Spaces:
Sleeping
Sleeping
File size: 58,203 Bytes
66c9c8a | 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 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 | Runtime Reference
=================
.. currentmodule:: warp
This section describes the Warp Python runtime API, how to manage memory, launch kernels, and high-level functionality
for dealing with objects such as meshes and volumes. The APIs described in this section are intended to be used at
the *Python Scope* and run inside the CPython interpreter. For a comprehensive list of functions available at
the *Kernel Scope*, please see the :doc:`functions` section.
Kernels
-------
Kernels are launched with the :func:`wp.launch() <launch>` function on a specific device (CPU/GPU)::
wp.launch(simple_kernel, dim=1024, inputs=[a, b, c], device="cuda")
Kernels may be launched with multi-dimensional grid bounds. In this case threads are not assigned a single index,
but a coordinate in an n-dimensional grid, e.g.::
wp.launch(complex_kernel, dim=(128, 128, 3), ...)
Launches a 3D grid of threads with dimension 128 x 128 x 3. To retrieve the 3D index for each thread use the following syntax::
i,j,k = wp.tid()
.. note::
Currently kernels launched on CPU devices will be executed in serial.
Kernels launched on CUDA devices will be launched in parallel with a fixed block-size.
.. note::
Note that all the kernel inputs must live on the target device, or a runtime exception will be raised.
.. autofunction:: launch
Large Kernel Launches
#####################
A limitation of Warp is that each dimension of the grid used to launch a kernel must be representable as a 32-bit
signed integer. Therefore, no single dimension of a grid should exceed :math:`2^{31}-1`.
Warp also currently uses a fixed block size of 256 (CUDA) threads per block.
By default, Warp will try to process one element from the Warp grid in one CUDA thread.
This is not always possible for kernels launched with multi-dimensional grid bounds, as there are
`hardware limitations <https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications-technical-specifications-per-compute-capability>`_
on CUDA block dimensions.
Warp will automatically resort to using
`grid-stride loops <https://developer.nvidia.com/blog/cuda-pro-tip-write-flexible-kernels-grid-stride-loops/>`_ when
it is not possible for a CUDA thread to process only one element from the Warp grid
When this happens, some CUDA threads may process more than one element from the Warp grid.
Users can also set the ``max_blocks`` parameter to fine-tune the grid-striding behavior of kernels, even for kernels that are otherwise
able to process one Warp-grid element per CUDA thread.
Runtime Kernel Specialization
#############################
It is often desirable to specialize kernels for different types, constants, or functions.
We can achieve this through the use of runtime kernel specialization using Python closures.
For example, we might require a variety of kernels that execute particular functions for each item in an array.
We might also want this function call to be valid for a variety of data types. Making use of closure and generics, we can generate
these kernels using a single kernel definition: ::
def make_kernel(func, dtype):
def closure_kernel_fn(data: wp.array(dtype=dtype), out: wp.array(dtype=dtype)):
tid = wp.tid()
out[tid] = func(data[tid])
return wp.Kernel(closure_kernel_fn)
In practice, we might use our kernel generator, ``make_kernel()`` as follows: ::
@wp.func
def sqr(x: Any) -> Any:
return x * x
@wp.func
def cube(x: Any) -> Any:
return sqr(x) * x
sqr_float = make_kernel(sqr, wp.float32)
cube_double = make_kernel(cube, wp.float64)
arr = [1.0, 2.0, 3.0]
N = len(arr)
data_float = wp.array(arr, dtype=wp.float32, device=device)
data_double = wp.array(arr, dtype=wp.float64, device=device)
out_float = wp.zeros(N, dtype=wp.float32, device=device)
out_double = wp.zeros(N, dtype=wp.float64, device=device)
wp.launch(sqr_float, dim=N, inputs=[data_float], outputs=[out_float], device=device)
wp.launch(cube_double, dim=N, inputs=[data_double], outputs=[out_double], device=device)
We can specialize kernel definitions over warp constants similarly. The following generates kernels that add a specified constant
to a generic-typed array value: ::
def make_add_kernel(key, constant):
def closure_kernel_fn(data: wp.array(dtype=Any), out: wp.array(dtype=Any)):
tid = wp.tid()
out[tid] = data[tid] + constant
return wp.Kernel(closure_kernel_fn, key=key)
add_ones_int = make_add_kernel("add_one", wp.constant(1))
add_ones_vec3 = make_add_kernel("add_ones_vec3", wp.constant(wp.vec3(1.0, 1.0, 1.0)))
a = wp.zeros(2, dtype=int)
b = wp.zeros(2, dtype=wp.vec3)
a_out = wp.zeros_like(a)
b_out = wp.zeros_like(b)
wp.launch(add_ones_int, dim=a.size, inputs=[a], outputs=[a_out], device=device)
wp.launch(add_ones_vec3, dim=b.size, inputs=[b], outputs=[b_out], device=device)
Arrays
------
Arrays are the fundamental memory abstraction in Warp; they are created through the following global constructors: ::
wp.empty(shape=1024, dtype=wp.vec3, device="cpu")
wp.zeros(shape=1024, dtype=float, device="cuda")
wp.full(shape=1024, value=10, dtype=int, device="cuda")
Arrays can also be constructed directly from ``numpy`` ndarrays as follows: ::
r = np.random.rand(1024)
# copy to Warp owned array
a = wp.array(r, dtype=float, device="cpu")
# return a Warp array wrapper around the NumPy data (zero-copy)
a = wp.array(r, dtype=float, copy=False, device="cpu")
# return a Warp copy of the array data on the GPU
a = wp.array(r, dtype=float, device="cuda")
Note that for multi-dimensional data the ``dtype`` parameter must be specified explicitly, e.g.: ::
r = np.random.rand((1024, 3))
# initialize as an array of vec3 objects
a = wp.array(r, dtype=wp.vec3, device="cuda")
If the shapes are incompatible, an error will be raised.
