Spaces:
Sleeping
Sleeping
File size: 38,003 Bytes
e45d058 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 |
/***************************************************************************************************
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* 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 copyright holder nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "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 THE COPYRIGHT HOLDER OR 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.
*
**************************************************************************************************/
/*! \file
\brief CUTLASS Library handle.
*/
#include <iostream>
#include <stdexcept>
#include <cstdint>
#include "cutlass/library/handle.h"
#include "cutlass/library/singleton.h"
#include "cutlass/library/util.h"
namespace cutlass {
namespace library {
///////////////////////////////////////////////////////////////////////////////////////////////////
/// Constructor
Handle::Handle(
cudaStream_t stream,
size_t workspace_size
):
provider_(Provider::kCUTLASS),
stream_(stream),
workspace_(nullptr),
workspace_size_(0),
scalar_pointer_mode_(ScalarPointerMode::kHost),
last_operation_(nullptr) {
int device_idx = -1;
cudaError_t error = cudaGetDevice(&device_idx);
if (error != cudaSuccess) {
throw std::runtime_error("cudaGetDevice() failed");
}
error = cudaGetDeviceProperties(&device_, device_idx);
if (error != cudaSuccess) {
throw std::runtime_error("cudaGetDeviceProperties() failed");
}
set_workspace_size(workspace_size);
Singleton::get();
}
/// Destructor
Handle::~Handle() {
if (workspace_) {
if (workspace_) {
cudaFree(workspace_);
}
workspace_ = nullptr;
workspace_size_ = 0;
}
}
/// Move constructor
Handle::Handle(Handle && handle) {
device_ = handle.device_;
workspace_size_ = handle.workspace_size_;
workspace_ = handle.workspace_;
stream_ = handle.stream_;
scalar_pointer_mode_ = handle.scalar_pointer_mode_;
handle.workspace_ = nullptr;
handle.workspace_size_ = 0;
}
/// Move assignment operator
Handle & Handle::operator=(Handle && handle) {
provider_ = handle.provider_;
device_ = handle.device_;
workspace_size_ = handle.workspace_size_;
workspace_ = handle.workspace_;
stream_ = handle.stream_;
scalar_pointer_mode_ = handle.scalar_pointer_mode_;
handle.workspace_ = nullptr;
handle.workspace_size_ = 0;
return *this;
}
int Handle::compute_capability() const {
return device_.major * 10 + device_.minor;
}
/// Sets the current CUDA stream
void Handle::set_stream(cudaStream_t stream) {
stream_ = stream;
}
/// Gets the current CUDA stream
cudaStream_t Handle::get_stream() const {
return stream_;
}
/// Gets the current provider
Provider Handle::get_provider() const {
return provider_;
}
/// Sets the provider of operations
void Handle::set_provider(Provider provider) {
provider_ = provider;
}
/// Gets the device workspace size
size_t Handle::get_workspace_size() const {
return workspace_size_;
}
/// Gets a pointer to the device workspace allocation in Global Memory
void *Handle::get_workspace() const {
return workspace_;
}
/// Sets the size of device workspace, invalidating previous calls to get_device_workspace()
void Handle::set_workspace_size(size_t bytes) {
if (bytes != workspace_size_) {
if (workspace_) {
cudaFree(workspace_);
}
workspace_ = nullptr;
workspace_size_ = bytes;
if (workspace_size_) {
cudaError_t error = cudaMalloc((void **)&workspace_, workspace_size_);
if (error != cudaSuccess) {
throw std::runtime_error("Failed to allocate workspace");
}
}
}
if (workspace_) {
cudaError_t error = cudaMemset(workspace_, 0, workspace_size_);
if (error != cudaSuccess) {
throw std::runtime_error("Failed to clear workspace");
}
}
}
/// Gets the scalar pointer mode
ScalarPointerMode Handle::get_scalar_pointer_mode() const {
return scalar_pointer_mode_;
}
/// Sets the scalar pointer mode
void Handle::set_scalar_pointer_mode(ScalarPointerMode mode) {
scalar_pointer_mode_ = mode;
}
/// Gets the last operation
Operation const *Handle::get_last_operation() const {
return last_operation_;
}
///////////////////////////////////////////////////////////////////////////////////////////////////
/// Returns the maximum required alignment for each operator
static int maximum_alignment_requirement(GemmDescription const &desc) {
return std::max(
std::max(desc.A.alignment, desc.B.alignment), desc.C.alignment);
}
/// Returns the largest alignment (in units of elements) the problem satisfies, starting from a
/// given upper limit.
