1 //===- SparseTensorRuntime.cpp - SparseTensor runtime support lib ---------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements a light-weight runtime support library for
10 // manipulating sparse tensors from MLIR. More specifically, it provides
11 // C-API wrappers so that MLIR-generated code can call into the C++ runtime
12 // support library. The functionality provided in this library is meant
13 // to simplify benchmarking, testing, and debugging of MLIR code operating
14 // on sparse tensors. However, the provided functionality is **not**
15 // part of core MLIR itself.
17 // The following memory-resident sparse storage schemes are supported:
19 // (a) A coordinate scheme for temporarily storing and lexicographically
20 // sorting a sparse tensor by coordinate (SparseTensorCOO).
22 // (b) A "one-size-fits-all" sparse tensor storage scheme defined by
23 // per-dimension sparse/dense annnotations together with a dimension
24 // ordering used by MLIR compiler-generated code (SparseTensorStorage).
26 // The following external formats are supported:
28 // (1) Matrix Market Exchange (MME): *.mtx
29 // https://math.nist.gov/MatrixMarket/formats.html
31 // (2) Formidable Repository of Open Sparse Tensors and Tools (FROSTT): *.tns
32 // http://frostt.io/tensors/file-formats.html
34 // Two public APIs are supported:
36 // (I) Methods operating on MLIR buffers (memrefs) to interact with sparse
37 // tensors. These methods should be used exclusively by MLIR
38 // compiler-generated code.
40 // (II) Methods that accept C-style data structures to interact with sparse
41 // tensors. These methods can be used by any external runtime that wants
42 // to interact with MLIR compiler-generated code.
44 // In both cases (I) and (II), the SparseTensorStorage format is externally
45 // only visible as an opaque pointer.
47 //===----------------------------------------------------------------------===//
49 #include "mlir/ExecutionEngine/SparseTensorRuntime.h"
51 #ifdef MLIR_CRUNNERUTILS_DEFINE_FUNCTIONS
53 #include "mlir/ExecutionEngine/SparseTensor/ArithmeticUtils.h"
54 #include "mlir/ExecutionEngine/SparseTensor/COO.h"
55 #include "mlir/ExecutionEngine/SparseTensor/ErrorHandling.h"
56 #include "mlir/ExecutionEngine/SparseTensor/File.h"
57 #include "mlir/ExecutionEngine/SparseTensor/Storage.h"
62 using namespace mlir::sparse_tensor
;
64 //===----------------------------------------------------------------------===//
66 // Utilities for manipulating `StridedMemRefType`.
68 //===----------------------------------------------------------------------===//
72 #define ASSERT_NO_STRIDE(MEMREF) \
74 assert((MEMREF) && "Memref is nullptr"); \
75 assert(((MEMREF)->strides[0] == 1) && "Memref has non-trivial stride"); \
78 #define MEMREF_GET_USIZE(MEMREF) \
79 detail::checkOverflowCast<uint64_t>((MEMREF)->sizes[0])
81 #define ASSERT_USIZE_EQ(MEMREF, SZ) \
82 assert(detail::safelyEQ(MEMREF_GET_USIZE(MEMREF), (SZ)) && \
83 "Memref size mismatch")
85 #define MEMREF_GET_PAYLOAD(MEMREF) ((MEMREF)->data + (MEMREF)->offset)
87 /// Initializes the memref with the provided size and data pointer. This
88 /// is designed for functions which want to "return" a memref that aliases
89 /// into memory owned by some other object (e.g., `SparseTensorStorage`),
90 /// without doing any actual copying. (The "return" is in scarequotes
91 /// because the `_mlir_ciface_` calling convention migrates any returned
92 /// memrefs into an out-parameter passed before all the other function
94 template <typename DataSizeT
, typename T
>
95 static inline void aliasIntoMemref(DataSizeT size
, T
*data
,
96 StridedMemRefType
<T
, 1> &ref
) {
97 ref
.basePtr
= ref
.data
= data
;
99 using MemrefSizeT
= std::remove_reference_t
<decltype(ref
.sizes
[0])>;
100 ref
.sizes
[0] = detail::checkOverflowCast
<MemrefSizeT
>(size
);
104 } // anonymous namespace
108 //===----------------------------------------------------------------------===//
110 // Public functions which operate on MLIR buffers (memrefs) to interact
111 // with sparse tensors (which are only visible as opaque pointers externally).
