Composable Kernel Overview

Composable Kernel Features

Hardware Agnostic “Tensor Coordinate Transformation” Primitives

• Enable composition of complex operators from basic ones, without overhead of tensor re-formatting
  • GEMM --> Implicit GEMM, Hybrid direct/implicit GEMM
  • Reduction --> Pooling, Batch-norm, etc
  • Data-Transfer --> Im2Col, Depth2Space, etc
• Automatically generated optimized logic for address calculation associated with coordinate transformation without developer's intervention.

Reusable Tensor Operators for AMD GPUs

• Grid/Block/wave/thread-level tensor operators implemented as C++ templated device functions/classes
  • GEMM-like operators
  • Reduction-like operators
  • Data-transfer-like operators

Prebuilt and Customized Operator Fusion

• Prebuilt fused operators include [Work in progress]
  • GEMM/Conv + pointwise Op
  • Conv + Pooling
  • GEMM/Conv + reduction-like operator

Unified Implementation of Tensor Operators