 |
Caffe2 - C++ API
A deep learning, cross platform ML framework
|
|
Caffe2 - C++ API
A deep learning, cross platform ML framework
|
Data Structures | |
| struct | Builder |
Public Types | |
| using | DimMask = std::bitset< 64 > |
| using | PtrVector = SmallVector< char *, 4 > |
| using | loop_t = std::function< void(int ntensors, char **data, const int64_t *strides, int64_t size)> |
| using | loop2d_t = std::function< void(int ntensors, char **data, const int64_t *strides, int64_t size0, int64_t size1)> |
| using | loop_subiter_t = std::function< void(TensorIterator &subiter)> |
Public Member Functions | |
| void | foreach_reduced_elt (const loop_subiter_t &loop, bool parallelize=true) |
| int | ndim () const |
| IntArrayRef | shape () const |
| int64_t | numel () const |
| int | ntensors () const |
| int64_t | num_output_elements () const |
| number of elements in the output operand. More... | |
| int | num_reduce_dims () const |
| number of reduced dimensions in a reduction operation | |
| bool | is_trivial_1d () const |
| 1-dimensional iteration and no buffering or type conversion | |
| bool | is_dim_reduced (int dim) const |
| IntArrayRef | strides (int arg) const |
| Accessors for each operand. | |
| void * | data_ptr (int arg) const |
| const Type & | type (int arg=0) const |
| ScalarType | dtype (int arg=0) const |
| DeviceType | device_type (int arg=0) const |
| int64_t | element_size (int arg) const |
| bool | is_scalar (int arg) const |
| bool | is_cpu_scalar (int arg) const |
| const Tensor & | tensor (int arg) const |
| Tensor & | tensor (int arg) |
| Tensor | output (int arg=0) const |
| void | remove_operand (int arg) |
| Removes an operand from this iterator. | |
| void | remove_dimension (int dim) |
| Removes a dimension from this iterator. | |
| void | narrow (int dim, int64_t start, int64_t size) |
| Shrinks an iterated dimension. | |
| void | select_all_keeping_dim (int start_dim, IntArrayRef starts) |
Narrows every dim after and including start_dim to size one. | |
| void | replace_operand (int arg, void *data, IntArrayRef stride) |
Replaces the data pointer and strides for the operand at index arg | |
| std::unique_ptr< TensorIterator > | split (int dim) |
| Splits this TensorIterator into two iterators. More... | |
| int | get_dim_to_split () const |
| Returns the dimension with the largest extent: (size[dim]-1) * stride[dim]. | |
| template<typename T > | |
| T | scalar_value (int arg) |
| void | for_each (const loop_t &loop) |
| void | for_each (const loop2d_t &loop) |
| void | parallel_reduce (const loop2d_t &loop) |
| void | serial_for_each (const loop_t &loop, Range range) const |
| void | serial_for_each (const loop2d_t &loop, Range range) const |
| DimVector | compatible_stride (int element_size) const |
| Create a strides array for a Tensor with shape of this iterator. More... | |
| DimVector | invert_perm (IntArrayRef input) const |
| Inverts the re-ordering done by reorder_dimensions. More... | |
| DimVector | get_dim_strides (int dim) const |
| Helper functions for CPU iteration. | |
| DimVector | get_strides () const |
| DimVector | get_inner_strides () const |
| PtrVector | get_data_ptrs (ArrayRef< char * > base, IntArrayRef counter) const |
| PtrVector | get_base_ptrs () const |
| bool | can_use_32bit_indexing () const |
| true if the stride computation can use 32-bit arithmetic. Used by GPU kernels | |
| SplitUntil32Bit | with_32bit_indexing () const |
| An "iteratable" object that recursively splits this iterator into sub-iterators that can use 32-bit indexing. More... | |
| bool | should_accumulate () const |
| If the kernel should accumulate into the output. More... | |
| bool | is_final_output () const |
| Whether this iterator produces the actual output, as opposed to something that will be accumulated further. More... | |
Static Public Member Functions | |
| static std::unique_ptr< TensorIterator > | binary_op (Tensor &out, const Tensor &a, const Tensor &b) |
| static std::unique_ptr< TensorIterator > | reduce_op (Tensor &out, const Tensor &a) |
Protected Member Functions | |
| void | mark_outputs () |
| void | compute_shape () |
| void | compute_strides () |
| void | reorder_dimensions () |
| void | permute_dimensions (IntArrayRef perm) |
| void | compute_types () |
| Type & | compute_common_type () |
| void | allocate_outputs () |
| void | coalesce_dimensions () |
Protected Attributes | |
| DimVector | shape_ |
| DimVector | perm_ |
| SmallVector< OperandInfo, 4 > | operands_ |
| int | num_outputs_ = 0 |
| bool | has_coalesced_dimensions_ = false |
| bool | accumulate_ = false |
| bool | resize_outputs_ = true |
| bool | is_reduction_ = false |
| bool | compute_common_dtype_ = true |
| bool | allow_cpu_scalars_ = false |
| bool | promote_gpu_output_dtypes_ = false |
| bool | final_output_ = true |
Friends | |
| struct | Builder |
Definition at line 101 of file TensorIterator.h.
| DimVector at::TensorIterator::compatible_stride | ( | int | element_size | ) | const |
Create a strides array for a Tensor with shape of this iterator.
The parameter element_size specifies the size of Tensor's data type in bytes (e.g. 4 for float)
Definition at line 160 of file TensorIterator.cpp.
| DimVector at::TensorIterator::invert_perm | ( | IntArrayRef | input | ) | const |
Inverts the re-ordering done by reorder_dimensions.
This can only be called before coalesce_dimensions() is called.
Definition at line 170 of file TensorIterator.cpp.
|
inline |
Whether this iterator produces the actual output, as opposed to something that will be accumulated further.
Only relevant for CUDA reductions.
Definition at line 230 of file TensorIterator.h.
| int64_t at::TensorIterator::num_output_elements | ( | ) | const |
number of elements in the output operand.
this is the same as numel() for operations that are not reductions.
Definition at line 317 of file TensorIterator.cpp.
|
inline |
If the kernel should accumulate into the output.
Only relevant for CUDA reductions.
Definition at line 225 of file TensorIterator.h.
| std::unique_ptr< TensorIterator > at::TensorIterator::split | ( | int | dim | ) |
Splits this TensorIterator into two iterators.
Together they iterate over the entire operation. Used by with_32bit_indexing().
Definition at line 590 of file TensorIterator.cpp.
| SplitUntil32Bit at::TensorIterator::with_32bit_indexing | ( | ) | const |
An "iteratable" object that recursively splits this iterator into sub-iterators that can use 32-bit indexing.
Definition at line 623 of file TensorIterator.cpp.
1.8.11
