/usr/include/trilinos/Kokkos_CrsMatrix_MP_Vector_Cuda.hpp is in libtrilinos-stokhos-dev 12.10.1-3.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 | // @HEADER
// ***********************************************************************
//
// Stokhos Package
// Copyright (2009) Sandia Corporation
//
// Under terms of Contract DE-AC04-94AL85000, there is a non-exclusive
// license for use of this work by or on behalf of the U.S. Government.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// Questions? Contact Eric T. Phipps (etphipp@sandia.gov).
//
// ***********************************************************************
// @HEADER
#ifndef KOKKOS_CRSMATRIX_MP_VECTOR_CUDA_HPP
#define KOKKOS_CRSMATRIX_MP_VECTOR_CUDA_HPP
#if defined( __CUDACC__)
#include "Kokkos_CrsMatrix_MP_Vector.hpp"
#include "Kokkos_Core.hpp"
#include "Stokhos_Cuda_DeviceProp.hpp"
//----------------------------------------------------------------------------
// Specializations of Kokkos::CrsMatrix for Sacado::MP::Vector scalar type
// and Cuda device
//----------------------------------------------------------------------------
namespace Stokhos {
namespace details {
// We can make things slightly faster for the ensemble multiply kernel with
// NumPerThread == 1 by using at most 32 registers per thread for 100%
// occupancy, which we get with these launch bounds on Kepler. For
// NumPerThread > 1 it is worse though due to too many spilled registers.
template <typename Kernel>
__global__ void
#if __CUDA_ARCH__ >= 300
__launch_bounds__(1024,2)
#endif
FullOccupancyKernelLaunch(Kernel kernel) {
kernel();
}
// Kernel implementing y = A * x for Cuda device where
// A == Kokkos::CrsMatrix< Sacado::MP::Vector<...>,...>,
// x, y == Kokkos::View< Sacado::MP::Vector<...>*,...>,
// x and y are rank 1
// We spell everything out here to make sure the ranks and devices match.
//
// This implementation uses the underlying 2-D view directly.
// Currently only works for statically sized MP::Vector
template <typename MatrixStorage,
typename MatrixOrdinal,
typename MatrixMemory,
typename MatrixSize,
typename InputStorage,
typename ... InputP,
typename OutputStorage,
typename ... OutputP,
typename Update>
class MPMultiply< Kokkos::CrsMatrix<Sacado::MP::Vector<MatrixStorage>,
MatrixOrdinal,
Kokkos::Cuda,
MatrixMemory,
MatrixSize>,
Kokkos::View< Sacado::MP::Vector<InputStorage>*,
InputP... >,
Kokkos::View< Sacado::MP::Vector<OutputStorage>*,
OutputP... >,
Update,
void
>
{
public:
typedef Sacado::MP::Vector<MatrixStorage> MatrixValue;
typedef Sacado::MP::Vector<InputStorage> InputVectorValue;
typedef Sacado::MP::Vector<OutputStorage> OutputVectorValue;
typedef typename Kokkos::Cuda Device;
typedef Device execution_space;
typedef typename execution_space::size_type size_type;
typedef Kokkos::CrsMatrix<MatrixValue,
MatrixOrdinal,
execution_space,
MatrixMemory,
MatrixSize> matrix_type;
typedef typename matrix_type::values_type matrix_values_type;
typedef typename matrix_type::StaticCrsGraphType matrix_graph_type;
