warp-lang 1.6.2__py3-none-win_amd64.whl → 1.7.1__py3-none-win_amd64.whl

warp/tests/test_mat.py

@@ -50,6 +50,22 @@ def get_select_kernel(dtype):
     return getkernel(output_select_kernel_fn, suffix=dtype.__name__)
 
 
+def test_shape_mismatch(test, device):
+    test.assertNotEqual(wp.mat33f(0.0), wp.mat22f(0.0))
+    test.assertNotEqual(wp.mat22f(0.0), wp.mat33f(0.0))
+
+    @wp.kernel
+    def kernel():
+        wp.expect_neq(wp.mat33f(0.0), wp.mat22f(0.0))
+        wp.expect_neq(wp.mat22f(0.0), wp.mat33f(0.0))
+
+    with test.assertRaisesRegex(
+        RuntimeError,
+        r"Can't test equality for objects with different types$",
+    ):
+        wp.launch(kernel, dim=1, inputs=[], device=device)
+
+
 def test_anon_constructor_error_shape_arg_missing(test, device):
     @wp.kernel
     def kernel():
@@ -127,30 +143,6 @@ def test_tpl_constructor_error_incompatible_sizes(test, device):
         wp.launch(kernel, dim=1, inputs=[], device=device)
 
 
-def test_tpl_constructor_error_invalid_vector_count(test, device):
-    @wp.kernel
-    def kernel():
-        wp.mat33(wp.vec3(1.0, 2.0, 3.0), wp.vec3(1.0, 2.0, 3.0))
-
-    with test.assertRaisesRegex(
-        RuntimeError,
-        r"incompatible number of column vectors given \(2\) when constructing a matrix of shape \(3, 3\)$",
-    ):
-        wp.launch(kernel, dim=1, inputs=[], device=device)
-
-
-def test_tpl_constructor_error_invalid_vector_shape(test, device):
-    @wp.kernel
-    def kernel():
-        wp.mat22(wp.vec3(1.0, 2.0, 3.0), wp.vec3(4.0, 5.0, 6.0))
-
-    with test.assertRaisesRegex(
-        RuntimeError,
-        r"incompatible column vector lengths given when constructing a matrix of shape \(2, 2\)$",
-    ):
-        wp.launch(kernel, dim=1, inputs=[], device=device)
-
-
 def test_tpl_constructor_error_invalid_arg_count(test, device):
     @wp.kernel
     def kernel():
@@ -234,7 +226,7 @@ def test_quat_constructor(test, device, dtype, register_kernels=False):
         c0 = s[0][0] * R[0]
         c1 = s[0][1] * R[1]
         c2 = s[0][2] * R[2]
-        m_alt = mat44(
+        m_alt = wp.matrix_from_cols(
             vec4(c0[0], c0[1], c0[2], wptype(0.0)),
             vec4(c1[0], c1[1], c1[2], wptype(0.0)),
             vec4(c2[0], c2[1], c2[2], wptype(0.0)),
@@ -1066,6 +1058,124 @@ def test_svd(test, device, dtype, register_kernels=False):
                 assert_np_equal((plusval - minusval) / (2 * dx), m3grads[ii, jj], tol=fdtol)
 
