Talk #1190
Use SIMD for maximal performance | The Vc library
100%
Description
Basically all currently used CPUs support (data-)parallel execution in a single thread (SIMD — i.e. parallel execution without using multi-threading). Depending on the CPU and problem the execution throughput differs by factors of 2 up to 32.
In this talk I will introduce what SIMD means on current CPUs. Then I will give a quick overview over the SIMD programming models provided by the industry. Finally, in the main part, I will talk about how you can use SIMD types and the Vc library, which encode data-parallel execution in the types and operations.
Files
Updated by Kretz, Matthias over 9 years ago
- Subject changed from express data-parallelism for maximal performance to Use SIMD for maximal performance | The Vc library
Updated by Kretz, Matthias over 9 years ago
- File I_Want_Performance!.pdf I_Want_Performance!.pdf added
- Status changed from New to Presented
- % Done changed from 0 to 100
I recommend to also watch: Chandler Carruth on Efficiency with Algorithms, Performance with Data Structures. He underlines my point about data-structures that have to do pointer chasing. He also makes a very strong case for optimizing your C++ codes...