DeepVision-RTX (Starter)

A CUDA C++ practice project designed for the RTX 4070 SUPER (Ada 8.9), demonstrating how to overlap data transfers with computation using streams and pinned memory, apply basic kernel optimizations with 1D and 2D grid configurations, and perform precise event timing for profiling in Nsight Systems and Nsight Compute.

What’s here?

• Pinned host memory + cudaMemcpyAsync to demonstrate overlap
• Multiple streams for concurrent copy/compute
• Timed sections with cudaEventRecord / cudaEventElapsedTime
• Kernels: saxpy (1D), blur3x3_naive (2D)

Build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j
./build/deepvision_rtx

Profile (examples)

# Nsight Systems GUI (on host with CUDA Toolkit installed)
nsys profile -o nsys_report ./build/deepvision_rtx

# Nsight Compute single-kernel collection
ncu --set full --target-processes all ./build/deepvision_rtx

Next steps

• Convert blur kernel to shared-memory tiled version
• Add half-precision path to prep for Tensor Cores
• Compare end-to-end with cuDNN and optionally TensorRT