113 //===----------------------------------------------------------------------===//
115 #define CASE(p, c, v, P, C, V) \
116 if (posTp == (p) && crdTp == (c) && valTp == (v)) { \
118 case Action::kEmpty: { \
119 return SparseTensorStorage<P, C, V>::newEmpty( \
120 dimRank, dimSizes, lvlRank, lvlSizes, lvlTypes, dim2lvl, lvl2dim, \
123 case Action::kEmptyForward: { \
124 return SparseTensorStorage<P, C, V>::newEmpty( \
125 dimRank, dimSizes, lvlRank, lvlSizes, lvlTypes, dim2lvl, lvl2dim, \
128 case Action::kFromCOO: { \
129 assert(ptr && "Received nullptr for SparseTensorCOO object"); \
130 auto &coo = *static_cast<SparseTensorCOO<V> *>(ptr); \
131 return SparseTensorStorage<P, C, V>::newFromCOO( \
132 dimRank, dimSizes, lvlRank, lvlSizes, lvlTypes, dim2lvl, lvl2dim, \
135 case Action::kFromReader: { \
136 assert(ptr && "Received nullptr for SparseTensorReader object"); \
137 SparseTensorReader &reader = *static_cast<SparseTensorReader *>(ptr); \
138 return static_cast<void *>(reader.readSparseTensor<P, C, V>( \
139 lvlRank, lvlSizes, lvlTypes, dim2lvl, lvl2dim)); \
141 case Action::kToCOO: { \
142 assert(ptr && "Received nullptr for SparseTensorStorage object"); \
143 auto &tensor = *static_cast<SparseTensorStorage<P, C, V> *>(ptr); \
144 return tensor.toCOO(); \
146 case Action::kPack: { \
147 assert(ptr && "Received nullptr for SparseTensorStorage object"); \
148 intptr_t *buffers = static_cast<intptr_t *>(ptr); \
149 return SparseTensorStorage<P, C, V>::packFromLvlBuffers( \
150 dimRank, dimSizes, lvlRank, lvlSizes, lvlTypes, dim2lvl, lvl2dim, \
153 case Action::kSortCOOInPlace: { \
154 assert(ptr && "Received nullptr for SparseTensorStorage object"); \
155 auto &tensor = *static_cast<SparseTensorStorage<P, C, V> *>(ptr); \
156 tensor.sortInPlace(); \
160 MLIR_SPARSETENSOR_FATAL("unknown action: %d\n", \
161 static_cast<uint32_t>(action)); \
164 #define CASE_SECSAME(p, v, P, V) CASE(p, p, v, P, P, V)
166 // Assume index_type is in fact uint64_t, so that _mlir_ciface_newSparseTensor
167 // can safely rewrite kIndex to kU64. We make this assertion to guarantee
168 // that this file cannot get out of sync with its header.
169 static_assert(std::is_same
<index_type
, uint64_t>::value
,
170 "Expected index_type == uint64_t");
172 // The Swiss-army-knife for sparse tensor creation.