 
+def test_svd_2D(test, device, dtype, register_kernels=False):
+    rng = np.random.default_rng(123)
+
+    tol = {
+        np.float16: 1.0e-3,
+        np.float32: 1.0e-6,
+        np.float64: 1.0e-12,
+    }.get(dtype, 0)
+
+    wptype = wp.types.np_dtype_to_warp_type[np.dtype(dtype)]
+    vec2 = wp.types.vector(length=2, dtype=wptype)
+    mat22 = wp.types.matrix(shape=(2, 2), dtype=wptype)
+
+    def check_mat_svd2(
+        m2: wp.array(dtype=mat22),
+        Uout: wp.array(dtype=mat22),
+        sigmaout: wp.array(dtype=vec2),
+        Vout: wp.array(dtype=mat22),
+        outcomponents: wp.array(dtype=wptype),
+    ):
+        U = mat22()
+        sigma = vec2()
+        V = mat22()
+
+        wp.svd2(m2[0], U, sigma, V)  # Assuming there's a 2D SVD kernel
+
+        Uout[0] = U
+        sigmaout[0] = sigma
+        Vout[0] = V
+
+        # multiply outputs by 2 so we've got something to backpropagate:
+        idx = 0
+        for i in range(2):
+            for j in range(2):
+                outcomponents[idx] = wptype(2) * U[i, j]
+                idx = idx + 1
+
+        for i in range(2):
+            outcomponents[idx] = wptype(2) * sigma[i]
+            idx = idx + 1
+
+        for i in range(2):
+            for j in range(2):
+                outcomponents[idx] = wptype(2) * V[i, j]
+                idx = idx + 1
+
+    kernel = getkernel(check_mat_svd2, suffix=dtype.__name__)
+
+    output_select_kernel = get_select_kernel(wptype)
+
+    if register_kernels:
+        return
+
+    m2 = wp.array(randvals(rng, [1, 2, 2], dtype) + np.eye(2), dtype=mat22, requires_grad=True, device=device)
+
+    outcomponents = wp.zeros(2 * 2 * 2 + 2, dtype=wptype, requires_grad=True, device=device)
+    Uout = wp.zeros(1, dtype=mat22, requires_grad=True, device=device)
+    sigmaout = wp.zeros(1, dtype=vec2, requires_grad=True, device=device)
+    Vout = wp.zeros(1, dtype=mat22, requires_grad=True, device=device)
+
+    wp.launch(kernel, dim=1, inputs=[m2], outputs=[Uout, sigmaout, Vout, outcomponents], device=device)
+
+    Uout_np = Uout.numpy()[0].astype(np.float64)
+    sigmaout_np = np.diag(sigmaout.numpy()[0].astype(np.float64))
+    Vout_np = Vout.numpy()[0].astype(np.float64)
+
+    assert_np_equal(
+        np.matmul(Uout_np, np.matmul(sigmaout_np, Vout_np.T)), m2.numpy()[0].astype(np.float64), tol=30 * tol
+    )
+
+    if dtype == np.float16:
+        # Skip gradient check for float16 due to rounding errors
+        return
+
+    # Check gradients:
+    out = wp.zeros(1, dtype=wptype, requires_grad=True, device=device)
+    idx = 0
+    for idx in range(2 * 2 + 2 + 2 * 2):
+        tape = wp.Tape()
+        with tape:
+            wp.launch(kernel, dim=1, inputs=[m2], outputs=[Uout, sigmaout, Vout, outcomponents], device=device)
+            wp.launch(output_select_kernel, dim=1, inputs=[outcomponents, idx], outputs=[out], device=device)
+        tape.backward(out)
+        m2grads = 1.0 * tape.gradients[m2].numpy()[0]
+
+        tape.zero()
+
+        dx = 0.0001
+        fdtol = 5.0e-4 if dtype == np.float64 else 2.0e-2
+        for ii in range(2):
+            for jj in range(2):
+                m2test = 1.0 * m2.numpy()
+                m2test[0, ii, jj] += dx
+                wp.launch(
+                    kernel,
+                    dim=1,
+                    inputs=[wp.array(m2test, dtype=mat22, device=device)],
+                    outputs=[Uout, sigmaout, Vout, outcomponents],
+                    device=device,
+                )
+                wp.launch(output_select_kernel, dim=1, inputs=[outcomponents, idx], outputs=[out], device=device)
+                plusval = out.numpy()[0]
+
+                m2test = 1.0 * m2.numpy()
+                m2test[0, ii, jj] -= dx
+                wp.launch(
+                    kernel,
+                    dim=1,
+                    inputs=[wp.array(m2test, dtype=mat22, device=device)],
+                    outputs=[Uout, sigmaout, Vout, outcomponents],
+                    device=device,
+                )
+                wp.launch(output_select_kernel, dim=1, inputs=[outcomponents, idx], outputs=[out], device=device)
+                minusval = out.numpy()[0]
+
+                assert_np_equal((plusval - minusval) / (2 * dx), m2grads[ii, jj], tol=fdtol)
+
+
 def test_qr(test, device, dtype, register_kernels=False):
     rng = np.random.default_rng(123)
 