Arrays can be moved between devices using the ``array.to()`` method: ::
host_array = wp.array(a, dtype=float, device="cpu")
# allocate and copy to GPU
device_array = host_array.to("cuda")
Additionally, arrays can be copied directly between memory spaces: ::
src_array = wp.array(a, dtype=float, device="cpu")
dest_array = wp.empty_like(host_array)
# copy from source CPU buffer to GPU
wp.copy(dest_array, src_array)
.. autoclass:: array
:members:
:undoc-members:
Multi-dimensional Arrays
########################
Multi-dimensional arrays can be constructed by passing a tuple of sizes for each dimension, e.g.: the following constructs a 2d array of size 1024x16::
wp.zeros(shape=(1024, 16), dtype=float, device="cuda")
When passing multi-dimensional arrays to kernels users must specify the expected array dimension inside the kernel signature,
e.g. to pass a 2d array to a kernel the number of dims is specified using the ``ndim=2`` parameter::
@wp.kernel
def test(input: wp.array(dtype=float, ndim=2)):
Type-hint helpers are provided for common array sizes, e.g.: ``array2d()``, ``array3d()``, which are equivalent to calling ``array(..., ndim=2)```, etc. To index a multi-dimensional array use a the following kernel syntax::
# returns a float from the 2d array
value = input[i,j]
To create an array slice use the following syntax, where the number of indices is less than the array dimensions::
# returns an 1d array slice representing a row of the 2d array
row = input[i]
Slice operators can be concatenated, e.g.: ``s = array[i][j][k]``. Slices can be passed to ``wp.func`` user functions provided
the function also declares the expected array dimension. Currently only single-index slicing is supported.
.. note::
Currently Warp limits arrays to 4 dimensions maximum. This is in addition to the contained datatype, which may be 1-2 dimensional for vector and matrix types such as ``vec3``, and ``mat33``.
The following construction methods are provided for allocating zero-initialized and empty (non-initialized) arrays:
.. autofunction:: zeros
.. autofunction:: zeros_like
.. autofunction:: full
.. autofunction:: full_like
.. autofunction:: empty
.. autofunction:: empty_like
.. autofunction:: copy
.. autofunction:: clone
Matrix Multiplication
#####################
Warp 2D array multiplication is built on NVIDIA's CUTLASS library, which enables fast matrix multiplication of large arrays on the GPU.
If no GPU is detected, matrix multiplication falls back to Numpy's implementation on the CPU.
Matrix multiplication is fully differentiable, and can be recorded on the tape like so::
tape = wp.Tape()
with tape:
wp.matmul(A, B, C, D, device=device)
wp.launch(loss_kernel, dim=(m, n), inputs=[D, loss], device=device)
tape.backward(loss=loss)
A_grad = A.grad.numpy()
Using the ``@`` operator (``D = A @ B``) will default to the same CUTLASS algorithm used in ``wp.matmul``.
.. autofunction:: matmul
Data Types
----------
Scalar Types
############
The following scalar storage types are supported for array structures:
+---------+------------------------+
| bool | boolean |
+---------+------------------------+
| int8 | signed byte |
+---------+------------------------+
| uint8 | unsigned byte |
+---------+------------------------+
| int16 | signed short |
+---------+------------------------+
| uint16 | unsigned short |
+---------+------------------------+
| int32 | signed integer |
+---------+------------------------+
| uint32 | unsigned integer |
+---------+------------------------+
| int64 | signed long integer |
+---------+------------------------+
| uint64 | unsigned long integer |
+---------+------------------------+
| float32 | single precision float |
+---------+------------------------+
| float64 | double precision float |
+---------+------------------------+
Warp supports ``float`` and ``int`` as aliases for ``wp.float32`` and ``wp.int32`` respectively.
.. _vec:
Vectors
#######
Warp provides built-in math and geometry types for common simulation and graphics problems.
A full reference for operators and functions for these types is available in the :doc:`/modules/functions`.
Warp supports vectors of numbers with an arbitrary length/numeric type. The built in concrete types are as follows:
+-----------------------+------------------------------------------------+
| vec2 vec3 vec4 | 2d, 3d, 4d vector of default precision floats |
+-----------------------+------------------------------------------------+
| vec2f vec3f vec4f | 2d, 3d, 4d vector of single precision floats |
+-----------------------+------------------------------------------------+
| vec2d vec3d vec4d | 2d, 3d, 4d vector of double precision floats |
+-----------------------+------------------------------------------------+
| vec2h vec3h vec4h | 2d, 3d, 4d vector of half precision floats |
+-----------------------+------------------------------------------------+
| vec2ub vec3ub vec4ub | 2d, 3d, 4d vector of half precision floats |
+-----------------------+------------------------------------------------+
| spatial_vector | 6d vector of single precision floats |
+-----------------------+------------------------------------------------+
| spatial_vectorf | 6d vector of single precision floats |
+-----------------------+------------------------------------------------+
| spatial_vectord | 6d vector of double precision floats |
+-----------------------+------------------------------------------------+
| spatial_vectorh | 6d vector of half precision floats |
+-----------------------+------------------------------------------------+
Vectors support most standard linear algebra operations, e.g.: ::
@wp.kernel
def compute( ... ):
# basis vectors
a = wp.vec3(1.0, 0.0, 0.0)
b = wp.vec3(0.0, 1.0, 0.0)
# take the cross product
c = wp.cross(a, b)
# compute
r = wp.dot(c, c)
...
It's possible to declare additional vector types with different lengths and data types. This is done in outside of kernels in *Python scope* using ``warp.types.vector()``, for example: ::
# declare a new vector type for holding 5 double precision floats:
vec5d = wp.types.vector(length=5, dtype=wp.float64)
Once declared, the new type can be used when allocating arrays or inside kernels: ::
# create an array of vec5d
arr = wp.zeros(10, dtype=vec5d)
# use inside a kernel
@wp.kernel
def compute( ... ):
# zero initialize a custom named vector type
v = vec5d()
...
# component-wise initialize a named vector type
v = vec5d(wp.float64(1.0),
wp.float64(2.0),
wp.float64(3.0),
wp.float64(4.0),
wp.float64(5.0))
...
In addition, it's possible to directly create *anonymously* typed instances of these vectors without declaring their type in advance. In this case the type will be inferred by the constructor arguments. For example: ::
@wp.kernel
def compute( ... ):
# zero initialize vector of 5 doubles:
v = wp.vector(dtype=wp.float64, length=5)
# scalar initialize a vector of 5 doubles to the same value:
v = wp.vector(wp.float64(1.0), length=5)
# component-wise initialize a vector of 5 doubles
v = wp.vector(wp.float64(1.0),
wp.float64(2.0),
wp.float64(3.0),
wp.float64(4.0),
wp.float64(5.0))
These can be used with all the standard vector arithmetic operators, e.g.: ``+``, ``-``, scalar multiplication, and can also be transformed using matrices with compatible dimensions, potentially returning vectors with a different length.
