warp-lang 0.9.0__py3-none-win_amd64.whl → 0.11.0__py3-none-win_amd64.whl

warp/tests/test_linear_solvers.py

@@ -0,0 +1,154 @@
+import unittest
+
+import numpy as np
+import warp as wp
+import unittest
+
+from warp.optim.linear import preconditioner, cg, bicgstab, gmres
+from warp.tests.unittest_utils import *
+
+
+wp.init()
+
+
+def _check_linear_solve(test, A, b, func, *args, **kwargs):
+    # test from zero
+    x = wp.zeros_like(b)
+    with wp.ScopedDevice(A.device):
+        niter, err, atol = func(A, b, x, *args, use_cuda_graph=True, **kwargs)
+
+    test.assertLessEqual(err, atol)
+
+    # test with warm start
+    with wp.ScopedDevice(A.device):
+        niter_warm, err, atol = func(A, b, x, *args, use_cuda_graph=False, **kwargs)
+
+    test.assertLessEqual(err, atol)
+    test.assertLess(niter_warm, niter)
+
+    # In CG and BiCGSTAB residual norm is evaluating from running residual
+    # rather then being computed from scratch as Ax - b
+    # This can lead to accumulated inaccuracies over iterations, esp in float32
+    residual = A.numpy() @ x.numpy() - b.numpy()
+    err_np = np.linalg.norm(residual)
+    if A.dtype == wp.float64:
+        test.assertLessEqual(err_np, 2.0 * atol)
+    else:
+        test.assertLessEqual(err_np, 32.0 * atol)
+
+
+def _least_square_system(rng, n: int):
+    C = rng.uniform(low=-100, high=100, size=(n, n))
+    f = rng.uniform(low=-100, high=100, size=(n,))
+
+    A = C @ C.T
+    b = C @ f
+
+    return A, b
+
+
+def _make_spd_system(n: int, seed: int, dtype, device):
+    rng = np.random.default_rng(seed)
+
+    A, b = _least_square_system(rng, n)
+
+    return wp.array(A, dtype=dtype, device=device), wp.array(b, dtype=dtype, device=device)
+
+
+def _make_nonsymmetric_system(n: int, seed: int, dtype, device, spd=False):
+    rng = np.random.default_rng(seed)
+    s = rng.uniform(low=0.1, high=10, size=(n,))
+
+    A, b = _least_square_system(rng, n)
+    A = A @ np.diag(s)
+
+    return wp.array(A, dtype=dtype, device=device), wp.array(b, dtype=dtype, device=device)
+
+
+def _make_indefinite_system(n: int, seed: int, dtype, device, spd=False):
+    rng = np.random.default_rng(seed)
+    s = rng.uniform(low=0.1, high=10, size=(n,))
+
+    A, b = _least_square_system(rng, n)
+    A = A @ np.diag(s)
+
+    return wp.array(A, dtype=dtype, device=device), wp.array(b, dtype=dtype, device=device)
+
+
+def test_cg(test, device):
+    A, b = _make_spd_system(n=64, seed=123, device=device, dtype=wp.float64)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, cg, maxiter=1000)
+    _check_linear_solve(test, A, b, cg, M=M, maxiter=1000)
+
+    A, b = _make_spd_system(n=16, seed=321, device=device, dtype=wp.float32)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, cg, maxiter=1000)
+    _check_linear_solve(test, A, b, cg, M=M, maxiter=1000)
+
+
+def test_bicgstab(test, device):
+    A, b = _make_nonsymmetric_system(n=64, seed=123, device=device, dtype=wp.float64)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, bicgstab, maxiter=1000)
+    _check_linear_solve(test, A, b, bicgstab, M=M, maxiter=1000)
+    _check_linear_solve(test, A, b, bicgstab, M=M, maxiter=1000, is_left_preconditioner=True)
+
+    A, b = _make_nonsymmetric_system(n=16, seed=321, device=device, dtype=wp.float32)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, bicgstab, maxiter=1000)
+    _check_linear_solve(test, A, b, bicgstab, M=M, maxiter=1000)
+    _check_linear_solve(test, A, b, bicgstab, M=M, maxiter=1000, is_left_preconditioner=True)
+
+    A, b = _make_indefinite_system(n=64, seed=121, device=device, dtype=wp.float64)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, bicgstab, maxiter=1000)
+    _check_linear_solve(test, A, b, bicgstab, M=M, maxiter=1000)
+    _check_linear_solve(test, A, b, bicgstab, M=M, maxiter=1000, is_left_preconditioner=True)
+
+
+def test_gmres(test, device):
+    A, b = _make_nonsymmetric_system(n=64, seed=456, device=device, dtype=wp.float64)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, gmres, maxiter=1000, tol=1.0e-3)
+    _check_linear_solve(test, A, b, gmres, M=M, maxiter=1000, tol=1.0e-5)
+    _check_linear_solve(test, A, b, gmres, M=M, maxiter=1000, tol=1.0e-5, is_left_preconditioner=True)
+
+    A, b = _make_nonsymmetric_system(n=64, seed=654, device=device, dtype=wp.float64)
+    M = preconditioner(A, "diag")
+
+    _check_linear_solve(test, A, b, gmres, maxiter=1000, tol=1.0e-3)
+    _check_linear_solve(test, A, b, gmres, M=M, maxiter=1000, tol=1.0e-5)
+    _check_linear_solve(test, A, b, gmres, M=M, maxiter=1000, tol=1.0e-5, is_left_preconditioner=True)
+
+
+class TestLinearSolvers(unittest.TestCase):
+    pass
+
+
+devices = get_test_devices()
+
+from warp.context import runtime
+
+if not runtime.core.is_cutlass_enabled():
+    devices = [d for d in devices if not d.is_cuda]
+    print("Skipping CUDA linear solver tests because CUTLASS is not supported in this build")
+
+if runtime.core.is_debug_enabled():
+    # cutlass-based matmul is *very* slow in debug mode -- skip
+    devices = [d for d in devices if not d.is_cuda]
+    print("Skipping CUDA linear solver tests in debug mode")
+
+add_function_test(TestLinearSolvers, "test_cg", test_cg, devices=devices)
+add_function_test(TestLinearSolvers, "test_bicgstab", test_bicgstab, devices=devices)
+add_function_test(TestLinearSolvers, "test_gmres", test_gmres, devices=devices)
+
+if __name__ == "__main__":
+    wp.build.clear_kernel_cache()
+    unittest.main(verbosity=2)

