https://github.com/kokkos/mdspangithub.com/griswaldbrooks/spanny2 goals ● deeper understanding of std::mdspan layouts and accessors ● how to write custom layouts and accessors ● dispel common misconceptions accessor ● improving memory access using asynchronicity 4 ● motivations for std::mdspan ● review std::mdspan declarationgithub.com/griswaldbrooks/spanny2 how ● layouts and their requirements ● accessor ● improving memory access using asynchronicity 5 ● motivations for std::mdspan ● review std::mdspan declarationgithub.com/griswaldbrooks/spanny2 how ● layouts and their requirements ●

Is std::mdspan a Zero-overhead Abstraction? Oleksandr Bacherikov Snap IncWhat is std::mdspan? It’s a view over a multi-dimensional array. It’s designed primarily to be used as a function parameter 19Designing the std::mdspan version Looks like we’re in luck, because there’s a pending proposal P1673 “A free function linear algebra interface based on the BLAS” which is based on std::mdspan. 20What do we add_rank_2( std::experimental::mdspan, Layout_x, Accessor_x> x, std::experimental::mdspan

0 码力 | 75 页 | 1.04 MB | 5 月前

3

Extents Going beyond Conclusion A standardization effort On mdspan P0332: Relaxed Incomplete Multidimensional Array Type Declaration P0009: mdspan: A Non-Owning Multidimensional Array Reference P1684: mdarray: mdarray: An Owning Multidimensional Array Analog of mdspan On linear algebra in general P1417: Historical lessons for C++ linear algebra library standardization P1166: What do we need from a linear Standardization Design EDSL Extents Going beyond Conclusion The mdspan proposal: synopsis Proposed mdspan synopsis as of P0009R10 1 namespace std { 2 // [mdspan.extents], class template extents 3 template

0 码力 | 62 页 | 1.38 MB | 5 月前

3

simulations Existing C++ features: parallel algorithms (C++17), mdspan (C++23), submdspan (C++26) Proposals in flight: mdarray, mdspan padded layouts, SIMD Classic domain of “numerical linear algebra std::linalg Existing C++ features: parallel algorithms (C++17), mdspan (C++23), submdspan (C++26) Proposals in flight: mdarray, mdspan padded layouts, SIMD Classic domain of “numerical linear algebra mdspan A{A_raw_pointer, m, k}; mdspan B{B_raw_pointer, n, k}; mdspan C{C_raw_pointer, m, n}; matrix_product(par_unseq, scaled(alpha, A), transposed(B), scaled(beta, C), C);13 mdspan represents

Formatting Improvements  Standard Library Modules  std::flat_(multi)map / std::flat_(multi)set  std::mdspan  std::generator  basic_string(_view)::contains()  Construct string(_view) From nullptr  Formatting Improvements  Standard Library Modules  std::flat_(multi)map / std::flat_(multi)set  std::mdspan  std::generator  basic_string(_view)::contains()  Construct string(_view) From nullptr  Formatting Improvements  Standard Library Modules  std::flat_(multi)map / std::flat_(multi)set  std::mdspan  std::generator  basic_string(_view)::contains()  Construct string(_view) From nullptr 

const noexcept; //... };} ● std::atomic_ref vs. sycl::atomic_refMultidimensional subscript & mdspan C++23 91Multidimensional subscript operator[] ● C++ 20 deprecated a comma expression within square to C arrays 92mdspan 93 template> struct mdspan { //... /* ... */ } //... }; ● std::mdspan ○ Non-owning multi-dimensional array view ○ First standard library type to use multidimensional subscript operator[]mdspan 94 template

0 码力 | 114 页 | 7.94 MB | 5 月前

3

included comparisons with Eigen A Little History: Open-Source Libraries Daniel Hanson CppCon 2024 7• mdspan (P0009), C++23 • A mutating multidimensional array reference (view)… • …of a container whose elements C++26 • A free function linear algebra interface based on the BLAS • Matrix represented by a 2-D mdspan • Default BLAS with major compilers/library vendors A Little History: C++23 and C++26 Daniel 𝝎𝑻(𝚺𝝎): // mdspan views of data (can use CTAD): stdex::mdspan md_cov_mtx{cov_mtx_data.data(), stdex::extents{n, n}}; stdex::mdspan md_wgts{omega.data(), stdex::extents{n}}; stdex::mdspan md_inner_m

wrapping 34P0122R0 ARRAY_VIEW: BOUNDS-SAFE VIEWS FOR SEQUENCES OF OBJECTS ▪ Precursor to span & mdspan ▪ “Any failure to meet array_view’s bounds-safety constraints will result in a call to std::terminate() 2015 35P0122R0 ARRAY_VIEW: BOUNDS-SAFE VIEWS FOR SEQUENCES OF OBJECTS ▪ Precursor to span & mdspan ▪ “Any failure to meet array_view’s bounds-safety constraints will result in a call to std::terminate() questions: how do different profiles in the same code base work together? Profiles 41LIBRARYMDSPAN ▪ mdspan is a non-owning multidimensional array view < ElementType, Extents, LayoutPolicy

ranges utilities general math support constexpr all the things concurrency simd linear algebra and mdspan 6Language and Library 7Debugging 8pass a string to static_assert build-time diagnostics P2741 . } 1 2 3 4 5 6 75Library - Linear Algebra and mdspan Adding Blas based Linear Algebra P2630 submdspan P3029 New CTAD for std::span and std::mdspan with integral constants 76elementwise vector sums A_vec(N*M); std::vector x_vec(M); std::array y_vec(N); mdspan A(A_vec.data(), N, M); mdspan x(x_vec.data(), M); mdspan y(y_vec.data(), N); // y = 0.5 * y + 2 * A * x linalg::matrix_vect

such as range algorithms, their parallel execution strategies, and multi-dimensional arrays through ‘mdspan’, it is interesting to explore what C++ can offer for lower level performance critical operations using ‘mdspan’ for true two-dimensional index- ing. We use range algorithms in the sense that we range over a cartesian product of iota views, giving us the two-dimensional index into the mdspan. Next