@@ -1513,83 +1623,6 @@ def test_transform_vector(test, device, dtype, register_kernels=False):
             tape.zero()
 
 
-def test_mat_array_type_indexing(test, device, dtype, register_kernels=False):
-    np_type = np.dtype(dtype)
-    wp_type = wp.types.np_dtype_to_warp_type[np_type]
-
-    vec2 = wp.types.vector(length=2, dtype=wp_type)
-    mat22 = wp.types.matrix(shape=(2, 2), dtype=wp_type)
-    mat33 = wp.types.matrix(shape=(3, 3), dtype=wp_type)
-
-    def mattest_read_write_store(x: wp.array(dtype=wp_type), a: wp.array(dtype=mat22)):
-        tid = wp.tid()
-
-        t = a[tid]
-        t[0, 0] = x[tid]
-        a[tid] = t
-
-    def mattest_in_register(x: wp.array2d(dtype=mat22), y: wp.array(dtype=vec2)):
-        i, j = wp.tid()
-
-        a = mat22(wp_type(0.0))
-        a[0] = y[i]
-        a[1, 1] = wp_type(3.0)
-        x[i, j] = a
-
-    def mattest_in_register_overwrite(x: wp.array2d(dtype=mat22), y: wp.array(dtype=vec2)):
-        i, j = wp.tid()
-
-        a = mat22(wp_type(0.0))
-        a[0] = y[i]
-        a[0, 1] = wp_type(3.0)
-        x[i, j] = a
-
-    kernel_read_write_store = getkernel(mattest_read_write_store, suffix=dtype.__name__)
-    kernel_in_register = getkernel(mattest_in_register, suffix=dtype.__name__)
-    kernel_in_register_overwrite = getkernel(mattest_in_register_overwrite, suffix=dtype.__name__)
-
-    if register_kernels:
-        return
-
-    a = wp.ones(1, dtype=mat22, device=device, requires_grad=True)
-    x = wp.full(1, value=2.0, dtype=wp_type, device=device, requires_grad=True)
-
-    tape = wp.Tape()
-    with tape:
-        wp.launch(kernel_read_write_store, dim=1, inputs=[x, a], device=device)
-
-    tape.backward(grads={a: wp.ones_like(a, requires_grad=False)})
-
-    assert_np_equal(a.numpy(), np.array([[[2.0, 1.0], [1.0, 1.0]]], dtype=np_type))
-    assert_np_equal(x.grad.numpy(), np.array([1.0], dtype=np_type))
-
-    tape.reset()
-
-    x = wp.zeros((1, 1), dtype=mat22, device=device, requires_grad=True)
-    y = wp.ones(1, dtype=vec2, device=device, requires_grad=True)
-
-    with tape:
-        wp.launch(kernel_in_register, dim=(1, 1), inputs=[x, y], device=device)
-
-    tape.backward(grads={x: wp.ones_like(x, requires_grad=False)})
-
-    assert_np_equal(x.numpy(), np.array([[[[1.0, 1.0], [0.0, 3.0]]]], dtype=np_type))
-    assert_np_equal(y.grad.numpy(), np.array([[1.0, 1.0]], dtype=np_type))
-
-    tape.reset()
-
-    x = wp.zeros((1, 1), dtype=mat22, device=device, requires_grad=True)
-    y = wp.ones(1, dtype=vec2, device=device, requires_grad=True)
