18a91bc7bSHarrietAkot //===- SparseTensorUtils.cpp - Sparse Tensor Utils for MLIR execution -----===// 28a91bc7bSHarrietAkot // 38a91bc7bSHarrietAkot // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 48a91bc7bSHarrietAkot // See https://llvm.org/LICENSE.txt for license information. 58a91bc7bSHarrietAkot // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 68a91bc7bSHarrietAkot // 78a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 88a91bc7bSHarrietAkot // 98a91bc7bSHarrietAkot // This file implements a light-weight runtime support library that is useful 108a91bc7bSHarrietAkot // for sparse tensor manipulations. The functionality provided in this library 118a91bc7bSHarrietAkot // is meant to simplify benchmarking, testing, and debugging MLIR code that 128a91bc7bSHarrietAkot // operates on sparse tensors. The provided functionality is **not** part 138a91bc7bSHarrietAkot // of core MLIR, however. 148a91bc7bSHarrietAkot // 158a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 168a91bc7bSHarrietAkot 17845561ecSwren romano #include "mlir/ExecutionEngine/SparseTensorUtils.h" 188a91bc7bSHarrietAkot #include "mlir/ExecutionEngine/CRunnerUtils.h" 198a91bc7bSHarrietAkot 208a91bc7bSHarrietAkot #ifdef MLIR_CRUNNERUTILS_DEFINE_FUNCTIONS 218a91bc7bSHarrietAkot 228a91bc7bSHarrietAkot #include <algorithm> 238a91bc7bSHarrietAkot #include <cassert> 24736c1b66SAart Bik #include <complex> 258a91bc7bSHarrietAkot #include <cctype> 268a91bc7bSHarrietAkot #include <cinttypes> 278a91bc7bSHarrietAkot #include <cstdio> 288a91bc7bSHarrietAkot #include <cstdlib> 298a91bc7bSHarrietAkot #include <cstring> 30efa15f41SAart Bik #include <fstream> 31753fe330Swren romano #include <functional> 32efa15f41SAart Bik #include <iostream> 334d0a18d0Swren romano #include <limits> 348a91bc7bSHarrietAkot #include <numeric> 358a91bc7bSHarrietAkot #include <vector> 368a91bc7bSHarrietAkot 37736c1b66SAart Bik using complex64 = std::complex<double>; 38736c1b66SAart Bik using complex32 = std::complex<float>; 39736c1b66SAart Bik 408a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 418a91bc7bSHarrietAkot // 428a91bc7bSHarrietAkot // Internal support for storing and reading sparse tensors. 438a91bc7bSHarrietAkot // 448a91bc7bSHarrietAkot // The following memory-resident sparse storage schemes are supported: 458a91bc7bSHarrietAkot // 468a91bc7bSHarrietAkot // (a) A coordinate scheme for temporarily storing and lexicographically 478a91bc7bSHarrietAkot // sorting a sparse tensor by index (SparseTensorCOO). 488a91bc7bSHarrietAkot // 498a91bc7bSHarrietAkot // (b) A "one-size-fits-all" sparse tensor storage scheme defined by 508a91bc7bSHarrietAkot // per-dimension sparse/dense annnotations together with a dimension 518a91bc7bSHarrietAkot // ordering used by MLIR compiler-generated code (SparseTensorStorage). 528a91bc7bSHarrietAkot // 538a91bc7bSHarrietAkot // The following external formats are supported: 548a91bc7bSHarrietAkot // 558a91bc7bSHarrietAkot // (1) Matrix Market Exchange (MME): *.mtx 568a91bc7bSHarrietAkot // https://math.nist.gov/MatrixMarket/formats.html 578a91bc7bSHarrietAkot // 588a91bc7bSHarrietAkot // (2) Formidable Repository of Open Sparse Tensors and Tools (FROSTT): *.tns 598a91bc7bSHarrietAkot // http://frostt.io/tensors/file-formats.html 608a91bc7bSHarrietAkot // 618a91bc7bSHarrietAkot // Two public APIs are supported: 628a91bc7bSHarrietAkot // 638a91bc7bSHarrietAkot // (I) Methods operating on MLIR buffers (memrefs) to interact with sparse 648a91bc7bSHarrietAkot // tensors. These methods should be used exclusively by MLIR 658a91bc7bSHarrietAkot // compiler-generated code. 668a91bc7bSHarrietAkot // 678a91bc7bSHarrietAkot // (II) Methods that accept C-style data structures to interact with sparse 688a91bc7bSHarrietAkot // tensors. These methods can be used by any external runtime that wants 698a91bc7bSHarrietAkot // to interact with MLIR compiler-generated code. 708a91bc7bSHarrietAkot // 718a91bc7bSHarrietAkot // In both cases (I) and (II), the SparseTensorStorage format is externally 728a91bc7bSHarrietAkot // only visible as an opaque pointer. 738a91bc7bSHarrietAkot // 748a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 758a91bc7bSHarrietAkot 768a91bc7bSHarrietAkot namespace { 778a91bc7bSHarrietAkot 7803fe15ceSAart Bik static constexpr int kColWidth = 1025; 7903fe15ceSAart Bik 8072ec2f76Swren romano /// A version of `operator*` on `uint64_t` which checks for overflows. 8172ec2f76Swren romano static inline uint64_t checkedMul(uint64_t lhs, uint64_t rhs) { 8272ec2f76Swren romano assert((lhs == 0 || rhs <= std::numeric_limits<uint64_t>::max() / lhs) && 8372ec2f76Swren romano "Integer overflow"); 8472ec2f76Swren romano return lhs * rhs; 8572ec2f76Swren romano } 8672ec2f76Swren romano 878cb33240Swren romano // TODO: adjust this so it can be used by `openSparseTensorCOO` too. 888cb33240Swren romano // That version doesn't have the permutation, and the `sizes` are 898cb33240Swren romano // a pointer/C-array rather than `std::vector`. 908cb33240Swren romano // 918cb33240Swren romano /// Asserts that the `sizes` (in target-order) under the `perm` (mapping 928cb33240Swren romano /// semantic-order to target-order) are a refinement of the desired `shape` 938cb33240Swren romano /// (in semantic-order). 948cb33240Swren romano /// 958cb33240Swren romano /// Precondition: `perm` and `shape` must be valid for `rank`. 968cb33240Swren romano static inline void 978cb33240Swren romano assertPermutedSizesMatchShape(const std::vector<uint64_t> &sizes, uint64_t rank, 988cb33240Swren romano const uint64_t *perm, const uint64_t *shape) { 998cb33240Swren romano assert(perm && shape); 1008cb33240Swren romano assert(rank == sizes.size() && "Rank mismatch"); 1018cb33240Swren romano for (uint64_t r = 0; r < rank; r++) 1028cb33240Swren romano assert((shape[r] == 0 || shape[r] == sizes[perm[r]]) && 1038cb33240Swren romano "Dimension size mismatch"); 1048cb33240Swren romano } 1058cb33240Swren romano 1068a91bc7bSHarrietAkot /// A sparse tensor element in coordinate scheme (value and indices). 1078a91bc7bSHarrietAkot /// For example, a rank-1 vector element would look like 1088a91bc7bSHarrietAkot /// ({i}, a[i]) 1098a91bc7bSHarrietAkot /// and a rank-5 tensor element like 1108a91bc7bSHarrietAkot /// ({i,j,k,l,m}, a[i,j,k,l,m]) 111ccd047cbSAart Bik /// We use pointer to a shared index pool rather than e.g. a direct 112ccd047cbSAart Bik /// vector since that (1) reduces the per-element memory footprint, and 113ccd047cbSAart Bik /// (2) centralizes the memory reservation and (re)allocation to one place. 1148a91bc7bSHarrietAkot template <typename V> 11576944420Swren romano struct Element final { 116ccd047cbSAart Bik Element(uint64_t *ind, V val) : indices(ind), value(val){}; 117ccd047cbSAart Bik uint64_t *indices; // pointer into shared index pool 1188a91bc7bSHarrietAkot V value; 1198a91bc7bSHarrietAkot }; 1208a91bc7bSHarrietAkot 121753fe330Swren romano /// The type of callback functions which receive an element. We avoid 122753fe330Swren romano /// packaging the coordinates and value together as an `Element` object 123753fe330Swren romano /// because this helps keep code somewhat cleaner. 124753fe330Swren romano template <typename V> 125753fe330Swren romano using ElementConsumer = 126753fe330Swren romano const std::function<void(const std::vector<uint64_t> &, V)> &; 127753fe330Swren romano 1288a91bc7bSHarrietAkot /// A memory-resident sparse tensor in coordinate scheme (collection of 1298a91bc7bSHarrietAkot /// elements). This data structure is used to read a sparse tensor from 1308a91bc7bSHarrietAkot /// any external format into memory and sort the elements lexicographically 1318a91bc7bSHarrietAkot /// by indices before passing it back to the client (most packed storage 1328a91bc7bSHarrietAkot /// formats require the elements to appear in lexicographic index order). 1338a91bc7bSHarrietAkot template <typename V> 13476944420Swren romano struct SparseTensorCOO final { 1358a91bc7bSHarrietAkot public: 1368a91bc7bSHarrietAkot SparseTensorCOO(const std::vector<uint64_t> &szs, uint64_t capacity) 137db6796dfSMehdi Amini : sizes(szs) { 138ccd047cbSAart Bik if (capacity) { 1398a91bc7bSHarrietAkot elements.reserve(capacity); 140ccd047cbSAart Bik indices.reserve(capacity * getRank()); 1418a91bc7bSHarrietAkot } 142ccd047cbSAart Bik } 143ccd047cbSAart Bik 1448a91bc7bSHarrietAkot /// Adds element as indices and value. 1458a91bc7bSHarrietAkot void add(const std::vector<uint64_t> &ind, V val) { 1468a91bc7bSHarrietAkot assert(!iteratorLocked && "Attempt to add() after startIterator()"); 147ccd047cbSAart Bik uint64_t *base = indices.data(); 148ccd047cbSAart Bik uint64_t size = indices.size(); 1498a91bc7bSHarrietAkot uint64_t rank = getRank(); 1508a91bc7bSHarrietAkot assert(rank == ind.size()); 151ccd047cbSAart Bik for (uint64_t r = 0; r < rank; r++) { 1528a91bc7bSHarrietAkot assert(ind[r] < sizes[r]); // within bounds 153ccd047cbSAart Bik indices.push_back(ind[r]); 1548a91bc7bSHarrietAkot } 155ccd047cbSAart Bik // This base only changes if indices were reallocated. In that case, we 156ccd047cbSAart Bik // need to correct all previous pointers into the vector. Note that this 157ccd047cbSAart Bik // only happens if we did not set the initial capacity right, and then only 158ccd047cbSAart Bik // for every internal vector reallocation (which with the doubling rule 159ccd047cbSAart Bik // should only incur an amortized linear overhead). 160298d2fa1SMehdi Amini uint64_t *newBase = indices.data(); 161298d2fa1SMehdi Amini if (newBase != base) { 162ccd047cbSAart Bik for (uint64_t i = 0, n = elements.size(); i < n; i++) 163298d2fa1SMehdi Amini elements[i].indices = newBase + (elements[i].indices - base); 164298d2fa1SMehdi Amini base = newBase; 165ccd047cbSAart Bik } 166ccd047cbSAart Bik // Add element as (pointer into shared index pool, value) pair. 167ccd047cbSAart Bik elements.emplace_back(base + size, val); 168ccd047cbSAart Bik } 169ccd047cbSAart Bik 1708a91bc7bSHarrietAkot /// Sorts elements lexicographically by index. 1718a91bc7bSHarrietAkot void sort() { 1728a91bc7bSHarrietAkot assert(!iteratorLocked && "Attempt to sort() after startIterator()"); 173cf358253Swren romano // TODO: we may want to cache an `isSorted` bit, to avoid 174cf358253Swren romano // unnecessary/redundant sorting. 175ccd047cbSAart Bik std::sort(elements.begin(), elements.end(), 176ccd047cbSAart Bik [this](const Element<V> &e1, const Element<V> &e2) { 177ccd047cbSAart Bik uint64_t rank = getRank(); 178ccd047cbSAart Bik for (uint64_t r = 0; r < rank; r++) { 179ccd047cbSAart Bik if (e1.indices[r] == e2.indices[r]) 180ccd047cbSAart Bik continue; 181ccd047cbSAart Bik return e1.indices[r] < e2.indices[r]; 1828a91bc7bSHarrietAkot } 183ccd047cbSAart Bik return false; 184ccd047cbSAart Bik }); 185ccd047cbSAart Bik } 186ccd047cbSAart Bik 1878a91bc7bSHarrietAkot /// Returns rank. 1888a91bc7bSHarrietAkot uint64_t getRank() const { return sizes.size(); } 189ccd047cbSAart Bik 1908a91bc7bSHarrietAkot /// Getter for sizes array. 1918a91bc7bSHarrietAkot const std::vector<uint64_t> &getSizes() const { return sizes; } 192ccd047cbSAart Bik 1938a91bc7bSHarrietAkot /// Getter for elements array. 1948a91bc7bSHarrietAkot const std::vector<Element<V>> &getElements() const { return elements; } 1958a91bc7bSHarrietAkot 1968a91bc7bSHarrietAkot /// Switch into iterator mode. 1978a91bc7bSHarrietAkot void startIterator() { 1988a91bc7bSHarrietAkot iteratorLocked = true; 1998a91bc7bSHarrietAkot iteratorPos = 0; 2008a91bc7bSHarrietAkot } 201ccd047cbSAart Bik 2028a91bc7bSHarrietAkot /// Get the next element. 2038a91bc7bSHarrietAkot const Element<V> *getNext() { 2048a91bc7bSHarrietAkot assert(iteratorLocked && "Attempt to getNext() before startIterator()"); 2058a91bc7bSHarrietAkot if (iteratorPos < elements.size()) 2068a91bc7bSHarrietAkot return &(elements[iteratorPos++]); 2078a91bc7bSHarrietAkot iteratorLocked = false; 2088a91bc7bSHarrietAkot return nullptr; 2098a91bc7bSHarrietAkot } 2108a91bc7bSHarrietAkot 2118a91bc7bSHarrietAkot /// Factory method. Permutes the original dimensions according to 2128a91bc7bSHarrietAkot /// the given ordering and expects subsequent add() calls to honor 2138a91bc7bSHarrietAkot /// that same ordering for the given indices. The result is a 2148a91bc7bSHarrietAkot /// fully permuted coordinate scheme. 2158d8b566fSwren romano /// 2168d8b566fSwren romano /// Precondition: `sizes` and `perm` must be valid for `rank`. 2178a91bc7bSHarrietAkot static SparseTensorCOO<V> *newSparseTensorCOO(uint64_t rank, 2188a91bc7bSHarrietAkot const uint64_t *sizes, 2198a91bc7bSHarrietAkot const uint64_t *perm, 2208a91bc7bSHarrietAkot uint64_t capacity = 0) { 2218a91bc7bSHarrietAkot std::vector<uint64_t> permsz(rank); 222d83a7068Swren romano for (uint64_t r = 0; r < rank; r++) { 223d83a7068Swren romano assert(sizes[r] > 0 && "Dimension size zero has trivial storage"); 2248a91bc7bSHarrietAkot permsz[perm[r]] = sizes[r]; 225d83a7068Swren romano } 2268a91bc7bSHarrietAkot return new SparseTensorCOO<V>(permsz, capacity); 2278a91bc7bSHarrietAkot } 2288a91bc7bSHarrietAkot 2298a91bc7bSHarrietAkot private: 2308a91bc7bSHarrietAkot const std::vector<uint64_t> sizes; // per-dimension sizes 231ccd047cbSAart Bik std::vector<Element<V>> elements; // all COO elements 232ccd047cbSAart Bik std::vector<uint64_t> indices; // shared index pool 233db6796dfSMehdi Amini bool iteratorLocked = false; 234db6796dfSMehdi Amini unsigned iteratorPos = 0; 2358a91bc7bSHarrietAkot }; 2368a91bc7bSHarrietAkot 2371313f5d3Swren romano // See <https://en.wikipedia.org/wiki/X_Macro> 2381313f5d3Swren romano // 2391313f5d3Swren romano // `FOREVERY_SIMPLEX_V` only specifies the non-complex `V` types, because 2401313f5d3Swren romano // the ABI for complex types has compiler/architecture dependent complexities 2411313f5d3Swren romano // we need to work around. Namely, when a function takes a parameter of 2421313f5d3Swren romano // C/C++ type `complex32` (per se), then there is additional padding that 2431313f5d3Swren romano // causes it not to match the LLVM type `!llvm.struct<(f32, f32)>`. This 2441313f5d3Swren romano // only happens with the `complex32` type itself, not with pointers/arrays 2451313f5d3Swren romano // of complex values. So far `complex64` doesn't exhibit this ABI 2461313f5d3Swren romano // incompatibility, but we exclude it anyways just to be safe. 2471313f5d3Swren romano #define FOREVERY_SIMPLEX_V(DO) \ 2481313f5d3Swren romano DO(F64, double) \ 2491313f5d3Swren romano DO(F32, float) \ 2501313f5d3Swren romano DO(I64, int64_t) \ 2511313f5d3Swren romano DO(I32, int32_t) \ 2521313f5d3Swren romano DO(I16, int16_t) \ 2531313f5d3Swren romano DO(I8, int8_t) 2541313f5d3Swren romano 2551313f5d3Swren romano #define FOREVERY_V(DO) \ 2561313f5d3Swren romano FOREVERY_SIMPLEX_V(DO) \ 2571313f5d3Swren romano DO(C64, complex64) \ 2581313f5d3Swren romano DO(C32, complex32) 2591313f5d3Swren romano 2608cb33240Swren romano // Forward. 