.. _mat:
Matrices
########
Matrices with arbitrary shapes/numeric types are also supported. The built in concrete matrix types are as follows:
+--------------------------+-----------------------------------------------------+
| mat22 mat33 mat44 | 2,3 and 4d square matrix of default precision |
+--------------------------+-----------------------------------------------------+
| mat22f mat33f mat44f | 2,3 and 4d square matrix of single precision floats |
+--------------------------+-----------------------------------------------------+
| mat22d mat33d mat44d | 2,3 and 4d square matrix of double precision floats |
+--------------------------+-----------------------------------------------------+
| mat22h mat33h mat44h | 2,3 and 4d square matrix of half precision floats |
+--------------------------+-----------------------------------------------------+
| spatial_matrix | 6x6 matrix of single precision floats |
+--------------------------+-----------------------------------------------------+
| spatial_matrixf | 6x6 matrix of single precision floats |
+--------------------------+-----------------------------------------------------+
| spatial_matrixd | 6x6 matrix of double precision floats |
+--------------------------+-----------------------------------------------------+
| spatial_matrixh | 6x6 matrix of half precision floats |
+--------------------------+-----------------------------------------------------+
Matrices are stored in row-major format and support most standard linear algebra operations: ::
@wp.kernel
def compute( ... ):
# initialize matrix
m = wp.mat22(1.0, 2.0,
3.0, 4.0)
# compute inverse
minv = wp.inverse(m)
# transform vector
v = minv * wp.vec2(0.5, 0.3)
...
In a similar manner to vectors, it's possible to declare new matrix types with arbitrary shapes and data types using ``wp.types.matrix()``, for example: ::
# declare a new 3x2 half precision float matrix type:
mat32h = wp.types.matrix(shape=(3,2), dtype=wp.float64)
# create an array of this type
a = wp.zeros(10, dtype=mat32h)
These can be used inside a kernel: ::
@wp.kernel
def compute( ... ):
...
# initialize a mat32h matrix
m = mat32h(wp.float16(1.0), wp.float16(2.0),
wp.float16(3.0), wp.float16(4.0),
wp.float16(5.0), wp.float16(6.0))
# declare a 2 component half precision vector
v2 = wp.vec2h(wp.float16(1.0), wp.float16(1.0))
# multiply by the matrix, returning a 3 component vector:
v3 = m * v2
...
It's also possible to directly create anonymously typed instances inside kernels where the type is inferred from constructor arguments as follows: ::
@wp.kernel
def compute( ... ):
...
# create a 3x2 half precision matrix from components (row major ordering):
m = wp.matrix(
wp.float16(1.0), wp.float16(2.0),
wp.float16(1.0), wp.float16(2.0),
wp.float16(1.0), wp.float16(2.0),
shape=(3,2))
# zero initialize a 3x2 half precision matrix:
m = wp.matrix(wp.float16(0.0),shape=(3,2))
# create a 5x5 double precision identity matrix:
m = wp.identity(n=5, dtype=wp.float64)
As with vectors, you can do standard matrix arithmetic with these variables, along with multiplying matrices with compatible shapes and potentially returning a matrix with a new shape.
.. _quat:
Quaternions
###########
Warp supports quaternions with the layout ``i, j, k, w`` where ``w`` is the real part. Here are the built in concrete quaternion types:
+-----------------+-----------------------------------------------+
| quat | Default precision floating point quaternion |
+-----------------+-----------------------------------------------+
| quatf | Single precision floating point quaternion |
+-----------------+-----------------------------------------------+
| quatd | Double precision floating point quaternion |
+-----------------+-----------------------------------------------+
| quath | Half precision floating point quaternion |
+-----------------+-----------------------------------------------+
Quaternions can be used to transform vectors as follows: ::
@wp.kernel
def compute( ... ):
...
# construct a 30 degree rotation around the x-axis
q = wp.quat_from_axis_angle(wp.vec3(1.0, 0.0, 0.0), wp.degrees(30.0))
# rotate an axis by this quaternion
v = wp.quat_rotate(q, wp.vec3(0.0, 1.0, 0.0))
As with vectors and matrices, you can declare quaternion types with an arbitrary numeric type like so: ::
quatd = wp.types.quaternion(dtype=wp.float64)
You can also create identity quaternion and anonymously typed instances inside a kernel like so: ::
@wp.kernel
def compute( ... ):
...
# create a double precision identity quaternion:
qd = wp.quat_identity(dtype=wp.float64)
# precision defaults to wp.float32 so this creates a single precision identity quaternion:
qf = wp.quat_identity()
# create a half precision quaternion from components, or a vector/scalar:
qh = wp.quaternion(wp.float16(0.0),
wp.float16(0.0),
wp.float16(0.0),
wp.float16(1.0))
qh = wp.quaternion(
wp.vector(wp.float16(0.0),wp.float16(0.0),wp.float16(0.0)),
wp.float16(1.0))
.. _transform:
Transforms
##########
Transforms are 7d vectors of floats representing a spatial rigid body transformation in format (p, q) where p is a 3d vector, and q is a quaternion.
+-----------------+--------------------------------------------+
| transform | Default precision floating point transform |
+-----------------+--------------------------------------------+
| transformf | Single precision floating point transform |
+-----------------+--------------------------------------------+
| transformd | Double precision floating point transform |
+-----------------+--------------------------------------------+
| transformh | Half precision floating point transform |
+-----------------+--------------------------------------------+
Transforms can be constructed inside kernels from translation and rotation parts: ::
@wp.kernel
def compute( ... ):
...
# create a transform from a vector/quaternion:
t = wp.transform(
wp.vec3(1.0, 2.0, 3.0),
wp.quat_from_axis_angle(0.0, 1.0, 0.0, wp.degrees(30.0)))
# transform a point
p = wp.transform_point(t, wp.vec3(10.0, 0.5, 1.0))
# transform a vector (ignore translation)
p = wp.transform_vector(t, wp.vec3(10.0, 0.5, 1.0))
As with vectors and matrices, you can declare transform types with an arbitrary numeric type using ``wp.types.transformation()``, for example: ::
transformd = wp.types.transformation(dtype=wp.float64)
You can also create identity transforms and anonymously typed instances inside a kernel like so: ::
@wp.kernel
def compute( ... ):
# create double precision identity transform:
qd = wp.transform_identity(dtype=wp.float64)
Structs
#######
Users can define custom structure types using the ``@wp.struct`` decorator as follows::
@wp.struct
class MyStruct:
param1: int
param2: float
param3: wp.array(dtype=wp.vec3)
Struct attributes must be annotated with their respective type. They can be constructed in Python scope and then passed to kernels as arguments::
@wp.kernel
def compute(args: MyStruct):
tid = wp.tid()
print(args.param1)
print(args.param2)
print(args.param3[tid])
# construct an instance of the struct in Python
s = MyStruct()
s.param1 = 10
s.param2 = 2.5
s.param3 = wp.zeros(shape=10, dtype=wp.vec3)
# pass to our compute kernel
wp.launch(compute, dim=10, inputs=[s])
An array of structs can be zero-initialized as follows::
a = wp.zeros(shape=10, dtype=MyStruct)
An array of structs can also be initialized from a list of struct objects::
a = wp.array([MyStruct(), MyStruct(), MyStruct()], dtype=MyStruct)
Example: Using a custom struct in gradient computation
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
.. code:: python
import numpy as np
import warp as wp
wp.init()
@wp.struct
class TestStruct:
x: wp.vec3
a: wp.array(dtype=wp.vec3)
b: wp.array(dtype=wp.vec3)
@wp.kernel
def test_kernel(s: TestStruct):
tid = wp.tid()
s.b[tid] = s.a[tid] + s.x
@wp.kernel
def loss_kernel(s: TestStruct, loss: wp.array(dtype=float)):
tid = wp.tid()
v = s.b[tid]
wp.atomic_add(loss, 0, float(tid + 1) * (v[0] + 2.0 * v[1] + 3.0 * v[2]))
# create struct
ts = TestStruct()
# set members
ts.x = wp.vec3(1.0, 2.0, 3.0)
ts.a = wp.array(np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]), dtype=wp.vec3, requires_grad=True)
ts.b = wp.zeros(2, dtype=wp.vec3, requires_grad=True)
loss = wp.zeros(1, dtype=float, requires_grad=True)
tape = wp.Tape()
with tape:
wp.launch(test_kernel, dim=2, inputs=[ts])
wp.launch(loss_kernel, dim=2, inputs=[ts, loss])
tape.backward(loss)
print(loss)
print(tape.gradients[ts].a)
Type Conversions
################
Warp is particularly strict regarding type conversions and does not perform *any* implicit conversion between numeric types.