173 void *_mlir_ciface_newSparseTensor( // NOLINT
174 StridedMemRefType
<index_type
, 1> *dimSizesRef
,
175 StridedMemRefType
<index_type
, 1> *lvlSizesRef
,
176 StridedMemRefType
<DimLevelType
, 1> *lvlTypesRef
,
177 StridedMemRefType
<index_type
, 1> *dim2lvlRef
,
178 StridedMemRefType
<index_type
, 1> *lvl2dimRef
, OverheadType posTp
,
179 OverheadType crdTp
, PrimaryType valTp
, Action action
, void *ptr
) {
180 ASSERT_NO_STRIDE(dimSizesRef
);
181 ASSERT_NO_STRIDE(lvlSizesRef
);
182 ASSERT_NO_STRIDE(lvlTypesRef
);
183 ASSERT_NO_STRIDE(dim2lvlRef
);
184 ASSERT_NO_STRIDE(lvl2dimRef
);
185 const uint64_t dimRank
= MEMREF_GET_USIZE(dimSizesRef
);
186 const uint64_t lvlRank
= MEMREF_GET_USIZE(lvlSizesRef
);
187 ASSERT_USIZE_EQ(lvlTypesRef
, lvlRank
);
188 ASSERT_USIZE_EQ(dim2lvlRef
, lvlRank
);
189 ASSERT_USIZE_EQ(lvl2dimRef
, dimRank
);
190 const index_type
*dimSizes
= MEMREF_GET_PAYLOAD(dimSizesRef
);
191 const index_type
*lvlSizes
= MEMREF_GET_PAYLOAD(lvlSizesRef
);
192 const DimLevelType
*lvlTypes
= MEMREF_GET_PAYLOAD(lvlTypesRef
);
193 const index_type
*dim2lvl
= MEMREF_GET_PAYLOAD(dim2lvlRef
);
194 const index_type
*lvl2dim
= MEMREF_GET_PAYLOAD(lvl2dimRef
);
196 // Rewrite kIndex to kU64, to avoid introducing a bunch of new cases.
197 // This is safe because of the static_assert above.
198 if (posTp
== OverheadType::kIndex
)
199 posTp
= OverheadType::kU64
;
200 if (crdTp
== OverheadType::kIndex
)
201 crdTp
= OverheadType::kU64
;
203 // Double matrices with all combinations of overhead storage.
204 CASE(OverheadType::kU64
, OverheadType::kU64
, PrimaryType::kF64
, uint64_t,
206 CASE(OverheadType::kU64
, OverheadType::kU32
, PrimaryType::kF64
, uint64_t,
208 CASE(OverheadType::kU64
, OverheadType::kU16
, PrimaryType::kF64
, uint64_t,
210 CASE(OverheadType::kU64
, OverheadType::kU8
, PrimaryType::kF64
, uint64_t,
212 CASE(OverheadType::kU32
, OverheadType::kU64
, PrimaryType::kF64
, uint32_t,
214 CASE(OverheadType::kU32
, OverheadType::kU32
, PrimaryType::kF64
, uint32_t,
216 CASE(OverheadType::kU32
, OverheadType::kU16
, PrimaryType::kF64
, uint32_t,
218 CASE(OverheadType::kU32
, OverheadType::kU8
, PrimaryType::kF64
, uint32_t,
220 CASE(OverheadType::kU16
, OverheadType::kU64
, PrimaryType::kF64
, uint16_t,
222 CASE(OverheadType::kU16
, OverheadType::kU32
, PrimaryType::kF64
, uint16_t,
224 CASE(OverheadType::kU16
, OverheadType::kU16
, PrimaryType::kF64
, uint16_t,
226 CASE(OverheadType::kU16
, OverheadType::kU8
, PrimaryType::kF64
, uint16_t,
228 CASE(OverheadType::kU8
, OverheadType::kU64
, PrimaryType::kF64
, uint8_t,
230 CASE(OverheadType::kU8
, OverheadType::kU32
, PrimaryType::kF64
, uint8_t,
232 CASE(OverheadType::kU8
, OverheadType::kU16
, PrimaryType::kF64
, uint8_t,
234 CASE(OverheadType::kU8
, OverheadType::kU8
, PrimaryType::kF64
, uint8_t,
237 // Float matrices with all combinations of overhead storage.