2618cb33240Swren romano template <typename V> 2628cb33240Swren romano class SparseTensorEnumeratorBase; 2638cb33240Swren romano 2648d8b566fSwren romano /// Abstract base class for `SparseTensorStorage<P,I,V>`. This class 2658d8b566fSwren romano /// takes responsibility for all the `<P,I,V>`-independent aspects 2668d8b566fSwren romano /// of the tensor (e.g., shape, sparsity, permutation). In addition, 2678d8b566fSwren romano /// we use function overloading to implement "partial" method 2688d8b566fSwren romano /// specialization, which the C-API relies on to catch type errors 2698d8b566fSwren romano /// arising from our use of opaque pointers. 2708a91bc7bSHarrietAkot class SparseTensorStorageBase { 2718a91bc7bSHarrietAkot public: 2728d8b566fSwren romano /// Constructs a new storage object. The `perm` maps the tensor's 2738d8b566fSwren romano /// semantic-ordering of dimensions to this object's storage-order. 2748d8b566fSwren romano /// The `szs` and `sparsity` arrays are already in storage-order. 2758d8b566fSwren romano /// 2768d8b566fSwren romano /// Precondition: `perm` and `sparsity` must be valid for `szs.size()`. 2778d8b566fSwren romano SparseTensorStorageBase(const std::vector<uint64_t> &szs, 2788d8b566fSwren romano const uint64_t *perm, const DimLevelType *sparsity) 2798d8b566fSwren romano : dimSizes(szs), rev(getRank()), 2808d8b566fSwren romano dimTypes(sparsity, sparsity + getRank()) { 281753fe330Swren romano assert(perm && sparsity); 2828d8b566fSwren romano const uint64_t rank = getRank(); 2838d8b566fSwren romano // Validate parameters. 2848d8b566fSwren romano assert(rank > 0 && "Trivial shape is unsupported"); 2858d8b566fSwren romano for (uint64_t r = 0; r < rank; r++) { 2868d8b566fSwren romano assert(dimSizes[r] > 0 && "Dimension size zero has trivial storage"); 2878d8b566fSwren romano assert((dimTypes[r] == DimLevelType::kDense || 2888d8b566fSwren romano dimTypes[r] == DimLevelType::kCompressed) && 2898d8b566fSwren romano "Unsupported DimLevelType"); 2908d8b566fSwren romano } 2918d8b566fSwren romano // Construct the "reverse" (i.e., inverse) permutation. 2928d8b566fSwren romano for (uint64_t r = 0; r < rank; r++) 2938d8b566fSwren romano rev[perm[r]] = r; 2948d8b566fSwren romano } 2958d8b566fSwren romano 2968d8b566fSwren romano virtual ~SparseTensorStorageBase() = default; 2978d8b566fSwren romano 2988d8b566fSwren romano /// Get the rank of the tensor. 2998d8b566fSwren romano uint64_t getRank() const { return dimSizes.size(); } 3008d8b566fSwren romano 3018d8b566fSwren romano /// Getter for the dimension-sizes array, in storage-order. 3028d8b566fSwren romano const std::vector<uint64_t> &getDimSizes() const { return dimSizes; } 3038d8b566fSwren romano 3048d8b566fSwren romano /// Safely lookup the size of the given (storage-order) dimension. 3058d8b566fSwren romano uint64_t getDimSize(uint64_t d) const { 3068d8b566fSwren romano assert(d < getRank()); 3078d8b566fSwren romano return dimSizes[d]; 3088d8b566fSwren romano } 3098d8b566fSwren romano 3108d8b566fSwren romano /// Getter for the "reverse" permutation, which maps this object's 3118d8b566fSwren romano /// storage-order to the tensor's semantic-order. 3128d8b566fSwren romano const std::vector<uint64_t> &getRev() const { return rev; } 3138d8b566fSwren romano 3148d8b566fSwren romano /// Getter for the dimension-types array, in storage-order. 3158d8b566fSwren romano const std::vector<DimLevelType> &getDimTypes() const { return dimTypes; } 3168d8b566fSwren romano 3178d8b566fSwren romano /// Safely check if the (storage-order) dimension uses compressed storage. 3188d8b566fSwren romano bool isCompressedDim(uint64_t d) const { 3198d8b566fSwren romano assert(d < getRank()); 3208d8b566fSwren romano return (dimTypes[d] == DimLevelType::kCompressed); 3218d8b566fSwren romano } 3228a91bc7bSHarrietAkot 3238cb33240Swren romano /// Allocate a new enumerator. 3241313f5d3Swren romano #define DECL_NEWENUMERATOR(VNAME, V) \ 3251313f5d3Swren romano virtual void newEnumerator(SparseTensorEnumeratorBase<V> **, uint64_t, \ 3261313f5d3Swren romano const uint64_t *) const { \ 3271313f5d3Swren romano fatal("newEnumerator" #VNAME); \ 3288cb33240Swren romano } 3291313f5d3Swren romano FOREVERY_V(DECL_NEWENUMERATOR) 3301313f5d3Swren romano #undef DECL_NEWENUMERATOR 3318cb33240Swren romano 3324f2ec7f9SAart Bik /// Overhead storage. 3338a91bc7bSHarrietAkot virtual void getPointers(std::vector<uint64_t> **, uint64_t) { fatal("p64"); } 3348a91bc7bSHarrietAkot virtual void getPointers(std::vector<uint32_t> **, uint64_t) { fatal("p32"); } 3358a91bc7bSHarrietAkot virtual void getPointers(std::vector<uint16_t> **, uint64_t) { fatal("p16"); } 3368a91bc7bSHarrietAkot virtual void getPointers(std::vector<uint8_t> **, uint64_t) { fatal("p8"); } 3378a91bc7bSHarrietAkot virtual void getIndices(std::vector<uint64_t> **, uint64_t) { fatal("i64"); } 3388a91bc7bSHarrietAkot virtual void getIndices(std::vector<uint32_t> **, uint64_t) { fatal("i32"); } 3398a91bc7bSHarrietAkot virtual void getIndices(std::vector<uint16_t> **, uint64_t) { fatal("i16"); } 3408a91bc7bSHarrietAkot virtual void getIndices(std::vector<uint8_t> **, uint64_t) { fatal("i8"); } 3418a91bc7bSHarrietAkot 3424f2ec7f9SAart Bik /// Primary storage. 3431313f5d3Swren romano #define DECL_GETVALUES(VNAME, V) \ 3441313f5d3Swren romano virtual void getValues(std::vector<V> **) { fatal("getValues" #VNAME); } 3451313f5d3Swren romano FOREVERY_V(DECL_GETVALUES) 3461313f5d3Swren romano #undef DECL_GETVALUES 3478a91bc7bSHarrietAkot 3484f2ec7f9SAart Bik /// Element-wise insertion in lexicographic index order. 3491313f5d3Swren romano #define DECL_LEXINSERT(VNAME, V) \ 3501313f5d3Swren romano virtual void lexInsert(const uint64_t *, V) { fatal("lexInsert" #VNAME); } 3511313f5d3Swren romano FOREVERY_V(DECL_LEXINSERT) 3521313f5d3Swren romano #undef DECL_LEXINSERT 3534f2ec7f9SAart Bik 3544f2ec7f9SAart Bik /// Expanded insertion. 3551313f5d3Swren romano #define DECL_EXPINSERT(VNAME, V) \ 3561313f5d3Swren romano virtual void expInsert(uint64_t *, V *, bool *, uint64_t *, uint64_t) { \ 3571313f5d3Swren romano fatal("expInsert" #VNAME); \ 3584f2ec7f9SAart Bik } 3591313f5d3Swren romano FOREVERY_V(DECL_EXPINSERT) 3601313f5d3Swren romano #undef DECL_EXPINSERT 3614f2ec7f9SAart Bik 3624f2ec7f9SAart Bik /// Finishes insertion. 363f66e5769SAart Bik virtual void endInsert() = 0; 364f66e5769SAart Bik 365753fe330Swren romano protected: 366753fe330Swren romano // Since this class is virtual, we must disallow public copying in 367753fe330Swren romano // order to avoid "slicing". Since this class has data members, 368753fe330Swren romano // that means making copying protected. 369753fe330Swren romano // <https://github.com/isocpp/CppCoreGuidelines/blob/master/CppCoreGuidelines.md#Rc-copy-virtual> 370753fe330Swren romano SparseTensorStorageBase(const SparseTensorStorageBase &) = default; 371753fe330Swren romano // Copy-assignment would be implicitly deleted (because `dimSizes` 372753fe330Swren romano // is const), so we explicitly delete it for clarity. 373753fe330Swren romano SparseTensorStorageBase &operator=(const SparseTensorStorageBase &) = delete; 374753fe330Swren romano 3758a91bc7bSHarrietAkot private: 37646bdacaaSwren romano static void fatal(const char *tp) { 3778a91bc7bSHarrietAkot fprintf(stderr, "unsupported %s\n", tp); 3788a91bc7bSHarrietAkot exit(1); 3798a91bc7bSHarrietAkot } 3808d8b566fSwren romano 3818d8b566fSwren romano const std::vector<uint64_t> dimSizes; 3828d8b566fSwren romano std::vector<uint64_t> rev; 3838d8b566fSwren romano const std::vector<DimLevelType> dimTypes; 3848a91bc7bSHarrietAkot }; 3858a91bc7bSHarrietAkot 386753fe330Swren romano // Forward. 387753fe330Swren romano template <typename P, typename I, typename V> 388753fe330Swren romano class SparseTensorEnumerator; 389753fe330Swren romano 3908a91bc7bSHarrietAkot /// A memory-resident sparse tensor using a storage scheme based on 3918a91bc7bSHarrietAkot /// per-dimension sparse/dense annotations. This data structure provides a 3928a91bc7bSHarrietAkot /// bufferized form of a sparse tensor type. In contrast to generating setup 3938a91bc7bSHarrietAkot /// methods for each differently annotated sparse tensor, this method provides 3948a91bc7bSHarrietAkot /// a convenient "one-size-fits-all" solution that simply takes an input tensor 3958a91bc7bSHarrietAkot /// and annotations to implement all required setup in a general manner. 3968a91bc7bSHarrietAkot template <typename P, typename I, typename V> 39776944420Swren romano class SparseTensorStorage final : public SparseTensorStorageBase { 3988cb33240Swren romano /// Private constructor to share code between the other constructors. 3998cb33240Swren romano /// Beware that the object is not necessarily guaranteed to be in a 4008cb33240Swren romano /// valid state after this constructor alone; e.g., `isCompressedDim(d)` 4018cb33240Swren romano /// doesn't entail `!(pointers[d].empty())`. 4028cb33240Swren romano /// 4038cb33240Swren romano /// Precondition: `perm` and `sparsity` must be valid for `szs.size()`. 4048cb33240Swren romano SparseTensorStorage(const std::vector<uint64_t> &szs, const uint64_t *perm, 4058cb33240Swren romano const DimLevelType *sparsity) 4068cb33240Swren romano : SparseTensorStorageBase(szs, perm, sparsity), pointers(getRank()), 4078cb33240Swren romano indices(getRank()), idx(getRank()) {} 4088cb33240Swren romano 4098a91bc7bSHarrietAkot public: 4108a91bc7bSHarrietAkot /// Constructs a sparse tensor storage scheme with the given dimensions, 4118a91bc7bSHarrietAkot /// permutation, and per-dimension dense/sparse annotations, using 4128a91bc7bSHarrietAkot /// the coordinate scheme tensor for the initial contents if provided. 4138d8b566fSwren romano /// 4148d8b566fSwren romano /// Precondition: `perm` and `sparsity` must be valid for `szs.size()`. 4158a91bc7bSHarrietAkot SparseTensorStorage(const std::vector<uint64_t> &szs, const uint64_t *perm, 4168cb33240Swren romano const DimLevelType *sparsity, SparseTensorCOO<V> *coo) 4178cb33240Swren romano : SparseTensorStorage(szs, perm, sparsity) { 4188a91bc7bSHarrietAkot // Provide hints on capacity of pointers and indices. 419175b9af4SAart Bik // TODO: needs much fine-tuning based on actual sparsity; currently 420175b9af4SAart Bik // we reserve pointer/index space based on all previous dense 421175b9af4SAart Bik // dimensions, which works well up to first sparse dim; but 422175b9af4SAart Bik // we should really use nnz and dense/sparse distribution. 423f66e5769SAart Bik bool allDense = true; 424f66e5769SAart Bik uint64_t sz = 1; 4258d8b566fSwren romano for (uint64_t r = 0, rank = getRank(); r < rank; r++) { 4268d8b566fSwren romano if (isCompressedDim(r)) { 4278d8b566fSwren romano // TODO: Take a parameter between 1 and `sizes[r]`, and multiply 4288d8b566fSwren romano // `sz` by that before reserving. (For now we just use 1.) 429f66e5769SAart Bik pointers[r].reserve(sz + 1); 4308d8b566fSwren romano pointers[r].push_back(0); 431f66e5769SAart Bik indices[r].reserve(sz); 432f66e5769SAart Bik sz = 1; 433f66e5769SAart Bik allDense = false; 4348d8b566fSwren romano } else { // Dense dimension. 4358d8b566fSwren romano sz = checkedMul(sz, getDimSizes()[r]); 4368a91bc7bSHarrietAkot } 4378a91bc7bSHarrietAkot } 4388a91bc7bSHarrietAkot // Then assign contents from coordinate scheme tensor if provided. 4398d8b566fSwren romano if (coo) { 4404d0a18d0Swren romano // Ensure both preconditions of `fromCOO`. 4418d8b566fSwren romano assert(coo->getSizes() == getDimSizes() && "Tensor size mismatch"); 4428d8b566fSwren romano coo->sort(); 4434d0a18d0Swren romano // Now actually insert the `elements`. 4448d8b566fSwren romano const std::vector<Element<V>> &elements = coo->getElements(); 445ceda1ae9Swren romano uint64_t nnz = elements.size(); 4468a91bc7bSHarrietAkot values.reserve(nnz); 447ceda1ae9Swren romano fromCOO(elements, 0, nnz, 0); 4481ce77b56SAart Bik } else if (allDense) { 449f66e5769SAart Bik values.resize(sz, 0); 4508a91bc7bSHarrietAkot } 4518a91bc7bSHarrietAkot } 4528a91bc7bSHarrietAkot 4538cb33240Swren romano /// Constructs a sparse tensor storage scheme with the given dimensions, 4548cb33240Swren romano /// permutation, and per-dimension dense/sparse annotations, using 4558cb33240Swren romano /// the given sparse tensor for the initial contents. 4568cb33240Swren romano /// 4578cb33240Swren romano /// Preconditions: 4588cb33240Swren romano /// * `perm` and `sparsity` must be valid for `szs.size()`. 4598cb33240Swren romano /// * The `tensor` must have the same value type `V`. 