args#4=..., args#3=..., args#2=..., args#1=..., args#0=..., this=0x7f046a780170 <at::_ops::rand::call(c10::ArrayRef<c10::SymInt>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>)::op>) at /root/Documents/pytorch/aten/src/ATen/core/dispatch/Dispatcher.h:487
#18 at::_ops::rand::call (size=..., dtype=..., layout=..., layout@entry=..., device=..., pin_memory=...) at /root/Documents/pytorch/build/aten/src/ATen/Operators_2.cpp:5213
#19 0x00007f046abdd595 in at::rand_symint (options=..., size=...) at /root/Documents/pytorch/build/aten/src/ATen/Functions.h:5770
#20 torch::rand_symint (options=..., size=...) at /root/Documents/pytorch/torch/csrc/autograd/generated/variable_factories.h:418
#21 operator() (__closure=<synthetic pointer>, options=..., size=...) at /root/Documents/pytorch/torch/csrc/autograd/generated/python_torch_functions_0.cpp:5256
#22 torch::autograd::THPVariable_rand (self_=<optimized out>, args=<optimized out>, kwargs=<optimized out>) at /root/Documents/pytorch/torch/csrc/autograd/generated/python_torch_functions_0.cpp:5258
#23 0x0000000000508127 in cfunction_call (func=0x7f048932aa90, args=<optimized out>, kwargs=<optimized out>) at /usr/local/src/conda/python-3.9.13/Objects/methodobject.c:543
#24 0x00000000004f0edc in _PyObject_MakeTpCall (tstate=0x1a610a0, callable=0x7f048932aa90, args=<optimized out>, nargs=<optimized out>, keywords=0x0) at /usr/local/src/conda/python-3.9.13/Objects/call.c:191
#25 0x00000000004ed255 in _PyObject_VectorcallTstate (kwnames=0x0, nargsf=<optimized out>, args=0x1abc308, callable=0x7f048932aa90, tstate=<optimized out>) at /usr/local/src/conda/python-3.9.13/Include/cpython/abstract.h:116
#26 _PyObject_VectorcallTstate (kwnames=0x0, nargsf=<optimized out>, args=0x1abc308, callable=0x7f048932aa90, tstate=<optimized out>) at /usr/local/src/conda/python-3.9.13/Include/cpython/abstract.h:103
#27 PyObject_Vectorcall (kwnames=0x0, nargsf=<optimized out>, args=0x1abc308, callable=0x7f048932aa90) at /usr/local/src/conda/python-3.9.13/Include/cpython/abstract.h:127
#28 call_function (kwnames=0x0, oparg=<optimized out>, pp_stack=<synthetic pointer>, tstate=0x1a610a0) at /usr/local/src/conda/python-3.9.13/Python/ceval.c:5077
#29 _PyEval_EvalFrameDefault (tstate=<optimized out>, f=0x1abc190, throwflag=<optimized out>) at /usr/local/src/conda/python-3.9.13/Python/ceval.c:3489
#30 0x00000000004e70ca in _PyEval_EvalFrame (throwflag=0, f=0x1abc190, tstate=0x1a610a0) at /usr/local/src/conda/python-3.9.13/Include/internal/pycore_ceval.h:40
#31 _PyEval_EvalCode (tstate=<optimized out>, _co=<optimized out>, globals=<optimized out>, locals=<optimized out>, args=<optimized out>, argcount=<optimized out>, kwnames=0x0, kwargs=0x0, kwcount=<optimized out>, kwstep=2, defs=0x0, defcount=<optimized out>, kwdefs=0x0, closure=0x0, name=0x0, qualname=0x0) at /usr/local/src/conda/python-3.9.13/Python/ceval.c:4329
#32 0x00000000004e6d57 in _PyEval_EvalCodeWithName (_co=<optimized out>, globals=<optimized out>, locals=<optimized out>, args=<optimized out>, argcount=<optimized out>, kwnames=<optimized out>, kwargs=0x0, kwcount=0, kwstep=2, defs=0x0, defcount=0, kwdefs=0x0, closure=0x0, name=0x0, qualname=0x0) at /usr/local/src/conda/python-3.9.13/Python/ceval.c:4361
#33 0x00000000004e6d09 in PyEval_EvalCodeEx (_co=<optimized out>, globals=<optimized out>, locals=<optimized out>, args=<optimized out>, argcount=<optimized out>, kws=<optimized out>, kwcount=0, defs=0x0, defcount=0, kwdefs=0x0, closure=0x0) at /usr/local/src/conda/python-3.9.13/Python/ceval.c:4377
#34 0x0000000000594e7b in PyEval_EvalCode (co=co@entry=0x7f049ec05870, globals=globals@entry=0x7f049ebfd780, locals=locals@entry=0x7f049ebfd780) at /usr/local/src/conda/python-3.9.13/Python/ceval.c:828
#35 0x00000000005c2307 in run_eval_code_obj (tstate=0x1a610a0, co=0x7f049ec05870, globals=0x7f049ebfd780, locals=0x7f049ebfd780) at /usr/local/src/conda/python-3.9.13/Python/pythonrun.c:1221
#36 0x00000000005be270 in run_mod (mod=<optimized out>, filename=<optimized out>, globals=0x7f049ebfd780, locals=0x7f049ebfd780, flags=<optimized out>, arena=<optimized out>) at /usr/local/src/conda/python-3.9.13/Python/pythonrun.c:1242
#37 0x00000000004563ed in pyrun_file (fp=0x1a5d440, filename=0x7f049eb2d930, start=<optimized out>, globals=0x7f049ebfd780, locals=0x7f049ebfd780, closeit=1, flags=0x7ffecfd9a7f8) at /usr/local/src/conda/python-3.9.13/Python/pythonrun.c:1140
#38 0x00000000005b8062 in pyrun_simple_file (flags=0x7ffecfd9a7f8, closeit=1, filename=0x7f049eb2d930, fp=0x1a5d440) at /usr/local/src/conda/python-3.9.13/Python/pythonrun.c:450
#39 PyRun_SimpleFileExFlags (fp=0x1a5d440, filename=<optimized out>, closeit=1, flags=0x7ffecfd9a7f8) at /usr/local/src/conda/python-3.9.13/Python/pythonrun.c:483
#40 0x00000000005b55ce in pymain_run_file (cf=0x7ffecfd9a7f8, config=0x1a5faa0) at /usr/local/src/conda/python-3.9.13/Modules/main.c:379
#41 pymain_run_python (exitcode=0x7ffecfd9a7f0) at /usr/local/src/conda/python-3.9.13/Modules/main.c:604
#42 Py_RunMain () at /usr/local/src/conda/python-3.9.13/Modules/main.c:683
#43 0x0000000000588ff9 in Py_BytesMain (argc=<optimized out>, argv=<optimized out>) at /usr/local/src/conda/python-3.9.13/Modules/main.c:1129
#44 0x00007f049ed56d90 in __libc_start_call_main (main=main@entry=0x588fb0 <main>, argc=argc@entry=2, argv=argv@entry=0x7ffecfd9aa28) at ../sysdeps/nptl/libc_start_call_main.h:58
#45 0x00007f049ed56e40 in __libc_start_main_impl (main=0x588fb0 <main>, argc=2, argv=0x7ffecfd9aa28, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>, stack_end=0x7ffecfd9aa18) at ../csu/libc-start.c:392
#46 0x0000000000588eae in _start ()

backtrace應由下往上看，一開始是_start，中間經過十幾個Python自身的函數，來到的第一個與PyTorch相關的函數為第22個節點：torch::autograd::THPVariable_rand。