static int gemm_problem_alignment(
int M,
int N,
int K,
NumericTypeID element_A,
void const *ptr_A,
int64_t lda,
int64_t batch_stride_A,
NumericTypeID element_B,
void const *ptr_B,
int64_t ldb,
int64_t batch_stride_B,
NumericTypeID element_C,
void const * ptr_C,
int64_t ldc,
int64_t batch_stride_C,
void const * ptr_D,
int64_t ldd,
int64_t batch_stride_D,
int max_alignment_in_bytes = 16
) {
void const *pointers[] = {
ptr_A, ptr_B, ptr_C, ptr_D
};
int64_t extents[] = {
M, N, K, lda, ldb, ldc, ldd, batch_stride_A, batch_stride_B, batch_stride_C, batch_stride_D
};
NumericTypeID elements[] = {
element_A, element_B, element_C
};
for (; max_alignment_in_bytes > 0; max_alignment_in_bytes /= 2) {
bool satisfied = true;
// Can pointers satisfy this?
for (void const *ptr : pointers) {
std::uintptr_t int_ptr = reinterpret_cast<std::uintptr_t>(ptr);
if (int_ptr % max_alignment_in_bytes) {
satisfied = false;
break;
}
}
if (!satisfied) {
continue;
}
// Compute the maximum alignment based on element data types
int max_element_alignment = 0;
for (NumericTypeID type_id : elements) {
int element_alignment = max_alignment_in_bytes * 8 / library::sizeof_bits(type_id);
max_element_alignment = std::max(max_element_alignment, element_alignment);
}
// Can the problem size and leading dimensions satisfy this?
for (int64_t extent : extents) {
if (extent % max_element_alignment) {
satisfied = false;
break;
}
}
if (!satisfied) {
continue;
}
// Yes
return max_element_alignment;
}
// No alignment satisfies this problem
return 0;
}
/// Find the best kernel in descending order of preference.
static Operation const * find_gemm_operation(
GemmOperationFunctionalMap::const_iterator operators_it,
GemmPreferenceKey const preference_key) {
auto cc_it = operators_it->second.upper_bound(preference_key);
if (cc_it == operators_it->second.begin()) {
return nullptr;
}
Operation const *operation = nullptr;
// Search in descending order of compute capability
do {
--cc_it;
// Search tile sizes in order, for now.
