AethronPhantom/pyc-kernels

PyC CUDA Kernels

Twenty CUDA .cu prototype files from the PyC kernel lab. The collection covers baseline and experimental matmul, Ada FP32/Tensor Core variants, Hopper WMMA BF16/FP16 variants, asynchronous staging, CTA/warp/K-stage experiments, and cuBLASLt control lanes.

These are source artifacts for inspection, compilation, and performance engineering—not precompiled model weights and not a claim of universal benchmark superiority. The 19 documented lab lanes are:

Kernel	Short description	Source
matrix_mult	Compile-only CUDA matmul baseline.	kernels/prototypes/baseline/matmul/kernel.cu
tokenizer_kernel	Experimental tokenizer/matrix kernel.	kernels/prototypes/experimental/tokenizer_matmul/kernel.cu
ada_gemm	Ada FP32 shared-memory GEMM.	kernels/prototypes/ada/gemm/kernel.cu
ada_gemm_k64_warp32_async	Ada FP32 GEMM with cp.async double buffering.	kernels/prototypes/ada/gemm_k64_warp32_async/kernel.cu
ada_tensor_core_fp16	Ada WMMA FP16 GEMM.	kernels/prototypes/ada/tensor_core/kernel.cu
ada_tensor_core_bf16	Ada WMMA BF16 GEMM.	kernels/prototypes/ada/tensor_core/kernel.cu
hopper_tensor_core_fp16	Hopper 64×64×16 WMMA FP16 GEMM.	kernels/prototypes/hopper/tensor_core/kernel.cu
hopper_tensor_core_bf16	Hopper 64×64×16 WMMA BF16 GEMM.	kernels/prototypes/hopper/tensor_core/kernel.cu
hopper_tensor_core_bf16_perf	Large-shape Hopper BF16 throughput lane.	kernels/prototypes/hopper/tensor_core/kernel.cu
hopper_tensor_core_bf16_warp2n	Hopper lane with two N fragments per warp.	kernels/prototypes/hopper/tensor_core/kernel.cu
hopper_tensor_core_bf16_k32	Hopper lane with 32-wide K staging.	kernels/prototypes/hopper/tensor_core/kernel.cu
hopper_tensor_core_bf16_warp2n_k32	Combined warp2n and K32 experiment.	kernels/prototypes/hopper/tensor_core/kernel.cu
hopper_cublaslt_bf16	cuBLASLt BF16 hardware-ceiling control.	kernels/prototypes/hopper/cublaslt_bf16/kernel.cu
hopper_cublaslt_bf16_check	cuBLASLt BF16 correctness lane.	kernels/prototypes/hopper/cublaslt_bf16/kernel.cu
hopper_tensor_core_bf16_async	Hopper BF16 WMMA with async double buffering.	kernels/prototypes/hopper/tensor_core_async/kernel.cu
hopper_tensor_core_bf16_async_wide	Async WMMA with a wider 64×128 CTA.	kernels/prototypes/hopper/tensor_core_async/kernel.cu
hopper_tensor_core_bf16_async_wide_k64	Wide async lane with K64 stages.	kernels/prototypes/hopper/tensor_core_async/kernel.cu
hopper_tensor_core_bf16_async_square	Async WMMA with a 128×128 CTA.	kernels/prototypes/hopper/tensor_core_async/kernel.cu
hopper_tensor_core_bf16_async_square_k64	Square async WMMA with 128×128×64 tiles.	kernels/prototypes/hopper/tensor_core_async/kernel.cu

The source files are organized by prototype family; several documented lanes share one parameterized .cu implementation. See the companion performance Space for the optimization story and selected measurements.