Software Methods for Product Virtualization ## Motivation: The Origin of the Talk ## The task: ■ Vectorization of time-consuming parts of a complex, existing code ■ no revolutionary approach, please ## The variables ☑ syntactically equalize scalar and vectorized code ## The talk: ■ share experience with vectorization using std::�md ■ introduce the SIMD_ACCESS library Nowadays, all your CPUs can compute four times Matthias Kretz' Cppcon talk about std::simd: https://youtu.be/LAJ_hywLtMA ## Background: Vectorization void add_array(double* x, double* y, double* z, int size) { for (int i = 0; i < size;

in the first step. The second step assigns a unique index to the words. This process is called vectorization. An embedding table with a row for each word is initialized in the third step. Finally, in the would have to pay the cost of a very large embedding table. ## Step 2: Dataset Preparation & Vectorization Once the window size $ (2k+1) $ is chosen, the dataset is preprocessed (lowercase, strip punctuation assigns them an index. This process of mapping free form inputs to integer sequences is known as vectorization, as introduced in the Word2Vec subsection. The TextVectorization layer takes the vocabulary size

is inefficient lots of time in interpreter ⇒ use native libraries low core utilization ⇒ use vectorization / MT / MP  inefficient lots of time in interpreter ⇒ use native libraries low core utilization ⇒ use vectorization / MT / MP no usage of GPU ![Image](/uploads/documents/a/7/a/f/a7afe81a9040067d79574ca170e1dde5/p55_2 is inefficient lots of time in interpreter ⇒ use native libraries low core utilization ⇒ use vectorization / MT / MP no usage of GPU ⇒ use GPU-optimized libraries ![Image](/uploads/documents/a/7/a

proposal P2897) • Expresses byte overalignment by using std::assume_aligned (C++20) • Useful for vectorization or special hardware ## #include constexpr size_t byte_alignment = 4 * sizeof(double); • run on thread(s) other than the calling thread, & • “interleave” operations (e.g., for vectorization) • 4 Standard policies, & vendors can add more • "On-ramp" to vendor-specific performance

math library in the vectorization phase. - SVE optimization: Significantly improves program running performance for ARM-based machines that support SVE instructions. - SLP vectorization optimization: Analyzes

(n--) { *dst++ = *src++; } } return dst; } Scalar base ISA only, no vectorization 0000000000001270 <_z13memcpy_scalarPcPKcm>: 1270: 48 89 f8 mov rax, rdi 1273: 48 85 d2 Performance characterization in terms of latency and throughput. • SIMD as a specific instantiation of vectorization approach. • Practical and Everyday => relies on the compiler. But we are not there yet. Not

757 46 Arrays 796 46.1 Constructors and Types 796 46.2 Basic functions 809 46.3 Broadcast and vectorization 814 46.4 Indexing and assignment 819 46.5 Views (SubArrays and other view types) 827 46.6 Concatenation used with broadcasting, as . |>, to provide a useful combination of the chaining/piping and dot vectorization syntax (described below). julia> [“a”, “list”, “of”, “strings”] .|> [uppercase, reverse, titlecase array via f(A). This kind of syntax is convenient for data processing, but in other languages vectorization is also often required for performance: if loops are slow, the “vectorized” version of a function

feedback-directed optimization (FDO), software and hardware collaboration, memory optimization, and automatic vectorization. GCC for openEuler is compatible with a wide range of hardware platforms such as Kunpeng, Phytium microarchitecture optimization to implement intelligent memory allocation, memory optimization, and automatic vectorization. Besides that, GCC for openEuler incorporates industry-leading FDO technologies to implement automatic

46 Arrays 860 46.1 Constructors and Types 860 46.2 Basic functions 874 46.3 Broadcast and vectorization 880 46.4 Indexing and assignment 885 46.5 Views (SubArrays and other view types) 893 46.6 Concatenation with broadcasting, as . |>, to provide a useful combination of the chaining/piping and dot vectorization syntax (described below). julia> ["a", "list", "of", "strings"] array via f(A). This kind of syntax is convenient for data processing, but in other languages vectorization is also often required for performance: if loops are slow, the "vectorized" version of

46 Arrays 860 46.1 Constructors and Types 860 46.2 Basic functions 874 46.3 Broadcast and vectorization 880 46.4 Indexing and assignment 885 46.5 Views (SubArrays and other view types) 893 46.6 Concatenation with broadcasting, as . |>, to provide a useful combination of the chaining/piping and dot vectorization syntax (described below). julia> ["a", "list", "of", "strings"] array via f(A). This kind of syntax is convenient for data processing, but in other languages vectorization is also often required for performance: if loops are slow, the "vectorized" version of