The user is responsible for ensuring types for most arithmetic operators match, e.g.: ``x = float(0.0) + int(4)`` will result in an error.
This can be surprising for users that are accustomed to C-style conversions but avoids a class of common bugs that result from implicit conversions.
.. note::
Warp does not currently perform implicit type conversions between numeric types.
Users should explicitly cast variables to compatible types using constructors like
``int()``, ``float()``, ``wp.float16()``, ``wp.uint8()``, etc.
Constants
---------
In general, Warp kernels cannot access variables in the global Python interpreter state. One exception to this is for compile-time constants, which may be declared globally (or as class attributes) and folded into the kernel definition.
Constants are defined using the ``wp.constant()`` function. An example is shown below::
TYPE_SPHERE = wp.constant(0)
TYPE_CUBE = wp.constant(1)
TYPE_CAPSULE = wp.constant(2)
@wp.kernel
def collide(geometry: wp.array(dtype=int)):
t = geometry[wp.tid()]
if (t == TYPE_SPHERE):
print("sphere")
if (t == TYPE_CUBE):
print("cube")
if (t == TYPE_CAPSULE):
print("capsule")
.. autoclass:: constant
Operators
----------
Boolean Operators
#################
+--------------+--------------------------------------+
| a and b | True if a and b are True |
+--------------+--------------------------------------+
| a or b | True if a or b is True |
+--------------+--------------------------------------+
| not a | True if a is False, otherwise False |
+--------------+--------------------------------------+
.. note::
Expressions such as ``if (a and b):`` currently do not perform short-circuit evaluation.
In this case ``b`` will also be evaluated even when ``a`` is ``False``.
Users should take care to ensure that secondary conditions are safe to evaluate (e.g.: do not index out of bounds) in all cases.
Comparison Operators
####################
+----------+---------------------------------------+
| a > b | True if a strictly greater than b |
+----------+---------------------------------------+
| a < b | True if a strictly less than b |
+----------+---------------------------------------+
| a >= b | True if a greater than or equal to b |
+----------+---------------------------------------+
| a <= b | True if a less than or equal to b |
+----------+---------------------------------------+
| a == b | True if a equals b |
+----------+---------------------------------------+
| a != b | True if a not equal to b |
+----------+---------------------------------------+
Arithmetic Operators
####################
+-----------+--------------------------+
| a + b | Addition |
+-----------+--------------------------+
| a - b | Subtraction |
+-----------+--------------------------+
| a * b | Multiplication |
+-----------+--------------------------+
| a / b | Floating point division |
+-----------+--------------------------+
| a // b | Floored division |
+-----------+--------------------------+
| a ** b | Exponentiation |
+-----------+--------------------------+
| a % b | Modulus |
+-----------+--------------------------+
.. note::
Since implicit conversions are not performed arguments types to operators should match.
Users should use type constructors, e.g.: ``float()``, ``int()``, ``wp.int64()``, etc. to cast variables
to the correct type. Also note that the multiplication expression ``a * b`` is used to represent scalar
multiplication and matrix multiplication. The ``@`` operator is not currently supported.
Graphs
-----------
Launching kernels from Python introduces significant additional overhead compared to C++ or native programs.
To address this, Warp exposes the concept of `CUDA graphs <https://developer.nvidia.com/blog/cuda-graphs/>`_
to allow recording large batches of kernels and replaying them with very little CPU overhead.
To record a series of kernel launches use the :func:`wp.capture_begin() <capture_begin>` and
:func:`wp.capture_end() <capture_end>` API as follows:
.. code:: python
# begin capture
wp.capture_begin()
try:
# record launches
for i in range(100):
wp.launch(kernel=compute1, inputs=[a, b], device="cuda")
finally:
# end capture and return a graph object
graph = wp.capture_end()
We strongly recommend the use of the the try-finally pattern when capturing graphs because
:func:`wp.capture_begin <capture_begin>` disables Python garbage collection so that Warp objects are not
garbage-collected during graph capture (CUDA does not allow memory to be deallocated during graph capture).
:func:`wp.capture_end <capture_end>` reenables garbage collection.
Once a graph has been constructed it can be executed: ::
wp.capture_launch(graph)
Note that only launch calls are recorded in the graph, any Python executed outside of the kernel code will not be recorded.
Typically it is only beneficial to use CUDA graphs when the graph will be reused or launched multiple times.
.. autofunction:: capture_begin
.. autofunction:: capture_end
.. autofunction:: capture_launch
Bounding Value Hierarchies (BVH)
--------------------------------
The :class:`wp.Bvh <Bvh>` class can be used to create a BVH for a group of bounding volumes. This object can then be traversed
to determine which parts are intersected by a ray using :func:`bvh_query_ray` and which parts are fully contained
within a certain bounding volume using :func:`bvh_query_aabb`.