for (auto const * op : cc_it->second) {
GemmDescription const &desc = static_cast<GemmDescription const &>(op->description());
int min_cc = desc.tile_description.minimum_compute_capability;
int max_cc = desc.tile_description.maximum_compute_capability;
int op_alignment = maximum_alignment_requirement(desc);
if ((min_cc <= preference_key.compute_capability) &&
(preference_key.compute_capability <= max_cc) &&
(op_alignment <= preference_key.alignment)) {
operation = op;
break;
}
}
} while (!operation && cc_it != operators_it->second.begin());
return operation;
}
///////////////////////////////////////////////////////////////////////////////////////////////////
/// Executes a GEMM computation: D <= alpha * A*B + beta * C
Status Handle::gemm(
int M, /// GEMM M dimension
int N, /// GEMM N dimension
int K, /// GEMM K dimension
NumericTypeID element_compute, /// Data type of internal accumulation
NumericTypeID element_scalar, /// Data type of alpha/beta scalars
void const *alpha, /// Pointer to alpha scalar
NumericTypeID element_A, /// Data type of A matrix elements
LayoutTypeID layout_A, /// Layout of A matrix
ComplexTransform transform_A, /// Complex transformation applied to A matrix - ignored for real-valued matrices
void const * ptr_A, /// Pointer to A matrix in Global Memory
int64_t lda, /// Leading dimension of A matrix
NumericTypeID element_B, /// Data type of B matrix elements
LayoutTypeID layout_B, /// Layout of B matrix
ComplexTransform transform_B, /// Complex transformation applied to B matrix - ignored for real-valued matrices
void const * ptr_B, /// Pointer to B matrix in Global Memory
int64_t ldb, /// Leading dimension of B matrix
void const * beta, /// Pointer to beta scalar
NumericTypeID element_C, /// Data type of C and D matrices
void const * ptr_C, /// Pointer to C matrix
int64_t ldc, /// Leading dimension of C matrix
void * ptr_D, /// Pointer to D matrix
int64_t ldd /// Leading dimension of D matrix
) {
//
// Find the operation
//
GemmFunctionalKey key(
provider_,
GemmKind::kGemm,
element_compute,
element_scalar,
element_A,
layout_A,
transform_A,
element_B,
layout_B,
transform_B,
element_C, // C/D are same type and col major default
LayoutTypeID::kColumnMajor,
element_C,
LayoutTypeID::kColumnMajor
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
if (operators_it == Singleton::get().operation_table.gemm_operations.end()) {
return cutlass::Status::kErrorNotSupported;
}
if (operators_it->second.empty()) {
return cutlass::Status::kErrorNotSupported;
}
//
// Compute the largest alignment restriction the kernel can satisfy.
//
// Maximum alignment expectation among all kernels (in units of bytes)
int const kMaximumAlignmentSize = 16;
int alignment = gemm_problem_alignment(
M, N, K,
element_A, ptr_A, lda, 0,
element_B, ptr_B, ldb, 0,
element_C, ptr_C, ldc, 0,
ptr_D, ldd, 0, kMaximumAlignmentSize
);
//
// Find the best kernel in descending order of preference.
//
GemmPreferenceKey preference_key(compute_capability(), alignment);
Operation const *operation = find_gemm_operation(operators_it, preference_key);
if (!operation) {
return cutlass::Status::kErrorNotSupported;
}
last_operation_ = operation;
//
// Configure operation
//
GemmConfiguration configuration{
{M, N, K},
lda,
ldb,
ldc,
ldd,
1
};
// Query host work space size
uint64_t host_workspace_size_needed = operation->get_host_workspace_size(&configuration);
if (uint64_t(kHostWorkspaceSize) < host_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
char host_workspace[kHostWorkspaceSize];
// Query device workspace size
uint64_t device_workspace_size_needed = operation->get_device_workspace_size(&configuration);
if (uint64_t(workspace_size_) < device_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
// Initialize host and device workspaces
Status status = operation->initialize(
&configuration,
host_workspace,
workspace_,
stream_);
if (status != cutlass::Status::kSuccess) {
return status;
}
// Run the operator
GemmArguments arguments{
ptr_A,
ptr_B,
ptr_C,
ptr_D,
alpha,
beta,
scalar_pointer_mode_
};
return operation->run(&arguments, host_workspace, workspace_, stream_);
}
///////////////////////////////////////////////////////////////////////////////////////////////////
/// Executes a GEMM computation: D <= alpha * A*B + beta * C.
//
// Supports batched-strided, batched array or split-K serial or split-K parallel.