typedef Kokkos::View< InputVectorValue*,
InputP... > input_vector_type;
typedef Kokkos::View< OutputVectorValue*,
OutputP... > output_vector_type;
typedef Update update_type;
// Multiply that uses shared memory for coalesced access of sparse
// graph row offsets and column indices and Rank-1 vectors
template <unsigned NumPerThread>
struct Kernel {
typedef typename output_vector_type::value_type output_vector_value;
typedef typename output_vector_value::value_type scalar_type;
const matrix_graph_type m_Agraph;
const matrix_values_type m_Avals;
const input_vector_type m_x;
const output_vector_type m_y;
const update_type m_update;
const size_type m_row_count;
Kernel( const matrix_type & A,
const input_vector_type & x,
const output_vector_type & y,
const update_type& update )
: m_Agraph( A.graph )
, m_Avals( A.values )
, m_x( x )
, m_y( y )
, m_update( update )
, m_row_count( A.graph.row_map.dimension_0()-1 )
{}
__device__
inline void operator()(void) const
{
volatile size_type * const sh_row =
kokkos_impl_cuda_shared_memory<size_type>();
volatile size_type * const sh_col = sh_row + blockDim.y+1;
const size_type tid = blockDim.x*threadIdx.y + threadIdx.x;
const size_type nid = blockDim.x*blockDim.y;
const size_type block_row = blockDim.y*blockIdx.x;
// Read blockDim.y+1 row offsets in coalesced manner
const size_type num_row =
block_row+blockDim.y+1 <= m_row_count+1 ? blockDim.y+1 :
m_row_count+1 - block_row;
for (size_type i=tid; i<num_row; i+=nid)
sh_row[i] = m_Agraph.row_map[block_row+i];
__syncthreads();
const size_type iRow = block_row + threadIdx.y;
if (iRow < m_row_count) {
scalar_type sum[NumPerThread];
const size_type iEntryBegin = sh_row[threadIdx.y];
const size_type iEntryEnd = sh_row[threadIdx.y+1];
for (size_type e=0; e<NumPerThread; ++e)
sum[e] = 0;
for (size_type col_block=iEntryBegin; col_block<iEntryEnd;
col_block+=blockDim.x) {
const size_type num_col =
col_block+blockDim.x <= iEntryEnd ?
blockDim.x : iEntryEnd-col_block;
// Read blockDim.x entries column indices at a time to maintain
// coalesced accesses (don't need __syncthreads() assuming
// blockDim.x <= warp_size
// Note: it might be a little faster if we ensured aligned access
// to m_A.graph.entries() and m_A.values() below.
if (threadIdx.x < num_col)
sh_col[tid] = m_Agraph.entries(col_block+threadIdx.x);
if (blockDim.x > Kokkos::Impl::CudaTraits::WarpSize)
__syncthreads();
for ( size_type col = 0; col < num_col; ++col ) {
size_type iCol = sh_col[blockDim.x*threadIdx.y + col];
for (size_type e=0, ee=threadIdx.x; e<NumPerThread;
++e, ee+=blockDim.x) {
sum[e] += m_Avals(col_block+col).fastAccessCoeff(ee) *
m_x(iCol).fastAccessCoeff(ee);
}
}
}
for (size_type e=0, ee=threadIdx.x; e<NumPerThread;
++e, ee+=blockDim.x) {
m_update( m_y(iRow).fastAccessCoeff(ee), sum[e] );
}
}
} // operator()
};
static void apply( const matrix_type & A,
const input_vector_type & x,
const output_vector_type & y,
const update_type & update )
{
// Compute CUDA block and grid sizes.