238 CASE(OverheadType::kU64
, OverheadType::kU64
, PrimaryType::kF32
, uint64_t,
240 CASE(OverheadType::kU64
, OverheadType::kU32
, PrimaryType::kF32
, uint64_t,
242 CASE(OverheadType::kU64
, OverheadType::kU16
, PrimaryType::kF32
, uint64_t,
244 CASE(OverheadType::kU64
, OverheadType::kU8
, PrimaryType::kF32
, uint64_t,
246 CASE(OverheadType::kU32
, OverheadType::kU64
, PrimaryType::kF32
, uint32_t,
248 CASE(OverheadType::kU32
, OverheadType::kU32
, PrimaryType::kF32
, uint32_t,
250 CASE(OverheadType::kU32
, OverheadType::kU16
, PrimaryType::kF32
, uint32_t,
252 CASE(OverheadType::kU32
, OverheadType::kU8
, PrimaryType::kF32
, uint32_t,
254 CASE(OverheadType::kU16
, OverheadType::kU64
, PrimaryType::kF32
, uint16_t,
256 CASE(OverheadType::kU16
, OverheadType::kU32
, PrimaryType::kF32
, uint16_t,
258 CASE(OverheadType::kU16
, OverheadType::kU16
, PrimaryType::kF32
, uint16_t,
260 CASE(OverheadType::kU16
, OverheadType::kU8
, PrimaryType::kF32
, uint16_t,
262 CASE(OverheadType::kU8
, OverheadType::kU64
, PrimaryType::kF32
, uint8_t,
264 CASE(OverheadType::kU8
, OverheadType::kU32
, PrimaryType::kF32
, uint8_t,
266 CASE(OverheadType::kU8
, OverheadType::kU16
, PrimaryType::kF32
, uint8_t,
268 CASE(OverheadType::kU8
, OverheadType::kU8
, PrimaryType::kF32
, uint8_t,
271 // Two-byte floats with both overheads of the same type.
272 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kF16
, uint64_t, f16
);
273 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kBF16
, uint64_t, bf16
);
274 CASE_SECSAME(OverheadType::kU32
, PrimaryType::kF16
, uint32_t, f16
);
275 CASE_SECSAME(OverheadType::kU32
, PrimaryType::kBF16
, uint32_t, bf16
);
276 CASE_SECSAME(OverheadType::kU16
, PrimaryType::kF16
, uint16_t, f16
);
277 CASE_SECSAME(OverheadType::kU16
, PrimaryType::kBF16
, uint16_t, bf16
);
278 CASE_SECSAME(OverheadType::kU8
, PrimaryType::kF16
, uint8_t, f16
);
279 CASE_SECSAME(OverheadType::kU8
, PrimaryType::kBF16
, uint8_t, bf16
);
281 // Integral matrices with both overheads of the same type.
282 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kI64
, uint64_t, int64_t);
283 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kI32
, uint64_t, int32_t);
284 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kI16
, uint64_t, int16_t);
285 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kI8
, uint64_t, int8_t);
286 CASE_SECSAME(OverheadType::kU32
, PrimaryType::kI64
, uint32_t, int64_t);
287 CASE_SECSAME(OverheadType::kU32
, PrimaryType::kI32
, uint32_t, int32_t);
288 CASE_SECSAME(OverheadType::kU32
, PrimaryType::kI16
, uint32_t, int16_t);
289 CASE_SECSAME(OverheadType::kU32
, PrimaryType::kI8
, uint32_t, int8_t);
290 CASE_SECSAME(OverheadType::kU16
, PrimaryType::kI64
, uint16_t, int64_t);
291 CASE_SECSAME(OverheadType::kU16
, PrimaryType::kI32
, uint16_t, int32_t);
292 CASE_SECSAME(OverheadType::kU16
, PrimaryType::kI16
, uint16_t, int16_t);
293 CASE_SECSAME(OverheadType::kU16
, PrimaryType::kI8
, uint16_t, int8_t);
294 CASE_SECSAME(OverheadType::kU8
, PrimaryType::kI64
, uint8_t, int64_t);
295 CASE_SECSAME(OverheadType::kU8
, PrimaryType::kI32
, uint8_t, int32_t);
296 CASE_SECSAME(OverheadType::kU8
, PrimaryType::kI16
, uint8_t, int16_t);
297 CASE_SECSAME(OverheadType::kU8
, PrimaryType::kI8
, uint8_t, int8_t);
299 // Complex matrices with wide overhead.
300 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kC64
, uint64_t, complex64
);
301 CASE_SECSAME(OverheadType::kU64
, PrimaryType::kC32
, uint64_t, complex32
);
303 // Unsupported case (add above if needed).
