MT-Agent

Provides MT-3000 AM vectorization interfaces and transformation templates. Use when converting scalar kernels to vector kernels, selecting vector templates, or fixing AM vectorization compile issues.

Defines compile-test-feedback loops for MT-3000 kernel generation, including retry polling, error extraction, and report output. Use when validating generated AM/SM code with compiler diagnostics and iterative fixes.

Captures hthreads kernel programming patterns for MT-3000, including thread partitioning, kernel structure, and device-side constraints. Use when generating or reviewing __global__ kernels and task partition logic.

Provides MT-3000 platform background, compute-memory model, and optimization constraints. Use when analyzing MT-3000 kernels, deciding AM/SM optimization boundaries, or reasoning about memory hierarchy and data movement.

Provides SM cache optimization strategy selection and templates for CACHEd, CACHEs, and CACHEb. Use when optimizing scalar kernels with on-chip cache or when choosing cache strategy by memory access pattern.