-
-    with tape:
-        wp.launch(kernel_in_register_overwrite, dim=(1, 1), inputs=[x, y], device=device)
-
-    tape.backward(grads={x: wp.ones_like(x, requires_grad=False)})
-
-    assert_np_equal(x.numpy(), np.array([[[[1.0, 3.0], [0.0, 0.0]]]], dtype=np_type))
-    assert_np_equal(y.grad.numpy(), np.array([[1.0, 0.0]], dtype=np_type))
-
-
 # Test matrix constructors using explicit type (float16)
 # note that these tests are specifically not using generics / closure
 # args to create kernels dynamically (like the rest of this file)
@@ -1623,10 +1656,62 @@ def test_matrix_constructor_value_func():
     c = mat32d()
     d = mat32d(c, shape=(3, 2))
     e = mat32d(wp.float64(1.0), wp.float64(2.0), wp.float64(1.0), wp.float64(2.0), wp.float64(1.0), wp.float64(2.0))
-    f = mat32d(
-        wp.vec3d(wp.float64(1.0), wp.float64(2.0), wp.float64(3.0)),
-        wp.vec3d(wp.float64(1.0), wp.float64(2.0), wp.float64(3.0)),
+    f = wp.matrix(1.0, 2.0, 3.0, 4.0, shape=(2, 2), dtype=float)
+
+
+@wp.kernel
+def test_matrix_from_vecs():
+    m1 = wp.matrix_from_cols(
+        wp.vec3(1.0, 2.0, 3.0),
+        wp.vec3(4.0, 5.0, 6.0),
+        wp.vec3(7.0, 8.0, 9.0),
+    )
+    wp.expect_eq(m1[0, 0], 1.0)
+    wp.expect_eq(m1[0, 1], 4.0)
+    wp.expect_eq(m1[0, 2], 7.0)
+    wp.expect_eq(m1[1, 0], 2.0)
+    wp.expect_eq(m1[1, 1], 5.0)
+    wp.expect_eq(m1[1, 2], 8.0)
+    wp.expect_eq(m1[2, 0], 3.0)
+    wp.expect_eq(m1[2, 1], 6.0)
+    wp.expect_eq(m1[2, 2], 9.0)
+
+    m2 = wp.matrix_from_rows(
+        wp.vec3(1.0, 2.0, 3.0),
+        wp.vec3(4.0, 5.0, 6.0),
+        wp.vec3(7.0, 8.0, 9.0),
+    )
+    wp.expect_eq(m2[0, 0], 1.0)
+    wp.expect_eq(m2[0, 1], 2.0)
+    wp.expect_eq(m2[0, 2], 3.0)
+    wp.expect_eq(m2[1, 0], 4.0)
+    wp.expect_eq(m2[1, 1], 5.0)
+    wp.expect_eq(m2[1, 2], 6.0)
+    wp.expect_eq(m2[2, 0], 7.0)
+    wp.expect_eq(m2[2, 1], 8.0)
+    wp.expect_eq(m2[2, 2], 9.0)
+
+    m3 = wp.matrix_from_cols(
+        wp.vec3(1.0, 2.0, 3.0),
+        wp.vec3(4.0, 5.0, 6.0),
     )
+    wp.expect_eq(m3[0, 0], 1.0)
+    wp.expect_eq(m3[0, 1], 4.0)
+    wp.expect_eq(m3[1, 0], 2.0)
+    wp.expect_eq(m3[1, 1], 5.0)
+    wp.expect_eq(m3[2, 0], 3.0)
+    wp.expect_eq(m3[2, 1], 6.0)
+
+    m4 = wp.matrix_from_rows(
+        wp.vec3(1.0, 2.0, 3.0),
+        wp.vec3(4.0, 5.0, 6.0),
+    )
+    wp.expect_eq(m4[0, 0], 1.0)
+    wp.expect_eq(m4[0, 1], 2.0)
+    wp.expect_eq(m4[0, 2], 3.0)
+    wp.expect_eq(m4[1, 0], 4.0)
+    wp.expect_eq(m4[1, 1], 5.0)
+    wp.expect_eq(m4[1, 2], 6.0)
 
