Merge pull request #157 from desy-ml/154-missing-batch-execution-compatibility-in-dipole-rotation-matrix-etc

Fix the batch execution for elements with tilts or misalignments

Author: jank324

CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### 🚨 Breaking Changes
 
-- Cheetah is now vectorised. This means that you can run multiple simulations in parallel by passing a batch of beams and settings, resulting a number of interfaces being changed. For Cheetah developers this means that you now have to account for an arbitrary-dimensional tensor of most of the properties of you element, rather than a single value, vector or whatever else a property was before. (see #116) (@jank324)
+- Cheetah is now vectorised. This means that you can run multiple simulations in parallel by passing a batch of beams and settings, resulting a number of interfaces being changed. For Cheetah developers this means that you now have to account for an arbitrary-dimensional tensor of most of the properties of you element, rather than a single value, vector or whatever else a property was before. (see #116, #157) (@jank324, @cr-xu)
 
 ### 🚀 Features
 

cheetah/accelerator.py

@@ -379,13 +379,11 @@ def transfer_map(self, energy: torch.Tensor) -> torch.Tensor:
             energy=energy,
         )
 
-        if torch.all(self.misalignment[:, 0] == 0) and torch.all(
-            self.misalignment[:, 1] == 0
-        ):
+        if torch.all(self.misalignment == 0):
             return R
         else:
-            R_exit, R_entry = misalignment_matrix(self.misalignment)
-            R = torch.matmul(R_exit, torch.matmul(R, R_entry))
+            R_entry, R_exit = misalignment_matrix(self.misalignment)
+            R = torch.einsum("...ij,...jk,...kl->...il", R_exit, R, R_entry)
             return R
 
     def broadcast(self, shape: Size) -> Element:
@@ -542,23 +540,21 @@ def transfer_map(self, energy: torch.Tensor) -> torch.Tensor:
                 hx=self.hx,
                 tilt=torch.zeros_like(self.length),
                 energy=energy,
-            )
+            )  # Tilt is applied after adding edges
         else:  # Reduce to Thin-Corrector
             R = torch.eye(7, device=device, dtype=dtype).repeat(
                 (*self.length.shape, 1, 1)
             )
-            R[:, 0, 1] = self.length
-            R[:, 2, 6] = self.angle
-            R[:, 2, 3] = self.length
+            R[..., 0, 1] = self.length
+            R[..., 2, 6] = self.angle
+            R[..., 2, 3] = self.length
 
         # Apply fringe fields
         R = torch.matmul(R_exit, torch.matmul(R, R_enter))
         # Apply rotation for tilted magnets
-        # TODO: Are we applying tilt twice (here and base_rmatrix)?
         R = torch.matmul(
             rotation_matrix(-self.tilt), torch.matmul(R, rotation_matrix(self.tilt))
         )
-
         return R
 
     def _transfer_map_enter(self) -> torch.Tensor:
@@ -576,8 +572,8 @@ def _transfer_map_enter(self) -> torch.Tensor:
         )
 
         tm = torch.eye(7, device=device, dtype=dtype).repeat(*phi.shape, 1, 1)
-        tm[:, 1, 0] = self.hx * torch.tan(self.e1)
-        tm[:, 3, 2] = -self.hx * torch.tan(self.e1 - phi)
+        tm[..., 1, 0] = self.hx * torch.tan(self.e1)
+        tm[..., 3, 2] = -self.hx * torch.tan(self.e1 - phi)
 
         return tm
 
@@ -596,8 +592,8 @@ def _transfer_map_exit(self) -> torch.Tensor:
         )
 
         tm = torch.eye(7, device=device, dtype=dtype).repeat(*phi.shape, 1, 1)
-        tm[:, 1, 0] = self.hx * torch.tan(self.e2)
-        tm[:, 3, 2] = -self.hx * torch.tan(self.e2 - phi)
+        tm[..., 1, 0] = self.hx * torch.tan(self.e2)
+        tm[..., 3, 2] = -self.hx * torch.tan(self.e2 - phi)
 
         return tm
 
@@ -1448,11 +1444,11 @@ def track(self, incoming: Beam) -> Beam:
             copy_of_incoming = deepcopy(incoming)
 
             if isinstance(incoming, ParameterBeam):
-                copy_of_incoming._mu[:, 0] -= self.misalignment[:, 0]
-                copy_of_incoming._mu[:, 2] -= self.misalignment[:, 1]
+                copy_of_incoming._mu[..., 0] -= self.misalignment[..., 0]
+                copy_of_incoming._mu[..., 2] -= self.misalignment[..., 1]
             elif isinstance(incoming, ParticleBeam):
-                copy_of_incoming.particles[:, :, 0] -= self.misalignment[:, 0]
-                copy_of_incoming.particles[:, :, 1] -= self.misalignment[:, 1]
+                copy_of_incoming.particles[..., :, 0] -= self.misalignment[..., 0]
+                copy_of_incoming.particles[..., :, 1] -= self.misalignment[..., 1]
 
             self.set_read_beam(copy_of_incoming)
 
