diffct 2.4.3__py3-none-any.whl → 2.5.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
diffct/__init__.py CHANGED
@@ -64,7 +64,7 @@ try:
64
64
  except ImportError:
65
65
  pass
66
66
 
67
- __version__ = '2.4.3'
67
+ __version__ = '2.5.0'
68
68
 
69
69
  __all__ = [
70
70
  'Projector',
diffct/_solver_state.py CHANGED
@@ -12,6 +12,7 @@ import torch.distributed as dist
12
12
  from .pipeline import _fragments
13
13
  from .storage import NpyStore, TensorStore
14
14
  from .distribution import control_status, preflight
15
+ from .utils import _work_dtype
15
16
 
16
17
 
17
18
  def count(value, name, positive=False):
@@ -65,20 +66,46 @@ def scalar_sum(operator, value, kind, *, return_tensor=False):
65
66
  dist.all_reduce(accumulator, group=group)
66
67
  return accumulator if return_tensor else accumulator.item()
67
68
  if return_tensor:
68
- return torch.as_tensor(value, dtype=torch.float64, device=result_device)
69
+ return torch.as_tensor(value, dtype=_work_dtype(result_device), device=result_device)
69
70
  return float(value)
70
71
 
71
72
 
72
73
  def squared_norm(operator, value, kind, elements=1048576, *, return_tensor=False):
73
74
  store = TensorStore(value) if isinstance(value, torch.Tensor) else value
74
- total = torch.zeros((), dtype=torch.float64, device=getattr(store, "device", "cpu"))
75
+ # Scale every block by its own peak before squaring and track the
76
+ # running maximum on the device: unscaled FP32 squares overflow for
77
+ # large measurements and underflow for tiny geometries.
78
+ device = getattr(store, "device", torch.device("cpu"))
79
+ if _work_dtype(device) == torch.float64:
80
+ total = torch.zeros((), dtype=torch.float64, device=device)
81
+ if owns_values(operator, kind):
82
+ for selected in _fragments(tuple(store.shape), elements):
83
+ # Clone even an FP64 view: norms must never alter caller measurements.
84
+ block = store.read(selected).to(dtype=torch.float64).clone()
85
+ # Preserve the block reductions and their sequential FP64 order.
86
+ total.add_(block.square_().sum().to(device=total.device))
87
+ return scalar_sum(operator, total, kind, return_tensor=return_tensor)
88
+ # FP32 devices (MPS): squares of large measurements overflow and tiny
89
+ # geometries underflow, so every block is rescaled by its peak and the
90
+ # running sum follows the largest peak seen. All block math stays on the
91
+ # device; the two closing scalar reads fit the per-iteration reduction
92
+ # budget, and their product leaves FP32 before it can overflow.
93
+ peak = torch.zeros((), device=device, dtype=torch.float32)
94
+ total = torch.zeros((), device=device, dtype=torch.float32)
75
95
  if owns_values(operator, kind):
76
96
  for selected in _fragments(tuple(store.shape), elements):
77
- # Clone even an FP64 view: norms must never alter caller measurements.
78
- block = store.read(selected).to(dtype=torch.float64).clone()
79
- # Preserve the block reductions and their sequential FP64 order.
80
- total.add_(block.square_().sum().to(device=total.device))
81
- return scalar_sum(operator, total, kind, return_tensor=return_tensor)
97
+ block = store.read(selected)
98
+ if not block.numel():
99
+ continue
100
+ block_peak = block.detach().abs().amax()
101
+ safe = torch.where(block_peak > 0, block_peak, torch.ones_like(block_peak))
102
+ combined = torch.maximum(peak, block_peak)
103
+ guard = torch.where(combined > 0, combined, torch.ones_like(combined))
104
+ total = (total * (peak / guard).square()
105
+ + (block.detach() / safe).square().sum() * (block_peak / guard).square())
106
+ peak = combined
107
+ return scalar_sum(operator, peak.item() ** 2 * total.item(), kind,
108
+ return_tensor=return_tensor)
82
109
 
83
110
 
84
111
  def copy_blocks(source, target, elements):
diffct/analytical.py CHANGED
@@ -27,7 +27,7 @@ import math
27
27
 
28
28
  import torch
29
29
 
30
- from .constants import _DTYPE
30
+ from .constants import _DTYPE, _native_spacing
31
31
  from .utils import (
32
32
  DeviceManager,
33
33
  TorchCUDABridge,
@@ -431,7 +431,7 @@ def _parallel_weighted_backproject_raw(sinogram, ray_dir, det_origin, det_u_vec,
431
431
  _resolve_kernel('_parallel_2d_fbp_backproject_kernel')[grid, tpb, numba_stream](
432
432
  d_sino, n_views, n_det, d_reco, Nx, Ny,
433
433
  _DTYPE(detector_spacing), d_ray_dir, d_det_origin, d_det_u_vec,
434
- cx, cy, _DTYPE(voxel_spacing),
434
+ cx, cy, _native_spacing(voxel_spacing),
435
435
  )
436
436
 