//
// For Kepler, the best block size appears to be 256 threads with
// 16 threads per vector for double precision, yielding 16 rows per
// block. Due to register usage, this gives 64 or 48 warps per SM
// and thus 8 or 6 blocks per SM. We use these values by default if
// the user-specified block dimensions are zero
const size_type value_dimension = Kokkos::dimension_scalar(x);
size_type threads_per_vector = A.dev_config.block_dim.x;
if (threads_per_vector == 0)
threads_per_vector = value_dimension ;
size_type rows_per_block = A.dev_config.block_dim.y;
if (rows_per_block == 0)
rows_per_block = 256 / threads_per_vector;
const size_type row_count = A.graph.row_map.dimension_0()-1;
const size_type num_blocks = (row_count+rows_per_block-1)/rows_per_block;
const dim3 block( threads_per_vector, rows_per_block, 1 );
const dim3 grid( num_blocks, 1 );
// Check threads_per_vector evenly divides number of vector entries
size_type num_per_thread = value_dimension / threads_per_vector;
TEUCHOS_TEST_FOR_EXCEPTION(
num_per_thread * threads_per_vector != value_dimension, std::logic_error,
"Entries/thread * threads/vector must equal number of vector entries");
// Check threads_per_vector is not greater than warp size (kernels assume
// this)
const size_type warp_size = Kokkos::Impl::CudaTraits::WarpSize;
TEUCHOS_TEST_FOR_EXCEPTION(
threads_per_vector > warp_size, std::logic_error,
"Threads/vector cannont exceed Cuda warp size");
// Launch kernel based on static number of entries per thread
if (num_per_thread == 1) {
launch_impl<1>( A, x, y, update, block, grid );
}
else if (num_per_thread == 2) {
launch_impl<2>( A, x, y, update, block, grid );
}
else if (num_per_thread == 3) {
launch_impl<3>( A, x, y, update, block, grid );
}
else if (num_per_thread == 4) {
launch_impl<4>( A, x, y, update, block, grid );
}
else
TEUCHOS_TEST_FOR_EXCEPTION(
true, std::logic_error, "Invalid num_per_thread == " << num_per_thread);
}
private:
// Function to launch our kernel, templated on number of entries per thread
// and shared memory choice
template <unsigned num_per_thread>
static void launch_impl( const matrix_type & A,
const input_vector_type & x,
const output_vector_type & y,
const update_type & update,
dim3 block,
dim3 grid)
{
typedef Kernel<num_per_thread> Krnl;
// Use this to check occupancy, 64 is 100% on Kepler
const bool occupancy_check = false;
if (occupancy_check) {
DeviceProp device_prop;
size_type warps_per_sm;
if (num_per_thread == 1)
warps_per_sm = device_prop.get_resident_warps_per_sm(
FullOccupancyKernelLaunch<Krnl>);
else
warps_per_sm = device_prop.get_resident_warps_per_sm(
Kokkos::Impl::cuda_parallel_launch_local_memory<Krnl>);
std::cout << "warps_per_sm = " << warps_per_sm
<< " max_warps_per_sm = " << device_prop.max_warps_per_sm
<< std::endl;
}
const size_t shared = (block.y+1 + block.x*block.y)*sizeof(size_type);
if (sizeof(size_type) <= 4)
cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeFourByte);
else
cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte);
// Launch
if (num_per_thread == 1)
FullOccupancyKernelLaunch<<< grid, block, shared >>>
( Krnl( A, x, y, update ) );
else
Kokkos::Impl::cuda_parallel_launch_local_memory<<< grid, block, shared >>>
( Krnl( A, x, y, update ) );
}
};
// Kernel implementing y = A * x for Cuda device where
// A == Kokkos::CrsMatrix< Sacado::MP::Vector<...>,...>,
// x, y == Kokkos::View< Sacado::MP::Vector<...>**,...>,
// x and y are rank 2
// We spell everything out here to make sure the ranks and devices match.
//
// This implementation uses the underlying 2-D view directly.
