DeepVision-RTX Starter

This is an enhanced README combining:

• GitHub-friendly formatting
• Technical deep dive
• ASCII diagrams
• Rendered PNG diagrams
• Profiling workflow
• Architecture explanations
• GPU hardware reasoning

Architecture Diagram (SVG)

Streams Overlap Diagram (SVG)

ASCII Architecture Diagram

   +-------------------+      PCIe / DMA      +---------------------+
   |      CPU Host     |  ----------------->  |     GPU Global      |
   |  (Pinned Memory)  |  <-----------------  |      Memory         |
   +-------------------+                      +---------------------+
            |                                           |
            | Launch Kernels                            |
            v                                           v
      +--------------+                          +------------------+
      | CUDA Driver  |                          |  SMs (Ada 8.9)   |
      | Runtime API  |                          |  Warps / Threads |
      +--------------+                          +------------------+

ASCII Streams Overlap Diagram

Time →
----------------------------------------------------------------------
H2D Stream 0: [========== copy =========]
H2D Stream 1:         [========== copy =========]
Compute Stream 2:             [==== kernel ====]
D2H Stream 0:                           [==== copy ====]
----------------------------------------------------------------------

Build Instructions

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

Profiling Instructions (Nsight Systems & Nsight Compute)

nsys profile -o nsys_report ./build/deepvision_rtx
ncu --set full ./build/deepvision_rtx

Project Structure

src/
   main.cpp            # Entry point
   main.cu             # Separate experimental demo
   conv_kernels.cu     # SAXPY + blur3x3 kernels
   conv_kernels.cuh    # Kernel declarations
   utils/
       check_cuda.hpp  # Error-checking utilities

GPU Architecture Notes (RTX 4070 SUPER - Ada 8.9)

• SM count: 46
• Warp size: 32
• Max threads/block: 1024
• Memory bandwidth: 504 GB/s
• Concurrent copy/compute supported
• Best performance achieved when:
  • You use pinned memory
  • H2D and D2H overlap with compute
  • Kernels maintain good occupancy

Roadmap

• Add shared-memory tiled blur
• Add constant-memory kernel variants
• Add half-precision path (FP16)
• Add Tensor Core WMMA version
• Add occupancy analysis + roofline plot
• Add Nsight Compute performance tables
• Add multi-kernel pipelines
• Compare against cuDNN for 3×3 conv

---