warp/tests/test_lvalue.py

@@ -0,0 +1,493 @@
+# Copyright (c) 2023 NVIDIA CORPORATION.  All rights reserved.
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import unittest
+
+import warp as wp
+from warp.tests.unittest_utils import *
+
+wp.init()
+
+
+@wp.kernel
+def rmw_array_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+
+    foos[i] += wp.uint32(1)
+
+
+def test_rmw_array(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=rmw_array_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.struct
+class RmwFoo:
+    field: wp.uint32
+
+
+@wp.kernel
+def rmw_array_struct_kernel(foos: wp.array(dtype=RmwFoo)):
+    i = wp.tid()
+    foos[i].field += wp.uint32(1)
+
+
+def test_rmw_array_struct(test, device):
+    foos = wp.zeros((10,), dtype=RmwFoo, device=device)
+
+    wp.launch(
+        kernel=rmw_array_struct_kernel,
+        dim=(10,),
+        inputs=[foos],
+        device=device,
+    )
+    wp.synchronize()
+
+    expected = RmwFoo()
+    expected.field = 1
+    for f in foos.list():
+        if f.field != expected.field:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+
+@wp.func
+def lookup(foos: wp.array(dtype=wp.uint32), index: int):
+    return foos[index]
+
+
+@wp.kernel
+def lookup_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+
+    x = lookup(foos, i)
+    foos[i] = x + wp.uint32(1)
+
+
+def test_lookup(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=lookup_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.func
+def lookup3(foos: wp.array(dtype=wp.float32), index: int):
+    return foos[index]
+
+
+@wp.kernel
+def grad_kernel(foos: wp.array(dtype=wp.float32), bars: wp.array(dtype=wp.float32)):
+    i = wp.tid()
+
+    x = lookup3(foos, i)
+    bars[i] = x * wp.float32(i) + 1.0
+
+
+def test_grad(test, device):
+    num = 10
+    data = np.linspace(20, 20 + num, num, endpoint=False, dtype=np.float32)
+    input = wp.array(data, device=device, requires_grad=True)
+    output = wp.zeros(num, dtype=wp.float32, device=device)
+
+    ones = wp.array(np.ones(len(output)), dtype=wp.float32, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(
+            kernel=grad_kernel,
+            dim=(num,),
+            inputs=[input],
+            outputs=[output],
+            device=device,
+        )
+
+    tape.backward(grads={output: ones})
+
+    wp.synchronize()
+
+    # test forward results
+    for i, f in enumerate(output.list()):
+        expected = data[i] * i + 1
+        if f != expected:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+    # test backward results
+    for i, f in enumerate(tape.gradients[input].list()):
+        expected = i
+        if f != expected:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+
+@wp.func
+def lookup2(foos: wp.array(dtype=wp.uint32), index: int):
+    if index % 2 == 0:
+        x = foos[index]
+        x = wp.uint32(0)
+        return x
+    else:
+        return foos[index]
+
+
+@wp.kernel
+def lookup2_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+
+    x = lookup2(foos, i)
+    foos[i] = x + wp.uint32(1)
+
+
+def test_lookup2(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=lookup2_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.kernel
+def unary_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+
+    foos[i] = wp.uint32(-1)
+    x = -foos[i]
+    foos[i] = x
+
+
+def test_unary(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=unary_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.kernel
+def rvalue_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+
+    if foos[i] < wp.uint32(1):
+        foos[i] = wp.uint32(1)
+
+
+def test_rvalue(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=rvalue_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+# Tests, among other things, that assigning a reference to a new variable does
+# not create a reference
+@wp.kernel
+def intermediate_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+
+    x = foos[i]
+    x = x + wp.uint32(1)
+    foos[i] = x
+
+
+def test_intermediate(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=intermediate_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.kernel
+def array_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+    foos[i] = wp.uint32(1)
+
+
+def test_array_assign(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=array_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.func
+def increment(arg: wp.uint32):
+    return arg + wp.uint32(1)
+
+
+@wp.kernel
+def array_call_kernel(foos: wp.array(dtype=wp.uint32)):
+    i = wp.tid()
+    foos[i] = increment(foos[i])
+
+
+def test_array_call_assign(test, device):
+    arr = wp.zeros((10,), dtype=wp.uint32, device=device)
+
+    wp.launch(
+        kernel=array_kernel,
+        dim=(10,),
+        inputs=[arr],
+        device=device,
+    )
+    wp.synchronize()
+
+    for f in arr.list():
+        if f != 1:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {1}")
