Computational Budget Tool from RidgeRun for NVIDIA GPU and CUDA Optimisation

RidgeRun CUDA Optimisation Guide
GPU Architecture
Execution process Memory hierarchy Communication & Concurrency
Optimisation Workflow
Tools CUDA Memcheck CUDA Profiler Computational Budget Tool
Optimisation Recipes
Types of optimisations Coarse optimisations Workload offloading Problem size Communication overlapping Correct memory access patterns Inter-thread communication Memory management Fine optimisations Increase arithmetic intensity Function approximation Condition and loops replacement Inlining
Common pitfalls when optimising
Examples
Empirical Experiments
Simple bounding test Multi-threaded bounding test
Contact Us

Introduction to Computational Budget Tool from RidgeRun

The following tool provides a reference model to analyze the computational budget for a platform with predefined characteristics, with a reference convolution algorithm with a configurable kernel size. It should be used just as a theoretical budget calculation as the real implementation of an algorithm varies in the computational resources used. The tool considers an ideal case where all the resources are available to the platform and does not consider any other high-level factors such as an operating system, therefore the results must be used with caution.

How to use

1. Select a profile (Jetson Xavier) or input the platform characteristics (CPU, GPU, Memory)
2. Select the default image and kernel characteristics.
3. Click Calculate.
4. You can also change the input data and press Enter to update the results.
5. Analyze the Roofline model and math computations.

Tool

The following tool will provide theoretical max computating capabilities for custom platforms for a sample convolution algorithm.

Select a default profile or provide platform characteristics.

Jetson Xavier

---

X Axis Init: X axis init value for the roofline function.

---

CPU Information

CPU Cores Number of cores available on the platform.

Vector Units Number of vector units available on CPU.

FMA Factor Fused-multiply-add operation factor available on CPU.

CPU Frequency (GHz) Base core clock CPU frequency.

GPU Information

GPU Throughput (GFLOP/s) GPU throughput available on the platform.

GPU-CPU-Memory Bandwidth (GB/s) Max theoretical shared bandwidth between GPU, CPU and memory.

Image Settings

Width Test image width.

Height Test image height.

Image bpp Bits per pixel for the test image.

Kernel Width Kernel width size for the convolution.

Kernel Height Kernel height size for the convolution.

Framerate Framerate value required for the test.

Calculate Reset

Suitability for GPU Compute the results and roofline model for GPU instead of CPU.

Roofline Model

---

Results

Application suitability on:

Resolution

Kernel Size

$O_{app} \;\text{(MFLOP)}$ Theoretical number of floating point operations needed by the application.

$T_{app} \;\text{(GFLOP/s)}$ Theoretical throughput with tolerance needed by the application.

$t_{CPU}$ Factor of the available computing time.
It must be < 1.

$B_{app} \;\text{(GB/s)}$ Bandwidth needed by the application.

Status Result if it is theoretically possible to run the application with current settings.