@@ -1476,18 +1472,18 @@ def reading(self) -> torch.Tensor:
             )
         elif isinstance(read_beam, ParameterBeam):
             transverse_mu = torch.stack(
-                [read_beam._mu[:, 0], read_beam._mu[:, 2]], dim=1
+                [read_beam._mu[..., 0], read_beam._mu[..., 2]], dim=-1
             )
             transverse_cov = torch.stack(
                 [
                     torch.stack(
-                        [read_beam._cov[:, 0, 0], read_beam._cov[:, 0, 2]], dim=1
+                        [read_beam._cov[..., 0, 0], read_beam._cov[..., 0, 2]], dim=-1
                     ),
                     torch.stack(
-                        [read_beam._cov[:, 2, 0], read_beam._cov[:, 2, 2]], dim=1
+                        [read_beam._cov[..., 2, 0], read_beam._cov[..., 2, 2]], dim=-1
                     ),
                 ],
-                dim=1,
+                dim=-1,
             )
             dist = [
                 MultivariateNormal(
@@ -1767,9 +1763,9 @@ def transfer_map(self, energy: torch.Tensor) -> torch.Tensor:
         )
 
         tm = torch.eye(7, device=device, dtype=dtype).repeat((*energy.shape, 1, 1))
-        tm[:, 0, 1] = self.length
-        tm[:, 2, 3] = self.length
-        tm[:, 4, 5] = self.length * igamma2
+        tm[..., 0, 1] = self.length
+        tm[..., 2, 3] = self.length
+        tm[..., 4, 5] = self.length * igamma2
 
         return tm
 
@@ -1866,38 +1862,38 @@ def transfer_map(self, energy: torch.Tensor) -> torch.Tensor:
             r56 -= self.length / (beta * beta * gamma2)
 
         R = torch.eye(7, device=device, dtype=dtype).repeat((*self.length.shape, 1, 1))
-        R[:, 0, 0] = c**2
-        R[:, 0, 1] = c * s_k
-        R[:, 0, 2] = s * c
-        R[:, 0, 3] = s * s_k
-        R[:, 1, 0] = -self.k * s * c
-        R[:, 1, 1] = c**2
-        R[:, 1, 2] = -self.k * s**2
-        R[:, 1, 3] = s * c
-        R[:, 2, 0] = -s * c
-        R[:, 2, 1] = -s * s_k
-        R[:, 2, 2] = c**2
-        R[:, 2, 3] = c * s_k
-        R[:, 3, 0] = self.k * s**2
-        R[:, 3, 1] = -s * c
-        R[:, 3, 2] = -self.k * s * c
-        R[:, 3, 3] = c**2
-        R[:, 4, 5] = r56
+        R[..., 0, 0] = c**2
+        R[..., 0, 1] = c * s_k
+        R[..., 0, 2] = s * c
+        R[..., 0, 3] = s * s_k
+        R[..., 1, 0] = -self.k * s * c
+        R[..., 1, 1] = c**2
+        R[..., 1, 2] = -self.k * s**2
+        R[..., 1, 3] = s * c
+        R[..., 2, 0] = -s * c
+        R[..., 2, 1] = -s * s_k
+        R[..., 2, 2] = c**2
+        R[..., 2, 3] = c * s_k
+        R[..., 3, 0] = self.k * s**2
+        R[..., 3, 1] = -s * c
+        R[..., 3, 2] = -self.k * s * c
+        R[..., 3, 3] = c**2
+        R[..., 4, 5] = r56
 
         R = R.real
 
-        if self.misalignment[0] == 0 and self.misalignment[1] == 0:
+        if torch.all(self.misalignment == 0):
             return R
         else:
-            R_exit, R_entry = misalignment_matrix(self.misalignment)
-            R = torch.matmul(R_exit, torch.matmul(R, R_entry))
+            R_entry, R_exit = misalignment_matrix(self.misalignment)
+            R = torch.einsum("...ij,...jk,...kl->...il", R_exit, R, R_entry)
             return R
 
     def broadcast(self, shape: Size) -> Element:
         return self.__class__(
             length=self.length.repeat(shape),
             k=self.k.repeat(shape),
-            misalignment=self.misalignment.repeat(shape),
+            misalignment=self.misalignment.repeat((*shape, 1)),
             name=self.name,
         )
 

cheetah/track_methods.py

@@ -20,15 +20,15 @@ def rotation_matrix(angle: torch.Tensor) -> torch.Tensor:
     cs = torch.cos(angle)
     sn = torch.sin(angle)
 
