• AutoGEMM
• I - Introduction to AutoGEMM
• II - Achieving Peak GEMM Performance
• 1. Global memory bandwidth as a limiting factor
• 2. Global memory latency as a limiting factor
• 3. Local memory bandwidth as a limiting factor
• 4. Local memory latency as a limiting factor
• 5. Kernel throughput is high
• 6. High percentage of MADD instructions
• 7. Special considerations for small MxN matrices
• 8. Special considerations for small K matrices
• 9. Special considerations for "skinny" matrices
• III - Architecture
• AutoGEMM Parameters
• Kernel selection data
• Generating the kernel selection data
• AutoGEMM generated files
• "Building" AutoGEMM with CMake for clBLAS
• IV - Customizing AutoGEMM to your needs
• Customize kernel assortment for an application
• Tuning kernel selection to new GPU