triton-ascend-case-elemwise-concat

What problem does it solve?

The Slice+Concat fusion case optimizes multi-input fusion kernels by loading only the required input slices and assembling the outputs through precise indexing, eliminating unnecessary intermediate storage and memory traffic.

Core Features & Use Cases

• Precise slicing: loads only the needed portions of each input (e.g., slices like 128, 32, 48) to form the final output without materializing full tensors.
• In-kernel concatenation: uses calculated output offsets to write inputs contiguously, avoiding cat-like operations and extra buffers.
• Use Case: accelerates fusion kernels where several inputs must be sliced and concatenated within a single kernel, reducing bandwidth and latency.

Quick Start

Run the provided fused kernel with the specified input slices to validate the elementwise concat optimization.