Python bindings

torch::autograd::THPVariable_rand

#22 torch::autograd::THPVariable_rand (self_=<optimized out>, args=<optimized out>, kwargs=<optimized out>) at /root/Documents/pytorch/torch/csrc/autograd/generated/python_torch_functions_0.cpp:5258

torch::autograd::THPVariable_rand函數定義於torch/csrc/autograd/generated/python_torch_functions_0.cpp，注意到該資料夾下還有其它名字類似但後綴不同的cpp檔。這些檔案（python_torch_functions_i.cpp）是在編譯PyTorch時由gen.py腳本依據native_functions.yaml及tools/autograd/templates/python_torch_functions.cpp這個模板來生成的，詳見PyTorch中的python_torch_functions_i.cpp檔案生成機制。

THPVariable_rand函數的宣告如下：

static PyObject * THPVariable_rand(PyObject* self_, PyObject* args, PyObject* kwargs);

定義PyMethodDef陣列，Python中的rand函數會被對應到C++的THPVariable_rand函數：

static PyMethodDef torch_functions_shard[] = {
  //...
  {"rand", castPyCFunctionWithKeywords(THPVariable_rand), METH_VARARGS | METH_KEYWORDS | METH_STATIC, NULL},
  //...
}

THPVariable_rand函數的具體實現如下：

// generated methods start here

// ...

// rand
static PyObject * THPVariable_rand(PyObject* self_, PyObject* args, PyObject* kwargs)
{
  HANDLE_TH_ERRORS
  static PythonArgParser parser({
    "rand(SymIntArrayRef size, *, Generator? generator, DimnameList? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",
    "rand(SymIntArrayRef size, *, Generator? generator, Tensor out=None, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",
    "rand(SymIntArrayRef size, *, Tensor out=None, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",
    "rand(SymIntArrayRef size, *, DimnameList? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",
  }, /*traceable=*/true);

  ParsedArgs<8> parsed_args;
  auto _r = parser.parse(nullptr, args, kwargs, parsed_args);
  if(_r.has_torch_function()) {
    return handle_torch_function(_r, nullptr, args, kwargs, THPVariableFunctionsModule, "torch");
  }
  switch (_r.idx) {
    case 0: {
      // aten::rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
      auto __names = _r.toDimnameListOptional(2);
      c10::optional<DimnameList> names = __names ? c10::make_optional(DimnameList(__names.value())) : c10::nullopt;
      const auto options = TensorOptions()
          .dtype(_r.scalartypeOptional(3))
          .device(_r.deviceWithDefault(5, torch::tensors::get_default_device()))
          .layout(_r.layoutOptional(4))
          .requires_grad(_r.toBool(7))
          .pinned_memory(_r.toBool(6));
      torch::utils::maybe_initialize_cuda(options);
      
      auto dispatch_rand = [](c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::TensorOptions options) -> at::Tensor {
        pybind11::gil_scoped_release no_gil;
        return torch::rand_symint(size, generator, names, options);
      };
      return wrap(dispatch_rand(_r.symintlist(0), _r.generator(1), names, options));
    }
    case 1: {
      if (_r.isNone(2)) {
        // aten::rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
        const auto options = TensorOptions()
            .dtype(_r.scalartypeOptional(3))
            .device(_r.deviceWithDefault(5, torch::tensors::get_default_device()))
            .layout(_r.layoutOptional(4))
            .requires_grad(_r.toBool(7))
            .pinned_memory(_r.toBool(6));
        torch::utils::maybe_initialize_cuda(options);
        
        auto dispatch_rand = [](c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::TensorOptions options) -> at::Tensor {
          pybind11::gil_scoped_release no_gil;
          return torch::rand_symint(size, generator, options);
        };
        return wrap(dispatch_rand(_r.symintlist(0), _r.generator(1), options));
      } else {
        // aten::rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)
        check_out_type_matches(_r.tensor(2), _r.scalartypeOptional(3),
                               _r.isNone(3), _r.layoutOptional(4),
                               _r.deviceWithDefault(5, torch::tensors::get_default_device()), _r.isNone(5));
        
        auto dispatch_rand_out = [](at::Tensor out, c10::SymIntArrayRef size, c10::optional<at::Generator> generator) -> at::Tensor {
          pybind11::gil_scoped_release no_gil;
          return at::rand_symint_out(out, size, generator);
        };
        return wrap(dispatch_rand_out(_r.tensor(2), _r.symintlist(0), _r.generator(1)).set_requires_grad(_r.toBool(7)));
      }
    }
    case 2: {
      if (_r.isNone(1)) {
        // aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
        const auto options = TensorOptions()
            .dtype(_r.scalartypeOptional(2))
            .device(_r.deviceWithDefault(4, torch::tensors::get_default_device()))
            .layout(_r.layoutOptional(3))
            .requires_grad(_r.toBool(6))
            .pinned_memory(_r.toBool(5));
        torch::utils::maybe_initialize_cuda(options);
        
        auto dispatch_rand = [](c10::SymIntArrayRef size, at::TensorOptions options) -> at::Tensor {
          pybind11::gil_scoped_release no_gil;
          return torch::rand_symint(size, options);
        };
        return wrap(dispatch_rand(_r.symintlist(0), options));
      } else {
        // aten::rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)
        check_out_type_matches(_r.tensor(1), _r.scalartypeOptional(2),
                               _r.isNone(2), _r.layoutOptional(3),
                               _r.deviceWithDefault(4, torch::tensors::get_default_device()), _r.isNone(4));
        