 
 # Same as above but with a default (float/int) type
@@ -1742,54 +1827,389 @@ def test_matrix_len(test, device):
 
 
 @wp.kernel
-def matrix_augassign_kernel(
-    a: wp.array(dtype=wp.mat22), b: wp.array(dtype=wp.mat22), c: wp.array(dtype=wp.mat22), d: wp.array(dtype=wp.mat22)
-):
+def mat_extract_element(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=float)):
+    tid = wp.tid()
+
+    a = x[tid]
+    b = a[0, 0] + 2.0 * a[0, 1] + 3.0 * a[1, 0] + 4.0 * a[1, 1]
+    y[tid] = b
+
+
+@wp.kernel
+def mat_extract_row(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.vec2)):
+    tid = wp.tid()
+
+    a = x[tid]
+    b = a[0] + 2.0 * a[1]
+    y[tid] = b
+
+
+def test_mat_extract(test, device):
+    # matrix element
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=float, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_extract_element, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([10.0], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[1.0, 2.0], [3.0, 4.0]]], dtype=float))
+
+    # matrix row
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.vec2, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_extract_row, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[3.0, 3.0]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[1.0, 1.0], [2.0, 2.0]]], dtype=float))
+
+
+@wp.kernel
+def mat_assign_element(x: wp.array(dtype=float), y: wp.array(dtype=wp.mat22)):
     i = wp.tid()
 
