/
sparse_blas_impl.cu.h
458 lines (414 loc) · 17.7 KB
/
sparse_blas_impl.cu.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/memory/malloc.h"
#include "paddle/phi/backends/dynload/cusparse.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/ddim.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/enforce.h"
#include "paddle/phi/core/sparse_coo_tensor.h"
#include "paddle/phi/core/sparse_csr_tensor.h"
#include "paddle/phi/core/visit_type.h"
namespace phi {
namespace funcs {
namespace sparse {
template <typename T>
cudaDataType_t GetGpuDataType() {
if (std::is_same<T, float>::value) {
return CUDA_R_32F;
} else if (std::is_same<T, double>::value) {
return CUDA_R_64F;
} else if (std::is_same<T, phi::dtype::float16>::value) {
return CUDA_R_16F;
}
}
inline cusparseOperation_t GetTransposeOperation(const bool trans) {
if (trans) {
return CUSPARSE_OPERATION_TRANSPOSE;
} else {
return CUSPARSE_OPERATION_NON_TRANSPOSE;
}
}
/************* SPARSE MATRIX DESCRIPTOR (COO/CSR) ************/
template <typename T, typename IntT>
inline void CreateCsrDescriptor(const phi::SparseCsrTensor& x,
const phi::GPUContext& dev_ctx,
cusparseSpMatDescr_t* descriptor) {
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
auto x_ndims = xdim_vec.size();
PADDLE_ENFORCE_GE(
x_ndims,
2,
phi::errors::InvalidArgument("the dim size of SparseCsrTensor must be "
"greater than or eaqual to 2."));
int64_t M = xdim_vec[x_ndims - 2];
int64_t N = xdim_vec[x_ndims - 1];
int batch_size = 1;
for (int i = 0; i < x_ndims - 2; i++) {
batch_size *= xdim_vec[i];
}
PADDLE_ENFORCE_EQ(x.non_zero_crows().numel(),
batch_size * (M + 1),
phi::errors::PreconditionNotMet(
"the length of SparseCsrTensor crows is not right."));
const IntT* crows_data = x.non_zero_crows().data<IntT>();
const IntT* cols_data = x.non_zero_cols().data<IntT>();
const T* values_data = x.non_zero_elements().data<T>();
int64_t batch_nnz = x.nnz() / batch_size;
cudaDataType_t gpu_type = GetGpuDataType<T>();
dev_ctx.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCreateCsr(descriptor,
M,
N,
batch_nnz,
const_cast<IntT*>(crows_data),
const_cast<IntT*>(cols_data),
const_cast<T*>(values_data),
CUSPARSE_INDEX_64I,
CUSPARSE_INDEX_64I,
CUSPARSE_INDEX_BASE_ZERO,
gpu_type);
});
if (batch_size > 1) {
#if CUDA_VERSION >= 11070
dev_ctx.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCsrSetStridedBatch(
*descriptor, batch_size, M + 1, batch_nnz);
});
#else
PADDLE_THROW(phi::errors::Unimplemented(
"Batch Sparse matmul use 'cusparseCsrSetStridedBatch', which is "
"supported from CUDA 11.7"));
#endif
}
}
template <typename T, typename IntT>
inline void CreateCooDescriptor(const phi::SparseCooTensor& x,
const phi::GPUContext& dev_ctx,
cusparseSpMatDescr_t* descriptor) {
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
auto x_ndims = xdim_vec.size();
PADDLE_ENFORCE_GE(
x_ndims,
2,
phi::errors::InvalidArgument("the dim size of SparseCsrTensor must be "
"greater than or eaqual to 2."));
int64_t M = xdim_vec[x_ndims - 2];
int64_t N = xdim_vec[x_ndims - 1];
int batch_size = 1;
for (int i = 0; i < x_ndims - 2; i++) {
batch_size *= xdim_vec[i];
}
int64_t nnz = x.nnz();
const IntT* indices_data = x.non_zero_indices().data<IntT>();
const T* values_data = x.non_zero_elements().data<T>();
auto rows_data = indices_data + (x_ndims - 2) * nnz;
auto cols_data = indices_data + (x_ndims - 1) * nnz;
int64_t batch_nnz = nnz / batch_size;
cudaDataType_t gpu_type = GetGpuDataType<T>();
dev_ctx.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCreateCoo(descriptor,
M,
N,
batch_nnz,
const_cast<IntT*>(rows_data),
const_cast<IntT*>(cols_data),
const_cast<T*>(values_data),
CUSPARSE_INDEX_64I,
CUSPARSE_INDEX_BASE_ZERO,
gpu_type);
});
if (batch_size > 1) {
#if CUDA_VERSION >= 11070