        auto dispatch_rand_out = [](at::Tensor out, c10::SymIntArrayRef size) -> at::Tensor {
          pybind11::gil_scoped_release no_gil;
          return at::rand_symint_out(out, size);
        };
        return wrap(dispatch_rand_out(_r.tensor(1), _r.symintlist(0)).set_requires_grad(_r.toBool(6)));
      }
    }
    case 3: {
      // aten::rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
      auto __names = _r.toDimnameListOptional(1);
      c10::optional<DimnameList> names = __names ? c10::make_optional(DimnameList(__names.value())) : c10::nullopt;
      const auto options = TensorOptions()
          .dtype(_r.scalartypeOptional(2))
          .device(_r.deviceWithDefault(4, torch::tensors::get_default_device()))
          .layout(_r.layoutOptional(3))
          .requires_grad(_r.toBool(6))
          .pinned_memory(_r.toBool(5));
      torch::utils::maybe_initialize_cuda(options);
      
      auto dispatch_rand = [](c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, at::TensorOptions options) -> at::Tensor {
        pybind11::gil_scoped_release no_gil;
        return torch::rand_symint(size, names, options);
      };
      return wrap(dispatch_rand(_r.symintlist(0), names, options));
    }
  }
  Py_RETURN_NONE;
  END_HANDLE_TH_ERRORS
}

在parser的定義中，rand函數依據簽名被分成了以下4大類，共6個函數：

第0個簽名需提供size, generator和names：

    "rand(SymIntArrayRef size, *, Generator? generator, DimnameList? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",

// aten::rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor

第1種簽名需提供size和generator：

    "rand(SymIntArrayRef size, *, Generator? generator, Tensor out=None, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",

搭配注釋，可以知道第1種簽名又可分別被細分為2種API，第1-1種不需提供out：

// aten::rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor

第1-2種是需提供out的變形：

// aten::rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)

第2種簽名只需提供size：

    "rand(SymIntArrayRef size, *, Tensor out=None, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",

搭配注釋，可以知道第2種簽名又可分別被細分為2種API，第2-1種不需提供out：

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor

第2-2種是需提供out的變形：

// aten::rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)

第3種簽名需提供size和names：

    "rand(SymIntArrayRef size, *, DimnameList? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False, bool? requires_grad=False)",

// aten::rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor

因為我們調用的是torch.rand(3,4)，只傳入size，所以此處會進入switch的case 2。

接著用_r.isNone(1)檢查第1個（0-based）參數，也就是out參數，是否為空。此處未提供out參數，所以進入if分支。對應到只提供size的aten::rand：

        // aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor

operator()

#21 operator() (__closure=<synthetic pointer>, options=..., size=...) at /root/Documents/pytorch/torch/csrc/autograd/generated/python_torch_functions_0.cpp:5256

torch/csrc/autograd/generated/python_torch_functions_0.cpp

這個節點其實就只是torch::autograd::THPVariable_rand中case 2的if分支裡定義的lambda函數：

        auto dispatch_rand = [](c10::SymIntArrayRef size, at::TensorOptions options) -> at::Tensor {
          pybind11::gil_scoped_release no_gil;
          return torch::rand_symint(size, options);
        };

其中pybind11::gil_scoped_release no_gil;釋放了GIL鎖，這使得接下來要調用的torch::rand_symint得以在其它的線程中執行（這裡有另外開thread？），相當於讓Python擁有了執行多線程的能力。詳見Python GIL及其釋放/獲取函數。

C++ API

torch::rand_symint

#20 torch::rand_symint (options=..., size=...) at /root/Documents/pytorch/torch/csrc/autograd/generated/variable_factories.h:418

torch/csrc/autograd/generated/variable_factories.h

inline at::Tensor rand_symint(c10::SymIntArrayRef size, at::TensorOptions options = {}) {
  at::AutoDispatchBelowADInplaceOrView guard;
  return autograd::make_variable(at::rand_symint(size, at::TensorOptions(options).requires_grad(c10::nullopt)), /*requires_grad=*/options.requires_grad());
}

torch::rand_symint同樣只接受size參數，其函數內容則是先用at::rand_symint得到一個at::Tensor，再用autograd::make_variable包裝後返回。

at::rand_symint回傳的是一個沒有自動微分功能的純張量，autograd::make_variable的作用則是設定該張量關於自動微分的metadata，設定完成之後，該張量便擁有了自動微分的功能。關於torch::rand_symint和at::rand_symint的差別，詳見torch::和at:: factory function的差別。

以下是同一個檔案中出現的rand系列函數，

inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::DimnameList> names, at::TensorOptions options = {}) {
  at::AutoDispatchBelowADInplaceOrView guard;
  return autograd::make_variable(at::rand(size, names, at::TensorOptions(options).requires_grad(c10::nullopt)), /*requires_grad=*/options.requires_grad());
}
//...
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::TensorOptions options = {}) {
  at::AutoDispatchBelowADInplaceOrView guard;
  return autograd::make_variable(at::rand(size, generator, names, at::TensorOptions(options).requires_grad(c10::nullopt)), /*requires_grad=*/options.requires_grad());
}
//...
inline at::Tensor rand(at::IntArrayRef size, at::TensorOptions options = {}) {
  at::AutoDispatchBelowADInplaceOrView guard;
  return autograd::make_variable(at::rand(size, at::TensorOptions(options).requires_grad(c10::nullopt)), /*requires_grad=*/options.requires_grad());
}
//...
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, at::TensorOptions options = {}) {
  at::AutoDispatchBelowADInplaceOrView guard;
  return autograd::make_variable(at::rand(size, generator, at::TensorOptions(options).requires_grad(c10::nullopt)), /*requires_grad=*/options.requires_grad());
}

分四種：

  • size, names
  • size, generator, names
  • size
  • size, generator

正好可與torch/_C/_VariableFucntions.pyi和torch::autograd::THPVariable_rand中的4個函數一一對應。

at::rand_symint

#19 0x00007f046abdd595 in at::rand_symint (options=..., size=...) at /root/Documents/pytorch/build/aten/src/ATen/Functions.h:5770

build/aten/src/ATen/Functions.h

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, at::TensorOptions options={}) {
    return at::_ops::rand::call(size, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}

at::rand_symint函數同樣只接受size參數。

注釋中表明了at::rand_symint與aten::rand的關係，待會會在native_functions.yaml看見aten::rand的身影。

下面還定義了at::symint::rand函數，其實現與at::rand_symint一致。

namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, at::TensorOptions options={}) {
    return at::_ops::rand::call(size, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

以下是同一個檔案中出現的rand系列函數，可分為aten::rand.names, aten::rand.generator_with_names, aten::rand, aten::rand.generator, aten::rand.out, aten::rand.generator_out六類，可與torch::autograd::THPVariable_rand注釋中的六個函數一一對應。