4608cb33240Swren romano SparseTensorStorage(const std::vector<uint64_t> &szs, const uint64_t *perm, 4618cb33240Swren romano const DimLevelType *sparsity, 4628cb33240Swren romano const SparseTensorStorageBase &tensor); 4638cb33240Swren romano 46476944420Swren romano ~SparseTensorStorage() final override = default; 4658a91bc7bSHarrietAkot 466f66e5769SAart Bik /// Partially specialize these getter methods based on template types. 46776944420Swren romano void getPointers(std::vector<P> **out, uint64_t d) final override { 4688a91bc7bSHarrietAkot assert(d < getRank()); 4698a91bc7bSHarrietAkot *out = &pointers[d]; 4708a91bc7bSHarrietAkot } 47176944420Swren romano void getIndices(std::vector<I> **out, uint64_t d) final override { 4728a91bc7bSHarrietAkot assert(d < getRank()); 4738a91bc7bSHarrietAkot *out = &indices[d]; 4748a91bc7bSHarrietAkot } 47576944420Swren romano void getValues(std::vector<V> **out) final override { *out = &values; } 4768a91bc7bSHarrietAkot 47703fe15ceSAart Bik /// Partially specialize lexicographical insertions based on template types. 47876944420Swren romano void lexInsert(const uint64_t *cursor, V val) final override { 4791ce77b56SAart Bik // First, wrap up pending insertion path. 4801ce77b56SAart Bik uint64_t diff = 0; 4811ce77b56SAart Bik uint64_t top = 0; 4821ce77b56SAart Bik if (!values.empty()) { 4831ce77b56SAart Bik diff = lexDiff(cursor); 4841ce77b56SAart Bik endPath(diff + 1); 4851ce77b56SAart Bik top = idx[diff] + 1; 4861ce77b56SAart Bik } 4871ce77b56SAart Bik // Then continue with insertion path. 4881ce77b56SAart Bik insPath(cursor, diff, top, val); 489f66e5769SAart Bik } 490f66e5769SAart Bik 4914f2ec7f9SAart Bik /// Partially specialize expanded insertions based on template types. 4924f2ec7f9SAart Bik /// Note that this method resets the values/filled-switch array back 4934f2ec7f9SAart Bik /// to all-zero/false while only iterating over the nonzero elements. 4944f2ec7f9SAart Bik void expInsert(uint64_t *cursor, V *values, bool *filled, uint64_t *added, 49576944420Swren romano uint64_t count) final override { 4964f2ec7f9SAart Bik if (count == 0) 4974f2ec7f9SAart Bik return; 4984f2ec7f9SAart Bik // Sort. 4994f2ec7f9SAart Bik std::sort(added, added + count); 5004f2ec7f9SAart Bik // Restore insertion path for first insert. 5013bf2ba3bSwren romano const uint64_t lastDim = getRank() - 1; 5024f2ec7f9SAart Bik uint64_t index = added[0]; 5033bf2ba3bSwren romano cursor[lastDim] = index; 5044f2ec7f9SAart Bik lexInsert(cursor, values[index]); 5054f2ec7f9SAart Bik assert(filled[index]); 5064f2ec7f9SAart Bik values[index] = 0; 5074f2ec7f9SAart Bik filled[index] = false; 5084f2ec7f9SAart Bik // Subsequent insertions are quick. 5094f2ec7f9SAart Bik for (uint64_t i = 1; i < count; i++) { 5104f2ec7f9SAart Bik assert(index < added[i] && "non-lexicographic insertion"); 5114f2ec7f9SAart Bik index = added[i]; 5123bf2ba3bSwren romano cursor[lastDim] = index; 5133bf2ba3bSwren romano insPath(cursor, lastDim, added[i - 1] + 1, values[index]); 5144f2ec7f9SAart Bik assert(filled[index]); 5153bf2ba3bSwren romano values[index] = 0; 5164f2ec7f9SAart Bik filled[index] = false; 5174f2ec7f9SAart Bik } 5184f2ec7f9SAart Bik } 5194f2ec7f9SAart Bik 520f66e5769SAart Bik /// Finalizes lexicographic insertions. 52176944420Swren romano void endInsert() final override { 5221ce77b56SAart Bik if (values.empty()) 52372ec2f76Swren romano finalizeSegment(0); 5241ce77b56SAart Bik else 5251ce77b56SAart Bik endPath(0); 5261ce77b56SAart Bik } 527f66e5769SAart Bik 5288cb33240Swren romano void newEnumerator(SparseTensorEnumeratorBase<V> **out, uint64_t rank, 52976944420Swren romano const uint64_t *perm) const final override { 5308cb33240Swren romano *out = new SparseTensorEnumerator<P, I, V>(*this, rank, perm); 5318cb33240Swren romano } 5328cb33240Swren romano 5338a91bc7bSHarrietAkot /// Returns this sparse tensor storage scheme as a new memory-resident 5348a91bc7bSHarrietAkot /// sparse tensor in coordinate scheme with the given dimension order. 5358d8b566fSwren romano /// 5368d8b566fSwren romano /// Precondition: `perm` must be valid for `getRank()`. 537753fe330Swren romano SparseTensorCOO<V> *toCOO(const uint64_t *perm) const { 5388cb33240Swren romano SparseTensorEnumeratorBase<V> *enumerator; 5398cb33240Swren romano newEnumerator(&enumerator, getRank(), perm); 540753fe330Swren romano SparseTensorCOO<V> *coo = 5418cb33240Swren romano new SparseTensorCOO<V>(enumerator->permutedSizes(), values.size()); 5428cb33240Swren romano enumerator->forallElements([&coo](const std::vector<uint64_t> &ind, V val) { 543753fe330Swren romano coo->add(ind, val); 544753fe330Swren romano }); 5458d8b566fSwren romano // TODO: This assertion assumes there are no stored zeros, 5468d8b566fSwren romano // or if there are then that we don't filter them out. 5478d8b566fSwren romano // Cf., <https://github.com/llvm/llvm-project/issues/54179> 5488d8b566fSwren romano assert(coo->getElements().size() == values.size()); 5498cb33240Swren romano delete enumerator; 5508d8b566fSwren romano return coo; 5518a91bc7bSHarrietAkot } 5528a91bc7bSHarrietAkot 5538a91bc7bSHarrietAkot /// Factory method. Constructs a sparse tensor storage scheme with the given 5548a91bc7bSHarrietAkot /// dimensions, permutation, and per-dimension dense/sparse annotations, 5558a91bc7bSHarrietAkot /// using the coordinate scheme tensor for the initial contents if provided. 5568a91bc7bSHarrietAkot /// In the latter case, the coordinate scheme must respect the same 5578a91bc7bSHarrietAkot /// permutation as is desired for the new sparse tensor storage. 5588d8b566fSwren romano /// 5598d8b566fSwren romano /// Precondition: `shape`, `perm`, and `sparsity` must be valid for `rank`. 5608a91bc7bSHarrietAkot static SparseTensorStorage<P, I, V> * 561d83a7068Swren romano newSparseTensor(uint64_t rank, const uint64_t *shape, const uint64_t *perm, 5628d8b566fSwren romano const DimLevelType *sparsity, SparseTensorCOO<V> *coo) { 5638a91bc7bSHarrietAkot SparseTensorStorage<P, I, V> *n = nullptr; 5648d8b566fSwren romano if (coo) { 5658d8b566fSwren romano const auto &coosz = coo->getSizes(); 5668cb33240Swren romano assertPermutedSizesMatchShape(coosz, rank, perm, shape); 5678d8b566fSwren romano n = new SparseTensorStorage<P, I, V>(coosz, perm, sparsity, coo); 5688a91bc7bSHarrietAkot } else { 5698a91bc7bSHarrietAkot std::vector<uint64_t> permsz(rank); 570d83a7068Swren romano for (uint64_t r = 0; r < rank; r++) { 571d83a7068Swren romano assert(shape[r] > 0 && "Dimension size zero has trivial storage"); 572d83a7068Swren romano permsz[perm[r]] = shape[r]; 573d83a7068Swren romano } 5748cb33240Swren romano // We pass the null `coo` to ensure we select the intended constructor. 5758cb33240Swren romano n = new SparseTensorStorage<P, I, V>(permsz, perm, sparsity, coo); 5768a91bc7bSHarrietAkot } 5778a91bc7bSHarrietAkot return n; 5788a91bc7bSHarrietAkot } 5798a91bc7bSHarrietAkot 5808cb33240Swren romano /// Factory method. Constructs a sparse tensor storage scheme with 5818cb33240Swren romano /// the given dimensions, permutation, and per-dimension dense/sparse 5828cb33240Swren romano /// annotations, using the sparse tensor for the initial contents. 5838cb33240Swren romano /// 5848cb33240Swren romano /// Preconditions: 5858cb33240Swren romano /// * `shape`, `perm`, and `sparsity` must be valid for `rank`. 5868cb33240Swren romano /// * The `tensor` must have the same value type `V`. 5878cb33240Swren romano static SparseTensorStorage<P, I, V> * 5888cb33240Swren romano newSparseTensor(uint64_t rank, const uint64_t *shape, const uint64_t *perm, 5898cb33240Swren romano const DimLevelType *sparsity, 5908cb33240Swren romano const SparseTensorStorageBase *source) { 5918cb33240Swren romano assert(source && "Got nullptr for source"); 5928cb33240Swren romano SparseTensorEnumeratorBase<V> *enumerator; 5938cb33240Swren romano source->newEnumerator(&enumerator, rank, perm); 5948cb33240Swren romano const auto &permsz = enumerator->permutedSizes(); 5958cb33240Swren romano assertPermutedSizesMatchShape(permsz, rank, perm, shape); 5968cb33240Swren romano auto *tensor = 5978cb33240Swren romano new SparseTensorStorage<P, I, V>(permsz, perm, sparsity, *source); 5988cb33240Swren romano delete enumerator; 5998cb33240Swren romano return tensor; 6008cb33240Swren romano } 6018cb33240Swren romano 6028a91bc7bSHarrietAkot private: 60372ec2f76Swren romano /// Appends an arbitrary new position to `pointers[d]`. This method 60472ec2f76Swren romano /// checks that `pos` is representable in the `P` type; however, it 60572ec2f76Swren romano /// does not check that `pos` is semantically valid (i.e., larger than 60672ec2f76Swren romano /// the previous position and smaller than `indices[d].capacity()`). 6078d8b566fSwren romano void appendPointer(uint64_t d, uint64_t pos, uint64_t count = 1) { 60872ec2f76Swren romano assert(isCompressedDim(d)); 60972ec2f76Swren romano assert(pos <= std::numeric_limits<P>::max() && 6104d0a18d0Swren romano "Pointer value is too large for the P-type"); 61172ec2f76Swren romano pointers[d].insert(pointers[d].end(), count, static_cast<P>(pos)); 6124d0a18d0Swren romano } 6134d0a18d0Swren romano 61472ec2f76Swren romano /// Appends index `i` to dimension `d`, in the semantically general 61572ec2f76Swren romano /// sense. For non-dense dimensions, that means appending to the 61672ec2f76Swren romano /// `indices[d]` array, checking that `i` is representable in the `I` 61772ec2f76Swren romano /// type; however, we do not verify other semantic requirements (e.g., 61872ec2f76Swren romano /// that `i` is in bounds for `sizes[d]`, and not previously occurring 61972ec2f76Swren romano /// in the same segment). For dense dimensions, this method instead 62072ec2f76Swren romano /// appends the appropriate number of zeros to the `values` array, 62172ec2f76Swren romano /// where `full` is the number of "entries" already written to `values` 62272ec2f76Swren romano /// for this segment (aka one after the highest index previously appended). 62372ec2f76Swren romano void appendIndex(uint64_t d, uint64_t full, uint64_t i) { 62472ec2f76Swren romano if (isCompressedDim(d)) { 6254d0a18d0Swren romano assert(i <= std::numeric_limits<I>::max() && 6264d0a18d0Swren romano "Index value is too large for the I-type"); 62772ec2f76Swren romano indices[d].push_back(static_cast<I>(i)); 62872ec2f76Swren romano } else { // Dense dimension. 62972ec2f76Swren romano assert(i >= full && "Index was already filled"); 63072ec2f76Swren romano if (i == full) 63172ec2f76Swren romano return; // Short-circuit, since it'll be a nop. 63272ec2f76Swren romano if (d + 1 == getRank()) 63372ec2f76Swren romano values.insert(values.end(), i - full, 0); 63472ec2f76Swren romano else 63572ec2f76Swren romano finalizeSegment(d + 1, 0, i - full); 63672ec2f76Swren romano } 6374d0a18d0Swren romano } 6384d0a18d0Swren romano 6398cb33240Swren romano /// Writes the given coordinate to `indices[d][pos]`. This method 6408cb33240Swren romano /// checks that `i` is representable in the `I` type; however, it 6418cb33240Swren romano /// does not check that `i` is semantically valid (i.e., in bounds 6428cb33240Swren romano /// for `sizes[d]` and not elsewhere occurring in the same segment). 6438cb33240Swren romano void writeIndex(uint64_t d, uint64_t pos, uint64_t i) { 6448cb33240Swren romano assert(isCompressedDim(d)); 6458cb33240Swren romano // Subscript assignment to `std::vector` requires that the `pos`-th 6468cb33240Swren romano // entry has been initialized; thus we must be sure to check `size()` 6478cb33240Swren romano // here, instead of `capacity()` as would be ideal. 6488cb33240Swren romano assert(pos < indices[d].size() && "Index position is out of bounds"); 6498cb33240Swren romano assert(i <= std::numeric_limits<I>::max() && 6508cb33240Swren romano "Index value is too large for the I-type"); 6518cb33240Swren romano indices[d][pos] = static_cast<I>(i); 6528cb33240Swren romano } 6538cb33240Swren romano 6548cb33240Swren romano /// Computes the assembled-size associated with the `d`-th dimension, 6558cb33240Swren romano /// given the assembled-size associated with the `(d-1)`-th dimension. 6568cb33240Swren romano /// "Assembled-sizes" correspond to the (nominal) sizes of overhead 6578cb33240Swren romano /// storage, as opposed to "dimension-sizes" which are the cardinality 6588cb33240Swren romano /// of coordinates for that dimension. 6598cb33240Swren romano /// 6608cb33240Swren romano /// Precondition: the `pointers[d]` array must be fully initialized 6618cb33240Swren romano /// before calling this method. 6628cb33240Swren romano uint64_t assembledSize(uint64_t parentSz, uint64_t d) const { 6638cb33240Swren romano if (isCompressedDim(d)) 6648cb33240Swren romano return pointers[d][parentSz]; 6658cb33240Swren romano // else if dense: 6668cb33240Swren romano return parentSz * getDimSizes()[d]; 6678cb33240Swren romano } 6688cb33240Swren romano 6698a91bc7bSHarrietAkot /// Initializes sparse tensor storage scheme from a memory-resident sparse 6708a91bc7bSHarrietAkot /// tensor in coordinate scheme. This method prepares the pointers and 6718a91bc7bSHarrietAkot /// indices arrays under the given per-dimension dense/sparse annotations. 6724d0a18d0Swren romano /// 6734d0a18d0Swren romano /// Preconditions: 6744d0a18d0Swren romano /// (1) the `elements` must be lexicographically sorted. 6754d0a18d0Swren romano /// (2) the indices of every element are valid for `sizes` (equal rank 6764d0a18d0Swren romano /// and pointwise less-than). 677ceda1ae9Swren romano void fromCOO(const std::vector<Element<V>> &elements, uint64_t lo, 678ceda1ae9Swren romano uint64_t hi, uint64_t d) { 679753fe330Swren romano uint64_t rank = getRank(); 680753fe330Swren romano assert(d <= rank && hi <= elements.size()); 6818a91bc7bSHarrietAkot // Once dimensions are exhausted, insert the numerical values. 