GPU High-Performance Operator Development Requirement to TileLang Code Generator

You are an expert GPU kernel developer proficient in TileLang (tile-lang), a Python-based DSL for writing high-performance GPU kernels. Help me implement a custom kernel. ## My Kernel Requirements - Operation type: [e.g., GEMM / FlashAttention / Dequant GEMM / Custom fused op] - Data types: [e.g., FP16 / BF16 / FP8 / INT8 with FP16 accumulate] - Matrix dimensions: [e.g., M=4096, N=4096, K=1024 / variable batch] - Target hardware: [e.g., NVIDIA A100 / H100 / Apple M-series / Huawei Ascend] - Performance target: [e.g., >90% of cuBLAS / match FlashAttention-2 throughput] ## Generate 1. **TileLang Kernel Code**: Complete, runnable TileLang kernel with: - Proper tile sizes for the target hardware - Shared memory usage and pipeline stages - T.gemm() or T.reduce() primitives as appropriate - Block and thread configuration 2. **Launch Configuration**: Host-side code to compile and invoke the kernel 3. **Correctness Test**: A simple test comparing against PyTorch reference 4. **Performance Benchmark**: Benchmark script with roofline analysis 5. **Optimization Notes**: Explain tile size choices, memory access patterns, and potential further optimizations Use TileLang v0.1.6+ API conventions. Include comments explaining each optimization decision.