Optimized Flash Attention with blocked matrix multiplication and OpenMP, achieving 3x speedup by improving cache locality and reducing DRAM access. Also, fused attention with softmax, reducing memory usage by 80% and boosting speed by 5x with SIMD and parallelization.For CUDA Image render, I used similar techniques in loop fusion to reduce DRAM accesses and did loop tiling for maximizing throughput and used thread pooling to minimize synchronization overhead for multithreading.