682753fe330Swren romano if (d == rank) { 683c4017f9dSwren romano assert(lo < hi); 6841ce77b56SAart Bik values.push_back(elements[lo].value); 6858a91bc7bSHarrietAkot return; 6868a91bc7bSHarrietAkot } 6878a91bc7bSHarrietAkot // Visit all elements in this interval. 6888a91bc7bSHarrietAkot uint64_t full = 0; 689c4017f9dSwren romano while (lo < hi) { // If `hi` is unchanged, then `lo < elements.size()`. 6908a91bc7bSHarrietAkot // Find segment in interval with same index elements in this dimension. 691f66e5769SAart Bik uint64_t i = elements[lo].indices[d]; 6928a91bc7bSHarrietAkot uint64_t seg = lo + 1; 693f66e5769SAart Bik while (seg < hi && elements[seg].indices[d] == i) 6948a91bc7bSHarrietAkot seg++; 6958a91bc7bSHarrietAkot // Handle segment in interval for sparse or dense dimension. 69672ec2f76Swren romano appendIndex(d, full, i); 69772ec2f76Swren romano full = i + 1; 698ceda1ae9Swren romano fromCOO(elements, lo, seg, d + 1); 6998a91bc7bSHarrietAkot // And move on to next segment in interval. 7008a91bc7bSHarrietAkot lo = seg; 7018a91bc7bSHarrietAkot } 7028a91bc7bSHarrietAkot // Finalize the sparse pointer structure at this dimension. 70372ec2f76Swren romano finalizeSegment(d, full); 7048a91bc7bSHarrietAkot } 7058a91bc7bSHarrietAkot 70672ec2f76Swren romano /// Finalize the sparse pointer structure at this dimension. 70772ec2f76Swren romano void finalizeSegment(uint64_t d, uint64_t full = 0, uint64_t count = 1) { 70872ec2f76Swren romano if (count == 0) 70972ec2f76Swren romano return; // Short-circuit, since it'll be a nop. 71072ec2f76Swren romano if (isCompressedDim(d)) { 71172ec2f76Swren romano appendPointer(d, indices[d].size(), count); 71272ec2f76Swren romano } else { // Dense dimension. 7138d8b566fSwren romano const uint64_t sz = getDimSizes()[d]; 71472ec2f76Swren romano assert(sz >= full && "Segment is overfull"); 7158d8b566fSwren romano count = checkedMul(count, sz - full); 71672ec2f76Swren romano // For dense storage we must enumerate all the remaining coordinates 71772ec2f76Swren romano // in this dimension (i.e., coordinates after the last non-zero 71872ec2f76Swren romano // element), and either fill in their zero values or else recurse 71972ec2f76Swren romano // to finalize some deeper dimension. 72072ec2f76Swren romano if (d + 1 == getRank()) 72172ec2f76Swren romano values.insert(values.end(), count, 0); 72272ec2f76Swren romano else 72372ec2f76Swren romano finalizeSegment(d + 1, 0, count); 7241ce77b56SAart Bik } 7251ce77b56SAart Bik } 7261ce77b56SAart Bik 7271ce77b56SAart Bik /// Wraps up a single insertion path, inner to outer. 7281ce77b56SAart Bik void endPath(uint64_t diff) { 7291ce77b56SAart Bik uint64_t rank = getRank(); 7301ce77b56SAart Bik assert(diff <= rank); 7311ce77b56SAart Bik for (uint64_t i = 0; i < rank - diff; i++) { 73272ec2f76Swren romano const uint64_t d = rank - i - 1; 73372ec2f76Swren romano finalizeSegment(d, idx[d] + 1); 7341ce77b56SAart Bik } 7351ce77b56SAart Bik } 7361ce77b56SAart Bik 7371ce77b56SAart Bik /// Continues a single insertion path, outer to inner. 738c03fd1e6Swren romano void insPath(const uint64_t *cursor, uint64_t diff, uint64_t top, V val) { 7391ce77b56SAart Bik uint64_t rank = getRank(); 7401ce77b56SAart Bik assert(diff < rank); 7411ce77b56SAart Bik for (uint64_t d = diff; d < rank; d++) { 7421ce77b56SAart Bik uint64_t i = cursor[d]; 74372ec2f76Swren romano appendIndex(d, top, i); 7441ce77b56SAart Bik top = 0; 7451ce77b56SAart Bik idx[d] = i; 7461ce77b56SAart Bik } 7471ce77b56SAart Bik values.push_back(val); 7481ce77b56SAart Bik } 7491ce77b56SAart Bik 7501ce77b56SAart Bik /// Finds the lexicographic differing dimension. 75146bdacaaSwren romano uint64_t lexDiff(const uint64_t *cursor) const { 7521ce77b56SAart Bik for (uint64_t r = 0, rank = getRank(); r < rank; r++) 7531ce77b56SAart Bik if (cursor[r] > idx[r]) 7541ce77b56SAart Bik return r; 7551ce77b56SAart Bik else 7561ce77b56SAart Bik assert(cursor[r] == idx[r] && "non-lexicographic insertion"); 7571ce77b56SAart Bik assert(0 && "duplication insertion"); 7581ce77b56SAart Bik return -1u; 7591ce77b56SAart Bik } 7601ce77b56SAart Bik 761753fe330Swren romano // Allow `SparseTensorEnumerator` to access the data-members (to avoid 762753fe330Swren romano // the cost of virtual-function dispatch in inner loops), without 763753fe330Swren romano // making them public to other client code. 764753fe330Swren romano friend class SparseTensorEnumerator<P, I, V>; 765753fe330Swren romano 7668a91bc7bSHarrietAkot std::vector<std::vector<P>> pointers; 7678a91bc7bSHarrietAkot std::vector<std::vector<I>> indices; 7688a91bc7bSHarrietAkot std::vector<V> values; 7698d8b566fSwren romano std::vector<uint64_t> idx; // index cursor for lexicographic insertion. 7708a91bc7bSHarrietAkot }; 7718a91bc7bSHarrietAkot 772753fe330Swren romano /// A (higher-order) function object for enumerating the elements of some 773753fe330Swren romano /// `SparseTensorStorage` under a permutation. That is, the `forallElements` 774753fe330Swren romano /// method encapsulates the loop-nest for enumerating the elements of 775753fe330Swren romano /// the source tensor (in whatever order is best for the source tensor), 776753fe330Swren romano /// and applies a permutation to the coordinates/indices before handing 777753fe330Swren romano /// each element to the callback. A single enumerator object can be 778753fe330Swren romano /// freely reused for several calls to `forallElements`, just so long 779753fe330Swren romano /// as each call is sequential with respect to one another. 780753fe330Swren romano /// 781753fe330Swren romano /// N.B., this class stores a reference to the `SparseTensorStorageBase` 782753fe330Swren romano /// passed to the constructor; thus, objects of this class must not 783753fe330Swren romano /// outlive the sparse tensor they depend on. 784753fe330Swren romano /// 785753fe330Swren romano /// Design Note: The reason we define this class instead of simply using 786753fe330Swren romano /// `SparseTensorEnumerator<P,I,V>` is because we need to hide/generalize 787753fe330Swren romano /// the `<P,I>` template parameters from MLIR client code (to simplify the 788753fe330Swren romano /// type parameters used for direct sparse-to-sparse conversion). And the 789753fe330Swren romano /// reason we define the `SparseTensorEnumerator<P,I,V>` subclasses rather 790753fe330Swren romano /// than simply using this class, is to avoid the cost of virtual-method 791753fe330Swren romano /// dispatch within the loop-nest. 792753fe330Swren romano template <typename V> 793753fe330Swren romano class SparseTensorEnumeratorBase { 794753fe330Swren romano public: 795753fe330Swren romano /// Constructs an enumerator with the given permutation for mapping 796753fe330Swren romano /// the semantic-ordering of dimensions to the desired target-ordering. 797753fe330Swren romano /// 798753fe330Swren romano /// Preconditions: 799753fe330Swren romano /// * the `tensor` must have the same `V` value type. 800753fe330Swren romano /// * `perm` must be valid for `rank`. 801753fe330Swren romano SparseTensorEnumeratorBase(const SparseTensorStorageBase &tensor, 802753fe330Swren romano uint64_t rank, const uint64_t *perm) 803753fe330Swren romano : src(tensor), permsz(src.getRev().size()), reord(getRank()), 804753fe330Swren romano cursor(getRank()) { 805753fe330Swren romano assert(perm && "Received nullptr for permutation"); 806753fe330Swren romano assert(rank == getRank() && "Permutation rank mismatch"); 807753fe330Swren romano const auto &rev = src.getRev(); // source stg-order -> semantic-order 808753fe330Swren romano const auto &sizes = src.getDimSizes(); // in source storage-order 809753fe330Swren romano for (uint64_t s = 0; s < rank; s++) { // `s` source storage-order 810753fe330Swren romano uint64_t t = perm[rev[s]]; // `t` target-order 811753fe330Swren romano reord[s] = t; 812753fe330Swren romano permsz[t] = sizes[s]; 813753fe330Swren romano } 814753fe330Swren romano } 815753fe330Swren romano 816753fe330Swren romano virtual ~SparseTensorEnumeratorBase() = default; 817753fe330Swren romano 818753fe330Swren romano // We disallow copying to help avoid leaking the `src` reference. 819753fe330Swren romano // (In addition to avoiding the problem of slicing.) 820753fe330Swren romano SparseTensorEnumeratorBase(const SparseTensorEnumeratorBase &) = delete; 821753fe330Swren romano SparseTensorEnumeratorBase & 822753fe330Swren romano operator=(const SparseTensorEnumeratorBase &) = delete; 823753fe330Swren romano 824753fe330Swren romano /// Returns the source/target tensor's rank. (The source-rank and 825753fe330Swren romano /// target-rank are always equal since we only support permutations. 826753fe330Swren romano /// Though once we add support for other dimension mappings, this 827753fe330Swren romano /// method will have to be split in two.) 828753fe330Swren romano uint64_t getRank() const { return permsz.size(); } 829753fe330Swren romano 830753fe330Swren romano /// Returns the target tensor's dimension sizes. 831753fe330Swren romano const std::vector<uint64_t> &permutedSizes() const { return permsz; } 832753fe330Swren romano 833753fe330Swren romano /// Enumerates all elements of the source tensor, permutes their 834753fe330Swren romano /// indices, and passes the permuted element to the callback. 835753fe330Swren romano /// The callback must not store the cursor reference directly, 836753fe330Swren romano /// since this function reuses the storage. Instead, the callback 837753fe330Swren romano /// must copy it if they want to keep it. 838753fe330Swren romano virtual void forallElements(ElementConsumer<V> yield) = 0; 839753fe330Swren romano 840753fe330Swren romano protected: 841753fe330Swren romano const SparseTensorStorageBase &src; 842753fe330Swren romano std::vector<uint64_t> permsz; // in target order. 843753fe330Swren romano std::vector<uint64_t> reord; // source storage-order -> target order. 844753fe330Swren romano std::vector<uint64_t> cursor; // in target order. 845753fe330Swren romano }; 846753fe330Swren romano 847753fe330Swren romano template <typename P, typename I, typename V> 848753fe330Swren romano class SparseTensorEnumerator final : public SparseTensorEnumeratorBase<V> { 849753fe330Swren romano using Base = SparseTensorEnumeratorBase<V>; 850753fe330Swren romano 851753fe330Swren romano public: 852753fe330Swren romano /// Constructs an enumerator with the given permutation for mapping 853753fe330Swren romano /// the semantic-ordering of dimensions to the desired target-ordering. 854753fe330Swren romano /// 855753fe330Swren romano /// Precondition: `perm` must be valid for `rank`. 856753fe330Swren romano SparseTensorEnumerator(const SparseTensorStorage<P, I, V> &tensor, 857753fe330Swren romano uint64_t rank, const uint64_t *perm) 858753fe330Swren romano : Base(tensor, rank, perm) {} 859753fe330Swren romano 860753fe330Swren romano ~SparseTensorEnumerator() final override = default; 861753fe330Swren romano 862753fe330Swren romano void forallElements(ElementConsumer<V> yield) final override { 863753fe330Swren romano forallElements(yield, 0, 0); 864753fe330Swren romano } 865753fe330Swren romano 866753fe330Swren romano private: 867753fe330Swren romano /// The recursive component of the public `forallElements`. 868753fe330Swren romano void forallElements(ElementConsumer<V> yield, uint64_t parentPos, 869753fe330Swren romano uint64_t d) { 870753fe330Swren romano // Recover the `<P,I,V>` type parameters of `src`. 871753fe330Swren romano const auto &src = 872753fe330Swren romano static_cast<const SparseTensorStorage<P, I, V> &>(this->src); 873753fe330Swren romano if (d == Base::getRank()) { 874753fe330Swren romano assert(parentPos < src.values.size() && 875753fe330Swren romano "Value position is out of bounds"); 876753fe330Swren romano // TODO: <https://github.com/llvm/llvm-project/issues/54179> 877753fe330Swren romano yield(this->cursor, src.values[parentPos]); 878753fe330Swren romano } else if (src.isCompressedDim(d)) { 879753fe330Swren romano // Look up the bounds of the `d`-level segment determined by the 880753fe330Swren romano // `d-1`-level position `parentPos`. 881753fe330Swren romano const std::vector<P> &pointers_d = src.pointers[d]; 882753fe330Swren romano assert(parentPos + 1 < pointers_d.size() && 883753fe330Swren romano "Parent pointer position is out of bounds"); 884753fe330Swren romano const uint64_t pstart = static_cast<uint64_t>(pointers_d[parentPos]); 885753fe330Swren romano const uint64_t pstop = static_cast<uint64_t>(pointers_d[parentPos + 1]); 886753fe330Swren romano // Loop-invariant code for looking up the `d`-level coordinates/indices. 887753fe330Swren romano const std::vector<I> &indices_d = src.indices[d]; 888753fe330Swren romano assert(pstop - 1 < indices_d.size() && "Index position is out of bounds"); 889753fe330Swren romano uint64_t &cursor_reord_d = this->cursor[this->reord[d]]; 890753fe330Swren romano for (uint64_t pos = pstart; pos < pstop; pos++) { 891753fe330Swren romano cursor_reord_d = static_cast<uint64_t>(indices_d[pos]); 892753fe330Swren romano forallElements(yield, pos, d + 1); 893753fe330Swren romano } 894753fe330Swren romano } else { // Dense dimension. 895753fe330Swren romano const uint64_t sz = src.getDimSizes()[d]; 896753fe330Swren romano const uint64_t pstart = parentPos * sz; 897753fe330Swren romano uint64_t &cursor_reord_d = this->cursor[this->reord[d]]; 898753fe330Swren romano for (uint64_t i = 0; i < sz; i++) { 899753fe330Swren romano cursor_reord_d = i; 900753fe330Swren romano forallElements(yield, pstart + i, d + 1); 901753fe330Swren romano } 902753fe330Swren romano } 903753fe330Swren romano } 904753fe330Swren romano }; 905753fe330Swren romano 9068cb33240Swren romano /// Statistics regarding the number of nonzero subtensors in 9078cb33240Swren romano /// a source tensor, for direct sparse=>sparse conversion a la 9088cb33240Swren romano /// <https://arxiv.org/abs/2001.02609>. 