304 MLIR_SPARSETENSOR_FATAL(
305 "unsupported combination of types: <P=%d, C=%d, V=%d>\n",
306 static_cast<int>(posTp
), static_cast<int>(crdTp
),
307 static_cast<int>(valTp
));
312 #define IMPL_SPARSEVALUES(VNAME, V) \
313 void _mlir_ciface_sparseValues##VNAME(StridedMemRefType<V, 1> *ref, \
315 assert(ref &&tensor); \
317 static_cast<SparseTensorStorageBase *>(tensor)->getValues(&v); \
319 aliasIntoMemref(v->size(), v->data(), *ref); \
321 MLIR_SPARSETENSOR_FOREVERY_V(IMPL_SPARSEVALUES
)
322 #undef IMPL_SPARSEVALUES
324 #define IMPL_GETOVERHEAD(NAME, TYPE, LIB) \
325 void _mlir_ciface_##NAME(StridedMemRefType<TYPE, 1> *ref, void *tensor, \
327 assert(ref &&tensor); \
328 std::vector<TYPE> *v; \
329 static_cast<SparseTensorStorageBase *>(tensor)->LIB(&v, lvl); \
331 aliasIntoMemref(v->size(), v->data(), *ref); \
333 #define IMPL_SPARSEPOSITIONS(PNAME, P) \
334 IMPL_GETOVERHEAD(sparsePositions##PNAME, P, getPositions)
335 MLIR_SPARSETENSOR_FOREVERY_O(IMPL_SPARSEPOSITIONS
)
336 #undef IMPL_SPARSEPOSITIONS
338 #define IMPL_SPARSECOORDINATES(CNAME, C) \
339 IMPL_GETOVERHEAD(sparseCoordinates##CNAME, C, getCoordinates)
340 MLIR_SPARSETENSOR_FOREVERY_O(IMPL_SPARSECOORDINATES
)
341 #undef IMPL_SPARSECOORDINATES
342 #undef IMPL_GETOVERHEAD
344 #define IMPL_FORWARDINGINSERT(VNAME, V) \
345 void _mlir_ciface_forwardingInsert##VNAME( \
346 void *t, StridedMemRefType<V, 0> *vref, \
347 StridedMemRefType<index_type, 1> *dimCoordsRef) { \
349 ASSERT_NO_STRIDE(dimCoordsRef); \
350 const index_type *dimCoords = MEMREF_GET_PAYLOAD(dimCoordsRef); \
352 const V *value = MEMREF_GET_PAYLOAD(vref); \
353 static_cast<SparseTensorStorageBase *>(t)->forwardingInsert(dimCoords, \
356 MLIR_SPARSETENSOR_FOREVERY_V(IMPL_FORWARDINGINSERT
)
357 #undef IMPL_FORWARDINGINSERT
359 #define IMPL_LEXINSERT(VNAME, V) \
360 void _mlir_ciface_lexInsert##VNAME( \
361 void *t, StridedMemRefType<index_type, 1> *lvlCoordsRef, \
362 StridedMemRefType<V, 0> *vref) { \
364 auto &tensor = *static_cast<SparseTensorStorageBase *>(t); \
365 ASSERT_NO_STRIDE(lvlCoordsRef); \
366 index_type *lvlCoords = MEMREF_GET_PAYLOAD(lvlCoordsRef); \
368 V *value = MEMREF_GET_PAYLOAD(vref); \
369 tensor.lexInsert(lvlCoords, *value); \
371 MLIR_SPARSETENSOR_FOREVERY_V(IMPL_LEXINSERT
)
372 #undef IMPL_LEXINSERT
374 #define IMPL_EXPINSERT(VNAME, V) \
375 void _mlir_ciface_expInsert##VNAME( \
376 void *t, StridedMemRefType<index_type, 1> *lvlCoordsRef, \
377 StridedMemRefType<V, 1> *vref, StridedMemRefType<bool, 1> *fref, \
378 StridedMemRefType<index_type, 1> *aref, index_type count) { \
380 auto &tensor = *static_cast<SparseTensorStorageBase *>(t); \
381 ASSERT_NO_STRIDE(lvlCoordsRef); \
382 ASSERT_NO_STRIDE(vref); \
383 ASSERT_NO_STRIDE(fref); \
384 ASSERT_NO_STRIDE(aref); \
385 ASSERT_USIZE_EQ(vref, MEMREF_GET_USIZE(fref)); \
