Skip to content

README.md

Latest commit

 

History

History
72 lines (55 loc) · 3.39 KB

README.md

File metadata and controls

72 lines (55 loc) · 3.39 KB
Raw

Algorithms

Pre-built algorithms for common dataflow patterns on AIE. These abstractions handle Workers, ObjectFifos, and data movement automatically.

Available Algorithms

Algorithm Description
transform Unary transformation with tiled processing
transform_binary Binary transformation with tiled processing
for_each In-place transformation (same tensor for input/output)
transform_parallel Parallel transform across multiple AIE tiles
transform_parallel_binary Parallel transform_binary across multiple AIE tiles

Using C++ Kernels (ExternalFunction)

For C++ kernels, the kernel signature and ExternalFunction must match the algorithm's expected format.

Algorithm C++ Kernel Signature ExternalFunction arg_types
transform void kernel(T* in, T* out, params..., int32_t tile_size) [tile_ty, tile_ty, *param_types, np.int32]
transform_binary void kernel(T* in1, T* in2, T* out, params..., int32_t tile_size) [tile_ty, tile_ty, tile_ty, *param_types, np.int32]
for_each void kernel(T* in, T* out, params..., int32_t tile_size) [tile_ty, tile_ty, *param_types, np.int32]
transform_parallel void kernel(T* in, T* out, params..., int32_t tile_size) [tile_ty, tile_ty, *param_types, np.int32]
transform_parallel_binary void kernel(T* in1, T* in2, T* out, params..., int32_t tile_size) [tile_ty, tile_ty, tile_ty, *param_types, np.int32]

Important: Tile Size Configuration

The algorithm library automatically tiles input/output tensors and streams them through ObjectFifos. However, C++ kernels require tile_size as a runtime parameter because they loop over tile data without compile-time knowledge of the size. As such the last argument of each kernel this library supports must be tile_size (np.int32)

For setup, this means:

  • ExternalFunction.arg_types must match the C++ kernel signature, using tile-sized array shapes
  • Algorithm call accepts full-size tensors and handles tile automatically in terms of ObjectFifo streaming.
  • tile_size must be passed explictly as a keyword argument because the kernel needs it at runtime.

TODO: Future work is needed to consolidate this setup process and combine with ExternalFunction declaration.

Example: transform with ExternalFunction

TILE_SIZE = 16
tile_ty = np.ndarray[(TILE_SIZE,), np.dtype[np.int16]]
scalar_ty = np.ndarray[(1,), np.dtype[np.int32]]

# Input and output should still be declared as tiled shapes in arg_types
my_kernel = ExternalFunction(
    "my_kernel",
    source_file="my_kernel.cc",
    arg_types=[tile_ty, tile_ty, np.int32, np.int32],  # [in, out, factor, tile_size]
)

# Pass in full size tensors, algorithm will tile it to tile_size
# tile_size must be passed as a keyword argument
iron.jit(transform)(
    my_kernel, input, output, factor, tile_size=TILE_SIZE
)

For a complete example, see vector_scalar_mul.py.

Ryzen™ AI Usage

Run and verify a design:

python3 transform.py