-    m1 = wp.mat22()
-    m2 = b[i]
+    a = wp.mat22()
+    a[0, 0] = 1.0 * x[i]
+    a[0, 1] = 2.0 * x[i]
+    a[1, 0] = 3.0 * x[i]
+    a[1, 1] = 4.0 * x[i]
+
+    y[i] = a
+
+
+@wp.kernel
+def mat_assign_row(x: wp.array(dtype=wp.vec2), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
+
+    a = wp.mat22()
+    a[0] = 1.0 * x[i]
+    a[1] = 2.0 * x[i]
+
+    y[i] = a
+
+
+def test_mat_assign(test, device):
+    # matrix element
+    x = wp.ones(1, dtype=float, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_assign_element, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[1.0, 2.0], [3.0, 4.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([10.0], dtype=float))
+
+    # matrix row
+    x = wp.ones(1, dtype=wp.vec2, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_assign_row, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[1.0, 1.0], [2.0, 2.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[3.0, 3.0]], dtype=float))
+
+
+def test_matrix_assign_copy(test, device):
+    saved_enable_vector_component_overwrites_setting = wp.config.enable_vector_component_overwrites
+    try:
+        wp.config.enable_vector_component_overwrites = True
+
+        @wp.kernel
+        def mat_in_register_overwrite(x: wp.array2d(dtype=wp.mat22), y: wp.array(dtype=wp.vec2)):
+            i, j = wp.tid()
+
+            a = wp.mat22()
+            a[0] = y[i]
+            a[0, 1] = 3.0
+            x[i, j] = a
+
+        x = wp.zeros((1, 1), dtype=wp.mat22, device=device, requires_grad=True)
+        y = wp.ones(1, dtype=wp.vec2, device=device, requires_grad=True)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(mat_in_register_overwrite, dim=(1, 1), inputs=[x, y], device=device)
+
+        tape.backward(grads={x: wp.ones_like(x, requires_grad=False)})
+
+        assert_np_equal(x.numpy(), np.array([[[[1.0, 3.0], [0.0, 0.0]]]], dtype=float))
+        assert_np_equal(y.grad.numpy(), np.array([[1.0, 0.0]], dtype=float))
+
+    finally:
+        wp.config.enable_vector_component_overwrites = saved_enable_vector_component_overwrites_setting
+
+
+@wp.kernel
+def mat_array_extract_element(x: wp.array2d(dtype=wp.mat22), y: wp.array2d(dtype=float)):
+    i, j = wp.tid()
+    a = x[i, j][0, 0]
+    b = x[i, j][0, 1]
+    c = x[i, j][1, 0]
+    d = x[i, j][1, 1]
+    y[i, j] = 1.0 * a + 2.0 * b + 3.0 * c + 4.0 * d
+
+
+@wp.kernel
+def mat_array_extract_row(x: wp.array2d(dtype=wp.mat22), y: wp.array2d(dtype=wp.vec2)):
+    i, j = wp.tid()
+    a = x[i, j][0]
+    b = x[i, j][1]
+    y[i, j] = 1.0 * a + 2.0 * b
+
+
+def test_mat_array_extract(test, device):
+    # matrix element
+    x = wp.ones((1, 1), dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros((1, 1), dtype=float, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_array_extract_element, (1, 1), inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[10.0]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[[1.0, 2.0], [3.0, 4.0]]]], dtype=float))
+
+    # matrix row
+    x = wp.ones((1, 1), dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros((1, 1), dtype=wp.vec2, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_array_extract_row, (1, 1), inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[3.0, 3.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[[1.0, 1.0], [2.0, 2.0]]]], dtype=float))
+
+
+""" TODO: gradient propagation for in-place array assignment
+@wp.kernel
+def mat_array_assign_element(x: wp.array2d(dtype=float), y: wp.array2d(dtype=wp.mat22)):
+    i, j = wp.tid()
+
+    y[i, j][0, 0] = 1.0 * x[i, j]
+    y[i, j][0, 1] = 2.0 * x[i, j]
+    y[i, j][1, 0] = 3.0 * x[i, j]
+    y[i, j][1, 1] = 4.0 * x[i, j]
+
+
+@wp.kernel
+def mat_array_assign_row(x: wp.array2d(dtype=wp.vec3), y: wp.array2d(dtype=wp.mat(shape=(2, 3), dtype=float))):
+    i, j = wp.tid()
+
+    y[i, j][0] = 1.0 * x[i, j]
+    y[i, j][1] = 2.0 * x[i, j]
+
+
+def test_mat_array_assign(test, device):
+    # matrix element
+    x = wp.ones((1, 1), dtype=float, requires_grad=True, device=device)
+    y = wp.zeros((1, 1), dtype=wp.mat22, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_array_assign_element, (1, 1), inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[[1.0, 2.0], [3.0, 4.0]]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[10.0]], dtype=float))
+
+    # matrix row
+    x = wp.ones((1, 1), dtype=wp.vec3, requires_grad=True, device=device)
+    y = wp.zeros((1, 1), dtype=wp.mat(shape=(2, 3), dtype=float), requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_array_assign_row, (1, 1), inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[[1.0, 1.0, 1.0], [2.0, 2.0, 2.0]]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[3.0, 3.0, 3.0]]], dtype=float))
+"""
+
+
+@wp.kernel
+def mat_add_inplace_element(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
+
+    a = wp.mat22()
+    b = x[i]
+
+    a[0, 0] += 1.0 * b[0, 0]
+    a[0, 1] += 2.0 * b[0, 1]
+    a[1, 0] += 3.0 * b[1, 0]
+    a[1, 1] += 4.0 * b[1, 1]
+
+    y[i] = a
+
+
+@wp.kernel
+def mat_add_inplace_row(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
+
+    a = wp.mat22()
+    b = x[i]
+
+    a[0] += 1.0 * b[0]
+    a[1] += 2.0 * b[1]
+
+    y[i] = a
+
+
+def test_mat_add_inplace(test, device):
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_add_inplace_element, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[1.0, 2.0], [3.0, 4.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[1.0, 2.0], [3.0, 4.0]]], dtype=float))
+
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
 
-    m1[0, 0] += m2[0, 0]
-    m1[0, 1] += m2[0, 1]
-    m1[1, 0] += m2[1, 0]
-    m1[1, 1] += m2[1, 1]
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_add_inplace_row, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
 
-    a[i] = m1
+    assert_np_equal(y.numpy(), np.array([[[1.0, 1.0], [2.0, 2.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[1.0, 1.0], [2.0, 2.0]]], dtype=float))
 
-    m3 = wp.mat22()
-    m4 = d[i]
 
