Silicon workbench for chip blueprints, roofline analysis, tile-memory flow, and AI accelerator co-design.
-
Updated
Oct 9, 2026 - HTML
Silicon workbench for chip blueprints, roofline analysis, tile-memory flow, and AI accelerator co-design.
Reproducible study of how tensor layout, reduction length, and pointer alignment drive vendor GEMM dispatch and latency cliffs.
Advanced static analysis and visual linting tool for Triton GPU IR (TTGIR). Automatically diagnoses shared-memory bank conflicts, global memory coalescing efficiency, register layouts, and SM occupancy limits. Includes an early-pruning autotune hook and an interactive HTML report engine.
Custom GPT-2 GPU inference engine with fused Triton kernels, KV-cache, and INT8 quantization. Hand-crafted & verified to FP32 correctness.
Diagnosing non-monotonic activation memory in torch.compile's memory-budget partitioner, with a use-site rematerialization fix.
To associate your repository with the deep-learning-compilers topic, visit your repo's landing page and select "manage topics."