9098cb33240Swren romano /// 9108cb33240Swren romano /// N.B., this class stores references to the parameters passed to 9118cb33240Swren romano /// the constructor; thus, objects of this class must not outlive 9128cb33240Swren romano /// those parameters. 91376944420Swren romano class SparseTensorNNZ final { 9148cb33240Swren romano public: 9158cb33240Swren romano /// Allocate the statistics structure for the desired sizes and 9168cb33240Swren romano /// sparsity (in the target tensor's storage-order). This constructor 9178cb33240Swren romano /// does not actually populate the statistics, however; for that see 9188cb33240Swren romano /// `initialize`. 9198cb33240Swren romano /// 9208cb33240Swren romano /// Precondition: `szs` must not contain zeros. 9218cb33240Swren romano SparseTensorNNZ(const std::vector<uint64_t> &szs, 9228cb33240Swren romano const std::vector<DimLevelType> &sparsity) 9238cb33240Swren romano : dimSizes(szs), dimTypes(sparsity), nnz(getRank()) { 9248cb33240Swren romano assert(dimSizes.size() == dimTypes.size() && "Rank mismatch"); 9258cb33240Swren romano bool uncompressed = true; 9268cb33240Swren romano uint64_t sz = 1; // the product of all `dimSizes` strictly less than `r`. 9278cb33240Swren romano for (uint64_t rank = getRank(), r = 0; r < rank; r++) { 9288cb33240Swren romano switch (dimTypes[r]) { 9298cb33240Swren romano case DimLevelType::kCompressed: 9308cb33240Swren romano assert(uncompressed && 9318cb33240Swren romano "Multiple compressed layers not currently supported"); 9328cb33240Swren romano uncompressed = false; 9338cb33240Swren romano nnz[r].resize(sz, 0); // Both allocate and zero-initialize. 9348cb33240Swren romano break; 9358cb33240Swren romano case DimLevelType::kDense: 9368cb33240Swren romano assert(uncompressed && 9378cb33240Swren romano "Dense after compressed not currently supported"); 9388cb33240Swren romano break; 9398cb33240Swren romano case DimLevelType::kSingleton: 9408cb33240Swren romano // Singleton after Compressed causes no problems for allocating 9418cb33240Swren romano // `nnz` nor for the yieldPos loop. This remains true even 9428cb33240Swren romano // when adding support for multiple compressed dimensions or 9438cb33240Swren romano // for dense-after-compressed. 9448cb33240Swren romano break; 9458cb33240Swren romano } 9468cb33240Swren romano sz = checkedMul(sz, dimSizes[r]); 9478cb33240Swren romano } 9488cb33240Swren romano } 9498cb33240Swren romano 9508cb33240Swren romano // We disallow copying to help avoid leaking the stored references. 9518cb33240Swren romano SparseTensorNNZ(const SparseTensorNNZ &) = delete; 9528cb33240Swren romano SparseTensorNNZ &operator=(const SparseTensorNNZ &) = delete; 9538cb33240Swren romano 9548cb33240Swren romano /// Returns the rank of the target tensor. 9558cb33240Swren romano uint64_t getRank() const { return dimSizes.size(); } 9568cb33240Swren romano 9578cb33240Swren romano /// Enumerate the source tensor to fill in the statistics. The 9588cb33240Swren romano /// enumerator should already incorporate the permutation (from 9598cb33240Swren romano /// semantic-order to the target storage-order). 9608cb33240Swren romano template <typename V> 9618cb33240Swren romano void initialize(SparseTensorEnumeratorBase<V> &enumerator) { 9628cb33240Swren romano assert(enumerator.getRank() == getRank() && "Tensor rank mismatch"); 9638cb33240Swren romano assert(enumerator.permutedSizes() == dimSizes && "Tensor size mismatch"); 9648cb33240Swren romano enumerator.forallElements( 9658cb33240Swren romano [this](const std::vector<uint64_t> &ind, V) { add(ind); }); 9668cb33240Swren romano } 9678cb33240Swren romano 9688cb33240Swren romano /// The type of callback functions which receive an nnz-statistic. 9698cb33240Swren romano using NNZConsumer = const std::function<void(uint64_t)> &; 9708cb33240Swren romano 9718cb33240Swren romano /// Lexicographically enumerates all indicies for dimensions strictly 9728cb33240Swren romano /// less than `stopDim`, and passes their nnz statistic to the callback. 9738cb33240Swren romano /// Since our use-case only requires the statistic not the coordinates 9748cb33240Swren romano /// themselves, we do not bother to construct those coordinates. 9758cb33240Swren romano void forallIndices(uint64_t stopDim, NNZConsumer yield) const { 9768cb33240Swren romano assert(stopDim < getRank() && "Stopping-dimension is out of bounds"); 9778cb33240Swren romano assert(dimTypes[stopDim] == DimLevelType::kCompressed && 9788cb33240Swren romano "Cannot look up non-compressed dimensions"); 9798cb33240Swren romano forallIndices(yield, stopDim, 0, 0); 9808cb33240Swren romano } 9818cb33240Swren romano 9828cb33240Swren romano private: 9838cb33240Swren romano /// Adds a new element (i.e., increment its statistics). We use 9848cb33240Swren romano /// a method rather than inlining into the lambda in `initialize`, 9858cb33240Swren romano /// to avoid spurious templating over `V`. And this method is private 9868cb33240Swren romano /// to avoid needing to re-assert validity of `ind` (which is guaranteed 9878cb33240Swren romano /// by `forallElements`). 9888cb33240Swren romano void add(const std::vector<uint64_t> &ind) { 9898cb33240Swren romano uint64_t parentPos = 0; 9908cb33240Swren romano for (uint64_t rank = getRank(), r = 0; r < rank; r++) { 9918cb33240Swren romano if (dimTypes[r] == DimLevelType::kCompressed) 9928cb33240Swren romano nnz[r][parentPos]++; 9938cb33240Swren romano parentPos = parentPos * dimSizes[r] + ind[r]; 9948cb33240Swren romano } 9958cb33240Swren romano } 9968cb33240Swren romano 9978cb33240Swren romano /// Recursive component of the public `forallIndices`. 9988cb33240Swren romano void forallIndices(NNZConsumer yield, uint64_t stopDim, uint64_t parentPos, 9998cb33240Swren romano uint64_t d) const { 10008cb33240Swren romano assert(d <= stopDim); 10018cb33240Swren romano if (d == stopDim) { 10028cb33240Swren romano assert(parentPos < nnz[d].size() && "Cursor is out of range"); 10038cb33240Swren romano yield(nnz[d][parentPos]); 10048cb33240Swren romano } else { 10058cb33240Swren romano const uint64_t sz = dimSizes[d]; 10068cb33240Swren romano const uint64_t pstart = parentPos * sz; 10078cb33240Swren romano for (uint64_t i = 0; i < sz; i++) 10088cb33240Swren romano forallIndices(yield, stopDim, pstart + i, d + 1); 10098cb33240Swren romano } 10108cb33240Swren romano } 10118cb33240Swren romano 10128cb33240Swren romano // All of these are in the target storage-order. 10138cb33240Swren romano const std::vector<uint64_t> &dimSizes; 10148cb33240Swren romano const std::vector<DimLevelType> &dimTypes; 10158cb33240Swren romano std::vector<std::vector<uint64_t>> nnz; 10168cb33240Swren romano }; 10178cb33240Swren romano 10188cb33240Swren romano template <typename P, typename I, typename V> 10198cb33240Swren romano SparseTensorStorage<P, I, V>::SparseTensorStorage( 10208cb33240Swren romano const std::vector<uint64_t> &szs, const uint64_t *perm, 10218cb33240Swren romano const DimLevelType *sparsity, const SparseTensorStorageBase &tensor) 10228cb33240Swren romano : SparseTensorStorage(szs, perm, sparsity) { 10238cb33240Swren romano SparseTensorEnumeratorBase<V> *enumerator; 10248cb33240Swren romano tensor.newEnumerator(&enumerator, getRank(), perm); 10258cb33240Swren romano { 10268cb33240Swren romano // Initialize the statistics structure. 10278cb33240Swren romano SparseTensorNNZ nnz(getDimSizes(), getDimTypes()); 10288cb33240Swren romano nnz.initialize(*enumerator); 10298cb33240Swren romano // Initialize "pointers" overhead (and allocate "indices", "values"). 10308cb33240Swren romano uint64_t parentSz = 1; // assembled-size (not dimension-size) of `r-1`. 10318cb33240Swren romano for (uint64_t rank = getRank(), r = 0; r < rank; r++) { 10328cb33240Swren romano if (isCompressedDim(r)) { 10338cb33240Swren romano pointers[r].reserve(parentSz + 1); 10348cb33240Swren romano pointers[r].push_back(0); 10358cb33240Swren romano uint64_t currentPos = 0; 10368cb33240Swren romano nnz.forallIndices(r, [this, ¤tPos, r](uint64_t n) { 10378cb33240Swren romano currentPos += n; 10388cb33240Swren romano appendPointer(r, currentPos); 10398cb33240Swren romano }); 10408cb33240Swren romano assert(pointers[r].size() == parentSz + 1 && 10418cb33240Swren romano "Final pointers size doesn't match allocated size"); 10428cb33240Swren romano // That assertion entails `assembledSize(parentSz, r)` 10438cb33240Swren romano // is now in a valid state. That is, `pointers[r][parentSz]` 10448cb33240Swren romano // equals the present value of `currentPos`, which is the 10458cb33240Swren romano // correct assembled-size for `indices[r]`. 10468cb33240Swren romano } 10478cb33240Swren romano // Update assembled-size for the next iteration. 10488cb33240Swren romano parentSz = assembledSize(parentSz, r); 10498cb33240Swren romano // Ideally we need only `indices[r].reserve(parentSz)`, however 10508cb33240Swren romano // the `std::vector` implementation forces us to initialize it too. 10518cb33240Swren romano // That is, in the yieldPos loop we need random-access assignment 10528cb33240Swren romano // to `indices[r]`; however, `std::vector`'s subscript-assignment 10538cb33240Swren romano // only allows assigning to already-initialized positions. 10548cb33240Swren romano if (isCompressedDim(r)) 10558cb33240Swren romano indices[r].resize(parentSz, 0); 10568cb33240Swren romano } 10578cb33240Swren romano values.resize(parentSz, 0); // Both allocate and zero-initialize. 10588cb33240Swren romano } 10598cb33240Swren romano // The yieldPos loop 10608cb33240Swren romano enumerator->forallElements([this](const std::vector<uint64_t> &ind, V val) { 10618cb33240Swren romano uint64_t parentSz = 1, parentPos = 0; 10628cb33240Swren romano for (uint64_t rank = getRank(), r = 0; r < rank; r++) { 10638cb33240Swren romano if (isCompressedDim(r)) { 10648cb33240Swren romano // If `parentPos == parentSz` then it's valid as an array-lookup; 10658cb33240Swren romano // however, it's semantically invalid here since that entry 10668cb33240Swren romano // does not represent a segment of `indices[r]`. Moreover, that 10678cb33240Swren romano // entry must be immutable for `assembledSize` to remain valid. 10688cb33240Swren romano assert(parentPos < parentSz && "Pointers position is out of bounds"); 10698cb33240Swren romano const uint64_t currentPos = pointers[r][parentPos]; 10708cb33240Swren romano // This increment won't overflow the `P` type, since it can't 10718cb33240Swren romano // exceed the original value of `pointers[r][parentPos+1]` 10728cb33240Swren romano // which was already verified to be within bounds for `P` 10738cb33240Swren romano // when it was written to the array. 10748cb33240Swren romano pointers[r][parentPos]++; 10758cb33240Swren romano writeIndex(r, currentPos, ind[r]); 10768cb33240Swren romano parentPos = currentPos; 10778cb33240Swren romano } else { // Dense dimension. 10788cb33240Swren romano parentPos = parentPos * getDimSizes()[r] + ind[r]; 10798cb33240Swren romano } 10808cb33240Swren romano parentSz = assembledSize(parentSz, r); 10818cb33240Swren romano } 10828cb33240Swren romano assert(parentPos < values.size() && "Value position is out of bounds"); 10838cb33240Swren romano values[parentPos] = val; 10848cb33240Swren romano }); 10858cb33240Swren romano // No longer need the enumerator, so we'll delete it ASAP. 10868cb33240Swren romano delete enumerator; 10878cb33240Swren romano // The finalizeYieldPos loop 10888cb33240Swren romano for (uint64_t parentSz = 1, rank = getRank(), r = 0; r < rank; r++) { 10898cb33240Swren romano if (isCompressedDim(r)) { 10908cb33240Swren romano assert(parentSz == pointers[r].size() - 1 && 10918cb33240Swren romano "Actual pointers size doesn't match the expected size"); 10928cb33240Swren romano // Can't check all of them, but at least we can check the last one. 10938cb33240Swren romano assert(pointers[r][parentSz - 1] == pointers[r][parentSz] && 10948cb33240Swren romano "Pointers got corrupted"); 10958cb33240Swren romano // TODO: optimize this by using `memmove` or similar. 10968cb33240Swren romano for (uint64_t n = 0; n < parentSz; n++) { 10978cb33240Swren romano const uint64_t parentPos = parentSz - n; 10988cb33240Swren romano pointers[r][parentPos] = pointers[r][parentPos - 1]; 10998cb33240Swren romano } 11008cb33240Swren romano pointers[r][0] = 0; 11018cb33240Swren romano } 11028cb33240Swren romano parentSz = assembledSize(parentSz, r); 11038cb33240Swren romano } 11048cb33240Swren romano } 11058cb33240Swren romano 11068a91bc7bSHarrietAkot /// Helper to convert string to lower case. 11078a91bc7bSHarrietAkot static char *toLower(char *token) { 11088a91bc7bSHarrietAkot for (char *c = token; *c; c++) 11098a91bc7bSHarrietAkot *c = tolower(*c); 11108a91bc7bSHarrietAkot return token; 11118a91bc7bSHarrietAkot } 11128a91bc7bSHarrietAkot 11138a91bc7bSHarrietAkot /// Read the MME header of a general sparse matrix of type real. 111403fe15ceSAart Bik static void readMMEHeader(FILE *file, char *filename, char *line, 111533e8ab8eSAart Bik uint64_t *idata, bool *isPattern, bool *isSymmetric) { 11168a91bc7bSHarrietAkot char header[64]; 11178a91bc7bSHarrietAkot char object[64]; 11188a91bc7bSHarrietAkot char format[64]; 11198a91bc7bSHarrietAkot char field[64]; 11208a91bc7bSHarrietAkot char symmetry[64]; 11218a91bc7bSHarrietAkot // Read header line. 11228a91bc7bSHarrietAkot if (fscanf(file, "%63s %63s %63s %63s %63s\n", header, object, format, field, 11238a91bc7bSHarrietAkot symmetry) != 5) { 112403fe15ceSAart Bik fprintf(stderr, "Corrupt header in %s\n", filename); 11258a91bc7bSHarrietAkot exit(1); 11268a91bc7bSHarrietAkot } 112733e8ab8eSAart Bik // Set properties 112833e8ab8eSAart Bik *isPattern = (strcmp(toLower(field), "pattern") == 0); 1129bb56c2b3SMehdi Amini *isSymmetric = (strcmp(toLower(symmetry), "symmetric") == 0); 11308a91bc7bSHarrietAkot // Make sure this is a general sparse matrix. 