dev_ctx.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCooSetStridedBatch(
*descriptor, batch_size, batch_nnz);
});
#else
PADDLE_THROW(phi::errors::Unimplemented(
"Batch Sparse matmul use 'cusparseCooSetStridedBatch', which is "
"supported from CUDA 11.7"));
#endif
}
}
template <typename T>
class CuSparseSpMatDescriptor {
public:
explicit CuSparseSpMatDescriptor(const phi::SparseCsrTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_crows().dtype(), "Csr CuSparseSpMatDescriptor", ([&] {
CreateCsrDescriptor<T, data_t>(x, dev_ctx_, &descriptor_);
}));
VLOG(6) << "Create csr cusparseSpMatDescr_t " << &descriptor_;
}
explicit CuSparseSpMatDescriptor(const phi::SparseCooTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_indices().dtype(), "Coo CuSparseSpMatDescriptor", ([&] {
CreateCooDescriptor<T, data_t>(x, dev_ctx_, &descriptor_);
}));
VLOG(6) << "Create coo cusparseSpMatDescr_t " << &descriptor_;
}
~CuSparseSpMatDescriptor() {
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseDestroySpMat(descriptor_);
});
VLOG(6) << "Destroy cusparseSpMatDescr_t " << &descriptor_;
}
const cusparseSpMatDescr_t& descriptor() const { return descriptor_; }
private:
const phi::GPUContext& dev_ctx_;
cusparseSpMatDescr_t descriptor_;
};
/************* DENSE MATRIX DESCRIPTOR ************/
template <typename T>
class CuSparseDnMatDescriptor {
public:
explicit CuSparseDnMatDescriptor(const phi::DenseTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
auto x_ndims = xdim_vec.size();
PADDLE_ENFORCE_GE(
x_ndims,
2,
phi::errors::InvalidArgument("the dim size of DenseTensor must be "
"greater than or eaqual to 2."));
int64_t M = xdim_vec[x_ndims - 2];
int64_t N = xdim_vec[x_ndims - 1];
int batch_size = 1;
for (int i = 0; i < x_ndims - 2; i++) {
batch_size *= xdim_vec[i];
}
const T* x_data = x.data<T>();
cudaDataType_t gpu_type = GetGpuDataType<T>();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCreateDnMat(&descriptor_,
M,
N,
N,
const_cast<T*>(x_data),
gpu_type,
CUSPARSE_ORDER_ROW);
});
PADDLE_ENFORCE_EQ(x.numel(), batch_size * M * N);
if (batch_size > 1) {
#if CUDA_VERSION >= 11030
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseDnMatSetStridedBatch(
descriptor_, batch_size, M * N);
});
#else
PADDLE_THROW(phi::errors::Unimplemented(
"Batch Sparse matmul use 'cusparseDnMatSetStridedBatch', which is "
"supported from CUDA 11.7"));
#endif
}
VLOG(6) << "Create cusparseDnMatDescr_t " << &descriptor_;
}
~CuSparseDnMatDescriptor() {
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseDestroyDnMat(descriptor_);
});
VLOG(6) << "Destroy cusparseDnMatDescr_t " << &descriptor_;
}
const cusparseDnMatDescr_t& descriptor() const { return descriptor_; }
private:
const phi::GPUContext& dev_ctx_;
cusparseDnMatDescr_t descriptor_;
};
/************* DENSE VECTOR DESCRIPTOR ************/
template <typename T>
class CuSparseDnVecDescriptor {
public:
explicit CuSparseDnVecDescriptor(const phi::DenseTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
auto x_ndims = xdim_vec.size();
PADDLE_ENFORCE_GE(x_ndims,
1,
phi::errors::InvalidArgument(
"the dim size of Vec must be eaqual to 1."));
const T* x_data = x.data<T>();
cudaDataType_t gpu_type = GetGpuDataType<T>();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCreateDnVec(
&descriptor_, x.numel(), const_cast<T*>(x_data), gpu_type);
});
VLOG(6) << "Create cusparseDnVecDescr_t " << &descriptor_;
}
~CuSparseDnVecDescriptor() {
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseDestroyDnVec(descriptor_);
});
VLOG(6) << "Destroy cusparseDnVecDescr_t " << &descriptor_;
}
const cusparseDnVecDescr_t& descriptor() const { return descriptor_; }
private:
const phi::GPUContext& dev_ctx_;
cusparseDnVecDescr_t descriptor_;
};
template <>
template <typename T, typename TensorType>
