CUDA Projects

A place for my personal CUDA tinkering to reside. The layout of the repository adheres to The Pitchfork Layout, specifically using seperate header placement and merged test placement.

Dependencies

• CMake
• CUDA Toolkit

Perform Out-Of-Source Build

From the repository root:

mkdir build && cd build && cmake ..

For example, to build Tests for Release on Windows:

msbuild.exe Tests.vcxproj /property:Platform=x64 /property:Configuration=Release

To then run the Tests executable:

Release\Tests.exe

---

Exclusive Scan

Work In Progress. Currently toying around with my exclusive scan implementation using the CUDA cooperative groups API.

Given a binary associative operator $\oplus$ with the identity element $\mathbb{i}$ and a contiguous array of n elements $\mathbf{A}_n = [a_1, a_2, \ldots, a_n]$:

$$\textrm{Exclusive Scan}(\mathbf{A}_n) = [\mathbb{i}, a_1, a_1 \oplus a_2, a_1 \oplus a_2 \oplus a_3, \ldots, a_1 \oplus a_2 \ldots \oplus a_n-1]$$

---

Saxpy

The canonical single-precision ax + y kernel: $$\mathbf{\overline{z}} = \alpha\mathbf{\overline{x}} + \mathbf{\overline{y}} \quad \textrm{where} \quad \mathbf{\overline{x}}, \mathbf{\overline{y}}, \mathbf{\overline{z}} \in \mathbb{R}^{n}, \textrm{ } \alpha \in \mathbb{R}$$ This is used as a prototype kernel for code and file organization.

---

Utilities

CUDA and host utility functions that are useful when working with the CUDA programming model.

---