11318a91bc7bSHarrietAkot if (strcmp(toLower(header), "%%matrixmarket") || 11328a91bc7bSHarrietAkot strcmp(toLower(object), "matrix") || 113333e8ab8eSAart Bik strcmp(toLower(format), "coordinate") || 113433e8ab8eSAart Bik (strcmp(toLower(field), "real") && !(*isPattern)) || 1135bb56c2b3SMehdi Amini (strcmp(toLower(symmetry), "general") && !(*isSymmetric))) { 113633e8ab8eSAart Bik fprintf(stderr, "Cannot find a general sparse matrix in %s\n", filename); 11378a91bc7bSHarrietAkot exit(1); 11388a91bc7bSHarrietAkot } 11398a91bc7bSHarrietAkot // Skip comments. 1140e5639b3fSMehdi Amini while (true) { 114103fe15ceSAart Bik if (!fgets(line, kColWidth, file)) { 114203fe15ceSAart Bik fprintf(stderr, "Cannot find data in %s\n", filename); 11438a91bc7bSHarrietAkot exit(1); 11448a91bc7bSHarrietAkot } 11458a91bc7bSHarrietAkot if (line[0] != '%') 11468a91bc7bSHarrietAkot break; 11478a91bc7bSHarrietAkot } 11488a91bc7bSHarrietAkot // Next line contains M N NNZ. 11498a91bc7bSHarrietAkot idata[0] = 2; // rank 11508a91bc7bSHarrietAkot if (sscanf(line, "%" PRIu64 "%" PRIu64 "%" PRIu64 "\n", idata + 2, idata + 3, 11518a91bc7bSHarrietAkot idata + 1) != 3) { 115203fe15ceSAart Bik fprintf(stderr, "Cannot find size in %s\n", filename); 11538a91bc7bSHarrietAkot exit(1); 11548a91bc7bSHarrietAkot } 11558a91bc7bSHarrietAkot } 11568a91bc7bSHarrietAkot 11578a91bc7bSHarrietAkot /// Read the "extended" FROSTT header. Although not part of the documented 11588a91bc7bSHarrietAkot /// format, we assume that the file starts with optional comments followed 11598a91bc7bSHarrietAkot /// by two lines that define the rank, the number of nonzeros, and the 11608a91bc7bSHarrietAkot /// dimensions sizes (one per rank) of the sparse tensor. 116103fe15ceSAart Bik static void readExtFROSTTHeader(FILE *file, char *filename, char *line, 116203fe15ceSAart Bik uint64_t *idata) { 11638a91bc7bSHarrietAkot // Skip comments. 1164e5639b3fSMehdi Amini while (true) { 116503fe15ceSAart Bik if (!fgets(line, kColWidth, file)) { 116603fe15ceSAart Bik fprintf(stderr, "Cannot find data in %s\n", filename); 11678a91bc7bSHarrietAkot exit(1); 11688a91bc7bSHarrietAkot } 11698a91bc7bSHarrietAkot if (line[0] != '#') 11708a91bc7bSHarrietAkot break; 11718a91bc7bSHarrietAkot } 11728a91bc7bSHarrietAkot // Next line contains RANK and NNZ. 11738a91bc7bSHarrietAkot if (sscanf(line, "%" PRIu64 "%" PRIu64 "\n", idata, idata + 1) != 2) { 117403fe15ceSAart Bik fprintf(stderr, "Cannot find metadata in %s\n", filename); 11758a91bc7bSHarrietAkot exit(1); 11768a91bc7bSHarrietAkot } 11778a91bc7bSHarrietAkot // Followed by a line with the dimension sizes (one per rank). 11788a91bc7bSHarrietAkot for (uint64_t r = 0; r < idata[0]; r++) { 11798a91bc7bSHarrietAkot if (fscanf(file, "%" PRIu64, idata + 2 + r) != 1) { 118003fe15ceSAart Bik fprintf(stderr, "Cannot find dimension size %s\n", filename); 11818a91bc7bSHarrietAkot exit(1); 11828a91bc7bSHarrietAkot } 11838a91bc7bSHarrietAkot } 118403fe15ceSAart Bik fgets(line, kColWidth, file); // end of line 11858a91bc7bSHarrietAkot } 11868a91bc7bSHarrietAkot 11878a91bc7bSHarrietAkot /// Reads a sparse tensor with the given filename into a memory-resident 11888a91bc7bSHarrietAkot /// sparse tensor in coordinate scheme. 11898a91bc7bSHarrietAkot template <typename V> 11908a91bc7bSHarrietAkot static SparseTensorCOO<V> *openSparseTensorCOO(char *filename, uint64_t rank, 1191d83a7068Swren romano const uint64_t *shape, 11928a91bc7bSHarrietAkot const uint64_t *perm) { 11938a91bc7bSHarrietAkot // Open the file. 11948a91bc7bSHarrietAkot FILE *file = fopen(filename, "r"); 11958a91bc7bSHarrietAkot if (!file) { 11963734c078Swren romano assert(filename && "Received nullptr for filename"); 11973734c078Swren romano fprintf(stderr, "Cannot find file %s\n", filename); 11988a91bc7bSHarrietAkot exit(1); 11998a91bc7bSHarrietAkot } 12008a91bc7bSHarrietAkot // Perform some file format dependent set up. 120103fe15ceSAart Bik char line[kColWidth]; 12028a91bc7bSHarrietAkot uint64_t idata[512]; 120333e8ab8eSAart Bik bool isPattern = false; 1204bb56c2b3SMehdi Amini bool isSymmetric = false; 12058a91bc7bSHarrietAkot if (strstr(filename, ".mtx")) { 120633e8ab8eSAart Bik readMMEHeader(file, filename, line, idata, &isPattern, &isSymmetric); 12078a91bc7bSHarrietAkot } else if (strstr(filename, ".tns")) { 120803fe15ceSAart Bik readExtFROSTTHeader(file, filename, line, idata); 12098a91bc7bSHarrietAkot } else { 12108a91bc7bSHarrietAkot fprintf(stderr, "Unknown format %s\n", filename); 12118a91bc7bSHarrietAkot exit(1); 12128a91bc7bSHarrietAkot } 12138a91bc7bSHarrietAkot // Prepare sparse tensor object with per-dimension sizes 12148a91bc7bSHarrietAkot // and the number of nonzeros as initial capacity. 12158a91bc7bSHarrietAkot assert(rank == idata[0] && "rank mismatch"); 12168a91bc7bSHarrietAkot uint64_t nnz = idata[1]; 12178a91bc7bSHarrietAkot for (uint64_t r = 0; r < rank; r++) 1218d83a7068Swren romano assert((shape[r] == 0 || shape[r] == idata[2 + r]) && 12198a91bc7bSHarrietAkot "dimension size mismatch"); 12208a91bc7bSHarrietAkot SparseTensorCOO<V> *tensor = 12218a91bc7bSHarrietAkot SparseTensorCOO<V>::newSparseTensorCOO(rank, idata + 2, perm, nnz); 12228a91bc7bSHarrietAkot // Read all nonzero elements. 12238a91bc7bSHarrietAkot std::vector<uint64_t> indices(rank); 12248a91bc7bSHarrietAkot for (uint64_t k = 0; k < nnz; k++) { 122503fe15ceSAart Bik if (!fgets(line, kColWidth, file)) { 122603fe15ceSAart Bik fprintf(stderr, "Cannot find next line of data in %s\n", filename); 12278a91bc7bSHarrietAkot exit(1); 12288a91bc7bSHarrietAkot } 122903fe15ceSAart Bik char *linePtr = line; 123003fe15ceSAart Bik for (uint64_t r = 0; r < rank; r++) { 123103fe15ceSAart Bik uint64_t idx = strtoul(linePtr, &linePtr, 10); 12328a91bc7bSHarrietAkot // Add 0-based index. 12338a91bc7bSHarrietAkot indices[perm[r]] = idx - 1; 12348a91bc7bSHarrietAkot } 12358a91bc7bSHarrietAkot // The external formats always store the numerical values with the type 12368a91bc7bSHarrietAkot // double, but we cast these values to the sparse tensor object type. 123733e8ab8eSAart Bik // For a pattern tensor, we arbitrarily pick the value 1 for all entries. 123833e8ab8eSAart Bik double value = isPattern ? 1.0 : strtod(linePtr, &linePtr); 12398a91bc7bSHarrietAkot tensor->add(indices, value); 124002710413SBixia Zheng // We currently chose to deal with symmetric matrices by fully constructing 124102710413SBixia Zheng // them. In the future, we may want to make symmetry implicit for storage 124202710413SBixia Zheng // reasons. 1243bb56c2b3SMehdi Amini if (isSymmetric && indices[0] != indices[1]) 124402710413SBixia Zheng tensor->add({indices[1], indices[0]}, value); 12458a91bc7bSHarrietAkot } 12468a91bc7bSHarrietAkot // Close the file and return tensor. 12478a91bc7bSHarrietAkot fclose(file); 12488a91bc7bSHarrietAkot return tensor; 12498a91bc7bSHarrietAkot } 12508a91bc7bSHarrietAkot 1251efa15f41SAart Bik /// Writes the sparse tensor to extended FROSTT format. 1252efa15f41SAart Bik template <typename V> 125346bdacaaSwren romano static void outSparseTensor(void *tensor, void *dest, bool sort) { 12546438783fSAart Bik assert(tensor && dest); 12556438783fSAart Bik auto coo = static_cast<SparseTensorCOO<V> *>(tensor); 12566438783fSAart Bik if (sort) 12576438783fSAart Bik coo->sort(); 12586438783fSAart Bik char *filename = static_cast<char *>(dest); 12596438783fSAart Bik auto &sizes = coo->getSizes(); 12606438783fSAart Bik auto &elements = coo->getElements(); 12616438783fSAart Bik uint64_t rank = coo->getRank(); 1262efa15f41SAart Bik uint64_t nnz = elements.size(); 1263efa15f41SAart Bik std::fstream file; 1264efa15f41SAart Bik file.open(filename, std::ios_base::out | std::ios_base::trunc); 1265efa15f41SAart Bik assert(file.is_open()); 1266efa15f41SAart Bik file << "; extended FROSTT format\n" << rank << " " << nnz << std::endl; 1267efa15f41SAart Bik for (uint64_t r = 0; r < rank - 1; r++) 1268efa15f41SAart Bik file << sizes[r] << " "; 1269efa15f41SAart Bik file << sizes[rank - 1] << std::endl; 1270efa15f41SAart Bik for (uint64_t i = 0; i < nnz; i++) { 1271efa15f41SAart Bik auto &idx = elements[i].indices; 1272efa15f41SAart Bik for (uint64_t r = 0; r < rank; r++) 1273efa15f41SAart Bik file << (idx[r] + 1) << " "; 1274efa15f41SAart Bik file << elements[i].value << std::endl; 1275efa15f41SAart Bik } 1276efa15f41SAart Bik file.flush(); 1277efa15f41SAart Bik file.close(); 1278efa15f41SAart Bik assert(file.good()); 12796438783fSAart Bik } 12806438783fSAart Bik 12816438783fSAart Bik /// Initializes sparse tensor from an external COO-flavored format. 12826438783fSAart Bik template <typename V> 128346bdacaaSwren romano static SparseTensorStorage<uint64_t, uint64_t, V> * 12846438783fSAart Bik toMLIRSparseTensor(uint64_t rank, uint64_t nse, uint64_t *shape, V *values, 128520eaa88fSBixia Zheng uint64_t *indices, uint64_t *perm, uint8_t *sparse) { 128620eaa88fSBixia Zheng const DimLevelType *sparsity = (DimLevelType *)(sparse); 128720eaa88fSBixia Zheng #ifndef NDEBUG 128820eaa88fSBixia Zheng // Verify that perm is a permutation of 0..(rank-1). 128920eaa88fSBixia Zheng std::vector<uint64_t> order(perm, perm + rank); 129020eaa88fSBixia Zheng std::sort(order.begin(), order.end()); 12911e47888dSAart Bik for (uint64_t i = 0; i < rank; ++i) { 129220eaa88fSBixia Zheng if (i != order[i]) { 1293988d4b0dSAart Bik fprintf(stderr, "Not a permutation of 0..%" PRIu64 "\n", rank); 129420eaa88fSBixia Zheng exit(1); 129520eaa88fSBixia Zheng } 129620eaa88fSBixia Zheng } 129720eaa88fSBixia Zheng 129820eaa88fSBixia Zheng // Verify that the sparsity values are supported. 12991e47888dSAart Bik for (uint64_t i = 0; i < rank; ++i) { 130020eaa88fSBixia Zheng if (sparsity[i] != DimLevelType::kDense && 130120eaa88fSBixia Zheng sparsity[i] != DimLevelType::kCompressed) { 130220eaa88fSBixia Zheng fprintf(stderr, "Unsupported sparsity value %d\n", 130320eaa88fSBixia Zheng static_cast<int>(sparsity[i])); 130420eaa88fSBixia Zheng exit(1); 130520eaa88fSBixia Zheng } 130620eaa88fSBixia Zheng } 130720eaa88fSBixia Zheng #endif 130820eaa88fSBixia Zheng 13096438783fSAart Bik // Convert external format to internal COO. 131063bdcaf9Swren romano auto *coo = SparseTensorCOO<V>::newSparseTensorCOO(rank, shape, perm, nse); 13116438783fSAart Bik std::vector<uint64_t> idx(rank); 13126438783fSAart Bik for (uint64_t i = 0, base = 0; i < nse; i++) { 13136438783fSAart Bik for (uint64_t r = 0; r < rank; r++) 1314d8b229a1SAart Bik idx[perm[r]] = indices[base + r]; 131563bdcaf9Swren romano coo->add(idx, values[i]); 13166438783fSAart Bik base += rank; 13176438783fSAart Bik } 13186438783fSAart Bik // Return sparse tensor storage format as opaque pointer. 131963bdcaf9Swren romano auto *tensor = SparseTensorStorage<uint64_t, uint64_t, V>::newSparseTensor( 132063bdcaf9Swren romano rank, shape, perm, sparsity, coo); 132163bdcaf9Swren romano delete coo; 132263bdcaf9Swren romano return tensor; 13236438783fSAart Bik } 13246438783fSAart Bik 13256438783fSAart Bik /// Converts a sparse tensor to an external COO-flavored format. 13266438783fSAart Bik template <typename V> 132746bdacaaSwren romano static void fromMLIRSparseTensor(void *tensor, uint64_t *pRank, uint64_t *pNse, 132846bdacaaSwren romano uint64_t **pShape, V **pValues, 132946bdacaaSwren romano uint64_t **pIndices) { 1330736c1b66SAart Bik assert(tensor); 13316438783fSAart Bik auto sparseTensor = 13326438783fSAart Bik static_cast<SparseTensorStorage<uint64_t, uint64_t, V> *>(tensor); 13336438783fSAart Bik uint64_t rank = sparseTensor->getRank(); 13346438783fSAart Bik std::vector<uint64_t> perm(rank); 13356438783fSAart Bik std::iota(perm.begin(), perm.end(), 0); 13366438783fSAart Bik SparseTensorCOO<V> *coo = sparseTensor->toCOO(perm.data()); 13376438783fSAart Bik 13386438783fSAart Bik const std::vector<Element<V>> &elements = coo->getElements(); 13396438783fSAart Bik uint64_t nse = elements.size(); 13406438783fSAart Bik 13416438783fSAart Bik uint64_t *shape = new uint64_t[rank]; 13426438783fSAart Bik for (uint64_t i = 0; i < rank; i++) 13436438783fSAart Bik shape[i] = coo->getSizes()[i]; 13446438783fSAart Bik 13456438783fSAart Bik V *values = new V[nse]; 13466438783fSAart Bik uint64_t *indices = new uint64_t[rank * nse]; 13476438783fSAart Bik 13486438783fSAart Bik for (uint64_t i = 0, base = 0; i < nse; i++) { 13496438783fSAart Bik values[i] = elements[i].value; 13506438783fSAart Bik for (uint64_t j = 0; j < rank; j++) 13516438783fSAart Bik indices[base + j] = elements[i].indices[j]; 13526438783fSAart Bik base += rank; 13536438783fSAart Bik } 13546438783fSAart Bik 13556438783fSAart Bik delete coo; 13566438783fSAart Bik *pRank = rank; 13576438783fSAart Bik *pNse = nse; 13586438783fSAart Bik *pShape = shape; 13596438783fSAart Bik *pValues = values; 13606438783fSAart Bik *pIndices = indices; 1361efa15f41SAart Bik } 1362efa15f41SAart Bik 1363be0a7e9fSMehdi Amini } // namespace 13648a91bc7bSHarrietAkot 13658a91bc7bSHarrietAkot extern "C" { 13668a91bc7bSHarrietAkot 13678a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 13688a91bc7bSHarrietAkot // 13698a91bc7bSHarrietAkot // Public API with methods that operate on MLIR buffers (memrefs) to interact 13708a91bc7bSHarrietAkot // with sparse tensors, which are only visible as opaque pointers externally. 