每類又依第一個參數型別為at::intArrayRef或c10::SymIntArrayRef各分兩種，再依第二個以後的參數為dtype, layout, device, pin_memory或直接用at::TensorOptions包起來各分兩種，所以共有6 * 2 * 2 ＝ 24個。

// aten::rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_names::call(c10::fromIntArrayRefSlow(size), names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_names::call(c10::fromIntArrayRefSlow(size), names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_names::call(c10::fromIntArrayRefSlow(size), names, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_names::call(c10::fromIntArrayRefSlow(size), names, dtype, layout, device, pin_memory);
  }
}

// aten::rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_names::call(size, names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_names::call(size, names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_names::call(size, names, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_names::call(size, names, dtype, layout, device, pin_memory);
  }
}

// aten::rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_generator_with_names::call(c10::fromIntArrayRefSlow(size), generator, names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_generator_with_names::call(c10::fromIntArrayRefSlow(size), generator, names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator_with_names::call(c10::fromIntArrayRefSlow(size), generator, names, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator_with_names::call(c10::fromIntArrayRefSlow(size), generator, names, dtype, layout, device, pin_memory);
  }
}

// aten::rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_generator_with_names::call(size, generator, names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::TensorOptions options={}) {
    return at::_ops::rand_generator_with_names::call(size, generator, names, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator_with_names::call(size, generator, names, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator_with_names::call(size, generator, names, dtype, layout, device, pin_memory);
  }
}

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, at::TensorOptions options={}) {
    return at::_ops::rand::call(c10::fromIntArrayRefSlow(size), optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, at::TensorOptions options={}) {
    return at::_ops::rand::call(c10::fromIntArrayRefSlow(size), optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand::call(c10::fromIntArrayRefSlow(size), dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand::call(c10::fromIntArrayRefSlow(size), dtype, layout, device, pin_memory);
  }
}

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, at::TensorOptions options={}) {
    return at::_ops::rand::call(size, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, at::TensorOptions options={}) {
    return at::_ops::rand::call(size, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand::call(size, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand::call(size, dtype, layout, device, pin_memory);
  }
}

// aten::rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, at::TensorOptions options={}) {
    return at::_ops::rand_generator::call(c10::fromIntArrayRefSlow(size), generator, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, at::TensorOptions options={}) {
    return at::_ops::rand_generator::call(c10::fromIntArrayRefSlow(size), generator, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator::call(c10::fromIntArrayRefSlow(size), generator, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor rand(at::IntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator::call(c10::fromIntArrayRefSlow(size), generator, dtype, layout, device, pin_memory);
  }
}

// aten::rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::TensorOptions options={}) {
    return at::_ops::rand_generator::call(size, generator, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::TensorOptions options={}) {
    return at::_ops::rand_generator::call(size, generator, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
  }
}

// aten::rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
inline at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator::call(size, generator, dtype, layout, device, pin_memory);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    return at::_ops::rand_generator::call(size, generator, dtype, layout, device, pin_memory);
  }
}

// aten::rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_out(at::Tensor & out, at::IntArrayRef size) {
    return at::_ops::rand_out::call(c10::fromIntArrayRefSlow(size), out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor & rand_out(at::Tensor & out, at::IntArrayRef size) {
    return at::_ops::rand_out::call(c10::fromIntArrayRefSlow(size), out);
  }
}

// aten::rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_outf(at::IntArrayRef size, at::Tensor & out) {
    return at::_ops::rand_out::call(c10::fromIntArrayRefSlow(size), out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor & rand_outf(at::IntArrayRef size, at::Tensor & out) {
    return at::_ops::rand_out::call(c10::fromIntArrayRefSlow(size), out);
  }
}

// aten::rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_symint_out(at::Tensor & out, c10::SymIntArrayRef size) {
    return at::_ops::rand_out::call(size, out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor & rand_out(at::Tensor & out, c10::SymIntArrayRef size) {
    return at::_ops::rand_out::call(size, out);
  }
}

// aten::rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_symint_outf(c10::SymIntArrayRef size, at::Tensor & out) {
    return at::_ops::rand_out::call(size, out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor & rand_outf(c10::SymIntArrayRef size, at::Tensor & out) {
    return at::_ops::rand_out::call(size, out);
  }
}

// aten::rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_out(at::Tensor & out, at::IntArrayRef size, c10::optional<at::Generator> generator) {
    return at::_ops::rand_generator_out::call(c10::fromIntArrayRefSlow(size), generator, out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor & rand_out(at::Tensor & out, at::IntArrayRef size, c10::optional<at::Generator> generator) {
    return at::_ops::rand_generator_out::call(c10::fromIntArrayRefSlow(size), generator, out);
  }
}

// aten::rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_outf(at::IntArrayRef size, c10::optional<at::Generator> generator, at::Tensor & out) {
    return at::_ops::rand_generator_out::call(c10::fromIntArrayRefSlow(size), generator, out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, int64_t>::value>>
  at::Tensor & rand_outf(at::IntArrayRef size, c10::optional<at::Generator> generator, at::Tensor & out) {
    return at::_ops::rand_generator_out::call(c10::fromIntArrayRefSlow(size), generator, out);
  }
}

// aten::rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_symint_out(at::Tensor & out, c10::SymIntArrayRef size, c10::optional<at::Generator> generator) {
    return at::_ops::rand_generator_out::call(size, generator, out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor & rand_out(at::Tensor & out, c10::SymIntArrayRef size, c10::optional<at::Generator> generator) {
    return at::_ops::rand_generator_out::call(size, generator, out);
  }
}

// aten::rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)
inline at::Tensor & rand_symint_outf(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::Tensor & out) {
    return at::_ops::rand_generator_out::call(size, generator, out);
}
namespace symint {
  template <typename T, typename = std::enable_if_t<std::is_same<T, c10::SymInt>::value>>
  at::Tensor & rand_outf(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::Tensor & out) {
    return at::_ops::rand_generator_out::call(size, generator, out);
  }
}

這些運算子是由native_functions.yaml生成，而由native_functions.yaml生成的函數的命名空間默認為aten，所以注釋裡所有函數都加上了aten::的前綴。

接下來會調用at::_ops::rand::call做分發（dispatch），但在進入該函數前，先來看一下at::_ops::rand的定義。

dispatch

at::_ops::rand

torch/include/ATen/Operators.h

build/aten/src/ATen/Operators.h

rand是一個結構體，有call和redispatch兩個static的成員函數，待會都會用到。

struct TORCH_API rand {
  using schema = at::Tensor (c10::SymIntArrayRef, c10::optional<at::ScalarType>, c10::optional<at::Layout>, c10::optional<at::Device>, c10::optional<bool>);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor")
  static at::Tensor call(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
  static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
};

STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA為at::_ops::rand這個結構體宣告了三個名字分別為name, overload_name和schema_str的static字串成員變數，並在此處或待會在build/aten/src/ATen/Operators_2.cpp處賦予它們初始值。

對照schema跟call成員函數，可以發現schema正是call函數的type signature。

同檔案下共有六個rand系列函數（rand_names, rand_generator_with_names, rand, rand_generator, rand_out, rand_generator_out），可與torch::autograd::THPVariable_rand注釋中的六個函數一一對應。

struct TORCH_API rand_names {
  using schema = at::Tensor (c10::SymIntArrayRef, c10::optional<at::DimnameList>, c10::optional<at::ScalarType>, c10::optional<at::Layout>, c10::optional<at::Device>, c10::optional<bool>);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "names")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor")
  static at::Tensor call(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
  static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
};

struct TORCH_API rand_generator_with_names {
  using schema = at::Tensor (c10::SymIntArrayRef, c10::optional<at::Generator>, c10::optional<at::DimnameList>, c10::optional<at::ScalarType>, c10::optional<at::Layout>, c10::optional<at::Device>, c10::optional<bool>);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "generator_with_names")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor")
  static at::Tensor call(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
  static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
};

struct TORCH_API rand {
  using schema = at::Tensor (c10::SymIntArrayRef, c10::optional<at::ScalarType>, c10::optional<at::Layout>, c10::optional<at::Device>, c10::optional<bool>);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor")
  static at::Tensor call(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
  static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
};

struct TORCH_API rand_generator {
  using schema = at::Tensor (c10::SymIntArrayRef, c10::optional<at::Generator>, c10::optional<at::ScalarType>, c10::optional<at::Layout>, c10::optional<at::Device>, c10::optional<bool>);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "generator")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor")
  static at::Tensor call(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
  static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory);
};

struct TORCH_API rand_out {
  using schema = at::Tensor & (c10::SymIntArrayRef, at::Tensor &);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "out")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)")
  static at::Tensor & call(c10::SymIntArrayRef size, at::Tensor & out);
  static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, at::Tensor & out);
};

struct TORCH_API rand_generator_out {
  using schema = at::Tensor & (c10::SymIntArrayRef, c10::optional<at::Generator>, at::Tensor &);
  using ptr_schema = schema*;
  // See Note [static constexpr char* members for windows NVCC]
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::rand")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "generator_out")
  STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)")
  static at::Tensor & call(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::Tensor & out);
  static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::Generator> generator, at::Tensor & out);
};

at::_ops::rand::call

#18 at::_ops::rand::call (size=..., dtype=..., layout=..., layout@entry=..., device=..., pin_memory=...) at /root/Documents/pytorch/build/aten/src/ATen/Operators_2.cpp:5213

build/aten/src/ATen/Operators_2.cpp

STATIC_CONST_STR_OUT_OF_LINE_FOR_WIN_CUDA(rand, name, "aten::rand")
STATIC_CONST_STR_OUT_OF_LINE_FOR_WIN_CUDA(rand, overload_name, "")
STATIC_CONST_STR_OUT_OF_LINE_FOR_WIN_CUDA(rand, schema_str, "rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor")

at::_ops::rand中使用STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA宣告了name, overload_name, schema_str等三個成員變數。這裡則用STATIC_CONST_STR_OUT_OF_LINE_FOR_WIN_CUDA將rand的name成員變數的值設為"aten::rand"；把overload_name成員設為""；把schema_str成員設為"rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"。如果build/aten/src/ATen/Operators.h處已經賦予初始值，那麼此處的代碼就是無效的。

at::_ops::rand::call如下：

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
static C10_NOINLINE c10::TypedOperatorHandle<rand::schema> create_rand_typed_handle() {
  return c10::Dispatcher::singleton()
      .findSchemaOrThrow(rand::name, rand::overload_name)
      .typed<rand::schema>();
}

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
at::Tensor rand::call(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    
    static auto op = create_rand_typed_handle();
    return op.call(size, dtype, layout, device, pin_memory);
}

此處調用create_rand_typed_handle，會查表做dispatch，而分發時查的表格則是由native_functions.yaml生成。

另外從注釋中可以看出at::_ops::rand::call和待會會在native_functions.yaml看到的aten::rand是一對一的關係。

可以看到c10::TypedOperatorHandle的模板參數rand::schema正是at::_ops::rand::call的type signature。

而c10::Dispatcher::singleton().findSchemaOrThrow(rand::name, rand::overload_name).typed<rand::schema>();這一段代碼當中又用rand::name和rand::overload_name去尋找對應的c10::TypedOperatorHandle。

native_functions.yaml

aten/src/ATen/native/native_functions.yaml

- func: rand.names(SymInt[] size, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
  device_check: NoCheck
  device_guard: False
  dispatch:
    CompositeExplicitAutograd: rand
  autogen: rand.names_out
  tags: nondeterministic_seeded

- func: rand.generator_with_names(SymInt[] size, *, Generator? generator, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
  device_check: NoCheck
  device_guard: False
  tags: nondeterministic_seeded
  dispatch:
    CompositeExplicitAutograd: rand
  autogen: rand.generator_with_names_out

- func: rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
  tags: [core, nondeterministic_seeded]
  dispatch:
    CompositeExplicitAutograd: rand

- func: rand.generator(SymInt[] size, *, Generator? generator, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
  tags: nondeterministic_seeded
  dispatch:
    CompositeExplicitAutograd: rand

- func: rand.out(SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)
  tags: nondeterministic_seeded
  dispatch:
    CompositeExplicitAutograd: rand_out

- func: rand.generator_out(SymInt[] size, *, Generator? generator, Tensor(a!) out) -> Tensor(a!)
  tags: nondeterministic_seeded

native_functions.yaml有六個版本的rand，分別是：

  • size, names
  • size, generator, names
  • size
  • size, generator
  • size, out
  • size, generator, out

torch/include/ATen/Operators.h（build/aten/src/ATen/Operators.h）中對應的四個rand系列函數正是由native_functions.yaml生成的。

回顧torch/_C/_VariableFunctions.pyi中的四個分類：“size, generator, names”，“size, generator（out可省略）”，“size（out可省略）”，“size, names”。這四個分類可以拆開來變成6個：“size, generator, names”，“size, generator, out”，“size, generator”，“size, out”，“size”，“size, names”，正好可以跟此處的六個條目一一對應。