//
Status Handle::gemm_universal(
GemmUniversalMode mode, /// indicates the mode in which the kUniversal GEMM is launched
int M, /// GEMM M dimension
int N, /// GEMM N dimension
int K, /// GEMM K dimension
NumericTypeID element_compute, /// Data type of internal accumulation
NumericTypeID element_scalar, /// Data type of alpha/beta scalars
void const *alpha, /// Pointer to alpha scalar
NumericTypeID element_A, /// Data type of A matrix elements
LayoutTypeID layout_A, /// Layout of A matrix
ComplexTransform transform_A, /// Complex transformation applied to A matrix - ignored for real-valued matrices
void const * ptr_A, /// Pointer to A matrix in Global Memory
int64_t lda, /// Leading dimension of A matrix
NumericTypeID element_B, /// Data type of B matrix elements
LayoutTypeID layout_B, /// Layout of B matrix
ComplexTransform transform_B, /// Complex transformation applied to B matrix - ignored for real-valued matrices
void const * ptr_B, /// Pointer to B matrix in Global Memory
int64_t ldb, /// Leading dimension of B matrix
void const * beta, /// Pointer to beta scalar
NumericTypeID element_C, /// Data type of C matrix
LayoutTypeID layout_C, /// Layout of D matrix
void const * ptr_C, /// Pointer to C matrix
int64_t ldc, /// Leading dimension of C matrix
NumericTypeID element_D, /// Data type of D matrix
LayoutTypeID layout_D, /// Layout of D matrix
void * ptr_D, /// Pointer to D matrix
int64_t ldd, /// Leading dimension of D matrix
int batch_count, /// Batch count or number of split-K slices
int64_t batch_stride_A, /// Batch stride of A operand
int64_t batch_stride_B, /// Batch stride of B operand
int64_t batch_stride_C, /// Batch stride of C operand
int64_t batch_stride_D /// Batch stride of D operand
) {
//
// Find the operation
//
GemmFunctionalKey key(
provider_,
GemmKind::kUniversal,
element_compute,
element_scalar,
element_A,
layout_A,
transform_A,
element_B,
layout_B,
transform_B,
element_C,
layout_C,
element_D,
layout_D
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
if (operators_it == Singleton::get().operation_table.gemm_operations.end()) {
return cutlass::Status::kErrorNotSupported;
}
if (operators_it->second.empty()) {
return cutlass::Status::kErrorNotSupported;
}
//
// Compute the largest alignment restriction the kernel can satisfy.
//
// Maximum alignment expectation among all kernels (in units of bytes)
int const kMaximumAlignmentSize = 16;
void const *ptr_A_check = ptr_A;
void const *ptr_B_check = ptr_B;
void const *ptr_C_check = ptr_C;
void * ptr_D_check = ptr_D;
// Ignore alignment of pointers to pointers. We can't check this from the host,
// as each batch index has its own pointer in device memory.
if (mode == GemmUniversalMode::kArray) {
ptr_A_check = nullptr;
ptr_B_check = nullptr;
ptr_C_check = nullptr;
ptr_D_check = nullptr;
}
int alignment = gemm_problem_alignment(
M, N, K,
element_A, ptr_A_check, lda, 0,
element_B, ptr_B_check, ldb, 0,
element_C, ptr_C_check, ldc, 0,
ptr_D_check, ldd, 0, kMaximumAlignmentSize
);
//
// Find the best kernel in descending order of preference.