+
+
+@wp.struct
+class Foo:
+    field: wp.uint32
+
+
+@wp.kernel
+def array_struct_kernel(foos: wp.array(dtype=Foo)):
+    i = wp.tid()
+    foos[i].field = wp.uint32(1)
+
+
+def test_array_struct_assign(test, device):
+    foos = wp.zeros((10,), dtype=Foo, device=device)
+
+    wp.launch(
+        kernel=array_struct_kernel,
+        dim=(10,),
+        inputs=[foos],
+        device=device,
+    )
+    wp.synchronize()
+
+    expected = Foo()
+    expected.field = 1
+    for f in foos.list():
+        if f.field != expected.field:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+
+@wp.struct
+class Bar:
+    field: wp.uint32
+
+
+@wp.struct
+class Baz:
+    bar: Bar
+
+
+@wp.kernel
+def array_struct_struct_kernel(foos: wp.array(dtype=Baz)):
+    i = wp.tid()
+    foos[i].bar.field = wp.uint32(1)
+
+
+def test_array_struct_struct_assign(test, device):
+    foos = wp.zeros((10,), dtype=Baz, device=device)
+
+    wp.launch(
+        kernel=array_struct_struct_kernel,
+        dim=(10,),
+        inputs=[foos],
+        device=device,
+    )
+    wp.synchronize()
+
+    expected = Baz()
+    expected.bar.field = 1
+    for f in foos.list():
+        if f.bar.field != expected.bar.field:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+
+@wp.struct
+class S:
+    a: wp.uint32
+    b: wp.float32
+
+
+@wp.struct
+class F:
+    x: wp.float32
+    s: S
+    y: wp.int32
+
+
+@wp.kernel
+def complex_kernel(foos: wp.array(dtype=F)):
+    i = wp.tid()
+    foos[i].x += wp.float32(1.0)
+    foos[i].y = wp.int32(2)
+    foos[i].s.b += wp.float32(3.0)
+    foos[i].s.a = wp.uint32(foos[i].y)
+
+
+def test_complex(test, device):
+    foos = wp.zeros((10,), dtype=F, device=device)
+
+    wp.launch(
+        kernel=complex_kernel,
+        dim=(10,),
+        inputs=[foos],
+        device=device,
+    )
+    wp.synchronize()
+
+    expected = F()
+    expected.x = 1.0
+    expected.y = 2
+    expected.s.b = 3.0
+    expected.s.a = expected.y
+    for f in foos.list():
+        if f.x != expected.x or f.y != expected.y or f.s.a != expected.s.a or f.s.b != expected.s.b:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+
+@wp.struct
+class Svec:
+    a: wp.uint32
+    b: wp.vec2f
+
+
+@wp.struct
+class Fvec:
+    x: wp.vec2f
+    s: Svec
+    y: wp.int32
+
+
+@wp.kernel
+def swizzle_kernel(foos: wp.array(dtype=Fvec)):
+    i = wp.tid()
+
+    foos[i].x += wp.vec2f(1.0, 2.0)
+    foos[i].y = wp.int32(3)
+    foos[i].s.b = wp.vec2f(4.0, 5.0)
+    foos[i].s.b.y = wp.float32(6.0)
+    foos[i].s.b.x = foos[i].x.y
+    foos[i].s.a = wp.uint32(foos[i].y)
+
+
+def test_swizzle(test, device):
+    foos = wp.zeros((10,), dtype=Fvec, device=device)
+
+    wp.launch(
+        kernel=swizzle_kernel,
+        dim=(10,),
+        inputs=[foos],
+        device=device,
+    )
+    wp.synchronize()
+
+    expected = Fvec()
+    expected.x = wp.vec2f(1.0, 2.0)
+    expected.y = 3
+    expected.s.b = wp.vec2f(4.0, 5.0)
+    expected.s.b.y = 6.0
+    expected.s.b.x = expected.x.y
+    expected.s.a = expected.y
+    for f in foos.list():
+        if f.x != expected.x or f.y != expected.y or f.s.a != expected.s.a or f.s.b != expected.s.b:
+            raise AssertionError(f"Unexpected result, got: {f} expected: {expected}")
+
+
+devices = get_test_devices()
+
+
+class TestLValue(unittest.TestCase):
+    def test_swizzle_error_invalid_attribute(self):
+        v = wp.vec3(1, 2, 3)
+        with self.assertRaisesRegex(
+            AttributeError,
+            r"'vec3f' object has no attribute 'foo'$",
+        ):
+            v.foo
+
+        try:
+            v.bar = 123
+        except AttributeError:
+            self.fail()
+
+
+add_function_test(TestLValue, "test_rmw_array", test_rmw_array, devices=devices)
+add_function_test(TestLValue, "test_rmw_array_struct", test_rmw_array_struct, devices=devices)
+add_function_test(TestLValue, "test_lookup", test_lookup, devices=devices)
+add_function_test(TestLValue, "test_lookup2", test_lookup2, devices=devices)
+add_function_test(TestLValue, "test_grad", test_grad, devices=devices)
+add_function_test(TestLValue, "test_unary", test_unary, devices=devices)
+add_function_test(TestLValue, "test_rvalue", test_rvalue, devices=devices)
+add_function_test(TestLValue, "test_intermediate", test_intermediate, devices=devices)
+add_function_test(TestLValue, "test_array_assign", test_array_assign, devices=devices)
+add_function_test(TestLValue, "test_array_struct_assign", test_array_struct_assign, devices=devices)
+add_function_test(TestLValue, "test_array_struct_struct_assign", test_array_struct_struct_assign, devices=devices)
+add_function_test(TestLValue, "test_complex", test_complex, devices=devices)
+add_function_test(TestLValue, "test_swizzle", test_swizzle, devices=devices)
+
+
+if __name__ == "__main__":
+    wp.build.clear_kernel_cache()
+    unittest.main(verbosity=2)