回顧THPVariable_rand注釋中的六個函數：aten::rand.generator_with_names, aten::rand.generator, aten::rand.generator_out, aten::rand, aten::rand.out, aten::rand.names。正可與此處的六個函數一一對應。另外因為由native_functions.yaml中func:後的函數的命名空間默認為aten，所以注釋裡所有函數都加上了aten::的前綴。

at::_ops::rand::call（對應到aten::rand）調用了c10::TypedOperatorHandle<rand::schema>::call函數做分發，從Python bindings處已知torch.rand(3, 4)是對應到此處第三個沒有generator,names和out的rand函數，來到native_functions.yaml查表：

- func: rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
  tags: [core, nondeterministic_seeded]
  dispatch:
    CompositeExplicitAutograd: rand

後來到at::(anonymous namespace)::rand。

at::(anonymous namespace)::rand

#11 0x00007f0462c05329 in at::(anonymous namespace)::rand (pin_memory=..., device=..., layout=..., dtype=..., size=...) at /root/Documents/pytorch/build/aten/src/ATen/RegisterBackendSelect.cpp:365

build/aten/src/ATen/RegisterBackendSelect.cpp

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
C10_ALWAYS_INLINE
at::Tensor rand(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
  DispatchKeySet _dk = c10::DispatchKeySet(c10::computeDispatchKey(dtype, layout, device));
  return at::_ops::rand::redispatch(
      _dk, size, dtype, layout, device, pin_memory);
}

進行再次分發（redispatch）。

在同一個檔案下有：

namespace at {

namespace {
//...
TORCH_LIBRARY_IMPL(aten, BackendSelect, m) {
  //...
  m.impl("aten::rand.names", TORCH_FN(rand_names));
  m.impl("aten::rand.generator_with_names", TORCH_FN(rand_generator_with_names));
  m.impl("aten::rand", TORCH_FN(rand));
  m.impl("aten::rand.generator", TORCH_FN(rand_generator));
  //...
}

} // namespace
} // at

在m.impl("aten::rand", TORCH_FN(rand));這個欄位中，rand全名為at::(anonymous namespace)::rand函數，它與aten::rand在這裡被關聯起來了。

而剛剛看到的at::_ops::rand::call的注釋中也有aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor，因此at::ops_rand::call與aten::rand就這樣間接地被關聯起來了。

redisptach

at::_ops::rand::redispatch

#10 at::_ops::rand::redispatch (dispatchKeySet=..., size=..., dtype=..., layout=..., layout@entry=..., device=..., pin_memory=...) at /root/Documents/pytorch/build/aten/src/ATen/Operators_2.cpp:5220

build/aten/src/ATen/Operators_2.cpp

// aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
at::Tensor rand::redispatch(c10::DispatchKeySet dispatchKeySet, c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    
    static auto op = create_rand_typed_handle();
    return op.redispatch(dispatchKeySet, size, dtype, layout, device, pin_memory);
}

native_functions.yaml

回到native_functions.yaml查表，關注aten::rand：

- func: rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
  tags: [core, nondeterministic_seeded]
  dispatch:
    CompositeExplicitAutograd: rand

查看dispatch欄位，下有一個名為CompositeExplicitAutograd的key，表示redispatch後會去到CompositeExplicitAutograd backend，即待會會看到的wrapper_CompositeExplicitAutograd__rand函數；value則是rand，因為dispatch:後預設的命名空間是at::native，所以redispatch的終點是at::native::rand函數。

CPU kernel

at::(anonymous namespace)::(anonymous namespace)::wrapper_CompositeExplicitAutograd__rand

#3  0x00007f0462dd2c28 in at::(anonymous namespace)::(anonymous namespace)::wrapper_CompositeExplicitAutograd__rand (pin_memory=..., device=..., layout=..., 
    dtype=..., size=...) at /root/Documents/pytorch/build/aten/src/ATen/RegisterCompositeExplicitAutograd.cpp:2214

build/aten/src/ATen/RegisterCompositeExplicitAutograd.cpp

namespace {
at::Tensor wrapper_CompositeExplicitAutograd__rand(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand(C10_AS_INTARRAYREF_SLOW(size), dtype, layout, device, pin_memory);
}
} // anonymous namespace

TORCH_LIBRARY_IMPL(aten, CPU, m) {
// ...
m.impl("rand",
TORCH_FN(wrapper_CompositeExplicitAutograd__rand));
m.impl("rand.out",
TORCH_FN(wrapper_CompositeExplicitAutograd_out_rand_out));
m.impl("rand.generator",
TORCH_FN(wrapper_CompositeExplicitAutograd_generator_rand));
// ...
};

此處將aten::rand與wrapper_CompositeExplicitAutograd__rand做關聯。

而從at::_ops::rand::redispatch的注釋中可以看到它與aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor也是相關聯的，所以at::_ops::rand::redispatch的目的地才會是這裡。

wrapper_CompositeExplicitAutograd__rand函數在同一個檔案下被調用：

at::Tensor rand(at::IntArrayRef size, at::TensorOptions options) {
return wrapper_CompositeExplicitAutograd__rand(c10::fromIntArrayRefSlow(size), optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
at::Tensor rand(at::IntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
return wrapper_CompositeExplicitAutograd__rand(c10::fromIntArrayRefSlow(size), dtype, layout, device, pin_memory);
}
at::Tensor rand_symint(c10::SymIntArrayRef size, at::TensorOptions options) {
return wrapper_CompositeExplicitAutograd__rand(size, optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt());
}
at::Tensor rand_symint(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
return wrapper_CompositeExplicitAutograd__rand(size, dtype, layout, device, pin_memory);
}

注：在PyTorch 1.14版中是build/aten/src/ATen/RegisterCompositeExplicitAutograd.cpp中的at::(anonymous namespace)::(anonymous namespace)::wrapper__rand函數。

