SPPU · Final Year Artificial Intelligence and Data Science · 2020 Course
GPU Programming and Architecture
417524A · Elective IV · Semester VII · 3 credits · In-semester 30 + end-semester 70
One of the Elective IV options.
Unit 1: GPU Architecture
6 hoursEvolution of GPUs; parallelism; CUDA hardware; threads, blocks, grids, warps and scheduling; shared, global, constant and texture memory.
Marks: 15 in the in-semester exam
- GPU architecture
- Threads, blocks, grids and warps
- CUDA memory types
Unit 2: CUDA Programming
6 hoursCUDA platform and programming model; kernels and thread and memory hierarchies; heterogeneous and asynchronous SIMT programming; compute capability; multi-CPU and multi-GPU; optimising decomposition, memory, transfers and threads.
Marks: 15 in the in-semester exam
- CUDA kernels and thread hierarchy
- Heterogeneous programming
- Multi-GPU programming
- Optimising CUDA applications
Unit 3: CUDA Programming Issues
6 hoursError handling; synchronisation and algorithmic issues; debugging; NVCC compilation workflow; binary and PTX compatibility; the CUDA runtime.
Marks: 17–18 in the end-semester exam
- CUDA error handling
- Synchronisation in CUDA
- NVCC compilation and PTX
Unit 4: Introduction to OpenCL
6 hoursOpenCL model and host programs; alternatives; kernels; host–device interaction; execution environment and memory model; examples.
Marks: 17–18 in the end-semester exam
- OpenCL programming model
- OpenCL kernels and host code
- OpenCL memory model
Unit 5: Algorithms on GPUs
6 hoursConvolution, prefix sum and sparse matrix multiplication; heterogeneous clusters; dynamic parallelism and its limits; memory allocation.
Marks: 17–18 in the end-semester exam
- Parallel convolution
- Parallel prefix sum
- Sparse matrix multiplication on GPUs
- Dynamic parallelism
Unit 6: OpenCL Application Design
6 hoursOpenCL platform and memory models; application-specific processors and TTA; standalone OpenCL; heterogeneous computing; neural network training and inference.
Marks: 17–18 in the end-semester exam
- OpenCL for heterogeneous computing
- GPU neural network training
Unit contents are summarised from the official SPPU syllabus . AiCanCode.org is not affiliated with Savitribai Phule Pune University.