386 index_type *lvlCoords = MEMREF_GET_PAYLOAD(lvlCoordsRef); \
387 V *values = MEMREF_GET_PAYLOAD(vref); \
388 bool *filled = MEMREF_GET_PAYLOAD(fref); \
389 index_type *added = MEMREF_GET_PAYLOAD(aref); \
390 uint64_t expsz = vref->sizes[0]; \
391 tensor.expInsert(lvlCoords, values, filled, added, count, expsz); \
393 MLIR_SPARSETENSOR_FOREVERY_V(IMPL_EXPINSERT
)
394 #undef IMPL_EXPINSERT
396 void *_mlir_ciface_createCheckedSparseTensorReader(
397 char *filename
, StridedMemRefType
<index_type
, 1> *dimShapeRef
,
399 ASSERT_NO_STRIDE(dimShapeRef
);
400 const uint64_t dimRank
= MEMREF_GET_USIZE(dimShapeRef
);
401 const index_type
*dimShape
= MEMREF_GET_PAYLOAD(dimShapeRef
);
402 auto *reader
= SparseTensorReader::create(filename
, dimRank
, dimShape
, valTp
);
403 return static_cast<void *>(reader
);
406 void _mlir_ciface_getSparseTensorReaderDimSizes(
407 StridedMemRefType
<index_type
, 1> *out
, void *p
) {
409 SparseTensorReader
&reader
= *static_cast<SparseTensorReader
*>(p
);
410 auto *dimSizes
= const_cast<uint64_t *>(reader
.getDimSizes());
411 aliasIntoMemref(reader
.getRank(), dimSizes
, *out
);
414 #define IMPL_GETNEXT(VNAME, V, CNAME, C) \
415 bool _mlir_ciface_getSparseTensorReaderReadToBuffers##CNAME##VNAME( \
416 void *p, StridedMemRefType<index_type, 1> *dim2lvlRef, \
417 StridedMemRefType<index_type, 1> *lvl2dimRef, \
418 StridedMemRefType<C, 1> *cref, StridedMemRefType<V, 1> *vref) { \
420 auto &reader = *static_cast<SparseTensorReader *>(p); \
421 ASSERT_NO_STRIDE(dim2lvlRef); \
422 ASSERT_NO_STRIDE(lvl2dimRef); \
423 ASSERT_NO_STRIDE(cref); \
424 ASSERT_NO_STRIDE(vref); \
425 const uint64_t dimRank = reader.getRank(); \
426 const uint64_t lvlRank = MEMREF_GET_USIZE(dim2lvlRef); \
427 const uint64_t cSize = MEMREF_GET_USIZE(cref); \
428 const uint64_t vSize = MEMREF_GET_USIZE(vref); \
429 ASSERT_USIZE_EQ(lvl2dimRef, dimRank); \
430 assert(cSize >= lvlRank * vSize); \
431 assert(vSize >= reader.getNSE() && "Not enough space in buffers"); \
435 index_type *dim2lvl = MEMREF_GET_PAYLOAD(dim2lvlRef); \
436 index_type *lvl2dim = MEMREF_GET_PAYLOAD(lvl2dimRef); \
437 C *lvlCoordinates = MEMREF_GET_PAYLOAD(cref); \
438 V *values = MEMREF_GET_PAYLOAD(vref); \
439 return reader.readToBuffers<C, V>(lvlRank, dim2lvl, lvl2dim, \
440 lvlCoordinates, values); \
442 MLIR_SPARSETENSOR_FOREVERY_V_O(IMPL_GETNEXT
)
445 void _mlir_ciface_outSparseTensorWriterMetaData(
446 void *p
, index_type dimRank
, index_type nse
,
447 StridedMemRefType
<index_type
, 1> *dimSizesRef
) {
449 ASSERT_NO_STRIDE(dimSizesRef
);
450 assert(dimRank
!= 0);
451 index_type
*dimSizes
= MEMREF_GET_PAYLOAD(dimSizesRef
);
452 std::ostream
&file
= *static_cast<std::ostream
*>(p
);
453 file
<< dimRank
<< " " << nse
<< std::endl
;
454 for (index_type d
= 0; d
< dimRank
- 1; d