同檔案下的rand系列函數：

namespace {
at::Tensor wrapper_CompositeExplicitAutograd_names_rand(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand(C10_AS_INTARRAYREF_SLOW(size), names, dtype, layout, device, pin_memory);
}
} // anonymous namespace
namespace {
at::Tensor & wrapper_CompositeExplicitAutograd_names_out_rand_out(c10::SymIntArrayRef size, c10::optional<at::DimnameList> names, at::Tensor & out) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand_names_out_symint(size, names, out);
}
} // anonymous namespace
namespace {
at::Tensor wrapper_CompositeExplicitAutograd_generator_with_names_rand(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand(C10_AS_INTARRAYREF_SLOW(size), generator, names, dtype, layout, device, pin_memory);
}
} // anonymous namespace
namespace {
at::Tensor & wrapper_CompositeExplicitAutograd_generator_with_names_out_rand_out(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::DimnameList> names, at::Tensor & out) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand_generator_with_names_out_symint(size, generator, names, out);
}
} // anonymous namespace
namespace {
at::Tensor wrapper_CompositeExplicitAutograd__rand(c10::SymIntArrayRef size, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand(C10_AS_INTARRAYREF_SLOW(size), dtype, layout, device, pin_memory);
}
} // anonymous namespace
namespace {
at::Tensor & wrapper_CompositeExplicitAutograd_out_rand_out(c10::SymIntArrayRef size, at::Tensor & out) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand_out(C10_AS_INTARRAYREF_SLOW(size), out);
}
} // anonymous namespace
namespace {
at::Tensor wrapper_CompositeExplicitAutograd_generator_rand(c10::SymIntArrayRef size, c10::optional<at::Generator> generator, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
    // No device check
  // DeviceGuard omitted
  return at::native::rand(C10_AS_INTARRAYREF_SLOW(size), generator, dtype, layout, device, pin_memory);
}
} // anonymous namespace

有names_rand，names_out_rand_out，generator_with_names_rand，generator_with_names_out_rand_out，_rand，out_rand_out和genrator_rand共七個（缺了有generator和out的版本？）。

at::native::rand

#2  0x00007f04620d61fc in at::native::rand (size=..., dtype=..., layout=..., layout@entry=..., device=..., pin_memory=...)
    at /root/Documents/pytorch/aten/src/ATen/native/TensorFactories.cpp:781

aten/src/ATen/native/TensorFactories.cpp

// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ rand ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Tensor rand(IntArrayRef size,
    c10::optional<ScalarType> dtype,
    c10::optional<Layout> layout,
    c10::optional<Device> device,
    c10::optional<bool> pin_memory) {
  return native::rand(size, static_cast<c10::optional<Generator>>(c10::nullopt), dtype, layout, device, pin_memory);
}

創建一個nullopt的generator後調用同一個檔案下簽名較完整的at::native::rand函數。

at::native::rand

#1  0x00007f04620d60b1 in at::native::rand (size=..., generator=..., dtype=..., layout=..., device=..., pin_memory=...)
    at /root/Documents/pytorch/c10/util/Optional.h:204

在backtrace中裡直接從at::native::rand跳來了c10/util/Optional.h，有點不明所以，以下是筆者猜測的完整調用路徑。

首先在at::native::rand中有一個static_cast<c10::optional<Generator>>(c10::nullopt)的參數，當中用到了c10::optional：

c10::optional

c10/util/Optional.h

template <class T>
class optional : private OptionalBase<T> {
    // ...
};

c10::OptionalBase

c10/util/Optional.h

template <class T>
using OptionalBase = std::conditional_t<
    detail_::is_arrayref<T>::value,
    arrayref_optional_base<T>,
    std::conditional_t<
        std::is_trivially_destructible<T>::value &&
            C10_IS_TRIVIALLY_COPYABLE(T) &&
            // Avoid using is_trivially_copy_{constructible,assignable}
            // because old GCC versions don't support them. Also,
            // is_trivially_copyable seems not to do what I expect, so check
            // trivially_copyable_optimization_optional_base directly.
            std::is_copy_constructible<
                trivially_copyable_optimization_optional_base<T>>::value &&
            std::is_copy_assignable<
                trivially_copyable_optimization_optional_base<T>>::value,
        trivially_copyable_optimization_optional_base<T>,
        std::conditional_t<
            std::is_trivially_destructible<T>::value, // if possible
            constexpr_optional_base<std::remove_const_t<T>>, // use base with
                                                             // trivial
                                                             // destructor
            optional_base<std::remove_const_t<T>>>>>;

當中用到了constexpr_optional_base。

c10::constexpr_optional_base

c10/util/Optional.h

template <class T>
struct constexpr_optional_base {
  bool init_;
  constexpr_storage_t<T> storage_;
  // ...
};

constexpr_storage_t

c10/util/Optional.h

template <class T>
union constexpr_storage_t {
  unsigned char dummy_;
  T value_;

#if __cplusplus >= 202002L
  // C++20 lifted the requirement to initialize a union member in order to be
  // constexpr.
  constexpr constexpr_storage_t(trivial_init_t) noexcept {
    new (&dummy_) unsigned char;
  }
#else
  constexpr constexpr_storage_t(trivial_init_t) noexcept : dummy_() {}
#endif

  template <class... Args>
  constexpr constexpr_storage_t(Args&&... args)
      : value_(constexpr_forward<Args>(args)...) {}

  ~constexpr_storage_t() = default;
};

從backtrace看此處是進入了else分支，然後直接進入at::empty函數，仍然不懂箇中奧妙。

at::native::rand

下面則是筆者根據調用處的函數簽名找到的對應的函數定義：

aten/src/ATen/native/TensorFactories.cpp

Tensor rand(IntArrayRef size, c10::optional<Generator> generator,
    c10::optional<ScalarType> dtype,
    c10::optional<Layout> layout,
    c10::optional<Device> device,
    c10::optional<bool> pin_memory) {
  // See [Note: hacky wrapper removal for TensorOptions]
  TensorOptions options = TensorOptions().dtype(dtype).layout(layout).device(device).pinned_memory(pin_memory);

  auto result = at::empty(size, options);
  return result.uniform_(0, 1, std::move(generator));
}

首先調用at::empty得到一個at::Tensor後，再使用at::Tensor::uniform_來讓張量裡的元素化為均勻分布。自此將進入at::empty或uniform_函數，結束rand函數的旅程。

