Tensor and DLPack

Tensor and DLPack#

At runtime, TVM-FFI often needs to accept tensors from many sources:

TVM-FFI standardizes on DLPack as the lingua franca: tensors are built on top of DLPack structs with additional C++ convenience methods and minimal extensions for ownership management.

Tip

Prefer tvm::ffi::TensorView or tvm::ffi::Tensor in C++ code; they provide safer and more convenient abstractions over raw DLPack structs.

This tutorial covers common usage patterns, tensor classes, and how tensors flow across ABI boundaries.

Glossary#

DLPack

A cross-library tensor interchange standard defined in the small C header dlpack.h. It defines pure C data structures for describing n-dimensional arrays and their memory layout, including DLTensor, DLManagedTensorVersioned, DLDataType, DLDevice, and related types.

View (non-owning)

A “header” that describes a tensor but does not own its memory. When a consumer receives a view, it must respect that the producer owns the underlying storage and controls its lifetime. The view is valid only while the producer guarantees it remains valid.

Managed object (owning)

An object that includes lifetime management, using reference counting or a cleanup callback mechanism. This establishes a contract between producer and consumer about when the consumer’s ownership ends.

Note

As a loose analogy, think of view vs. managed as similar to T* (raw pointer) vs. std::shared_ptr<T> (reference-counted pointer) in C++.

Common Usage#

This section introduces the most important APIs for day-to-day use in C++ and Python.

Kernel Signatures#

A typical kernel implementation accepts TensorView parameters, validates metadata (dtype, shape, device), and then accesses the data pointer for computation:

#include <tvm/ffi/tvm_ffi.h>

void MyKernel(tvm::ffi::TensorView input, tvm::ffi::TensorView output) {
  // Validate dtype & device
  if (input.dtype() != DLDataType{kDLFloat, 32, 1})
    TVM_FFI_THROW(TypeError) << "Expect float32 input, but got " << input.dtype();
  if (input.device() != DLDevice{kDLCUDA, 0})
    TVM_FFI_THROW(ValueError) << "Expect input on CUDA:0, but got " << input.device();
  // Access data pointer
  float* input_data_ptr = static_cast<float*>(input.data_ptr());
  float* output_data_ptr = static_cast<float*>(output.data_ptr());
  Kernel<<<...>>>(..., input_data_ptr, output_data_ptr, ...);
}

On the C++ side, the following APIs are available to query a tensor’s metadata:

TensorView::shape() and Tensor::shape()

shape array

TensorView::dtype() and Tensor::dtype()

element data type

TensorView::data_ptr() and Tensor::data_ptr()

base pointer to the tensor’s data

TensorView::device() and Tensor::device()

device type and id

TensorView::byte_offset() and Tensor::byte_offset()

byte offset to the first element

TensorView::ndim() and Tensor::ndim()

number of dimensions (ShapeView::size)

TensorView::numel() and Tensor::numel()

total number of elements (ShapeView::Product)

PyTorch Interop#

On the Python side, tvm_ffi.Tensor is a managed n-dimensional array that:

The following example demonstrates a typical round-trip pattern:

import tvm_ffi
import torch

x_torch = torch.randn(1024, device="cuda")
x_tvm_ffi = tvm_ffi.from_dlpack(x_torch, require_contiguous=True)
x_torch_again = torch.from_dlpack(x_tvm_ffi)

In this example, tvm_ffi.from_dlpack() creates x_tvm_ffi, which views the same memory as x_torch. Similarly, torch.from_dlpack() creates x_torch_again, which shares the underlying buffer with both x_tvm_ffi and x_torch. No data is copied in either direction.

C++ Allocation#

TVM-FFI is not a kernel library and is not linked to any specific device memory allocator or runtime. However, it provides standardized allocation entry points for kernel library developers by interfacing with the surrounding framework’s allocator - for example, using PyTorch’s allocator when running inside a PyTorch environment.

Env Allocator. Use Tensor::FromEnvAlloc() along with C API TVMFFIEnvTensorAlloc() to allocate a tensor using the framework’s allocator.

Tensor tensor = Tensor::FromEnvAlloc(
  TVMFFIEnvTensorAlloc,
  /*shape=*/{1, 2, 3},
  /*dtype=*/DLDataType({kDLFloat, 32, 1}),
  /*device=*/DLDevice({kDLCPU, 0})
);

In a PyTorch environment, this is equivalent to torch.empty().

Warning

While allocation APIs are available, it is generally recommended to avoid allocating tensors inside kernels. Instead, prefer pre-allocating outputs and passing them as tvm::ffi::TensorView parameters. This approach:

  • avoids memory fragmentation and performance pitfalls,

  • prevents CUDA graph incompatibilities on GPU, and

  • allows the outer framework to control allocation policy (pools, device strategies, etc.).

Custom Allocator. Use Tensor::FromNDAlloc(custom_alloc, ...), or its advanced variant Tensor::FromNDAllocStrided(custom_alloc, ...), to allocate a tensor with a user-provided allocation callback.