void SparseBlas<phi::GPUContext>::SPMM(bool transa,
bool transb,
T alpha,
const TensorType& mat_a,
const phi::DenseTensor& mat_b,
T beta,
phi::DenseTensor* mat_out) const {
auto a_descriptor = CuSparseSpMatDescriptor<T>(mat_a, dev_ctx_);
auto b_descriptor = CuSparseDnMatDescriptor<T>(mat_b, dev_ctx_);
auto out_descriptor = CuSparseDnMatDescriptor<T>(*mat_out, dev_ctx_);
cudaDataType_t gpu_type = GetGpuDataType<T>();
size_t buffer_size = 0;
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSpMM_bufferSize(handle,
GetTransposeOperation(transa),
GetTransposeOperation(transb),
&alpha,
a_descriptor.descriptor(),
b_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_SPMM_ALG_DEFAULT,
&buffer_size);
});
paddle::memory::allocation::AllocationPtr tmp_buffer =
paddle::memory::Alloc(dev_ctx_, buffer_size);
void* tmp_buffer_ptr = tmp_buffer->ptr();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSpMM(handle,
GetTransposeOperation(transa),
GetTransposeOperation(transb),
&alpha,
a_descriptor.descriptor(),
b_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_SPMM_ALG_DEFAULT,
tmp_buffer_ptr);
});
}
template <>
template <typename T, typename TensorType>
void SparseBlas<phi::GPUContext>::SPMV(bool transa,
T alpha,
const TensorType& mat_a,
const phi::DenseTensor& vec_x,
T beta,
phi::DenseTensor* vec_out) const {
auto a_descriptor = CuSparseSpMatDescriptor<T>(mat_a, dev_ctx_);
auto x_descriptor = CuSparseDnVecDescriptor<T>(vec_x, dev_ctx_);
auto out_descriptor = CuSparseDnVecDescriptor<T>(*vec_out, dev_ctx_);
cudaDataType_t gpu_type = GetGpuDataType<T>();
size_t buffer_size = 0;
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSpMV_bufferSize(handle,
GetTransposeOperation(transa),
&alpha,
a_descriptor.descriptor(),
x_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_MV_ALG_DEFAULT,
&buffer_size);
});
paddle::memory::allocation::AllocationPtr tmp_buffer =
paddle::memory::Alloc(dev_ctx_, buffer_size);
void* tmp_buffer_ptr = tmp_buffer->ptr();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSpMV(handle,
GetTransposeOperation(transa),
&alpha,
a_descriptor.descriptor(),
x_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_MV_ALG_DEFAULT,
tmp_buffer_ptr);
});
}
#if CUDA_VERSION >= 11030
template <>
template <typename T, typename TensorType>
void SparseBlas<phi::GPUContext>::SDDMM(bool transa,
bool transb,
T alpha,
const phi::DenseTensor& mat_a,
const phi::DenseTensor& mat_b,
T beta,
TensorType* mat_out) const {
auto a_descriptor = CuSparseDnMatDescriptor<T>(mat_a, dev_ctx_);
auto b_descriptor = CuSparseDnMatDescriptor<T>(mat_b, dev_ctx_);
auto out_descriptor = CuSparseSpMatDescriptor<T>(*mat_out, dev_ctx_);
cudaDataType_t gpu_type = GetGpuDataType<T>();
size_t buffer_size = 0;
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSDDMM_bufferSize(handle,
GetTransposeOperation(transa),
GetTransposeOperation(transb),
&alpha,
a_descriptor.descriptor(),
b_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_SDDMM_ALG_DEFAULT,
&buffer_size);
});
paddle::memory::allocation::AllocationPtr tmp_buffer =
paddle::memory::Alloc(dev_ctx_, buffer_size);
void* tmp_buffer_ptr = tmp_buffer->ptr();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSDDMM_preprocess(handle,
GetTransposeOperation(transa),
GetTransposeOperation(transb),
&alpha,
a_descriptor.descriptor(),
b_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_SDDMM_ALG_DEFAULT,
tmp_buffer_ptr);
});
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSDDMM(handle,
GetTransposeOperation(transa),
GetTransposeOperation(transb),
&alpha,
a_descriptor.descriptor(),
b_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_SDDMM_ALG_DEFAULT,
tmp_buffer_ptr);
});
}
#endif
} // namespace sparse
} // namespace funcs
} // namespace phi