Popular repositories Loading

1. cuda-sgemm-optimization cuda-sgemm-optimization Public

   CUDA SGEMM kernel optimization: naive vs shared-memory tiled implementation with roofline analysis on Tesla T4

   Cuda

2. cuda-flash-attention cuda-flash-attention Public

   CUDA Flash Attention kernel with online softmax — naive vs tiled implementation, verified on Tesla T4

   Cuda

3. cuda-batched-attention cuda-batched-attention Public

   CUDA Batched Multi-Head Attention — warp-cooperative flash attention with online softmax, verified on Tesla T4 across 6 benchmark configurations

   Cuda

4. cuda-fused-softmax-layernorm cuda-fused-softmax-layernorm Public

   High-performance CUDA implementation of a fused Softmax + LayerNorm kernel with shared memory optimization, warp-level reductions, and 1.62x latency improvement on NVIDIA Tesla T4.

   Cuda

5. cuda-int8-gemm-optimization cuda-int8-gemm-optimization Public

   INT8 Quantized GEMM kernel for NVIDIA T4 using Tensor Cores. Zero register spills, 6.75 TOPS sustained throughput, full correctness validation. Production-ready CUDA optimization.

   Cuda

6. cuda-parallel-prefix-scan cuda-parallel-prefix-scan Public

   Optimized parallel prefix sum (scan) for NVIDIA T4 with zero register spills, bank-conflict-free shared memory, and hierarchical recursion. Achieves ~119 GB/s sustained throughput.

   Cuda