warp/tests/test_marching_cubes.py

@@ -5,14 +5,12 @@
 # distribution of this software and related documentation without an express
 # license agreement from NVIDIA CORPORATION is strictly prohibited.
 
-# include parent path
+import unittest
+
 import numpy as np
-import math
 
 import warp as wp
-from warp.tests.test_base import *
-
-import unittest
+from warp.tests.unittest_utils import *
 
 wp.init()
 
@@ -39,7 +37,7 @@ def test_marching_cubes(test, device):
 
     radius = dim / 4.0
 
-    wp.launch(make_field, dim=field.shape, inputs=[field, wp.vec3(dim / 2, dim / 2, dim / 2), radius], device="cuda")
+    wp.launch(make_field, dim=field.shape, inputs=[field, wp.vec3(dim / 2, dim / 2, dim / 2), radius], device=device)
 
     iso.surface(field=field, threshold=0.0)
 
@@ -49,18 +47,19 @@ def test_marching_cubes(test, device):
 
     test.assertTrue(np.max(error) < 1.0)
 
+    iso.resize(nx=dim * 2, ny=dim * 2, nz=dim * 2, max_verts=max_verts, max_tris=max_tris)
+
+
+devices = get_unique_cuda_test_devices()
 
-def register(parent):
-    devices = ["cuda"]
 
-    class TestMarchingCubes(parent):
-        pass
+class TestMarchingCubes(unittest.TestCase):
+    pass
 
-    add_function_test(TestMarchingCubes, "test_marching_cubes", test_marching_cubes, devices=devices)
 
-    return TestMarchingCubes
+add_function_test(TestMarchingCubes, "test_marching_cubes", test_marching_cubes, devices=devices)
 
 
 if __name__ == "__main__":
-    c = register(unittest.TestCase)
+    wp.build.clear_kernel_cache()
     unittest.main(verbosity=2)