//
GemmPreferenceKey preference_key(compute_capability(), alignment);
Operation const *operation = find_gemm_operation(operators_it, preference_key);
if (!operation) {
return cutlass::Status::kErrorNotSupported;
}
last_operation_ = operation;
//
// Configure operation
//
GemmUniversalConfiguration configuration{
mode,
{M, N, K},
batch_count,
lda,
ldb,
ldc,
ldd
};
// Query host work space size
uint64_t host_workspace_size_needed = operation->get_host_workspace_size(&configuration);
if (uint64_t(kHostWorkspaceSize) < host_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
char host_workspace[kHostWorkspaceSize];
GemmUniversalArguments arguments{
{M, N, K},
batch_count,
ptr_A,
ptr_B,
ptr_C,
ptr_D,
alpha,
beta,
scalar_pointer_mode_,
lda,
ldb,
ldc,
ldd,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_stride_D
};
// Query device workspace size
uint64_t device_workspace_size_needed = operation->get_device_workspace_size(&configuration, &arguments);
if (uint64_t(workspace_size_) < device_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
// Initialize host and device workspaces
Status status = operation->initialize(
&configuration,
host_workspace,
workspace_,
stream_);
if (status != cutlass::Status::kSuccess) {
return status;
}
// Run the operator
return operation->run(&arguments, host_workspace, workspace_, stream_);
}
///////////////////////////////////////////////////////////////////////////////////////////////////
/// Planar complex GEMM
Status Handle::gemm_planar_complex(
int M, /// GEMM M dimension
int N, /// GEMM N dimension
int K, /// GEMM K dimension
NumericTypeID element_compute, /// Data type of internal accumulation
NumericTypeID element_scalar, /// Data type of alpha/beta scalars
void const *alpha, /// Pointer to alpha scalar
NumericTypeID element_A, /// Data type of A matrix elements
LayoutTypeID layout_A, /// Layout of A matrix
ComplexTransform transform_A, /// Complex transformation applied to A matrix
void const * ptr_A_real, /// Pointer to real part of A matrix
void const * ptr_A_imag, /// Pointer to imaginary part of A matrix
int64_t lda_real, /// Leading dimension of real part of A matrix
int64_t lda_imag, /// Leading dimension of imaginary part of A matrix
NumericTypeID element_B, /// Data type of B matrix elements
LayoutTypeID layout_B, /// Layout of B matrix
ComplexTransform transform_B, /// Complex transformation applied to B matrix
void const * ptr_B_real, /// Pointer to real part of B matrix
void const * ptr_B_imag, /// Pointer to imaginary part of B matrix
int64_t ldb_real, /// Leading dimension of real part of B matrix
int64_t ldb_imag, /// Leading dimension of imaginary part of B matrix
void const * beta, /// Pointer to beta scalar
NumericTypeID element_C, /// Data type of C and D matrix
void const * ptr_C_real, /// Pointer to real part of C matrix
void const * ptr_C_imag, /// Pointer to imaginary part of C matrix
int64_t ldc_real, /// Leading dimension of real part of C matrix
int64_t ldc_imag, /// Leading dimension of imaginary part of C matrix
void * ptr_D_real, /// Pointer to real part of D matrix
void * ptr_D_imag, /// Pointer to imaginary part of D matrix
int64_t ldd_real, /// Leading dimension of real part of D matrix
int64_t ldd_imag, /// Leading dimension of imaginary part of D matrix
int batch_count, /// Number of batched GEMMs to execute
int64_t batch_stride_A_real,
int64_t batch_stride_A_imag,
int64_t batch_stride_B_real,
int64_t batch_stride_B_imag,
int64_t batch_stride_C_real,
int64_t batch_stride_C_imag,
int64_t batch_stride_D_real,
int64_t batch_stride_D_imag
) {
//
// Find the operation
//
GemmFunctionalKey key(
provider_,
GemmKind::kPlanarComplex,
element_compute,
element_scalar,
element_A,
layout_A,
transform_A,
element_B,
layout_B,
transform_B,
element_C, // C/D are same type
LayoutTypeID::kColumnMajor,
element_C,
LayoutTypeID::kColumnMajor
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
if (operators_it == Singleton::get().operation_table.gemm_operations.end()) {
return cutlass::Status::kErrorNotSupported;
}
if (operators_it->second.empty()) {
return cutlass::Status::kErrorNotSupported;
}
//
// Compute the largest alignment restriction the kernel can satisfy.
//
// Maximum alignment expectation among all kernels (in units of bytes)
int const kMaximumAlignmentSize = 16;
int alignment = std::max(
gemm_problem_alignment(
M, N, K,
element_A, ptr_A_real, lda_real, batch_stride_A_real,
element_B, ptr_B_real, ldb_real, batch_stride_B_real,
element_C, ptr_C_real, ldc_real, batch_stride_C_real,
ptr_D_real, ldd_real, batch_stride_D_real, kMaximumAlignmentSize
),
gemm_problem_alignment(
M, N, K,
element_A, ptr_A_imag, lda_imag, batch_stride_A_imag,
element_B, ptr_B_imag, ldb_imag, batch_stride_B_imag,
element_C, ptr_C_imag, ldc_imag, batch_stride_C_imag,
ptr_D_imag, ldd_imag, batch_stride_D_imag, kMaximumAlignmentSize
)
);
//
// Find the best kernel in descending order of preference.
