🎬 Custom CUDA Kernel with Stream Concurrency

A fused image processing pipeline using CUDA, Cupy, and NVTX—profiling kernel performance with Nsight Systems and visualizing execution in a dashboard-style notebook.

---

---

📸 Input vs. Fused Output

Input Image

Fused Output

The fused output combines blur, brightness correction, and Sobel edge detection in a single GPU pass.

---

⚙️ Pipeline Overview

• Preprocessing: Load grayscale image and transfer to GPU

• Fused Kernel: Apply blur → brightness → Sobel → blend

• Postprocessing: Transfer result back to CPU and save

All operations are wrapped in NVTX markers and launched inside a non-blocking CUDA stream for concurrency.

---

🔧 Kernel Summary

// CUDA kernel pseudocode blur = average(3x3 neighborhood) corrected = clamp(blur + brightness) edge = sobel_x + sobel_y output = (corrected + edge) / 2

Implemented using cp.RawKernel and launched with (16, 16) blocks.

---

📊 Nsight Systems Timeline

• 🟩 Preprocessing: Image load and GPU transfer

• 🟪 Fused Kernel in Stream: CUDA kernel execution

• 🟥 Postprocessing: Result transfer and save

NVTX markers segment the timeline for clarity. Kernel execution time: ~X.XX ms

---

🚀 Performance Insights

• Fused kernel reduced launch overhead and memory latency

• Stream concurrency enabled overlap between compute and memory ops

• NVTX markers provided clear segmentation for profiling

---

🧠 Next Steps

• Add batch processing and multi-stream orchestration

• Expand to RGB image support

• Integrate Nsight Compute for SM occupancy and memory throughput

• Build a visual dashboard to compare fused vs. modular kernels

---

🧾 Notebook Preview

• Explore the full pipeline in GPU_Profiling_Storyboard.ipynb, including:

• Input/output previews

• Kernel summary

• Timeline analysis

• Performance commentary

---

Install instructions

conda create -n gpu-lab --file requirements.txt

---

🧠 Author

Dartayous

Creative Technologist & AI/ML Engineer Specializing in GPU profiling, modular pipelines, and cinematic documentation