-    m3[0, 0] -= m4[0, 0]
-    m3[0, 1] -= m4[0, 1]
-    m3[1, 0] -= m4[1, 0]
-    m3[1, 1] -= m4[1, 1]
+@wp.kernel
+def mat_sub_inplace_element(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
 
-    c[i] = m3
+    a = wp.mat22()
+    b = x[i]
 
+    a[0, 0] -= 1.0 * b[0, 0]
+    a[0, 1] -= 2.0 * b[0, 1]
+    a[1, 0] -= 3.0 * b[1, 0]
+    a[1, 1] -= 4.0 * b[1, 1]
 
-def test_matrix_augassign(test, device):
-    N = 3
+    y[i] = a
 
-    a = wp.zeros(N, dtype=wp.mat22, requires_grad=True)
-    b = wp.ones(N, dtype=wp.mat22, requires_grad=True)
 
-    c = wp.zeros(N, dtype=wp.mat22, requires_grad=True)
-    d = wp.ones(N, dtype=wp.mat22, requires_grad=True)
+@wp.kernel
+def mat_sub_inplace_row(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
+
+    a = wp.mat22()
+    b = x[i]
+
+    a[0] -= 1.0 * b[0]
+    a[1] -= 2.0 * b[1]
+
+    y[i] = a
+
+
+def test_mat_sub_inplace(test, device):
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
 
     tape = wp.Tape()
     with tape:
-        wp.launch(matrix_augassign_kernel, N, inputs=[a, b, c, d])
+        wp.launch(mat_sub_inplace_element, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[-1.0, -2.0], [-3.0, -4.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[-1.0, -2.0], [-3.0, -4.0]]], dtype=float))
 
-    tape.backward(grads={a: wp.ones_like(a), c: wp.ones_like(c)})
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
 
-    assert_np_equal(a.numpy(), wp.ones_like(a).numpy())
-    assert_np_equal(a.grad.numpy(), wp.ones_like(a).numpy())
-    assert_np_equal(b.grad.numpy(), wp.ones_like(a).numpy())
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_sub_inplace_row, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
 
-    assert_np_equal(c.numpy(), -wp.ones_like(c).numpy())
-    assert_np_equal(c.grad.numpy(), wp.ones_like(c).numpy())
-    assert_np_equal(d.grad.numpy(), -wp.ones_like(d).numpy())
+    assert_np_equal(y.numpy(), np.array([[[-1.0, -1.0], [-2.0, -2.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[-1.0, -1.0], [-2.0, -2.0]]], dtype=float))
+
+
+@wp.kernel
+def mat_array_add_inplace(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
+
+    y[i] += x[i]
+
+
+def test_mat_array_add_inplace(test, device):
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_array_add_inplace, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[1.0, 1.0], [1.0, 1.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[1.0, 1.0], [1.0, 1.0]]], dtype=float))
+
+
+@wp.kernel
+def mat_array_sub_inplace(x: wp.array(dtype=wp.mat22), y: wp.array(dtype=wp.mat22)):
+    i = wp.tid()
+
+    y[i] -= x[i]
+
+
+def test_mat_array_sub_inplace(test, device):
+    x = wp.ones(1, dtype=wp.mat22, requires_grad=True, device=device)
+    y = wp.zeros(1, dtype=wp.mat22, requires_grad=True, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(mat_array_sub_inplace, 1, inputs=[x], outputs=[y], device=device)
+
+    y.grad = wp.ones_like(y)
+    tape.backward()
+
+    assert_np_equal(y.numpy(), np.array([[[-1.0, -1.0], [-1.0, -1.0]]], dtype=float))
+    assert_np_equal(x.grad.numpy(), np.array([[[-1.0, -1.0], [-1.0, -1.0]]], dtype=float))
 
 
 devices = get_test_devices()
@@ -1814,6 +2234,7 @@ add_kernel_test(TestMat, test_constructors_explicit_precision, dim=1, devices=de
 add_kernel_test(TestMat, test_constructors_default_precision, dim=1, devices=devices)
 add_kernel_test(TestMat, test_constructors_constant_shape, dim=1, devices=devices)
 add_kernel_test(TestMat, test_matrix_constructor_value_func, dim=1, devices=devices)
+add_kernel_test(TestMat, test_matrix_from_vecs, dim=1, devices=devices)
 
 mat103 = wp.types.matrix(shape=(10, 3), dtype=float)
 add_kernel_test(
@@ -1848,6 +2269,12 @@ for dtype in np_signed_int_types + np_float_types:
         TestMat, f"test_matmul_{dtype.__name__}", test_matmul, devices=devices, dtype=dtype
     )
 