// Currently only works for statically sized MP::Vector
template <typename MatrixStorage,
typename MatrixOrdinal,
typename MatrixMemory,
typename MatrixSize,
typename InputStorage,
typename ... InputP,
typename OutputStorage,
typename ... OutputP,
typename Update>
class MPMultiply< Kokkos::CrsMatrix<Sacado::MP::Vector<MatrixStorage>,
MatrixOrdinal,
Kokkos::Cuda,
MatrixMemory,
MatrixSize>,
Kokkos::View< Sacado::MP::Vector<InputStorage>**,
InputP... >,
Kokkos::View< Sacado::MP::Vector<OutputStorage>**,
OutputP... >,
Update,
void
>
{
public:
typedef Sacado::MP::Vector<MatrixStorage> MatrixValue;
typedef Sacado::MP::Vector<InputStorage> InputVectorValue;
typedef Sacado::MP::Vector<OutputStorage> OutputVectorValue;
typedef typename Kokkos::Cuda Device;
typedef Device execution_space;
typedef typename execution_space::size_type size_type;
typedef Kokkos::CrsMatrix<MatrixValue,
MatrixOrdinal,
execution_space,
MatrixMemory,
MatrixSize> matrix_type;
typedef typename matrix_type::values_type matrix_values_type;
typedef typename matrix_type::StaticCrsGraphType matrix_graph_type;
typedef Kokkos::View< InputVectorValue**,
InputP... > input_vector_type;
typedef Kokkos::View< OutputVectorValue**,
OutputP... > output_vector_type;
typedef Update update_type;
// Specialization that uses shared memory for coalesced access of sparse
// graph row offsets and column indices and Rank-2 vectors
template <unsigned NumPerThread>
struct Kernel {
typedef typename output_vector_type::value_type output_vector_value;
typedef typename output_vector_value::value_type scalar_type;
const matrix_graph_type m_Agraph;
const matrix_values_type m_Avals;
const input_vector_type m_x;
const output_vector_type m_y;
const update_type m_update;
const size_type m_row_count;
const size_type m_num_vec_cols;
Kernel( const matrix_type & A,
const input_vector_type & x,
const output_vector_type & y,
const update_type& update )
: m_Agraph( A.graph )
, m_Avals( A.values )
, m_x( x )
, m_y( y )
, m_update( update )
, m_row_count( A.graph.row_map.dimension_0()-1 )
, m_num_vec_cols( x.dimension_1() )
{}
__device__
inline void operator()(void) const
{
volatile size_type * const sh_row =
kokkos_impl_cuda_shared_memory<size_type>();
volatile size_type * const sh_col = sh_row + blockDim.y+1;
const size_type tid = blockDim.x*threadIdx.y + threadIdx.x;
const size_type nid = blockDim.x*blockDim.y;
const size_type block_row = blockDim.y*blockIdx.x;
// Read blockDim.y+1 row offsets in coalesced manner
const size_type num_row =
block_row+blockDim.y+1 <= m_row_count+1 ? blockDim.y+1 :
m_row_count+1 - block_row;
for (size_type i=tid; i<num_row; i+=nid)
sh_row[i] = m_Agraph.row_map[block_row+i];
__syncthreads();
const size_type iRow = block_row + threadIdx.y;
if (iRow < m_row_count) {
scalar_type sum[NumPerThread];
const size_type iEntryBegin = sh_row[threadIdx.y];
const size_type iEntryEnd = sh_row[threadIdx.y+1];
// We could potentially improve performance by putting this loop
// inside the matrix column loop, however that would require storing
// an array for sum over vector columns, which would require shared
// memory.
for (size_type vec_col=0; vec_col<m_num_vec_cols; vec_col++) {
for (size_type e=0; e<NumPerThread; ++e)
sum[e] = 0;
for (size_type col_block=iEntryBegin; col_block<iEntryEnd;
col_block+=blockDim.x) {
const size_type num_col =
col_block+blockDim.x <= iEntryEnd ?
blockDim.x : iEntryEnd-col_block;
// Read blockDim.x entries column indices at a time to maintain
// coalesced accesses (don't need __syncthreads() assuming
// blockDim.x <= warp_size
// Note: it might be a little faster if we ensured aligned access
// to m_A.graph.entries() and m_A.values() below.
if (threadIdx.x < num_col)
sh_col[tid] = m_Agraph.entries(col_block+threadIdx.x);
if (blockDim.x > Kokkos::Impl::CudaTraits::WarpSize)
__syncthreads();
for ( size_type col = 0; col < num_col; ++col ) {
size_type iCol = sh_col[blockDim.x*threadIdx.y + col];
for (size_type e=0, ee=threadIdx.x; e<NumPerThread;
++e, ee+=blockDim.x) {
sum[e] += m_Avals(col_block+col).fastAccessCoeff(ee) *
m_x(iCol, vec_col).fastAccessCoeff(ee);
}
}
}
for (size_type e=0, ee=threadIdx.x; e<NumPerThread;
++e, ee+=blockDim.x) {
m_update( m_y(iRow, vec_col).fastAccessCoeff(ee), sum[e] );
}
}
}
} // operator()
};
static void apply( const matrix_type & A,
const input_vector_type & x,
const output_vector_type & y,
const update_type & update )
{
// Compute CUDA block and grid sizes.