//
GemmPreferenceKey preference_key(compute_capability(), alignment);
Operation const *operation = find_gemm_operation(operators_it, preference_key);
if (!operation) {
return cutlass::Status::kErrorNotSupported;
}
last_operation_ = operation;
//
// Configure operation
//
GemmPlanarComplexConfiguration configuration{
GemmUniversalMode::kBatched,
{M, N, K},
batch_count,
lda_real,
lda_imag,
ldb_real,
ldb_imag,
ldc_real,
ldc_imag,
ldd_real,
ldd_imag
};
// Query host work space size
uint64_t host_workspace_size_needed = operation->get_host_workspace_size(&configuration);
if (uint64_t(kHostWorkspaceSize) < host_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
char host_workspace[kHostWorkspaceSize];
// Query device workspace size
uint64_t device_workspace_size_needed = operation->get_device_workspace_size(&configuration);
if (uint64_t(workspace_size_) < device_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
// Initialize host and device workspaces
Status status = operation->initialize(
&configuration,
host_workspace,
workspace_,
stream_);
if (status != cutlass::Status::kSuccess) {
return status;
}
// Run the operator
GemmPlanarComplexArguments arguments{
ptr_A_real,
ptr_A_imag,
ptr_B_real,
ptr_B_imag,
ptr_C_real,
ptr_C_imag,
ptr_D_real,
ptr_D_imag,
alpha,
beta,
scalar_pointer_mode_,
batch_stride_A_real,
batch_stride_A_imag,
batch_stride_B_real,
batch_stride_B_imag,
batch_stride_C_real,
batch_stride_C_imag,
batch_stride_D_real,
batch_stride_D_imag
};
return operation->run(&arguments, host_workspace, workspace_, stream_);
}
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Planar complex batched GEMM loading pointers from arrays in global memory
Status Handle::gemm_planar_complex_array(
int expected_M, /// Expected GEMM M dimension (used for sizing CUDA grid)
int expected_N, /// Expected GEMM N dimension (used for sizing CUDA grid)
int expected_K, /// Expected GEMM K dimension
int batch_count, /// Number of independent GEMM computations to execute
int const *M, /// Array containing the GEMM M dimension for each batch index
int const *N, /// Array containing the GEMM N dimension for each batch index
int const *K, /// Array containing the GEMM K dimension for each batch index
NumericTypeID element_compute, /// Data type of internal accumulation
NumericTypeID element_scalar, /// Data type of alpha/beta scalars
void const *alpha, /// Pointer to alpha scalar
NumericTypeID element_A, /// Data type of A matrix elements
LayoutTypeID layout_A, /// Layout of A matrix
ComplexTransform transform_A, /// Complex transformation applied to A matrix
void const * const * ptr_A_real, /// Pointer to array containing pointers to real part of A matrices
void const * const * ptr_A_imag, /// Pointer to array containing pointers to imaginary part of A matrices
int64_t lda_real, /// Leading dimension of real part of A matrix
int64_t lda_imag, /// Leading dimension of imaginary part of A matrix
NumericTypeID element_B, /// Data type of B matrix elements
LayoutTypeID layout_B, /// Layout of B matrix
ComplexTransform transform_B, /// Complex transformation applied to B matrix
void const * const * ptr_B_real, /// Pointer to array containing pointers to real part of B matrices
void const * const * ptr_B_imag, /// Pointer to array containing pointers to imaginary part of B matrices
int64_t ldb_real, /// Leading dimension of real part of B matrix
int64_t ldb_imag, /// Leading dimension of imaginary part of B matrix
void const * beta, /// Pointer to beta scalar
NumericTypeID element_C, /// Data type of C and D matrix
void const * const * ptr_C_real, /// Pointer to array containing pointers to real part of C matrices
void const * const * ptr_C_imag, /// Pointer to array containing pointers to imaginary part of C matrices
int64_t ldc_real, /// Leading dimension of real part of C matrix
int64_t ldc_imag, /// Leading dimension of imaginary part of C matrix
void * const * ptr_D_real, /// Pointer to array containing pointers to real part of D matrices
