@mkretz@floss.social github.com/mattkretzMotivation std::simd Overview Example: Image Processing Programming Models Outlook Summary Goals and non-goals for this talk • This is not a tutorial! You won’t really know Center for Heavy Ion ResearchMotivation std::simd Overview Example: Image Processing Programming Models Outlook Summary Motivation Motivation © by Matthias Kretz Matthias Kretz CppCon ’23 GSI Helmholtz Programming Models Outlook Summary std::simd is for you! Matthias Kretz CppCon ’23 4 GSI Helmholtz Center for Heavy Ion ResearchMotivation std::simd Overview Example: Image Processing Programming Models Outlook

the preprocessor • Use the preprocessor to determine if you are using the models via a compiler flag, then: #ifdef USE_MODELS typedef iouint32_t HookableRegister; #else typedef iouint32_t uint32_t; #endif must be written in C, this gives an easy way to not modify the C code but use the C++ coroutine models.Intel Confidential Department or Event Name 37 © 2023 Intel Corporation and Jeffrey E. Erickson cppcon easier to write • You can create complex ‘parallel’ models with relative ease • Your more HW focused friends will probably find coroutine based models easier to read than thread-based ones • This is a

efficiency to stop scaling at a certain core count. It is an interesting question whether some parallelism models have other limitations that make them scale worse than others. Computational structure Our mini-app some profiling to analyze the indi- vidual operations. Parallelism Models In scientific computing, the two dominant parallelism models are MPI (Messsage Passing Interface) for distributed memory, and OpenMP will compare to a more ordinary desktop, where bandwidth limitations will be much more severe. Most models are capable of running – with varying amounts of effort – on GPUs. The software / hardware integration

library aims to provide a simple yet powerful framework for building and training machine learning models. By leveraging the performance efficiency of C++, micro- grad++ offers a robust solution for integrating Backpropagation: The implementation of backpropa- gation in micrograd++ allows for efficient training of models through gradient descent. • Gradient Clipping: To prevent the issue of exploding gradients, micrograd++ class represents a multi-layer percep- tron, composed of multiple layers. It supports the training of models using backpropagation and gradient descent, allowing for the efficient optimization of network parameters

capabilities for your extension using new APIs Customize Copilot to your needs using custom models, instructions, and more…Build custom Copilot Chat features Build your own custom features with your extension using new APIs Customize Copilot to your needs using custom models, instructions, and more…Custom Models for tailored Copilot suggestions Receive more personalized and precise code

• Iterates backward from the end of a sequence to the beginning • Models a bidirectional iterator when Iter is bidirectional • Models a random-access iterator when Iter is random-access • Insert iterators template front_insert_iterator; • template insert_iterator; • Models an output iterator that inserts elements at the back / front / interior of a container 45CppCon

Example WDF Circuit Model chowdsp_wdf wd_models RT-WDF LPF-2 0.3319 0.7685 3.141 FF-2 2.035 2.083 11.538 Diode Clipper 1.905 5.756 N/A Bassman Tone Stack 0.6576 0.7411 7.92 Baxandall EQ 1.184 1.021 run-time performance of chowdsp_wdf with two other open-source WDF libraries: RT-WDF (C++) and wd_models (Faust). For all tested circuits chowdsp_wdf achieved the best or 2nd-best score. Code Repository

r dialog = some_object; if (cond) dialog = some_other_object; process(dialog); 11 CppCon 2020models rebindable reference 12 CppCon 2020closely matches lvalue references • it may refer to objects f(args); 39 CppCon 2020Function pointers – a rebindable reference in the language • its usage models after references • you don’t need to write • just , as if is a reference to function (e.g., an

50Conclusions ◮ Code isn’t enough... data matters ◮ More diverse data leads to better models ◮ Building accurate models is an expert job ◮ Running on-device inference is straightforward ◮ Running on-device

object is a . For a range adaptor closure object C and an expression R such that decltype((R)) models viewable_range, the following expressions are equivalent and yield a view: unary function object object is a . For a range adaptor closure object C and an expression R such that decltype((R)) models viewable_range, the following expressions are equivalent and yield a view: unary function object object is a . For a range adaptor closure object C and an expression R such that decltype((R)) models viewable_range, the following expressions are equivalent and yield a view: unary function object