//
// For Kepler, the best block size appears to be 256 threads with
// 16 threads per vector for double precision, yielding 16 rows per
// block. Due to register usage, this gives 64 or 48 warps per SM
// and thus 8 or 6 blocks per SM. We use these values by default if
// the user-specified block dimensions are zero
const size_type value_dimension = dimension_scalar(x);
size_type threads_per_vector = A.dev_config.block_dim.x;
if (threads_per_vector == 0)
threads_per_vector = value_dimension ;
size_type rows_per_block = A.dev_config.block_dim.y;
if (rows_per_block == 0)
rows_per_block = 256 / threads_per_vector;
const size_type row_count = A.graph.row_map.dimension_0()-1;
const size_type num_blocks = (row_count+rows_per_block-1)/rows_per_block;
const dim3 block( threads_per_vector, rows_per_block, 1 );
const dim3 grid( num_blocks, 1 );
// Check threads_per_vector evenly divides number of vector entries
size_type num_per_thread = value_dimension / threads_per_vector;
TEUCHOS_TEST_FOR_EXCEPTION(
num_per_thread * threads_per_vector != value_dimension, std::logic_error,
"Entries/thread * threads/vector must equal number of vector entries");
// Launch kernel based on static number of entries per thread
if (num_per_thread == 1) {
launch_impl<1>( A, x, y, update, block, grid );
}
else if (num_per_thread == 2) {
launch_impl<2>( A, x, y, update, block, grid );
}
else if (num_per_thread == 3) {
launch_impl<3>( A, x, y, update, block, grid );
}
else if (num_per_thread == 4) {
launch_impl<4>( A, x, y, update, block, grid );
}
else
TEUCHOS_TEST_FOR_EXCEPTION(
true, std::logic_error, "Invalid num_per_thread == " << num_per_thread);
}
private:
// Function to launch our kernel, templated on number of entries per thread
// and shared memory choice
template <unsigned num_per_thread>
static void launch_impl( const matrix_type & A,
const input_vector_type & x,
const output_vector_type & y,
const update_type & update,
dim3 block,
dim3 grid)
{
typedef Kernel<num_per_thread> Krnl;
// Use this to check occupancy, 64 is 100% on Kepler
const bool occupancy_check = false;
if (occupancy_check) {
DeviceProp device_prop;
size_type warps_per_sm;
if (num_per_thread == 1)
warps_per_sm = device_prop.get_resident_warps_per_sm(
FullOccupancyKernelLaunch<Krnl>);
else
warps_per_sm = device_prop.get_resident_warps_per_sm(
Kokkos::Impl::cuda_parallel_launch_local_memory<Krnl>);
std::cout << "warps_per_sm = " << warps_per_sm
<< " max_warps_per_sm = " << device_prop.max_warps_per_sm
<< std::endl;
}
const size_t shared = (block.y+1 + block.x*block.y)*sizeof(size_type);
if (sizeof(size_type) <= 4)
cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeFourByte);
else
cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte);
// Launch
if (num_per_thread == 1)
FullOccupancyKernelLaunch<<< grid, block, shared >>>
( Krnl( A, x, y, update ) );
else
Kokkos::Impl::cuda_parallel_launch_local_memory<<< grid, block, shared >>>
( Krnl( A, x, y, update ) );
}
};
} // namespace details
} // namespace Stokhos
#endif /* #if defined( __CUDACC__) */
#endif /* #ifndef KOKKOS_CRSMATRIX_MP_VECTOR_CUDA_HPP */
|