void * const * ptr_D_imag, /// Pointer to array containing pointers to imaginary part of D matrices
int64_t ldd_real, /// Leading dimension of real part of D matrix
int64_t ldd_imag /// Leading dimension of imaginary part of D matrix
) {
//
// Find the operation
//
GemmFunctionalKey key(
provider_,
GemmKind::kPlanarComplexArray,
element_compute,
element_scalar,
element_A,
layout_A,
transform_A,
element_B,
layout_B,
transform_B,
element_C, // C/D are same type
LayoutTypeID::kColumnMajor,
element_C,
LayoutTypeID::kColumnMajor
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
if (operators_it == Singleton::get().operation_table.gemm_operations.end()) {
return cutlass::Status::kErrorNotSupported;
}
if (operators_it->second.empty()) {
return cutlass::Status::kErrorNotSupported;
}
//
// Compute the largest alignment restriction the kernel can satisfy.
//
// Maximum alignment expectation among all kernels (in units of bytes)
int const kMaximumAlignmentSize = 16;
int alignment = std::max(
gemm_problem_alignment(
expected_M, expected_N, expected_K,
element_A, nullptr, lda_real, 0,
element_B, nullptr, ldb_real, 0,
element_C, nullptr, ldc_real, 0,
nullptr, ldd_real, 0, kMaximumAlignmentSize
),
gemm_problem_alignment(
expected_M, expected_N, expected_K,
element_A, nullptr, lda_imag, 0,
element_B, nullptr, ldb_imag, 0,
element_C, nullptr, ldc_imag, 0,
nullptr, ldd_imag, 0, kMaximumAlignmentSize
)
);
//
// Find the best kernel in descending order of preference.
//
GemmPreferenceKey preference_key(compute_capability(), alignment);
Operation const *operation = find_gemm_operation(operators_it, preference_key);
if (!operation) {
return cutlass::Status::kErrorNotSupported;
}
last_operation_ = operation;
//
// Configure operation
//
GemmPlanarComplexArrayConfiguration configuration{
{expected_M, expected_N, expected_K},
batch_count,
lda_real,
lda_imag,
ldb_real,
ldb_imag,
ldc_real,
ldc_imag,
ldd_real,
ldd_imag
};
// Query host work space size
uint64_t host_workspace_size_needed = operation->get_host_workspace_size(&configuration);
if (uint64_t(kHostWorkspaceSize) < host_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
char host_workspace[kHostWorkspaceSize];
// Query device workspace size
uint64_t device_workspace_size_needed = operation->get_device_workspace_size(&configuration);
if (uint64_t(workspace_size_) < device_workspace_size_needed) {
return cutlass::Status::kErrorNotSupported;
}
// Initialize host and device workspaces
Status status = operation->initialize(
&configuration,
host_workspace,
workspace_,
stream_);
if (status != cutlass::Status::kSuccess) {
return status;
}
// Run the operator
GemmPlanarComplexArrayArguments arguments{
M, N, K,
ptr_A_real,
ptr_A_imag,
ptr_B_real,
ptr_B_imag,
ptr_C_real,
ptr_C_imag,
ptr_D_real,
ptr_D_imag,
alpha,
beta,
scalar_pointer_mode_
};
return operation->run(&arguments, host_workspace, workspace_, stream_);
}
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Finds conv operation instances with Conv::ElementC = Reduction::ElementWorkspace
Operation const* find_conv_operation_for_parallel_reduction(Operation const *operation) {
ConvDescription const &conv_desc =
static_cast<ConvDescription const &>(operation->description());
// if the curren conv operation accumulator and output data type match return operation
if(conv_desc.tile_description.math_instruction.element_accumulator == conv_desc.C.element) {
return operation;
}
// find conv operation to match conv output and reduction workspace data type
ConvFunctionalKey key(
library::Provider::kCUTLASS,
conv_desc.conv_kind,
conv_desc.A.element,
conv_desc.A.layout,
conv_desc.B.element,
conv_desc.B.layout,
conv_desc.tile_description.math_instruction.element_accumulator,
conv_desc.C.layout,
conv_desc.tile_description.math_instruction.element_accumulator,
conv_desc.element_epilogue);
// conv operation table for conv2d or conv3d
auto conv_operations = (conv_desc.kind == OperationKind::kConv2d) ?
