CUDA C Best Practices Guide

CUDA_C_Best_Practices_Guide

User Manual:

Open the PDF directly: View PDF .
Page Count: 90

Scroll down to view the document on your mobile browser.

Table of Contents
List of Figures
List of Tables
Preface
- What Is This Document?
- Who Should Read This Guide?
- Assess, Parallelize, Optimize, Deploy
- Recommendations and Best Practices
Assessing Your Application
Heterogeneous Computing
- 2.1. Differences between Host and Device
- 2.2. What Runs on a CUDA-Enabled Device?
Application Profiling
- 3.1. Profile
Parallelizing Your Application
Getting Started
- 5.1. Parallel Libraries
- 5.2. Parallelizing Compilers
- 5.3. Coding to Expose Parallelism
Getting the Right Answer
- 6.1. Verification
  - 6.1.1. Reference Comparison
  - 6.1.2. Unit Testing
- 6.2. Debugging
- 6.3. Numerical Accuracy and Precision
Optimizing CUDA Applications
Performance Metrics
- 8.1. Timing
  - 8.1.1. Using CPU Timers
  - 8.1.2. Using CUDA GPU Timers
- 8.2. Bandwidth
Memory Optimizations
- 9.1. Data Transfer Between Host and Device
- 9.2. Device Memory Spaces
- 9.3. Allocation
- 9.4. NUMA Best Practices
Execution Configuration Optimizations
- 10.1. Occupancy
  - 10.1.1. Calculating Occupancy
- 10.2. Concurrent Kernel Execution
- 10.3. Multiple contexts
- 10.4. Hiding Register Dependencies
- 10.5. Thread and Block Heuristics
- 10.6. Effects of Shared Memory
Instruction Optimization
- 11.1. Arithmetic Instructions
- 11.2. Memory Instructions
Control Flow
- 12.1. Branching and Divergence
- 12.2. Branch Predication
- 12.3. Loop Counters Signed vs. Unsigned
- 12.4. Synchronizing Divergent Threads in a Loop
Deploying CUDA Applications
Understanding the Programming Environment
- 14.1. CUDA Compute Capability
- 14.2. Additional Hardware Data
- 14.3. Which Compute Capability Target
- 14.4. CUDA Runtime
CUDA Compatibility and Upgrades
- 15.1. CUDA Runtime and Driver API Version
- 15.2. Standard Upgrade Path
- 15.3. Flexible Upgrade Path
- 15.4. CUDA Compatibility Platform Package
- 15.5. Extended nvidia-smi
Preparing for Deployment
- 16.1. Testing for CUDA Availability
- 16.2. Error Handling
- 16.3. Building for Maximum Compatibility
- 16.4. Distributing the CUDA Runtime and Libraries
  - 16.4.1. CUDA Toolkit Library Redistribution
    - 16.4.1.1. Which Files to Redistribute
    - 16.4.1.2. Where to Install Redistributed CUDA Libraries
Deployment Infrastructure Tools
- 17.1. Nvidia-SMI
  - 17.1.1. Queryable state
  - 17.1.2. Modifiable state
- 17.2. NVML
- 17.3. Cluster Management Tools
- 17.4. Compiler JIT Cache Management Tools
- 17.5. CUDA_VISIBLE_DEVICES
Recommendations and Best Practices
- A.1. Overall Performance Optimization Strategies
nvcc Compiler Switches
- B.1. nvcc

Navigation menu