The following snippet demonstrates how to create a :class:`wp.Bvh <Bvh>` object from 100 random bounding volumes:
.. code:: python
rng = np.random.default_rng(123)
num_bounds = 100
lowers = rng.random(size=(num_bounds, 3)) * 5.0
uppers = lowers + rng.random(size=(num_bounds, 3)) * 5.0
device_lowers = wp.array(lowers, dtype=wp.vec3, device="cuda:0")
device_uppers = wp.array(uppers, dtype=wp.vec3, device="cuda:0")
bvh = wp.Bvh(device_lowers, device_uppers)
.. autoclass:: Bvh
:members:
Example: BVH Ray Traversal
##########################
An example of performing a ray traversal on the data structure is as follows:
.. code:: python
@wp.kernel
def bvh_query_ray(
bvh_id: wp.uint64,
start: wp.vec3,
dir: wp.vec3,
bounds_intersected: wp.array(dtype=wp.bool),
):
query = wp.bvh_query_ray(bvh_id, start, dir)
bounds_nr = wp.int32(0)
while wp.bvh_query_next(query, bounds_nr):
# The ray intersects the volume with index bounds_nr
bounds_intersected[bounds_nr] = True
bounds_intersected = wp.zeros(shape=(num_bounds), dtype=wp.bool, device="cuda:0")
query_start = wp.vec3(0.0, 0.0, 0.0)
query_dir = wp.normalize(wp.vec3(1.0, 1.0, 1.0))
wp.launch(
kernel=bvh_query_ray,
dim=1,
inputs=[bvh.id, query_start, query_dir, bounds_intersected],
device="cuda:0",
)
The Warp kernel ``bvh_query_ray`` is launched with a single thread, provided the unique :class:`uint64`
identifier of the :class:`wp.Bvh <Bvh>` object, parameters describing the ray, and an array to store the results.
In ``bvh_query_ray``, :func:`wp.bvh_query_ray() <bvh_query_ray>` is called once to obtain an object that is stored in the
variable ``query``. An integer is also allocated as ``bounds_nr`` to store the volume index of the traversal.
A while statement is used for the actual traversal using :func:`wp.bvh_query_next() <bvh_query_next>`,
which returns ``True`` as long as there are intersecting bounds.
Example: BVH Volume Traversal
#############################
Similar to the ray-traversal example, we can perform volume traversal to find the volumes that are fully contained
within a specified bounding box.
.. code:: python
@wp.kernel
def bvh_query_aabb(
bvh_id: wp.uint64,
lower: wp.vec3,
upper: wp.vec3,
bounds_intersected: wp.array(dtype=wp.bool),
):
query = wp.bvh_query_aabb(bvh_id, lower, upper)
bounds_nr = wp.int32(0)
while wp.bvh_query_next(query, bounds_nr):
# The volume with index bounds_nr is fully contained
# in the (lower,upper) bounding box
bounds_intersected[bounds_nr] = True
bounds_intersected = wp.zeros(shape=(num_bounds), dtype=wp.bool, device="cuda:0")
query_lower = wp.vec3(4.0, 4.0, 4.0)
query_upper = wp.vec3(6.0, 6.0, 6.0)
wp.launch(
kernel=bvh_query_aabb,
dim=1,
inputs=[bvh.id, query_lower, query_upper, bounds_intersected],
device="cuda:0",
)
The kernel is nearly identical to the ray-traversal example, except we obtain ``query`` using
:func:`wp.bvh_query_aabb() <bvh_query_aabb>`.
Meshes
------
Warp provides a ``wp.Mesh`` class to manage triangle mesh data. To create a mesh users provide a points, indices and optionally a velocity array::
mesh = wp.Mesh(points, indices, velocities)
.. note::
Mesh objects maintain references to their input geometry buffers. All buffers should live on the same device.
Meshes can be passed to kernels using their ``id`` attribute which uniquely identifies the mesh by a unique ``uint64`` value.
Once inside a kernel you can perform geometric queries against the mesh such as ray-casts or closest point lookups::
@wp.kernel
def raycast(mesh: wp.uint64,
ray_origin: wp.array(dtype=wp.vec3),
ray_dir: wp.array(dtype=wp.vec3),
ray_hit: wp.array(dtype=wp.vec3)):
tid = wp.tid()
t = float(0.0) # hit distance along ray
u = float(0.0) # hit face barycentric u
v = float(0.0) # hit face barycentric v
sign = float(0.0) # hit face sign
n = wp.vec3() # hit face normal
f = int(0) # hit face index
color = wp.vec3()
# ray cast against the mesh
if wp.mesh_query_ray(mesh, ray_origin[tid], ray_dir[tid], 1.e+6, t, u, v, sign, n, f):
# if we got a hit then set color to the face normal
color = n*0.5 + wp.vec3(0.5, 0.5, 0.5)
ray_hit[tid] = color
Users may update mesh vertex positions at runtime simply by modifying the points buffer.
After modifying point locations users should call ``Mesh.refit()`` to rebuild the bounding volume hierarchy (BVH) structure and ensure that queries work correctly.
.. note::
Updating Mesh topology (indices) at runtime is not currently supported. Users should instead recreate a new Mesh object.
.. autoclass:: Mesh
:members:
Hash Grids
----------
Many particle-based simulation methods such as the Discrete Element Method (DEM), or Smoothed Particle Hydrodynamics (SPH), involve iterating over spatial neighbors to compute force interactions. Hash grids are a well-established data structure to accelerate these nearest neighbor queries, and particularly well-suited to the GPU.
To support spatial neighbor queries Warp provides a ``HashGrid`` object that may be created as follows::
grid = wp.HashGrid(dim_x=128, dim_y=128, dim_z=128, device="cuda")
grid.build(points=p, radius=r)
``p`` is an array of ``wp.vec3`` point positions, and ``r`` is the radius to use when building the grid.
Neighbors can then be iterated over inside the kernel code using :func:`wp.hash_grid_query() <hash_grid_query>`
and :func:`wp.hash_grid_query_next() <hash_grid_query_next>` as follows:
.. code:: python
@wp.kernel
def sum(grid : wp.uint64,
points: wp.array(dtype=wp.vec3),
output: wp.array(dtype=wp.vec3),
radius: float):
tid = wp.tid()
# query point
p = points[tid]
# create grid query around point
query = wp.hash_grid_query(grid, p, radius)
index = int(0)
sum = wp.vec3()
while(wp.hash_grid_query_next(query, index)):
neighbor = points[index]
# compute distance to neighbor point
dist = wp.length(p-neighbor)
if (dist <= radius):
sum += neighbor
output[tid] = sum
.. note::
The ``HashGrid`` query will give back all points in *cells* that fall inside the query radius.
When there are hash conflicts it means that some points outside of query radius will be returned, and users should
check the distance themselves inside their kernels. The reason the query doesn't do the check itself for each
returned point is because it's common for kernels to compute the distance themselves, so it would redundant to
check/compute the distance twice.