Singleton::get().operation_table.conv2d_operations :
Singleton::get().operation_table.conv3d_operations;
// find ConvFunctionalKey in convolution operation table
auto operators_it = conv_operations.find(key);
if (operators_it == conv_operations.end()) {
return nullptr;
}
if (operators_it->second.empty()) {
return nullptr;
}
// conv operation for same compute capability and iterator algorithm
ConvPreferenceKey preference_key(
conv_desc.tile_description.minimum_compute_capability,
conv_desc.iterator_algorithm);
auto it = operators_it->second.find(preference_key);
if(it == operators_it->second.end()) {
return nullptr;
}
// return matching conv opertion (same tile sizes and instruction)
for (auto op : it->second) {
if (op->description().tile_description == operation->description().tile_description) {
return op;
}
}
return nullptr;
}
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Finds gemm operation instances with Gemm::ElementC = Reduction::ElementWorkspace
Operation const* find_gemm_operation_for_parallel_reduction(Operation const *operation) {
GemmDescription const &gemm_desc =
static_cast<GemmDescription const &>(operation->description());
// if the curren gemm operation accumulator and output data type match return operation
if(gemm_desc.tile_description.math_instruction.element_accumulator == gemm_desc.D.element) {
return operation;
}
// find gemm operation to match gemm output and reduction workspace data type
GemmFunctionalKey key(
library::Provider::kCUTLASS,
gemm_desc.gemm_kind,
gemm_desc.tile_description.math_instruction.element_accumulator,
gemm_desc.element_epilogue,
gemm_desc.A.element,
gemm_desc.A.layout,
gemm_desc.transform_A,
gemm_desc.B.element,
gemm_desc.B.layout,
gemm_desc.transform_B,
gemm_desc.tile_description.math_instruction.element_accumulator, // C/D are same type
LayoutTypeID::kColumnMajor,
gemm_desc.tile_description.math_instruction.element_accumulator,
LayoutTypeID::kColumnMajor);
// gemm operation table
auto gemm_operations = Singleton::get().operation_table.gemm_operations;
// find ConvFunctionalKey in gemm operation table
auto operators_it = gemm_operations.find(key);
if (operators_it == gemm_operations.end()) {
return nullptr;
}
if (operators_it->second.empty()) {
return nullptr;
}
// gemm operation for same compute capability and max operand alignment
int alignment = std::max(
gemm_desc.A.alignment,
gemm_desc.B.alignment);
GemmPreferenceKey preference_key(
gemm_desc.tile_description.minimum_compute_capability,
alignment);
auto it = operators_it->second.find(preference_key);
if(it == operators_it->second.end()) {
return nullptr;
}
// return matching gemm opertion (same tile shape, stages, warp count, and instruction)
for (auto op : it->second) {
if (op->description().tile_description == operation->description().tile_description) {
return op;
}
}
// return nullptr if no matching gemm operation found for parallel split-k reduction
return nullptr;
}
/////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace library
} // namespace cutlass
/////////////////////////////////////////////////////////////////////////////////////////////////
|