Welcome to soft-cuda Discussions!

[souls-syntax]

soft.h: The Lightweight Neural Runtime

Welcome to the optimization lab.

This repository hosts the development of soft.h, a headless, high-performance deep learning framework built from scratch in C++ and CUDA.

Unlike commercial "rocket launcher" frameworks (PyTorch, TensorFlow) that abstract away the hardware, soft.h is a "scalpel." We are here to expose the raw computational cost of neural networks and determine the precise crossover point where GPU parallelism beats CPU latency.

Research Objectives

This discussion board is the primary channel for:

• Kernel Strategy: Discussing tiled memory access, warp divergence, and reduction algorithms.
• Memory Management: Debating the cost of cudaMemcpy vs. pinned memory (HostRegister).
• Benchmarking: Sharing execution times in microseconds. We measure, we do not guess.

Engineering Standards

To ensure soft.h remains lightweight and performant:

1. Zero Bloat: No external heavy dependencies. If we need a matrix multiplication, we write the kernel or use pure BLAS.
2. Explicit Ownership: malloc must have a free. cudaMalloc must have cudaFree. Memory leaks will invalidate our benchmarks.
3. Data Over Feelings: Do not say "it feels slow." Post the nvprof or nsight logs.
4. Hardware Context: When reporting results, state the specific GPU model (e.g., T4, RTX 3060) and CUDA version.

---

Environment Verification

To confirm your workstation is combat-ready for soft.h development:

1. Run nvidia-smi in your terminal.
2. Paste the output below to register your Compute Capability.
3. If you are on a CPU-only node, specify your CPU model and AVX/AVX2 support status.