Module runtime

The runtime objects the IR refers to.

A MemoryScope, such as the scope of a PtrType, is an enum object that python/tvm_tirx/runtime/memory_scope.py defines. The runtime functions the Python package wraps are here too: tensor allocation and views, device attributes and runtime queries.

Structs

• MemoryScope
• MemoryScopeObj
  Opaque: parent ‘ffi.Enum’ is opaque (no-mirror). Fields are read through the C ABI getters.

Functions

• create_view
  A view of the storage of source, relative_byte_offset bytes past its data, with another shape and data type (runtime.TensorCreateView).
• empty
  An uninitialized tensor on device, in the memory scope mem_scope or the global one (runtime.TensorAllocWithScope).
• enabled
  Whether the runtime of the device or target kind target ("cpu", "cuda", "llvm"…) is enabled (runtime.RuntimeEnabled).
• get_cuda_device_count
  The number of CUDA devices (runtime.GetCudaDeviceCount).
• get_cuda_free_memory
  The free and total memory of the current CUDA device, as a message (runtime.GetCudaFreeMemory).
• get_device_attr
  The attribute name of device, such as exist or max_threads_per_block; None when the device does not report it (runtime.GetDeviceAttr).
• num_threads
  The number of threads the runtime thread pool uses (runtime.NumThreads).