DML_BUFFER_TENSOR_DESC structure

Describes a tensor that will be stored in a Direct3D 12 buffer resource. The corresponding tensor type is DML_TENSOR_TYPE_BUFFER, and the corresponding binding type is DML_BINDING_TYPE_BUFFER.

Syntax

struct DML_BUFFER_TENSOR_DESC {
  DML_TENSOR_DATA_TYPE DataType;
  DML_TENSOR_FLAGS     Flags;
  UINT                 DimensionCount;
  const UINT           *Sizes;
  const UINT           *Strides;
  UINT64               TotalTensorSizeInBytes;
  UINT                 GuaranteedBaseOffsetAlignment;
};

Members

DataType

Type: DML_TENSOR_DATA_TYPE

The type of the values in the tensor.

Flags

Type: DML_TENSOR_FLAGS

Specifies additional options for the tensor.

DimensionCount

Type: UINT

The number of dimensions of the tensor. This member determines the size of the Sizes and Strides arrays (if provided). In DirectML, all buffer tensors must have a DimensionCount of either 4 or 5. Not all operators support a DimensionCount of 5.

Sizes

Type: const UINT*

The size, in elements, of each dimension in the tensor. Specifying a size of zero in any dimension is invalid, and will result in an error. The Sizes member is always specified in the order {N, C, H, W} if DimensionCount is 4, and {N, C, D, H, W} if DimensionCount is 5.

Strides

Type: const UINT*

Optional. Determines the number of elements (not bytes) to linearly traverse in order to reach the next element in that dimension. For example, a stride of 5 in dimension 1 means that the distance between elements (n) and (n+1) in that dimension is 5 elements when traversing the buffer linearly. The Strides member is always specified in the order {N, C, H, W} if DimensionCount is 4, and {N, C, D, H, W} if DimensionCount is 5.

Strides can be used to express broadcasting (by specifying a stride of 0) as well as padding (for example, by using a stride larger than the physical size of a row, to pad the end of a row).

If Strides is not specified, each dimension in the tensor is considered to be contiguously packed, with no additional padding.

TotalTensorSizeInBytes

Type: UINT64

Defines a minimum size in bytes for the buffer that will contain this tensor. TotalTensorSizeInBytes must be at least as large as the minimum implied size given the sizes, strides, and data type of the tensor. You can calculate the minimum implied size by calling the DMLCalcBufferTensorSize utility free function.

Providing a TotalTensorSizeInBytes that is larger than the minimum implied size may enable additional optimizations by allowing DirectML to elide bounds checking in some cases if the TotalTensorSizeInBytes defines sufficient padding beyond the end of the tensor data.

When binding this tensor, the size of the buffer range must be at least as large as the TotalTensorSizeInBytes. For output tensors, this has the additional effect of permitting DirectML to write to any memory within the TotalTensorSizeInBytes. That is, your application mustn't assume that DirectML will preserve any padding bytes inside output tensors that are inside the TotalTensorSizeInBytes.

The total size of a buffer tensor may not exceed (2^32 - 1) elements—for example, 16GB for a FLOAT32 tensor.

GuaranteedBaseOffsetAlignment

Type: UINT

Optional. Defines a minimum guaranteed alignment in bytes for the base offset of the buffer range that will contain this tensor, or 0 to provide no minimum guaranteed alignment. If specified, this value must be a power of two that is at least as large as the element size.

When binding this tensor, the offset in bytes of the buffer range from the start of the buffer must be a multiple of the GuaranteedBaseOffsetAlignment, if provided.

Buffer tensors always have a minimum alignment of 16 bytes. However, providing a larger value for the GuaranteedBaseOffsetAlignment may allow DirectML to achieve better performance, because a larger alignment enables the use of vectorized load/store instructions.

Although this member is optional, for best performance we recommend that you align tensors to boundaries of 32 bytes or more, where possible.

Requirements

   
Header directml.h

See Also

Binding in DirectML