++)
455 file
<< dimSizes
[d
] << " ";
456 file
<< dimSizes
[dimRank
- 1] << std::endl
;
459 #define IMPL_OUTNEXT(VNAME, V) \
460 void _mlir_ciface_outSparseTensorWriterNext##VNAME( \
461 void *p, index_type dimRank, \
462 StridedMemRefType<index_type, 1> *dimCoordsRef, \
463 StridedMemRefType<V, 0> *vref) { \
465 ASSERT_NO_STRIDE(dimCoordsRef); \
466 const index_type *dimCoords = MEMREF_GET_PAYLOAD(dimCoordsRef); \
467 std::ostream &file = *static_cast<std::ostream *>(p); \
468 for (index_type d = 0; d < dimRank; d++) \
469 file << (dimCoords[d] + 1) << " "; \
470 V *value = MEMREF_GET_PAYLOAD(vref); \
471 file << *value << std::endl; \
473 MLIR_SPARSETENSOR_FOREVERY_V(IMPL_OUTNEXT
)
476 //===----------------------------------------------------------------------===//
478 // Public functions which accept only C-style data structures to interact
479 // with sparse tensors (which are only visible as opaque pointers externally).
481 //===----------------------------------------------------------------------===//
483 index_type
sparseLvlSize(void *tensor
, index_type l
) {
484 return static_cast<SparseTensorStorageBase
*>(tensor
)->getLvlSize(l
);
487 index_type
sparseDimSize(void *tensor
, index_type d
) {
488 return static_cast<SparseTensorStorageBase
*>(tensor
)->getDimSize(d
);
491 void endForwardingInsert(void *tensor
) {
492 return static_cast<SparseTensorStorageBase
*>(tensor
)->endForwardingInsert();
495 void endLexInsert(void *tensor
) {
496 return static_cast<SparseTensorStorageBase
*>(tensor
)->endLexInsert();
499 void delSparseTensor(void *tensor
) {
500 delete static_cast<SparseTensorStorageBase
*>(tensor
);
503 #define IMPL_DELCOO(VNAME, V) \
504 void delSparseTensorCOO##VNAME(void *coo) { \
505 delete static_cast<SparseTensorCOO<V> *>(coo); \
507 MLIR_SPARSETENSOR_FOREVERY_V(IMPL_DELCOO
)
510 char *getTensorFilename(index_type id
) {
511 constexpr size_t BUF_SIZE
= 80;
513 snprintf(var
, BUF_SIZE
, "TENSOR%" PRIu64
, id
);
514 char *env
= getenv(var
);
516 MLIR_SPARSETENSOR_FATAL("Environment variable %s is not set\n", var
);
520 index_type
getSparseTensorReaderNSE(void *p
) {
521 return static_cast<SparseTensorReader
*>(p
)->getNSE();
524 void delSparseTensorReader(void *p
) {
525 delete static_cast<SparseTensorReader
*>(p
);
528 void *createSparseTensorWriter(char *filename
) {
530 (filename
[0] == 0) ? &std::cout
: new std::ofstream(filename
);
531 *file
<< "# extended FROSTT format\n";
532 return static_cast<void *>(file
);
535 void delSparseTensorWriter(void *p
) {
536 std::ostream
*file
= static_cast<std::ostream
*>(p
);
538 assert(file
->good());
539 if (file
!= &std::cout
)
545 #undef MEMREF_GET_PAYLOAD
546 #undef ASSERT_USIZE_EQ
547 #undef MEMREF_GET_USIZE
548 #undef ASSERT_NO_STRIDE
550 #endif // MLIR_CRUNNERUTILS_DEFINE_FUNCTIONS