+add_function_test(
+    TestMat,
+    "test_shape_mismatch",
+    test_shape_mismatch,
+    devices=devices,
+)
 add_function_test(
     TestMat,
     "test_anon_constructor_error_shape_arg_missing",
@@ -1878,18 +2305,6 @@ add_function_test(
     test_tpl_constructor_error_incompatible_sizes,
     devices=devices,
 )
-add_function_test(
-    TestMat,
-    "test_tpl_constructor_error_invalid_vector_count",
-    test_tpl_constructor_error_invalid_vector_count,
-    devices=devices,
-)
-add_function_test(
-    TestMat,
-    "test_tpl_constructor_error_invalid_vector_shape",
-    test_tpl_constructor_error_invalid_vector_shape,
-    devices=devices,
-)
 add_function_test(
     TestMat,
     "test_tpl_constructor_error_invalid_arg_count",
@@ -1908,6 +2323,9 @@ for dtype in np_float_types:
         TestMat, f"test_inverse_{dtype.__name__}", test_inverse, devices=devices, dtype=dtype
     )
     add_function_test_register_kernel(TestMat, f"test_svd_{dtype.__name__}", test_svd, devices=devices, dtype=dtype)
+    add_function_test_register_kernel(
+        TestMat, f"test_svd_2D{dtype.__name__}", test_svd_2D, devices=devices, dtype=dtype
+    )
     add_function_test_register_kernel(TestMat, f"test_qr_{dtype.__name__}", test_qr, devices=devices, dtype=dtype)
     add_function_test_register_kernel(TestMat, f"test_eig_{dtype.__name__}", test_eig, devices=devices, dtype=dtype)
     add_function_test_register_kernel(
@@ -1920,15 +2338,18 @@ for dtype in np_float_types:
         TestMat, f"test_determinant_{dtype.__name__}", test_determinant, devices=devices, dtype=dtype
     )
     add_function_test_register_kernel(TestMat, f"test_skew_{dtype.__name__}", test_skew, devices=devices, dtype=dtype)
-    add_function_test_register_kernel(
-        TestMat,
-        f"test_mat_array_type_indexing_{dtype.__name__}",
-        test_mat_array_type_indexing,
-        devices=devices,
-        dtype=dtype,
-    )
+
 add_function_test(TestMat, "test_matrix_len", test_matrix_len, devices=devices)
-add_function_test(TestMat, "test_matrix_augassign", test_matrix_augassign, devices=devices)
+add_function_test(TestMat, "test_mat_extract", test_mat_extract, devices=devices)
+add_function_test(TestMat, "test_mat_assign", test_mat_assign, devices=devices)
+add_function_test(TestMat, "test_matrix_assign_copy", test_matrix_assign_copy, devices=devices)
+add_function_test(TestMat, "test_mat_array_extract", test_mat_array_extract, devices=devices)
+# add_function_test(TestMat, "test_mat_array_assign", test_mat_array_assign, devices=devices)
+add_function_test(TestMat, "test_mat_add_inplace", test_mat_add_inplace, devices=devices)
+add_function_test(TestMat, "test_mat_sub_inplace", test_mat_sub_inplace, devices=devices)
+add_function_test(TestMat, "test_mat_array_add_inplace", test_mat_array_add_inplace, devices=devices)
+add_function_test(TestMat, "test_mat_array_sub_inplace", test_mat_array_sub_inplace, devices=devices)
+
 
 if __name__ == "__main__":
     wp.clear_kernel_cache()