Data is all you need for fusion N 1int x = 4; callee(x); // do work } #include #include #include "benchmark.h" #include "matrix_lib.h" int main(..){ std::vector a; a.reserve(100); 1, &R(0, j), R.m); cblas_saxpy(m * ColTile, 1.0f, &A(0, j), inc_x_y, &R(0, j), inc_x_y); } } Fusion!A Simple Computation Chunky AccelerateKey Observation #2: Naive function composition results in Algebra Neural Nets Database Trees 89 x: 0, y: 0, len_x:2, len_y:2Write a Pipeline Lightweight Fusion of Black Box- Function Interfaces /* Make a pipeline of functions using the classes inhertied from

nested loops in a textbook sequential implementation 3. Increasing potential data reuse, loop fusion, & parallelism24 1980’s: Evolving computer architectures expanded BLAS Architecture Mainframe Vector nested loops in a textbook sequential implementation 3. Increasing potential data reuse, loop fusion, & parallelism Competed in ‘80s & early ‘90s; see 1991 New York Times article “Killer Micros”25

. * - this_talk * - this_talk C++ C++ boost.mpl boost.mpl boost.mp11 boost.mp11 boost.fusion boost.fusion boost.hana boost.hana mp mp https://wg21.link/p2996 https://wg21.link/p2996 Circle-lang Circle-lang

accelerating larger C++-based engines and applications with performance portability C++ Kernel Fusion can give better performance on complex apps and libs than hand-coding AI/Tensor HW GPU FPGA Parameterized Code and dynamically compose the algorithms (C++ templates, parallel STL, inlining and fusion, abstractions) Math, ML, Data Libraries; C++ Std, C, Python Libraries Libraries augment compiler

applications 8 years of professional experience Maintaining mission-critical software systems Ex VMware Carbon Black, Canon Medical Informatics Former founder of a developer tools startup Pejman Ghorbanzade

Positions • Flight Software Engineer (Image Processing/Algorithms) • Flight Software Engineer (Sensor Fusion/Algorithms) • Ground Systems Software Engineer • Embedded Software Engineer • GNC/Software Manager

return types: Consistent with serial range algorithms for easy migration. ● ● Enable single-call fusion of multiple operations ● Preserve the expressiveness of rangesKey Design Decisions 1. Return types

%o0, %o0 15 8d 04 03 37 af 26 11 78 62 10 82 f0 22 12 90 20 70 28 83 08 40 00 90Code density - fusion uint64_t v = 0x123456789abcdef0; 51 x86 movabs r10, 0x123456789abcdef0 49 ba f0 de bc 9a 78 56