vectorized code The talk: ▪ share experience with vectorization using std::simd ▪ introduce the SIMD_ACCESS library 3 Olaf Krzikalla, DLR SP, 2024-09-17Background: Vectorization Nowadays, all your CPUs can loop body No vector type information in x, y, z or i →Explicit load and store operations Explicit access provided by std::simd ▪ Requires rearranging and rewriting the loop body ▪ Different scalar version rearrangement) necessary ▪ Same code for scalar and vector versions SIMD_ACCESS(z, i) = SIMD_ACCESS(x, i) + SIMD_ACCESS(y, i);SIMD_ACCESS Library https://github.com/dlr-sp/simdize Supplements std::experimental::simd

here we will focus on buffers The buffers memory management model separates the storage and access of data and automatically handles data dependencies USM provides a more explicit memory management buffer manage data across the host and one or more devices An accessor represents a request to access the data on a particular device with a particular set of properties #include to create dependencies to the data managed by their respective buffers They are also used to access the data once on the device #include int main(int argc, char *argv[]) { std::vector

container's iterators understand (and abstract) that container's internal structure • Iterators • Provide access to container elements through well-defined interfaces with strict guarantees • Algorithms • Employ Classic STL has five iterator categories • Output • Input • Forward • Bidirectional • Random-access • Arranged in a hierarchy of requirements • Not public inheritance 23 Output Input Forward Operation Access element *p Access member of element p->mem Compare for equality of postion p == q, p != q Move forward by 1 ++p, p++ Move backward by 1 --p, p-- Make a copy (assign) q = p Access arbitrary

SFINAE Overload Resolution (Viable) Candidate Functions Ranking Ambiguous Function Calls Access Labels Function Template Specializations Virtual Dispatch Deleting FunctionsContent 4 Overview SFINAE Overload Resolution (Viable) Candidate Functions Ranking Ambiguous Function Calls Access Labels Function Template Specializations Virtual Dispatch Deleting FunctionsDisclaimer 5 This Resolution Access Labels Access Labels Check if the best match is accessible from the given call site.Overview 11 Template Argument Deduction Name Lookup Overload Resolution Access Labels Function

values - It has a shape (number of rows, cols) - It has a size (number of stored values) - Can access individual locations - Can iterate over values 69 Matrix[DISTRIBUTION STATEMENT A] This material values - It has a shape (number of rows, cols) - It has a size (number of stored values) - Can access individual locations - Can iterate over values 70 Matrix[DISTRIBUTION STATEMENT A] This material values - It has a shape (number of rows, cols) - It has a size (number of stored values) - Can access individual locations - Can iterate over values 71 Matrix[DISTRIBUTION STATEMENT A] This material

offset_policy :: pointer 9 offset(pointer p, ptrdiff_t i) const noexcept; 10 11 constexpr reference access(pointer p, ptrdiff_t i) const noexcept; 12 13 constexpr pointer decay(pointer p) const noexcept; noexcept; 11 12 // [mdspan.basic.mapping], basic_mdspan mapping domain multidimensional index to access codomain element 13 constexpr reference operator []( index_type) const noexcept; 14 template access 17 public: 18 constexpr indexed_extender operator =( std:: size_t size) const noexcept { 19 return

one-sided access - Maps well to data structure ops CPU NIC CPU NIC DRAM DRAMAdvantages of PGAS - Asynchronous - RDMA operations executed by NIC - Allows irregular, one-sided access - Maps well one-sided access - Maps well to data structure ops CPU NIC CPU NIC DRAM DRAMAdvantages of PGAS - Asynchronous - RDMA operations executed by NIC - Allows irregular, one-sided access - Maps well one-sided access - Maps well to data structure ops CPU NIC CPU NIC DRAM DRAMAdvantages of PGAS - Asynchronous - RDMA operations executed by NIC - Allows irregular, one-sided access - Maps well

size() - Random access: access any element at random in constant time - Contiguous: a contiguous block of memory begin end begin end Contiguous Range begin end Range Random Access Range begin memory that lives on another node - They still fulfill the concept random_access_iterator - We can use them to build a random_access_range containers/views using namespace std::ranges; remote_ptr memory that lives on another node - They still fulfill the concept random_access_iterator - We can use them to build a random_access_range containers/views using namespace std::ranges; remote_ptr

types ● roboticists and their lack of beer ● external state memory accessor ● improving memory access using asynchronicity 4 ● motivations for std::mdspan ● review std::mdspan declarationgithub types ● roboticists and their lack of beer ● external state memory accessor ● improving memory access using asynchronicity 5 ● motivations for std::mdspan ● review std::mdspan declarationgithub types ● roboticists and their lack of beer ● external state memory accessor ● improving memory access using asynchronicity 6 ● motivations for std::mdspan ● review std::mdspan declarationgithub