Introduction Standardization Design EDSL Extents Going beyond Conclusion The Shapes of Multi-Dimensional Arrays Vincent Reverdy September 17th, 2020 CppCon - Vincent Reverdy - September 17th, 2020 - Online Dedication 3Introduction Standardization Design EDSL Extents Going beyond Conclusion Multidimensional arrays and linear algebra        0.5 1.2 4.1        1D        0.5 4.5 −0.3 EDSL Extents Going beyond Conclusion Personal use case with optimal transport: up to 6D dynamic arrays with billions of elements CppCon - Vincent Reverdy - September 17th, 2020 - Online CC0 1.0 Universal

of raw pointers for creating data structures, passing data into functions, dynamically allocated arrays, and function pointers. This portion will cover capabilities of raw pointers and syntax: * (asterisk) Pointers as parameters...check Let’s now talk about pointers versus arrays (and dynamically allocated arrays) 71Pointers and Arrays 72Visualizing Memory - Linear array of addresses (1/4) 73 ● Let’s allocated) *Assume 1 byte per box 7 ‘a’ 3.1415Visualizing Arrays (1/8) 77 ● Okay, so what happens when we create an array of data?Visualizing Arrays (2/8) 78 ● Okay, so what happens when we create an

vector of future Eigen arrays) std::vector population_result; population_result.reserve(sample_tumor_par.size()); // get results (a vector of future Eigen arrays) std::ranges::transform(sample_tumor_par vector of future Eigen arrays) std::vector population_result; population_result.reserve(sample_tumor_par.size()); // get results (a vector of future Eigen arrays) std::ranges::transform(sample_tumor_par vector of future Eigen arrays) std::vector population_result; population_result.reserve(sample_tumor_par.size()); // get results (a vector of future Eigen arrays) std::ranges::transform(sample_tumor_par

11 12 *Copied from P2996 35Array of structs to struct of arrays struct point { float x; float y; }; using points = struct_of_arrays; // equivalent to: // struct points { // std::arrayarrays template struct struct_of_arrays_impl; consteval auto make_struct_of_arrays(std::meta::info type, std::meta::define_class( substitute(^struct_of_arrays_impl, {type, N}), new_members); } template using struct_of_arrays = [: make_struct_of_arrays(^T, ^N) :]; 1 2 3 4 5 6 7 8 9 10 11

is sync primitive - spins rather than sleeping Flat map is associative container Stores items in arrays - keys in sorted array - giving O(logn) access Speaker notesJOB SYSTEM • Schedules jobs on many evenly ▶ Long functions should be split into jobs Independent sections processed in parallel Large arrays using parallel algorithms Speaker notesBUILDING LARGE WORLDS • Create read-only world cache for 24GridCell stores sorted arrays of GridNodes • Array of owned nodes - cell contains bounding box origin • Array of intersecting nodes - cell intersects bounding box Different arrays are useful for different

structure with arrays of homogeneous data types (going from what in databases is called a row-representation to a columnar representation) also called “scattering” or achieving structures of arrays • …word entities: • Very simple binding of components to entities: The index into the global structure of arrays that contains the componentsHandles • Underneath they are an integer (the index of the components) accessor to the virtual table *INDEX* as a member • Each subtype has its own “state” struct of arrays or memory arena for their specific members (components) • Subtype handles inherit from the supertype

data and expose results in a memory layout that python understands ○ Write DSL to return numpy arrays for analysis Enter python bindingsWe’ve seen this before import binary_reader_bpy as mod librariesnumpy Optimized numeric operations for python using typed arrays● Abstraction over contiguous and multi-dimensional arrays exposed in Python (ndarray) ● Contains lots of optimized functions operates on numpy arrays? ● Your library needs to interop with some ML library: ○ return a std::vector and convert it to numpy array in python? ● Why not populate numpy arrays in your library?

user's • Example: lookup tables • Everyone uses lookup tables • Arrays are maximally space-efficient, very time-efficient • Sorted arrays: binary_search, lower_bound, equal_range • Now you can use constexpr • span is like a pair of T* and size_t, but better • spans are implicitly constructible from arrays and vectors • Avoids ownership confusion, accidental pointer arithmetic • span ➡ span (bit_cast, rotating/counting, power-of-2) • Several PRs by barcharcraz • make_shared() For Arrays • GH-309 by AdamBucior, tested by Weheineman • midpoint(), lerp() • Paolo Torres and BillyONeal

user's • Example: lookup tables • Everyone uses lookup tables • Arrays are maximally space-efficient, very time-efficient • Sorted arrays: binary_search, lower_bound, equal_range • Now you can use constexpr • span is like a pair of T* and size_t, but better • spans are implicitly constructible from arrays and vectors • Avoids ownership confusion, accidental pointer arithmetic • span  span; (bit_cast, rotating/counting, power-of-2) • Several PRs by barcharcraz • make_shared() For Arrays • GH-309 by AdamBucior, tested by Weheineman • midpoint(), lerp() • Paolo Torres and BillyONeal

equal to another. ■ Goal: ■ Code, as shown on the right. A Simple comparison of two equal­sized arrays. B Compare all fields of an array to a specific value. 1 struct MACAddress 2 { 3 unsigned char (macA.value, 2)); 14 static_assert(!Compare(macB.value, 2)); 15 16 B Compare two equally sized arrays 17 static_assert(!Compare(macA.value, macB.value)); 18 static_assert(Compare(macA.value, macC.value)); namespace details::array_compare */ 22 23 template B Compare two equally sized arrays 24 constexpr bool Compare(const T (&a)[N], const T (&b)[N]) 25 { 26 return details::array_compare::Compare(a