13718a91bc7bSHarrietAkot // These methods should be used exclusively by MLIR compiler-generated code. 13728a91bc7bSHarrietAkot // 13738a91bc7bSHarrietAkot // Some macro magic is used to generate implementations for all required type 13748a91bc7bSHarrietAkot // combinations that can be called from MLIR compiler-generated code. 13758a91bc7bSHarrietAkot // 13768a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 13778a91bc7bSHarrietAkot 13788a91bc7bSHarrietAkot #define CASE(p, i, v, P, I, V) \ 13798a91bc7bSHarrietAkot if (ptrTp == (p) && indTp == (i) && valTp == (v)) { \ 138063bdcaf9Swren romano SparseTensorCOO<V> *coo = nullptr; \ 1381845561ecSwren romano if (action <= Action::kFromCOO) { \ 1382845561ecSwren romano if (action == Action::kFromFile) { \ 13838a91bc7bSHarrietAkot char *filename = static_cast<char *>(ptr); \ 138463bdcaf9Swren romano coo = openSparseTensorCOO<V>(filename, rank, shape, perm); \ 1385845561ecSwren romano } else if (action == Action::kFromCOO) { \ 138663bdcaf9Swren romano coo = static_cast<SparseTensorCOO<V> *>(ptr); \ 13878a91bc7bSHarrietAkot } else { \ 1388845561ecSwren romano assert(action == Action::kEmpty); \ 13898a91bc7bSHarrietAkot } \ 139063bdcaf9Swren romano auto *tensor = SparseTensorStorage<P, I, V>::newSparseTensor( \ 139163bdcaf9Swren romano rank, shape, perm, sparsity, coo); \ 139263bdcaf9Swren romano if (action == Action::kFromFile) \ 139363bdcaf9Swren romano delete coo; \ 139463bdcaf9Swren romano return tensor; \ 1395bb56c2b3SMehdi Amini } \ 13968cb33240Swren romano if (action == Action::kSparseToSparse) { \ 13978cb33240Swren romano auto *tensor = static_cast<SparseTensorStorageBase *>(ptr); \ 13988cb33240Swren romano return SparseTensorStorage<P, I, V>::newSparseTensor(rank, shape, perm, \ 13998cb33240Swren romano sparsity, tensor); \ 14008cb33240Swren romano } \ 1401bb56c2b3SMehdi Amini if (action == Action::kEmptyCOO) \ 1402d83a7068Swren romano return SparseTensorCOO<V>::newSparseTensorCOO(rank, shape, perm); \ 140363bdcaf9Swren romano coo = static_cast<SparseTensorStorage<P, I, V> *>(ptr)->toCOO(perm); \ 1404845561ecSwren romano if (action == Action::kToIterator) { \ 140563bdcaf9Swren romano coo->startIterator(); \ 14068a91bc7bSHarrietAkot } else { \ 1407845561ecSwren romano assert(action == Action::kToCOO); \ 14088a91bc7bSHarrietAkot } \ 140963bdcaf9Swren romano return coo; \ 14108a91bc7bSHarrietAkot } 14118a91bc7bSHarrietAkot 1412845561ecSwren romano #define CASE_SECSAME(p, v, P, V) CASE(p, p, v, P, P, V) 14134f2ec7f9SAart Bik 1414d2215e79SRainer Orth // Assume index_type is in fact uint64_t, so that _mlir_ciface_newSparseTensor 1415bc04a470Swren romano // can safely rewrite kIndex to kU64. We make this assertion to guarantee 1416bc04a470Swren romano // that this file cannot get out of sync with its header. 1417d2215e79SRainer Orth static_assert(std::is_same<index_type, uint64_t>::value, 1418d2215e79SRainer Orth "Expected index_type == uint64_t"); 1419bc04a470Swren romano 14208a91bc7bSHarrietAkot /// Constructs a new sparse tensor. This is the "swiss army knife" 14218a91bc7bSHarrietAkot /// method for materializing sparse tensors into the computation. 14228a91bc7bSHarrietAkot /// 1423845561ecSwren romano /// Action: 14248a91bc7bSHarrietAkot /// kEmpty = returns empty storage to fill later 14258a91bc7bSHarrietAkot /// kFromFile = returns storage, where ptr contains filename to read 14268a91bc7bSHarrietAkot /// kFromCOO = returns storage, where ptr contains coordinate scheme to assign 14278a91bc7bSHarrietAkot /// kEmptyCOO = returns empty coordinate scheme to fill and use with kFromCOO 14288a91bc7bSHarrietAkot /// kToCOO = returns coordinate scheme from storage in ptr to use with kFromCOO 1429845561ecSwren romano /// kToIterator = returns iterator from storage in ptr (call getNext() to use) 14308a91bc7bSHarrietAkot void * 1431845561ecSwren romano _mlir_ciface_newSparseTensor(StridedMemRefType<DimLevelType, 1> *aref, // NOLINT 1432d2215e79SRainer Orth StridedMemRefType<index_type, 1> *sref, 1433d2215e79SRainer Orth StridedMemRefType<index_type, 1> *pref, 1434845561ecSwren romano OverheadType ptrTp, OverheadType indTp, 1435845561ecSwren romano PrimaryType valTp, Action action, void *ptr) { 14368a91bc7bSHarrietAkot assert(aref && sref && pref); 14378a91bc7bSHarrietAkot assert(aref->strides[0] == 1 && sref->strides[0] == 1 && 14388a91bc7bSHarrietAkot pref->strides[0] == 1); 14398a91bc7bSHarrietAkot assert(aref->sizes[0] == sref->sizes[0] && sref->sizes[0] == pref->sizes[0]); 1440845561ecSwren romano const DimLevelType *sparsity = aref->data + aref->offset; 1441d83a7068Swren romano const index_type *shape = sref->data + sref->offset; 1442d2215e79SRainer Orth const index_type *perm = pref->data + pref->offset; 14438a91bc7bSHarrietAkot uint64_t rank = aref->sizes[0]; 14448a91bc7bSHarrietAkot 1445bc04a470Swren romano // Rewrite kIndex to kU64, to avoid introducing a bunch of new cases. 1446bc04a470Swren romano // This is safe because of the static_assert above. 1447bc04a470Swren romano if (ptrTp == OverheadType::kIndex) 1448bc04a470Swren romano ptrTp = OverheadType::kU64; 1449bc04a470Swren romano if (indTp == OverheadType::kIndex) 1450bc04a470Swren romano indTp = OverheadType::kU64; 1451bc04a470Swren romano 14528a91bc7bSHarrietAkot // Double matrices with all combinations of overhead storage. 1453845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU64, PrimaryType::kF64, uint64_t, 1454845561ecSwren romano uint64_t, double); 1455845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU32, PrimaryType::kF64, uint64_t, 1456845561ecSwren romano uint32_t, double); 1457845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU16, PrimaryType::kF64, uint64_t, 1458845561ecSwren romano uint16_t, double); 1459845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU8, PrimaryType::kF64, uint64_t, 1460845561ecSwren romano uint8_t, double); 1461845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU64, PrimaryType::kF64, uint32_t, 1462845561ecSwren romano uint64_t, double); 1463845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU32, PrimaryType::kF64, uint32_t, 1464845561ecSwren romano uint32_t, double); 1465845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU16, PrimaryType::kF64, uint32_t, 1466845561ecSwren romano uint16_t, double); 1467845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU8, PrimaryType::kF64, uint32_t, 1468845561ecSwren romano uint8_t, double); 1469845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU64, PrimaryType::kF64, uint16_t, 1470845561ecSwren romano uint64_t, double); 1471845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU32, PrimaryType::kF64, uint16_t, 1472845561ecSwren romano uint32_t, double); 1473845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU16, PrimaryType::kF64, uint16_t, 1474845561ecSwren romano uint16_t, double); 1475845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU8, PrimaryType::kF64, uint16_t, 1476845561ecSwren romano uint8_t, double); 1477845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU64, PrimaryType::kF64, uint8_t, 1478845561ecSwren romano uint64_t, double); 1479845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU32, PrimaryType::kF64, uint8_t, 1480845561ecSwren romano uint32_t, double); 1481845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU16, PrimaryType::kF64, uint8_t, 1482845561ecSwren romano uint16_t, double); 1483845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU8, PrimaryType::kF64, uint8_t, 1484845561ecSwren romano uint8_t, double); 14858a91bc7bSHarrietAkot 14868a91bc7bSHarrietAkot // Float matrices with all combinations of overhead storage. 1487845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU64, PrimaryType::kF32, uint64_t, 1488845561ecSwren romano uint64_t, float); 1489845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU32, PrimaryType::kF32, uint64_t, 1490845561ecSwren romano uint32_t, float); 1491845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU16, PrimaryType::kF32, uint64_t, 1492845561ecSwren romano uint16_t, float); 1493845561ecSwren romano CASE(OverheadType::kU64, OverheadType::kU8, PrimaryType::kF32, uint64_t, 1494845561ecSwren romano uint8_t, float); 1495845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU64, PrimaryType::kF32, uint32_t, 1496845561ecSwren romano uint64_t, float); 1497845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU32, PrimaryType::kF32, uint32_t, 1498845561ecSwren romano uint32_t, float); 1499845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU16, PrimaryType::kF32, uint32_t, 1500845561ecSwren romano uint16_t, float); 1501845561ecSwren romano CASE(OverheadType::kU32, OverheadType::kU8, PrimaryType::kF32, uint32_t, 1502845561ecSwren romano uint8_t, float); 1503845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU64, PrimaryType::kF32, uint16_t, 1504845561ecSwren romano uint64_t, float); 1505845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU32, PrimaryType::kF32, uint16_t, 1506845561ecSwren romano uint32_t, float); 1507845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU16, PrimaryType::kF32, uint16_t, 1508845561ecSwren romano uint16_t, float); 1509845561ecSwren romano CASE(OverheadType::kU16, OverheadType::kU8, PrimaryType::kF32, uint16_t, 1510845561ecSwren romano uint8_t, float); 1511845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU64, PrimaryType::kF32, uint8_t, 1512845561ecSwren romano uint64_t, float); 1513845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU32, PrimaryType::kF32, uint8_t, 1514845561ecSwren romano uint32_t, float); 1515845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU16, PrimaryType::kF32, uint8_t, 1516845561ecSwren romano uint16_t, float); 1517845561ecSwren romano CASE(OverheadType::kU8, OverheadType::kU8, PrimaryType::kF32, uint8_t, 1518845561ecSwren romano uint8_t, float); 15198a91bc7bSHarrietAkot 1520845561ecSwren romano // Integral matrices with both overheads of the same type. 1521845561ecSwren romano CASE_SECSAME(OverheadType::kU64, PrimaryType::kI64, uint64_t, int64_t); 1522845561ecSwren romano CASE_SECSAME(OverheadType::kU64, PrimaryType::kI32, uint64_t, int32_t); 1523845561ecSwren romano CASE_SECSAME(OverheadType::kU64, PrimaryType::kI16, uint64_t, int16_t); 1524845561ecSwren romano CASE_SECSAME(OverheadType::kU64, PrimaryType::kI8, uint64_t, int8_t); 1525845561ecSwren romano CASE_SECSAME(OverheadType::kU32, PrimaryType::kI32, uint32_t, int32_t); 1526845561ecSwren romano CASE_SECSAME(OverheadType::kU32, PrimaryType::kI16, uint32_t, int16_t); 1527845561ecSwren romano CASE_SECSAME(OverheadType::kU32, PrimaryType::kI8, uint32_t, int8_t); 1528845561ecSwren romano CASE_SECSAME(OverheadType::kU16, PrimaryType::kI32, uint16_t, int32_t); 1529845561ecSwren romano CASE_SECSAME(OverheadType::kU16, PrimaryType::kI16, uint16_t, int16_t); 1530845561ecSwren romano CASE_SECSAME(OverheadType::kU16, PrimaryType::kI8, uint16_t, int8_t); 1531845561ecSwren romano CASE_SECSAME(OverheadType::kU8, PrimaryType::kI32, uint8_t, int32_t); 1532845561ecSwren romano CASE_SECSAME(OverheadType::kU8, PrimaryType::kI16, uint8_t, int16_t); 1533845561ecSwren romano CASE_SECSAME(OverheadType::kU8, PrimaryType::kI8, uint8_t, int8_t); 15348a91bc7bSHarrietAkot 1535736c1b66SAart Bik // Complex matrices with wide overhead. 1536736c1b66SAart Bik CASE_SECSAME(OverheadType::kU64, PrimaryType::kC64, uint64_t, complex64); 1537736c1b66SAart Bik CASE_SECSAME(OverheadType::kU64, PrimaryType::kC32, uint64_t, complex32); 1538736c1b66SAart Bik 15398a91bc7bSHarrietAkot // Unsupported case (add above if needed). 15408a91bc7bSHarrietAkot fputs("unsupported combination of types\n", stderr); 15418a91bc7bSHarrietAkot exit(1); 15428a91bc7bSHarrietAkot } 15438a91bc7bSHarrietAkot #undef CASE 15441313f5d3Swren romano #undef CASE_SECSAME 15456438783fSAart Bik 1546*bfadd13dSwren romano /// Methods that provide direct access to values. 1547*bfadd13dSwren romano #define IMPL_SPARSEVALUES(VNAME, V) \ 1548*bfadd13dSwren romano void _mlir_ciface_sparseValues##VNAME(StridedMemRefType<V, 1> *ref, \ 1549*bfadd13dSwren romano void *tensor) { \ 1550*bfadd13dSwren romano assert(ref &&tensor); \ 1551*bfadd13dSwren romano std::vector<V> *v; \ 1552*bfadd13dSwren romano static_cast<SparseTensorStorageBase *>(tensor)->getValues(&v); \ 1553*bfadd13dSwren romano ref->basePtr = ref->data = v->data(); \ 1554*bfadd13dSwren romano ref->offset = 0; \ 1555*bfadd13dSwren romano ref->sizes[0] = v->size(); \ 1556*bfadd13dSwren romano ref->strides[0] = 1; \ 1557*bfadd13dSwren romano } 1558*bfadd13dSwren romano FOREVERY_V(IMPL_SPARSEVALUES) 1559*bfadd13dSwren romano #undef IMPL_SPARSEVALUES 1560*bfadd13dSwren romano 1561*bfadd13dSwren romano #define IMPL_GETOVERHEAD(NAME, TYPE, LIB) \ 1562*bfadd13dSwren romano void _mlir_ciface_##NAME(StridedMemRefType<TYPE, 1> *ref, void *tensor, \ 1563*bfadd13dSwren romano index_type d) { \ 1564*bfadd13dSwren romano assert(ref &&tensor); \ 1565*bfadd13dSwren romano std::vector<TYPE> *v; \ 1566*bfadd13dSwren romano static_cast<SparseTensorStorageBase *>(tensor)->LIB(&v, d); \ 1567*bfadd13dSwren romano ref->basePtr = ref->data = v->data(); \ 1568*bfadd13dSwren romano ref->offset = 0; \ 1569*bfadd13dSwren romano ref->sizes[0] = v->size(); \ 1570*bfadd13dSwren romano ref->strides[0] = 1; \ 1571*bfadd13dSwren romano } 1572*bfadd13dSwren romano /// Methods that provide direct access to pointers. 1573*bfadd13dSwren romano IMPL_GETOVERHEAD(sparsePointers, index_type, getPointers) 1574*bfadd13dSwren romano IMPL_GETOVERHEAD(sparsePointers64, uint64_t, getPointers) 1575*bfadd13dSwren romano IMPL_GETOVERHEAD(sparsePointers32, uint32_t, getPointers) 1576*bfadd13dSwren romano IMPL_GETOVERHEAD(sparsePointers16, uint16_t, getPointers) 1577*bfadd13dSwren romano IMPL_GETOVERHEAD(sparsePointers8, uint8_t, getPointers) 1578*bfadd13dSwren romano 1579*bfadd13dSwren romano /// Methods that provide direct access to indices. 