.. autoclass:: HashGrid
:members:
Volumes
-------
Sparse volumes are incredibly useful for representing grid data over large domains, such as signed distance fields
(SDFs) for complex objects, or velocities for large-scale fluid flow. Warp supports reading sparse volumetric grids
stored using the `NanoVDB <https://developer.nvidia.com/nanovdb>`_ standard. Users can access voxels directly
or use built-in closest-point or trilinear interpolation to sample grid data from world or local space.
Volume objects can be created directly from Warp arrays containing a NanoVDB grid, from the contents of a
standard ``.nvdb`` file using :func:`load_from_nvdb() <warp.Volume.load_from_nvdb>`,
or from a dense 3D NumPy array using :func:`load_from_numpy() <warp.Volume.load_from_numpy>`.
Volumes can also be created using :func:`allocate() <warp.Volume.allocate>` or
:func:`allocate_by_tiles() <warp.Volume.allocate_by_tiles>`. The values for a Volume object can be modified in a Warp
kernel using :func:`wp.volume_store_f() <warp.volume_store_f>`, :func:`wp.volume_store_v() <warp.volume_store_v>`, and
:func:`wp.volume_store_i() <warp.volume_store_i>`.
.. note::
Warp does not currently support modifying the topology of sparse volumes at runtime.
Below we give an example of creating a Volume object from an existing NanoVDB file::
# open NanoVDB file on disk
file = open("mygrid.nvdb", "rb")
# create Volume object
volume = wp.Volume.load_from_nvdb(file, device="cpu")
.. note::
Files written by the NanoVDB library, commonly marked by the ``.nvdb`` extension, can contain multiple grids with
various compression methods, but a :class:`Volume` object represents a single NanoVDB grid therefore only files with
a single grid are supported. NanoVDB's uncompressed and zip-compressed file formats are supported.
To sample the volume inside a kernel we pass a reference to it by ID, and use the built-in sampling modes::
@wp.kernel
def sample_grid(volume: wp.uint64,
points: wp.array(dtype=wp.vec3),
samples: wp.array(dtype=float)):
tid = wp.tid()
# load sample point in world-space
p = points[tid]
# transform position to the volume's local-space
q = wp.volume_world_to_index(volume, p)
# sample volume with trilinear interpolation
f = wp.volume_sample_f(volume, q, wp.Volume.LINEAR)
# write result
samples[tid] = f
.. autoclass:: Volume
:members:
:undoc-members:
.. seealso:: `Reference <functions.html#volumes>`__ for the volume functions available in kernels.
Differentiability
-----------------
By default, Warp generates a forward and backward (adjoint) version of each kernel definition.
Buffers that participate in the chain of computation should be created with ``requires_grad=True``, for example::
a = wp.zeros(1024, dtype=wp.vec3, device="cuda", requires_grad=True)
The ``wp.Tape`` class can then be used to record kernel launches, and replay them to compute the gradient of a scalar loss function with respect to the kernel inputs::
tape = wp.Tape()
# forward pass
with tape:
wp.launch(kernel=compute1, inputs=[a, b], device="cuda")
wp.launch(kernel=compute2, inputs=[c, d], device="cuda")
wp.launch(kernel=loss, inputs=[d, l], device="cuda")
# reverse pass
tape.backward(l)
After the backward pass has completed, the gradients with respect to the inputs are available via a mapping in the :class:`wp.Tape <Tape>` object::
# gradient of loss with respect to input a
print(tape.gradients[a])
Note that gradients are accumulated on the participating buffers, so if you wish to reuse the same buffers for multiple backward passes you should first zero the gradients::
tape.zero()
.. note::
Warp uses a source-code transformation approach to auto-differentiation.
In this approach, the backwards pass must keep a record of intermediate values computed during the forward pass.
This imposes some restrictions on what kernels can do and still be differentiable:
* Dynamic loops should not mutate any previously declared local variable. This means the loop must be side-effect free.
A simple way to ensure this is to move the loop body into a separate function.
Static loops that are unrolled at compile time do not have this restriction and can perform any computation.
* Kernels should not overwrite any previously used array values except to perform simple linear add/subtract operations (e.g. via :func:`wp.atomic_add() <atomic_add>`)
.. autoclass:: Tape
:members:
Jacobians
#########
To compute the Jacobian matrix :math:`J\in\mathbb{R}^{m\times n}` of a multi-valued function :math:`f: \mathbb{R}^n \to \mathbb{R}^m`, we can evaluate an entire row of the Jacobian in parallel by finding the Jacobian-vector product :math:`J^\top \mathbf{e}`. The vector :math:`\mathbf{e}\in\mathbb{R}^m` selects the indices in the output buffer to differentiate with respect to.
In Warp, instead of passing a scalar loss buffer to the ``tape.backward()`` method, we pass a dictionary ``grads`` mapping from the function output array to the selection vector :math:`\mathbf{e}` having the same type::
# compute the Jacobian for a function of single output
jacobian = np.empty((output_dim, input_dim), dtype=np.float32)
tape = wp.Tape()
with tape:
output_buffer = launch_kernels_to_be_differentiated(input_buffer)
for output_index in range(output_dim):
# select which row of the Jacobian we want to compute
select_index = np.zeros(output_dim)
select_index[output_index] = 1.0
e = wp.array(select_index, dtype=wp.float32)
# pass input gradients to the output buffer to apply selection
tape.backward(grads={output_buffer: e})
q_grad_i = tape.gradients[input_buffer]
jacobian[output_index, :] = q_grad_i.numpy()
tape.zero()
When we run simulations independently in parallel, the Jacobian corresponding to the entire system dynamics is a block-diagonal matrix. In this case, we can compute the Jacobian in parallel for all environments by choosing a selection vector that has the output indices active for all environment copies. For example, to get the first rows of the Jacobians of all environments, :math:`\mathbf{e}=[\begin{smallmatrix}1 & 0 & 0 & \dots & 1 & 0 & 0 & \dots\end{smallmatrix}]^\top`, to compute the second rows, :math:`\mathbf{e}=[\begin{smallmatrix}0 & 1 & 0 & \dots & 0 & 1 & 0 & \dots\end{smallmatrix}]^\top`, etc.::
# compute the Jacobian for a function over multiple environments in parallel
jacobians = np.empty((num_envs, output_dim, input_dim), dtype=np.float32)
tape = wp.Tape()
with tape:
output_buffer = launch_kernels_to_be_differentiated(input_buffer)
for output_index in range(output_dim):
# select which row of the Jacobian we want to compute
select_index = np.zeros(output_dim)
select_index[output_index] = 1.0
# assemble selection vector for all environments (can be precomputed)
e = wp.array(np.tile(select_index, num_envs), dtype=wp.float32)
tape.backward(grads={output_buffer: e})
q_grad_i = tape.gradients[input_buffer]
jacobians[:, output_index, :] = q_grad_i.numpy().reshape(num_envs, input_dim)
tape.zero()
Custom Gradient Functions
#########################
Warp supports custom gradient function definitions for user-defined Warp functions.