-    tm = torch.eye(7, dtype=angle.dtype, device=angle.device)
-    tm[0, 0] = cs
-    tm[0, 2] = sn
-    tm[1, 1] = cs
-    tm[1, 3] = sn
-    tm[2, 0] = -sn
-    tm[2, 2] = cs
-    tm[3, 1] = -sn
-    tm[3, 3] = cs
+    tm = torch.eye(7, dtype=angle.dtype, device=angle.device).repeat(*angle.shape, 1, 1)
+    tm[..., 0, 0] = cs
+    tm[..., 0, 2] = sn
+    tm[..., 1, 1] = cs
+    tm[..., 1, 3] = sn
+    tm[..., 2, 0] = -sn
+    tm[..., 2, 2] = cs
+    tm[..., 3, 1] = -sn
+    tm[..., 3, 3] = cs
 
     return tm
 
@@ -98,7 +98,9 @@ def base_rmatrix(
 
     # Rotate the R matrix for skew / vertical magnets
     if torch.any(tilt != 0):
-        R = torch.matmul(torch.matmul(rotation_matrix(-tilt), R), rotation_matrix(tilt))
+        R = torch.einsum(
+            "...ij,...jk,...kl->...il", rotation_matrix(-tilt), R, rotation_matrix(tilt)
+        )
     return R
 
 
@@ -108,13 +110,14 @@ def misalignment_matrix(
     """Shift the beam for tracking beam through misaligned elements"""
     device = misalignment.device
     dtype = misalignment.dtype
+    batch_shape = misalignment.shape[:-1]
 
-    R_exit = torch.eye(7, device=device, dtype=dtype)
-    R_exit[0, 6] = misalignment[0]
-    R_exit[2, 6] = misalignment[1]
+    R_exit = torch.eye(7, device=device, dtype=dtype).repeat(*batch_shape, 1, 1)
+    R_exit[..., 0, 6] = misalignment[..., 0]
+    R_exit[..., 2, 6] = misalignment[..., 1]
 
-    R_entry = torch.eye(7, device=device, dtype=dtype)
-    R_entry[0, 6] = -misalignment[0]
-    R_entry[2, 6] = -misalignment[1]
+    R_entry = torch.eye(7, device=device, dtype=dtype).repeat(*batch_shape, 1, 1)
+    R_entry[..., 0, 6] = -misalignment[..., 0]
+    R_entry[..., 2, 6] = -misalignment[..., 1]
 
-    return R_exit, R_entry  # TODO: This order is confusing, should be entry, exit
+    return R_entry, R_exit

tests/test_compare_ocelot.py

@@ -101,6 +101,52 @@ def test_dipole_with_fringe_field():
     )
 
 
+def test_dipole_with_fringe_field_and_tilt():
+    """
+    Test that the tracking results through a Cheeath `Dipole` element match those
+    through an Oclet `Bend` element when there are fringe fields and tilt, and the
+    e1 and e2 angles are set.
+    """
+    # Cheetah
+    bend_angle = np.pi / 6
+    tilt_angle = np.pi / 4
+    incoming_beam = cheetah.ParticleBeam.from_astra(
+        "tests/resources/ACHIP_EA1_2021.1351.001"
+    )
+    cheetah_dipole = cheetah.Dipole(
+        length=torch.tensor([1.0]),
+        angle=torch.tensor([bend_angle]),
+        fringe_integral=torch.tensor([0.1]),
+        gap=torch.tensor([0.2]),
+        tilt=torch.tensor([tilt_angle]),
+        e1=torch.tensor([bend_angle / 2]),
+        e2=torch.tensor([bend_angle / 2]),
+    )
+    outgoing_beam = cheetah_dipole(incoming_beam)
+
+    # Ocelot
+    incoming_p_array = ocelot.astraBeam2particleArray(
+        "tests/resources/ACHIP_EA1_2021.1351.001"
+    )
+    ocelot_bend = ocelot.Bend(
+        l=1.0,
+        angle=bend_angle,
+        fint=0.1,
+        gap=0.2,
+        tilt=tilt_angle,
+        e1=bend_angle / 2,
+        e2=bend_angle / 2,
+    )
+    lattice = ocelot.MagneticLattice([ocelot_bend, ocelot.Marker()])
+    navigator = ocelot.Navigator(lattice)
+    _, outgoing_p_array = ocelot.track(lattice, deepcopy(incoming_p_array), navigator)
+
+    assert np.allclose(
+        outgoing_beam.particles[0, :, :6].cpu().numpy(),
+        outgoing_p_array.rparticles.transpose(),
+    )
+
+
 def test_aperture():
     """
     Test that the tracking results through a Cheeath `Aperture` element match those