keineahnung2345
关注
  • 8
    点赞
  • 踩
  • 9
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
pytorch 实现张量tensor,图片,CPU,GPU,数组等的转换
12-23
1, 创建pytorch 的Tensor张量： torch.rand((3,224,224)) #创建随机值的三维张量，大小为（3,224,224） torch.Tensor([3,2]) #创建张量，[3,2] 2, cpu上的tensor和GPU即pytorch创建的tensor的相互转化 b = a.cpu() # GPU → CPU a = b.cuda() #CPU → GPU 3, tensor和numpy的转化 b = a.numpy() # tensor转化为 numpy数组 a = b.from_numpy() # numpy数组转化为tensor 4, torch的
pytorch 生成随机数Tensor的方法 torch.rand torch.randn torch.normal torch.linespace
Ibelievesunshine的博客
08-15 3032
在使用PyTorch做实验时经常会用到生成随机数Tensor的方法，比如： torch.rand() torch.randn() torch.normal() torch.linespace() 在很长一段时间里我都没有区分这些方法生成的随机数究竟有什么不同，由此在做实验的时候经常会引起一些莫名其妙的麻烦。 所以在此做一个总结，以供大家阅读区分，不要重蹈我的覆辙。 均匀分布 ...
【pytorch】torch.Tensor详解和常用操作
热门推荐
sazass的博客
10-29 6万+
学习教材： 动手学深度学习 PYTORCH 版(DEMO) (https://github.com/ShusenTang/Dive-into-DL-PyTorch) PDF 制作by [Marcus Yang](https://github.com/chenyang1999) 1.tensor简介 在PyTorch中，torch.Tensor是存储和变换数据的主要工具。 Tensor与Numpy的多维数组非常相似。 Tensor还提供了GPU计算和自动求梯度等更多功能，这些使Tensor更适合深度学习。 2
【快速理解张量】通过torch.rand和举例通俗解释张量tensor
qq_45193872的博客
04-07 1万+
如何理解张量（tensor）成了深度学习路途中不可缺少的一步，讲真的，刚学习深度学习那会儿张量实在是困惑了自己很长一段时间，而这篇文章根据自己学习深度学习的历程给出了一个清晰通俗的解释，相比于官方教材给出了更容易让初学者更能理解的逻辑举例。 如果你的张量理解程度还停留在只能想象出三维的张量维度的话，相信这篇文章一定能让你彻底理解各种维度的张量！
PyTorch随机数生成：torch.rand，torch.randn，torch.randind，torch.rand_like
牧羊女
12-22 2356
pytorch中几个随机函数的定义和用法整理。
pytorch cuda上tensor的定义 以及减少cpu的操作详解
01-19
cuda上tensor的定义 a = torch.ones(1000,1000,3).cuda() 某一gpu上定义 cuda1 = torch.device(‘cuda:1’) b = torch.randn((1000,1000,1000),device=cuda1) 删除某一变量 del a 在cpu定义tensor然后转到gpu ...
Pytorch Tensor基本数学运算详解
12-23
# 这两个Tensor加减乘除会对b自动进行Broadcasting a = torch.rand(3, 4) b = torch.rand(4) c1 = a + b c2 = torch.add(a, b) print(c1.shape, c2.shape) print(torch.all(torch.eq(c1, c2))) 输出结果： torch....
Pytorch生成随机数Tensor的方法汇总
12-17
返回一个张量，包含了从区间[0, 1)的均匀分布中抽取的一组随机数。张量的形状由参数sizes定义。 参数: sizes (int…) – 整数序列，定义了输出张量的形状 out (Tensor, optinal) – 结果张量 torch.rand(2, 3) [[0...
pytorch 实现删除tensor中的指定行列
12-23
在pytorch中， 想删除tensor中的指定行列，原本以为有个函数或者直接把某一行赋值为[]就可以，结果发现没这么简单，因此用了一个曲线救国方法，希望如果有更直接的方法，请大家指出。 code 本质上是利用mask删除了...
Ansible 指定受控端使用Python的版本
shijin741231的博客
04-20 588
最近在装Ansible，有一台受控端Ubuntu16的服务器，安装了Python2.7.12和Pyhon3.5。当用Ansible连接它时，显示使用的是Python3.5。最后看文档，发现Ansible可以在hosts的文件中指定受控服上运行的Python。
【编译程序介绍】
武帝为此的博客
04-22 1198
简单来说，编译程序是一个软件，它读取用某种编程语言编写的源代码，分析并转换成等效的、可执行的机器语言代码。这个过程涉及多个复杂的步骤，包括语法分析、语义分析、代码优化和代码生成等。
[python3] 读取一个正在更新的日志文件
最新发布
言之。
04-24 226
等同于–follow=name --retry，根据文件名进行追踪，并保持重试，即该文件被删除或改名后，如果再次创建相同的文件名，会继续追踪。这两种方法都会持续监听日志文件的变化，并实时读取新增的日志内容。你可以根据实际需求选择其中一种方法。要读取一个正在更新的日志文件（即实时写入的日志文件），你可以使用 Python 的。的技巧来实现实时读取。函数打开文件，并使用。
实现自定义注解、实现自定义幂等性注解
qq_44721738的博客
04-23 582
添加 Spring AOP 依赖。创建自定义注解。创建一个新的 Java 注解类，通过@interface关键字来定义，并可以添加元注解以及属性。@Target(ElementType.METHOD) //表示作用于方法上@Retention(RetentionPolicy.RUNTIME) // 表示这个注解在运行时是可见的，这样 AOP 代理才能在运行时读取到这个注解编写 AOP 拦截（自定义注解）的逻辑代码。@Aspect@Component// 方法执行前的处理。
【InternLM实战营---第五节课作业】
weixin_45609124的博客
04-22 797
LMDeploy环境配置及基础使用
Pandas 2.2 中文官方教程和指南（十五）
龙哥盟
04-24 702
原文：pandas.pydata.org/docs/ 处理文本数据 原文：pandas.pydata.org/docs/user_guide/text.html 文本数据类型 在 pandas 中有两种存储文本数据的方式： object -dtype NumPy 数组。 StringDtype 扩展类型。 我们建议使用StringDtype来存储文本数据。 在 pandas 1.0 之前，object dtype 是唯一的选项。这在很多方面都是不幸的： 你可能会在object dty
[SWPUCTF 2021 新生赛]re2（不同字符加密相同，逆向修改范围）
2301_79326813的博客
04-23 154
无壳直接看ida。
python数据分析pyecharts【饼状图、直方图、词云、地图】
QAZWCC的博客
04-20 438
【代码】python数据分析pyecharts【饼状图、直方图、词云、地图】
Semibin2 GTDB 路径手动配置
Dylan's Blog
04-23 177
在 main.py 里的这部分代码的default 替换成你自己的 gtdb 的路径即可，当然，你也可以用参数输入对应的路径。将解压后的文件放到一个文件夹里，default 里填写这个文件夹的绝对路径。
y3 = torch.rand_like(tensor) torch.matmul(tensor, tensor.T, out=y3) 什么意思
08-10
代码中的y3 = torch.rand_like(tensor)是使用torch.rand_like()函数生成一个与给定张量tensor具有相同形状的随机张量y3。该函数会生成一个在123 #### 引用[.reference_title] - *1* *2* *3* [torch.as_tensor()、torch.Tensor() 、 torch.tensor() 、transforms.ToTensor()的区别](https://blog.csdn.net/qimo601/article/details/128014195)[target="_blank" data-report-click={"spm":"1018.2226.3001.9630","extra":{"utm_source":"vip_chatgpt_common_search_pc_result","utm_medium":"distribute.pc_search_result.none-task-cask-2~all~insert_cask~default-1-null.142^v92^chatsearchT3_1"}} ] [.reference_item] [ .reference_list ]

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值