This allows users to define code that should replace the automatically generated derivatives.
To differentiate a function :math:`h(x) = f(g(x))` that has a nested call to function :math:`g(x)`, the chain rule is evaluated in the automatic differentiation of :math:`h(x)`:
.. math::
h^\prime(x) = f^\prime({\color{green}{\underset{\textrm{replay}}{g(x)}}}) {\color{blue}{\underset{\textrm{grad}}{g^\prime(x)}}}
This implies that a function to be compatible with the autodiff engine needs to provide an implementation of its forward version
:math:`\color{green}{g(x)}`, which we refer to as "replay" function (that matches the original function definition by default),
and its derivative :math:`\color{blue}{g^\prime(x)}`, referred to as "grad".
Both the replay and the grad implementations can be customized by the user. They are defined as follows:
.. list-table:: Customizing the replay and grad versions of function ``myfunc``
:widths: 100
:header-rows: 0
* - Forward Function
* - .. code-block:: python
@wp.func
def myfunc(in1: InType1, ..., inN: InTypeN) -> OutType1, ..., OutTypeM:
return out1, ..., outM
* - Custom Replay Function
* - .. code-block:: python
@wp.func_replay(myfunc)
def replay_myfunc(in1: InType1, ..., inN: InTypeN) -> OutType1, ..., OutTypeM:
# Custom forward computations to be executed in the backward pass of a
# function calling `myfunc` go here
# Ensure the output variables match the original forward definition
return out1, ..., outM
* - Custom Grad Function
* - .. code-block:: python
@wp.func_grad(myfunc)
def adj_myfunc(in1: InType1, ..., inN: InTypeN, adj_out1: OutType1, ..., adj_outM: OutTypeM):
# Custom adjoint code goes here
# Update the partial derivatives for the inputs as follows:
wp.adjoint[in1] += ...
...
wp.adjoint[inN] += ...
.. note:: It is currently not possible to define custom replay or grad functions for functions that
have generic arguments, e.g. ``Any`` or ``wp.array(dtype=Any)``. Replay or grad functions that
themselves use generic arguments are also not yet supported.
Example 1: Custom Grad Function
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
In the following, we define a Warp function ``safe_sqrt`` that computes the square root of a number::
@wp.func
def safe_sqrt(x: float):
return wp.sqrt(x)
To evaluate this function, we define a kernel that applies ``safe_sqrt`` to an array of input values::
@wp.kernel
def run_safe_sqrt(xs: wp.array(dtype=float), output: wp.array(dtype=float)):
i = wp.tid()
output[i] = safe_sqrt(xs[i])
Calling the kernel for an array of values ``[1.0, 2.0, 0.0]`` yields the expected outputs, the gradients are finite except for the zero input::
xs = wp.array([1.0, 2.0, 0.0], dtype=wp.float32, requires_grad=True)
ys = wp.zeros_like(xs)
tape = wp.Tape()
with tape:
wp.launch(run_safe_sqrt, dim=len(xs), inputs=[xs], outputs=[ys])
tape.backward(grads={ys: wp.array(np.ones(len(xs)), dtype=wp.float32)})
print("ys ", ys)
print("xs.grad", xs.grad)
# ys [1. 1.4142135 0. ]
# xs.grad [0.5 0.35355338 inf]
It is often desired to catch nonfinite gradients in the computation graph as they may cause the entire gradient computation to be nonfinite.
To do so, we can define a custom gradient function that replaces the adjoint function for ``safe_sqrt`` which is automatically generated by
decorating the custom gradient code via ``@wp.func_grad(safe_sqrt)``::
@wp.func_grad(safe_sqrt)
def adj_safe_sqrt(x: float, adj_ret: float):
if x > 0.0:
wp.adjoint[x] += 1.0 / (2.0 * wp.sqrt(x)) * adj_ret
.. note:: The function signature of the custom grad code consists of the input arguments of the forward function plus the adjoint variables of the
forward function outputs. To access and modify the partial derivatives of the input arguments, we use the ``wp.adjoint`` dictionary.
The keys of this dictionary are the input arguments of the forward function, and the values are the partial derivatives of the forward function
output with respect to the input argument.
Example 2: Custom Replay Function
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
In the following, we increment an array index in each thread via :func:`wp.atomic_add() <atomic_add>` and compute
the square root of an input array at the incremented index::
@wp.kernel
def test_add(counter: wp.array(dtype=int), input: wp.array(dtype=float), output: wp.array(dtype=float)):
idx = wp.atomic_add(counter, 0, 1)
output[idx] = wp.sqrt(input[idx])
def main():
dim = 16
use_reversible_increment = False
input = wp.array(np.arange(1, dim + 1), dtype=wp.float32, requires_grad=True)
counter = wp.zeros(1, dtype=wp.int32)
thread_ids = wp.zeros(dim, dtype=wp.int32)
output = wp.zeros(dim, dtype=wp.float32, requires_grad=True)
tape = wp.Tape()
with tape:
if use_reversible_increment:
wp.launch(test_add_diff, dim, inputs=[counter, thread_ids, input], outputs=[output])
else:
wp.launch(test_add, dim, inputs=[counter, input], outputs=[output])
print("counter: ", counter.numpy())
print("thread_ids: ", thread_ids.numpy())
print("input: ", input.numpy())
print("output: ", output.numpy())
tape.backward(grads={
output: wp.array(np.ones(dim), dtype=wp.float32)
})
print("input.grad: ", input.grad.numpy())
if __name__ == "__main__":
main()
The output of the above code is:
.. code-block:: js
counter: [8]
thread_ids: [0 0 0 0 0 0 0 0]
input: [1. 2. 3. 4. 5. 6. 7. 8.]
output: [1. 1.4142135 1.7320508 2. 2.236068 2.4494898 2.6457512 2.828427]
input.grad: [4. 0. 0. 0. 0. 0. 0. 0.]
The gradient of the input is incorrect because the backward pass involving the atomic operation ``wp.atomic_add()`` does not know which thread ID corresponds
to which input value.
The index returned by the adjoint of ``wp.atomic_add()`` is always zero so that the gradient the first entry of the input array,
i.e. :math:`\frac{1}{2\sqrt{1}} = 0.5`, is accumulated ``dim`` times (hence ``input.grad[0] == 4.0`` and all other entries zero).