The following example uses cudaMalloc/cudaFree as custom allocators for GPU tensors:

struct CUDANDAlloc {
  void AllocData(DLTensor* tensor) {
    size_t data_size = ffi::GetDataSize(*tensor);
    void* ptr = nullptr;
    cudaError_t err = cudaMalloc(&ptr, data_size);
    TVM_FFI_ICHECK_EQ(err, cudaSuccess) << "cudaMalloc failed: " << cudaGetErrorString(err);
    tensor->data = ptr;
  }

  void FreeData(DLTensor* tensor) {
    if (tensor->data != nullptr) {
      cudaError_t err = cudaFree(tensor->data);
      TVM_FFI_ICHECK_EQ(err, cudaSuccess) << "cudaFree failed: " << cudaGetErrorString(err);
      tensor->data = nullptr;
    }
  }
};

ffi::Tensor cuda_tensor = ffi::Tensor::FromNDAlloc(
  CUDANDAlloc(),
  /*shape=*/{3, 4, 5},
  /*dtype=*/DLDataType({kDLFloat, 32, 1}),
  /*device=*/DLDevice({kDLCUDA, 0})
);

C++ Stream Handling#

Stream context is essential for GPU kernel execution. While CUDA does not have a global context for default streams, frameworks like PyTorch maintain a “current stream” per device (torch.cuda.current_stream()), and kernel libraries must read this stream from the embedding environment.

As a hardware-agnostic abstraction layer, TVM-FFI is not linked to any specific stream management library. However, to ensure GPU kernels launch on the correct stream, it provides standardized APIs to obtain the stream context from the host framework (e.g., PyTorch).

Obtain Stream Context. Use the C API TVMFFIEnvGetStream() to obtain the current stream for a given device:

void func(ffi::TensorView input, ...) {
  ffi::DLDevice device = input.device();
  cudaStream_t stream = reinterpret_cast<cudaStream_t>(
      TVMFFIEnvGetStream(device.device_type, device.device_id));
}

This is equivalent to the following PyTorch C++ code:

void func(at::Tensor input, ...) {
  c10::Device device = input.device();
  cudaStream_t stream = reinterpret_cast<cudaStream_t>(
      c10::cuda::getCurrentCUDAStream(device.index()).stream());
}

Auto-Update Stream Context. When converting framework tensors via tvm_ffi.from_dlpack(), TVM-FFI automatically updates the stream context to match the device of the converted tensor. For example, when converting a PyTorch tensor on torch.device('cuda:3'), TVM-FFI automatically captures the stream from torch.cuda.current_stream(device='cuda:3')().

Set Stream Context. Use tvm_ffi.use_torch_stream() or tvm_ffi.use_raw_stream() to manually set the stream context when automatic detection is insufficient.

Tensor Classes#

This section defines each tensor type in the TVM-FFI C++ API and explains its intended usage. Exact C layout details are covered in Tensor Layouts.

Tip

On the Python side, only tvm_ffi.Tensor exists. It strictly follows DLPack semantics for interop and can be converted to PyTorch via torch.from_dlpack().

DLPack Tensors#

DLPack tensors come in two main flavors:

Non-owning object, DLTensor

The tensor descriptor is a view of the underlying data. It describes the device the tensor lives on, its shape, dtype, and data pointer. It does not own the underlying data.

Owning object, DLManagedTensorVersioned, or its legacy counterpart DLManagedTensor

It is a managed variant that wraps a DLTensor descriptor with additional fields. Notably, it includes a deleter callback that releases ownership when the consumer is done with the tensor, and an opaque manager_ctx handle used by the producer to store additional context.

TVM-FFI Tensors#

Similarly, TVM-FFI defines two main tensor types in C++:

Non-owning object, tvm::ffi::TensorView

A thin C++ wrapper around DLTensor for inspecting metadata and accessing the data pointer. It is designed for kernel authors to inspect metadata and access the underlying data pointer during a call, without taking ownership of the tensor’s memory. Being a view also means you must ensure the backing tensor remains valid while you use it.

Owning object, tvm::ffi::TensorObj and tvm::ffi::Tensor

Tensor, similar to std::shared_ptr<TensorObj>, is the managed class to hold heap-allocated TensorObj. Once the reference count drops to zero, the cleanup logic deallocates the descriptor and releases ownership of the underlying data buffer.

Note

  • For handwritten C++, always use TVM-FFI tensors over DLPack’s raw C tensors.

  • For compiler development, DLPack’s raw C tensors are recommended because C is easier to target from codegen.

The owning Tensor is the recommended interface for passing around managed tensors. Use owning tensors when you need one or more of the following:

  • return a tensor from a function across ABI, which will be converted to tvm::ffi::Any;

  • allocate an output tensor as the producer, and hand it to a kernel consumer;

  • store a tensor in a long-lived object.

TensorObj vs Tensor

Tensor is an intrusive pointer of a heap-allocated TensorObj. As an analogy to std::shared_ptr, think of

using Tensor = std::shared_ptr<TensorObj>;

You can convert between the two types:

Tensor Layouts#

Figure 1 summarizes the layout relationships among DLPack tensors and TVM-FFI tensors. All tensor classes are POD-like; tvm::ffi::TensorObj is also a standard TVM-FFI object, typically heap-allocated and reference-counted.

Layout of DLPack Tensors and TVM-FFI Tensors

Figure 1. Layout specification of DLPack tensors and TVM-FFI tensors. All the tensor types share DLTensor as the common descriptor, while carrying different metadata and ownership semantics.#

As demonstrated in the figure, all tensor classes share DLTensor as the common descriptor. In particular,

What Tensor Is Not#

TVM-FFI is not a tensor library. While it provides a unified representation for tensors, it does not include:

  • kernels (e.g., vector addition, matrix multiplication),

  • host-device copy or synchronization primitives,

  • advanced indexing or slicing, or

  • automatic differentiation or computational graph support.

Conversion between TVMFFIAny#

At the stable C ABI boundary, TVM-FFI passes values using Any (owning) or AnyView (non-owning). Tensors have two possible representations:

When extracting a tensor from TVMFFIAny, check the type_index to determine the representation before conversion.

Important

An owning tensor can be converted to a non-owning view, but not vice versa.

See Tensor for C code examples demonstrating:

Further Reading#

Contents