1580*bfadd13dSwren romano IMPL_GETOVERHEAD(sparseIndices, index_type, getIndices) 1581*bfadd13dSwren romano IMPL_GETOVERHEAD(sparseIndices64, uint64_t, getIndices) 1582*bfadd13dSwren romano IMPL_GETOVERHEAD(sparseIndices32, uint32_t, getIndices) 1583*bfadd13dSwren romano IMPL_GETOVERHEAD(sparseIndices16, uint16_t, getIndices) 1584*bfadd13dSwren romano IMPL_GETOVERHEAD(sparseIndices8, uint8_t, getIndices) 1585*bfadd13dSwren romano #undef IMPL_GETOVERHEAD 1586*bfadd13dSwren romano 1587*bfadd13dSwren romano /// Helper to add value to coordinate scheme, one per value type. 1588*bfadd13dSwren romano #define IMPL_ADDELT(VNAME, V) \ 1589*bfadd13dSwren romano void *_mlir_ciface_addElt##VNAME(void *coo, V value, \ 1590*bfadd13dSwren romano StridedMemRefType<index_type, 1> *iref, \ 1591*bfadd13dSwren romano StridedMemRefType<index_type, 1> *pref) { \ 1592*bfadd13dSwren romano assert(coo &&iref &&pref); \ 1593*bfadd13dSwren romano assert(iref->strides[0] == 1 && pref->strides[0] == 1); \ 1594*bfadd13dSwren romano assert(iref->sizes[0] == pref->sizes[0]); \ 1595*bfadd13dSwren romano const index_type *indx = iref->data + iref->offset; \ 1596*bfadd13dSwren romano const index_type *perm = pref->data + pref->offset; \ 1597*bfadd13dSwren romano uint64_t isize = iref->sizes[0]; \ 1598*bfadd13dSwren romano std::vector<index_type> indices(isize); \ 1599*bfadd13dSwren romano for (uint64_t r = 0; r < isize; r++) \ 1600*bfadd13dSwren romano indices[perm[r]] = indx[r]; \ 1601*bfadd13dSwren romano static_cast<SparseTensorCOO<V> *>(coo)->add(indices, value); \ 1602*bfadd13dSwren romano return coo; \ 1603*bfadd13dSwren romano } 1604*bfadd13dSwren romano FOREVERY_SIMPLEX_V(IMPL_ADDELT) 1605*bfadd13dSwren romano // `complex64` apparently doesn't encounter any ABI issues (yet). 1606*bfadd13dSwren romano IMPL_ADDELT(C64, complex64) 1607*bfadd13dSwren romano // TODO: cleaner way to avoid ABI padding problem? 1608*bfadd13dSwren romano IMPL_ADDELT(C32ABI, complex32) 1609*bfadd13dSwren romano void *_mlir_ciface_addEltC32(void *coo, float r, float i, 1610*bfadd13dSwren romano StridedMemRefType<index_type, 1> *iref, 1611*bfadd13dSwren romano StridedMemRefType<index_type, 1> *pref) { 1612*bfadd13dSwren romano return _mlir_ciface_addEltC32ABI(coo, complex32(r, i), iref, pref); 1613*bfadd13dSwren romano } 1614*bfadd13dSwren romano #undef IMPL_ADDELT 1615*bfadd13dSwren romano 1616*bfadd13dSwren romano /// Helper to enumerate elements of coordinate scheme, one per value type. 1617*bfadd13dSwren romano #define IMPL_GETNEXT(VNAME, V) \ 1618*bfadd13dSwren romano bool _mlir_ciface_getNext##VNAME(void *coo, \ 1619*bfadd13dSwren romano StridedMemRefType<index_type, 1> *iref, \ 1620*bfadd13dSwren romano StridedMemRefType<V, 0> *vref) { \ 1621*bfadd13dSwren romano assert(coo &&iref &&vref); \ 1622*bfadd13dSwren romano assert(iref->strides[0] == 1); \ 1623*bfadd13dSwren romano index_type *indx = iref->data + iref->offset; \ 1624*bfadd13dSwren romano V *value = vref->data + vref->offset; \ 1625*bfadd13dSwren romano const uint64_t isize = iref->sizes[0]; \ 1626*bfadd13dSwren romano const Element<V> *elem = \ 1627*bfadd13dSwren romano static_cast<SparseTensorCOO<V> *>(coo)->getNext(); \ 1628*bfadd13dSwren romano if (elem == nullptr) \ 1629*bfadd13dSwren romano return false; \ 1630*bfadd13dSwren romano for (uint64_t r = 0; r < isize; r++) \ 1631*bfadd13dSwren romano indx[r] = elem->indices[r]; \ 1632*bfadd13dSwren romano *value = elem->value; \ 1633*bfadd13dSwren romano return true; \ 1634*bfadd13dSwren romano } 1635*bfadd13dSwren romano FOREVERY_V(IMPL_GETNEXT) 1636*bfadd13dSwren romano #undef IMPL_GETNEXT 1637*bfadd13dSwren romano 1638*bfadd13dSwren romano /// Insert elements in lexicographical index order, one per value type. 1639*bfadd13dSwren romano #define IMPL_LEXINSERT(VNAME, V) \ 1640*bfadd13dSwren romano void _mlir_ciface_lexInsert##VNAME( \ 1641*bfadd13dSwren romano void *tensor, StridedMemRefType<index_type, 1> *cref, V val) { \ 1642*bfadd13dSwren romano assert(tensor &&cref); \ 1643*bfadd13dSwren romano assert(cref->strides[0] == 1); \ 1644*bfadd13dSwren romano index_type *cursor = cref->data + cref->offset; \ 1645*bfadd13dSwren romano assert(cursor); \ 1646*bfadd13dSwren romano static_cast<SparseTensorStorageBase *>(tensor)->lexInsert(cursor, val); \ 1647*bfadd13dSwren romano } 1648*bfadd13dSwren romano FOREVERY_SIMPLEX_V(IMPL_LEXINSERT) 1649*bfadd13dSwren romano // `complex64` apparently doesn't encounter any ABI issues (yet). 1650*bfadd13dSwren romano IMPL_LEXINSERT(C64, complex64) 1651*bfadd13dSwren romano // TODO: cleaner way to avoid ABI padding problem? 1652*bfadd13dSwren romano IMPL_LEXINSERT(C32ABI, complex32) 1653*bfadd13dSwren romano void _mlir_ciface_lexInsertC32(void *tensor, 1654*bfadd13dSwren romano StridedMemRefType<index_type, 1> *cref, float r, 1655*bfadd13dSwren romano float i) { 1656*bfadd13dSwren romano _mlir_ciface_lexInsertC32ABI(tensor, cref, complex32(r, i)); 1657*bfadd13dSwren romano } 1658*bfadd13dSwren romano #undef IMPL_LEXINSERT 1659*bfadd13dSwren romano 1660*bfadd13dSwren romano /// Insert using expansion, one per value type. 1661*bfadd13dSwren romano #define IMPL_EXPINSERT(VNAME, V) \ 1662*bfadd13dSwren romano void _mlir_ciface_expInsert##VNAME( \ 1663*bfadd13dSwren romano void *tensor, StridedMemRefType<index_type, 1> *cref, \ 1664*bfadd13dSwren romano StridedMemRefType<V, 1> *vref, StridedMemRefType<bool, 1> *fref, \ 1665*bfadd13dSwren romano StridedMemRefType<index_type, 1> *aref, index_type count) { \ 1666*bfadd13dSwren romano assert(tensor &&cref &&vref &&fref &&aref); \ 1667*bfadd13dSwren romano assert(cref->strides[0] == 1); \ 1668*bfadd13dSwren romano assert(vref->strides[0] == 1); \ 1669*bfadd13dSwren romano assert(fref->strides[0] == 1); \ 1670*bfadd13dSwren romano assert(aref->strides[0] == 1); \ 1671*bfadd13dSwren romano assert(vref->sizes[0] == fref->sizes[0]); \ 1672*bfadd13dSwren romano index_type *cursor = cref->data + cref->offset; \ 1673*bfadd13dSwren romano V *values = vref->data + vref->offset; \ 1674*bfadd13dSwren romano bool *filled = fref->data + fref->offset; \ 1675*bfadd13dSwren romano index_type *added = aref->data + aref->offset; \ 1676*bfadd13dSwren romano static_cast<SparseTensorStorageBase *>(tensor)->expInsert( \ 1677*bfadd13dSwren romano cursor, values, filled, added, count); \ 1678*bfadd13dSwren romano } 1679*bfadd13dSwren romano FOREVERY_V(IMPL_EXPINSERT) 1680*bfadd13dSwren romano #undef IMPL_EXPINSERT 1681*bfadd13dSwren romano 16826438783fSAart Bik /// Output a sparse tensor, one per value type. 16831313f5d3Swren romano #define IMPL_OUTSPARSETENSOR(VNAME, V) \ 16841313f5d3Swren romano void outSparseTensor##VNAME(void *coo, void *dest, bool sort) { \ 16851313f5d3Swren romano return outSparseTensor<V>(coo, dest, sort); \ 16866438783fSAart Bik } 16871313f5d3Swren romano FOREVERY_V(IMPL_OUTSPARSETENSOR) 16881313f5d3Swren romano #undef IMPL_OUTSPARSETENSOR 16898a91bc7bSHarrietAkot 16908a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 16918a91bc7bSHarrietAkot // 16928a91bc7bSHarrietAkot // Public API with methods that accept C-style data structures to interact 16938a91bc7bSHarrietAkot // with sparse tensors, which are only visible as opaque pointers externally. 16948a91bc7bSHarrietAkot // These methods can be used both by MLIR compiler-generated code as well as by 16958a91bc7bSHarrietAkot // an external runtime that wants to interact with MLIR compiler-generated code. 16968a91bc7bSHarrietAkot // 16978a91bc7bSHarrietAkot //===----------------------------------------------------------------------===// 16988a91bc7bSHarrietAkot 16998a91bc7bSHarrietAkot /// Helper method to read a sparse tensor filename from the environment, 17008a91bc7bSHarrietAkot /// defined with the naming convention ${TENSOR0}, ${TENSOR1}, etc. 1701d2215e79SRainer Orth char *getTensorFilename(index_type id) { 17028a91bc7bSHarrietAkot char var[80]; 17038a91bc7bSHarrietAkot sprintf(var, "TENSOR%" PRIu64, id); 17048a91bc7bSHarrietAkot char *env = getenv(var); 17053734c078Swren romano if (!env) { 17063734c078Swren romano fprintf(stderr, "Environment variable %s is not set\n", var); 17073734c078Swren romano exit(1); 17083734c078Swren romano } 17098a91bc7bSHarrietAkot return env; 17108a91bc7bSHarrietAkot } 17118a91bc7bSHarrietAkot 17128a91bc7bSHarrietAkot /// Returns size of sparse tensor in given dimension. 1713d2215e79SRainer Orth index_type sparseDimSize(void *tensor, index_type d) { 17148a91bc7bSHarrietAkot return static_cast<SparseTensorStorageBase *>(tensor)->getDimSize(d); 17158a91bc7bSHarrietAkot } 17168a91bc7bSHarrietAkot 1717f66e5769SAart Bik /// Finalizes lexicographic insertions. 1718f66e5769SAart Bik void endInsert(void *tensor) { 1719f66e5769SAart Bik return static_cast<SparseTensorStorageBase *>(tensor)->endInsert(); 1720f66e5769SAart Bik } 1721f66e5769SAart Bik 17228a91bc7bSHarrietAkot /// Releases sparse tensor storage. 17238a91bc7bSHarrietAkot void delSparseTensor(void *tensor) { 17248a91bc7bSHarrietAkot delete static_cast<SparseTensorStorageBase *>(tensor); 17258a91bc7bSHarrietAkot } 17268a91bc7bSHarrietAkot 172763bdcaf9Swren romano /// Releases sparse tensor coordinate scheme. 172863bdcaf9Swren romano #define IMPL_DELCOO(VNAME, V) \ 172963bdcaf9Swren romano void delSparseTensorCOO##VNAME(void *coo) { \ 173063bdcaf9Swren romano delete static_cast<SparseTensorCOO<V> *>(coo); \ 173163bdcaf9Swren romano } 17321313f5d3Swren romano FOREVERY_V(IMPL_DELCOO) 173363bdcaf9Swren romano #undef IMPL_DELCOO 173463bdcaf9Swren romano 17358a91bc7bSHarrietAkot /// Initializes sparse tensor from a COO-flavored format expressed using C-style 17368a91bc7bSHarrietAkot /// data structures. The expected parameters are: 17378a91bc7bSHarrietAkot /// 17388a91bc7bSHarrietAkot /// rank: rank of tensor 17398a91bc7bSHarrietAkot /// nse: number of specified elements (usually the nonzeros) 17408a91bc7bSHarrietAkot /// shape: array with dimension size for each rank 17418a91bc7bSHarrietAkot /// values: a "nse" array with values for all specified elements 17428a91bc7bSHarrietAkot /// indices: a flat "nse x rank" array with indices for all specified elements 174320eaa88fSBixia Zheng /// perm: the permutation of the dimensions in the storage 174420eaa88fSBixia Zheng /// sparse: the sparsity for the dimensions 17458a91bc7bSHarrietAkot /// 17468a91bc7bSHarrietAkot /// For example, the sparse matrix 17478a91bc7bSHarrietAkot /// | 1.0 0.0 0.0 | 17488a91bc7bSHarrietAkot /// | 0.0 5.0 3.0 | 17498a91bc7bSHarrietAkot /// can be passed as 17508a91bc7bSHarrietAkot /// rank = 2 17518a91bc7bSHarrietAkot /// nse = 3 17528a91bc7bSHarrietAkot /// shape = [2, 3] 17538a91bc7bSHarrietAkot /// values = [1.0, 5.0, 3.0] 17548a91bc7bSHarrietAkot /// indices = [ 0, 0, 1, 1, 1, 2] 17558a91bc7bSHarrietAkot // 175620eaa88fSBixia Zheng // TODO: generalize beyond 64-bit indices. 17578a91bc7bSHarrietAkot // 17581313f5d3Swren romano #define IMPL_CONVERTTOMLIRSPARSETENSOR(VNAME, V) \ 17591313f5d3Swren romano void *convertToMLIRSparseTensor##VNAME( \ 17601313f5d3Swren romano uint64_t rank, uint64_t nse, uint64_t *shape, V *values, \ 17611313f5d3Swren romano uint64_t *indices, uint64_t *perm, uint8_t *sparse) { \ 17621313f5d3Swren romano return toMLIRSparseTensor<V>(rank, nse, shape, values, indices, perm, \ 17631313f5d3Swren romano sparse); \ 17648a91bc7bSHarrietAkot } 17651313f5d3Swren romano FOREVERY_V(IMPL_CONVERTTOMLIRSPARSETENSOR) 17661313f5d3Swren romano #undef IMPL_CONVERTTOMLIRSPARSETENSOR 17678a91bc7bSHarrietAkot 17682f49e6b0SBixia Zheng /// Converts a sparse tensor to COO-flavored format expressed using C-style 17692f49e6b0SBixia Zheng /// data structures. The expected output parameters are pointers for these 17702f49e6b0SBixia Zheng /// values: 17712f49e6b0SBixia Zheng /// 17722f49e6b0SBixia Zheng /// rank: rank of tensor 17732f49e6b0SBixia Zheng /// nse: number of specified elements (usually the nonzeros) 17742f49e6b0SBixia Zheng /// shape: array with dimension size for each rank 17752f49e6b0SBixia Zheng /// values: a "nse" array with values for all specified elements 17762f49e6b0SBixia Zheng /// indices: a flat "nse x rank" array with indices for all specified elements 17772f49e6b0SBixia Zheng /// 17782f49e6b0SBixia Zheng /// The input is a pointer to SparseTensorStorage<P, I, V>, typically returned 17792f49e6b0SBixia Zheng /// from convertToMLIRSparseTensor. 17802f49e6b0SBixia Zheng /// 17812f49e6b0SBixia Zheng // TODO: Currently, values are copied from SparseTensorStorage to 17822f49e6b0SBixia Zheng // SparseTensorCOO, then to the output. We may want to reduce the number of 17832f49e6b0SBixia Zheng // copies. 17842f49e6b0SBixia Zheng // 17856438783fSAart Bik // TODO: generalize beyond 64-bit indices, no dim ordering, all dimensions 17866438783fSAart Bik // compressed 17872f49e6b0SBixia Zheng // 17881313f5d3Swren romano #define IMPL_CONVERTFROMMLIRSPARSETENSOR(VNAME, V) \ 17891313f5d3Swren romano void convertFromMLIRSparseTensor##VNAME(void *tensor, uint64_t *pRank, \ 17901313f5d3Swren romano uint64_t *pNse, uint64_t **pShape, \ 17911313f5d3Swren romano V **pValues, uint64_t **pIndices) { \ 17921313f5d3Swren romano fromMLIRSparseTensor<V>(tensor, pRank, pNse, pShape, pValues, pIndices); \ 17932f49e6b0SBixia Zheng } 17941313f5d3Swren romano FOREVERY_V(IMPL_CONVERTFROMMLIRSPARSETENSOR) 17951313f5d3Swren romano #undef IMPL_CONVERTFROMMLIRSPARSETENSOR 1796efa15f41SAart Bik 17978a91bc7bSHarrietAkot } // extern "C" 17988a91bc7bSHarrietAkot 17998a91bc7bSHarrietAkot #endif // MLIR_CRUNNERUTILS_DEFINE_FUNCTIONS 1800