To fix this, we define a new Warp function ``reversible_increment()`` with a custom *replay* definition that stores the thread ID in a separate array::
@wp.func
def reversible_increment(
buf: wp.array(dtype=int),
buf_index: int,
value: int,
thread_values: wp.array(dtype=int),
tid: int
):
next_index = wp.atomic_add(buf, buf_index, value)
# store which thread ID corresponds to which index for the backward pass
thread_values[tid] = next_index
return next_index
@wp.func_replay(reversible_increment)
def replay_reversible_increment(
buf: wp.array(dtype=int),
buf_index: int,
value: int,
thread_values: wp.array(dtype=int),
tid: int
):
return thread_values[tid]
Instead of running ``reversible_increment()``, the custom replay code in ``replay_reversible_increment()`` is now executed
during forward phase in the backward pass of the function calling ``reversible_increment()``.
We first stored the array index to each thread ID in the forward pass, and now we retrieve the array index for each thread ID in the backward pass.
That way, the backward pass can reproduce the same addition operation as in the forward pass with exactly the same operands per thread.
.. warning:: The function signature of the custom replay code must match the forward function signature.
To use our function we write the following kernel::
@wp.kernel
def test_add_diff(
counter: wp.array(dtype=int),
thread_ids: wp.array(dtype=int),
input: wp.array(dtype=float),
output: wp.array(dtype=float)
):
tid = wp.tid()
idx = reversible_increment(counter, 0, 1, thread_ids, tid)
output[idx] = wp.sqrt(input[idx])
Running the ``test_add_diff`` kernel via the previous ``main`` function with ``use_reversible_increment = True``, we now compute correct gradients
for the input array:
.. code-block:: js
counter: [8]
thread_ids: [0 1 2 3 4 5 6 7]
input: [1. 2. 3. 4. 5. 6. 7. 8.]
output: [1. 1.4142135 1.7320508 2. 2.236068 2.4494898 2.6457512 2.828427 ]
input.grad: [0.5 0.35355338 0.28867513 0.25 0.2236068 0.20412414 0.18898225 0.17677669]
Custom Native Functions
-----------------------
Users may insert native C++/CUDA code in Warp kernels using ``@func_native`` decorated functions.
These accept native code as strings that get compiled after code generation, and are called within ``@wp.kernel`` functions.
For example::
snippet = """
__shared__ int s[128];
s[tid] = d[tid];
__syncthreads();
d[tid] = s[N - tid - 1];
"""
@wp.func_native(snippet)
def reverse(d: wp.array(dtype=int), N: int, tid: int):
...
@wp.kernel
def reverse_kernel(d: wp.array(dtype=int), N: int):
tid = wp.tid()
reverse(d, N, tid)
N = 128
x = wp.array(np.arange(N, dtype=int), dtype=int, device=device)
wp.launch(kernel=reverse_kernel, dim=N, inputs=[x, N], device=device)
Notice the use of shared memory here: the Warp library does not expose shared memory as a feature, but the CUDA compiler will
readily accept the above snippet. This means CUDA features not exposed in Warp are still accessible in Warp scripts.
Warp kernels meant for the CPU won't be able to leverage CUDA features of course, but this same mechanism supports pure C++ snippets as well.
Please bear in mind the following: the thread index in your snippet should be computed in a ``@wp.kernel`` and passed to your snippet,
as in the above example. This means your ``@wp.func_native`` function signature should include the variables used in your snippet,
as well as a thread index of type ``int``. The function body itself should be stubbed with ``...`` (the snippet will be inserted during compilation).
Should you wish to record your native function on the tape and then subsequently rewind the tape, you must include an adjoint snippet
alongside your snippet as an additional input to the decorator, as in the following example::
snippet = """
out[tid] = a * x[tid] + y[tid];
"""
adj_snippet = """
adj_a = x[tid] * adj_out[tid];
adj_x[tid] = a * adj_out[tid];
adj_y[tid] = adj_out[tid];
"""
@wp.func_native(snippet, adj_snippet)
def saxpy(
a: wp.float32,
x: wp.array(dtype=wp.float32),
y: wp.array(dtype=wp.float32),
out: wp.array(dtype=wp.float32),
tid: int,
):
...
@wp.kernel
def saxpy_kernel(
a: wp.float32,
x: wp.array(dtype=wp.float32),
y: wp.array(dtype=wp.float32),
out: wp.array(dtype=wp.float32)
):
tid = wp.tid()
saxpy(a, x, y, out, tid)
N = 128
a = 2.0
x = wp.array(np.arange(N, dtype=np.float32), dtype=wp.float32, device=device, requires_grad=True)
y = wp.zeros_like(x1)
out = wp.array(np.arange(N, dtype=np.float32), dtype=wp.float32, device=device)
adj_out = wp.array(np.ones(N, dtype=np.float32), dtype=wp.float32, device=device)
tape = wp.Tape()
with tape:
wp.launch(kernel=saxpy_kernel, dim=N, inputs=[a, x, y], outputs=[out], device=device)
tape.backward(grads={out: adj_out})
Profiling
---------
``wp.ScopedTimer`` objects can be used to gain some basic insight into the performance of Warp applications:
.. code:: python
with wp.ScopedTimer("grid build"):
self.grid.build(self.x, self.point_radius)
This results in a printout at runtime to the standard output stream like:
.. code:: console
grid build took 0.06 ms
The ``wp.ScopedTimer`` object does not synchronize (e.g. by calling ``wp.synchronize()``)
upon exiting the ``with`` statement, so this can lead to misleading numbers if the body
of the ``with`` statement launches device kernels.
When a ``wp.ScopedTimer`` object is passed ``use_nvtx=True`` as an argument, the timing functionality is replaced by calls
to ``nvtx.start_range()`` and ``nvtx.end_range()``:
.. code:: python
with wp.ScopedTimer("grid build", use_nvtx=True, color="cyan"):
self.grid.build(self.x, self.point_radius)
These range annotations can then be collected by a tool like `NVIDIA Nsight Systems <https://developer.nvidia.com/nsight-systems>`_
for visualization on a timeline, e.g.:
.. code:: console
$ nsys profile python warp_application.py
This code snippet also demonstrates the use of the ``color`` argument to specify a color
for the range, which may be a number representing the ARGB value or a recognized string
(refer to `colors.py <https://github.com/NVIDIA/NVTX/blob/release-v3/python/nvtx/colors.py>`_ for
additional color examples).
The `nvtx module <https://github.com/NVIDIA/NVTX>`_ must be
installed in the Python environment for this capability to work.
An equivalent way to create an NVTX range without using ``wp.ScopedTimer`` is:
.. code:: python
import nvtx
with nvtx.annotate("grid build", color="cyan"):
self.grid.build(self.x, self.point_radius)
This form may be more convenient if the user does not need to frequently switch
between timer and NVTX capabilities of ``wp.ScopedTimer``.
.. autoclass:: warp.ScopedTimer
|