437
437
  return reco * (1.0 / (2.0 * math.pi))
@@ -522,8 +522,9 @@ def _fan_weighted_backproject_raw(sinogram, src_pos, det_center, det_u_vec,
522
522
  _resolve_kernel('_fan_2d_fbp_backproject_kernel')[grid, tpb, numba_stream](
523
523
  d_sino, n_views, n_det, d_reco, Nx, Ny,
524
524
  _DTYPE(detector_spacing), d_src_pos, d_det_center, d_det_u_vec,
525
- cx, cy, _DTYPE(voxel_spacing),
526
- *(_DTYPE(value / voxel_spacing) for value in isocenter.tolist()),
525
+ cx, cy, _native_spacing(voxel_spacing),
526
+ *(_DTYPE(value if isinstance(voxel_spacing, (tuple, list)) else value / voxel_spacing)
527
+ for value in isocenter.tolist()),
527
528
  )
528
529
 
529
530
  return reco
@@ -584,8 +585,9 @@ def _cone_weighted_backproject_raw(sinogram, src_pos, det_center, det_u_vec, det
584
585
  d_sino, n_views, n_u, n_v, d_reco, Nx, Ny, Nz,
585
586
  _DTYPE(du), _DTYPE(dv),
586
587
  d_src_pos, d_det_center, d_det_u_vec, d_det_v_vec,
587
- cx, cy, cz, _DTYPE(voxel_spacing),
588
- *(_DTYPE(value / voxel_spacing) for value in isocenter.tolist()),
588
+ cx, cy, cz, _native_spacing(voxel_spacing),
589
+ *(_DTYPE(value if isinstance(voxel_spacing, (tuple, list)) else value / voxel_spacing)
590
+ for value in isocenter.tolist()),
589
591
  )
590
592
 
591
593
  # Convert back to (D, H, W) layout for the user.
@@ -637,7 +639,8 @@ def _analytical_transpose_raw(image, beam, geometry, parameters, isocenter, sino
637
639
  scale = (sdd_n / (2.0 * math.pi * sid_n)).view(
638
640
  (-1,) + (1,) * (len(sinogram_shape) - 1)
639
641
  )
640
- iso = tuple(_DTYPE(value / parameters[-1]) for value in isocenter.tolist())
642
+ iso = tuple(_DTYPE(value if isinstance(parameters[-1], (tuple, list)) else value / parameters[-1])
643
+ for value in isocenter.tolist())
641
644
 
642
645
  if beam == "cone":
643
646
  D, H, W, du, dv, spacing = parameters
@@ -647,7 +650,7 @@ def _analytical_transpose_raw(image, beam, geometry, parameters, isocenter, sino
647
650
  _resolve_kernel('_cone_3d_fdk_transpose_kernel')[grid, tpb, stream](
648
651
  d_sino, *sinogram_shape, d_image, W, H, D, _DTYPE(du), _DTYPE(dv),
649
652
  *d_geometry, _DTYPE(W * 0.5), _DTYPE(H * 0.5), _DTYPE(D * 0.5),
650
- _DTYPE(spacing), *iso,
653
+ _native_spacing(spacing), *iso,
651
654
  )
652
655
  else:
653
656
  pitch, H, W, spacing = parameters
@@ -657,7 +660,7 @@ def _analytical_transpose_raw(image, beam, geometry, parameters, isocenter, sino
657
660
  else '_fan_2d_fbp_transpose_kernel')
658
661
  kernel[grid, tpb, stream](
659
662
  d_sino, *sinogram_shape, d_image, W, H, _DTYPE(pitch), *d_geometry,
660
- _DTYPE(W * 0.5), _DTYPE(H * 0.5), _DTYPE(spacing), *iso,
663
+ _DTYPE(W * 0.5), _DTYPE(H * 0.5), _native_spacing(spacing), *iso,
661
664
  )
662
665
  return sinogram * scale
663
666
 
@@ -16,6 +16,7 @@ from .analytical import (
16
16
  fan_weighted_backproject, parallel_weighted_backproject,
17
17
  )
18
18
  from .backends import _zero_dependency, select_backend, voxel_spacing_tuple
19
+ from .constants import _native_spacing
19
20
  from .chunking import _memory_budget, _pixel_tiles
20
21
  from .distribution import Partition, preflight
21
22
  from .operators import _detector_spacing, _normalize_devices, _positive_int, _shape, _validate_trajectory
@@ -143,6 +144,9 @@ class AnalyticalExecution:
143
144
  self.rank = 3 if beam == "cone" else 2
144
145
  self.global_volume_shape = _shape(volume_shape, self.rank, "volume_shape")
145
146
  self.spacing = voxel_spacing_tuple(voxel_spacing, self.rank)
147
+ self.voxel_spacing = self.spacing if isinstance(voxel_spacing, (tuple, list)) else self.spacing[0]
148
+ if self.backend == "numba" and isinstance(self.voxel_spacing, tuple):
149
+ _native_spacing(self.voxel_spacing)
146
150
  self.pitch = _detector_spacing(detector_spacing, beam)
147
151
  self.devices = _normalize_devices(devices)
148
152
  if not isinstance(source, torch.Tensor):
@@ -156,6 +160,8 @@ class AnalyticalExecution:
156
160
  raise TypeError("sinogram store must have dtype float32 or float64")
157
161
  self.source = source
158
162
  self.output_device = torch.device(getattr(source, "device", "cpu"))
163
+ if self.dtype == torch.float64 and self.output_device.type == "mps":
164
+ raise ValueError("precision must be 'fast' or 'float32' on MPS; MPS provides no float64")
159
165
  self.detector_shape = tuple(source.shape[1:])
160
166
  for component in trajectory:
161
167
  _fixed(component, "trajectory")
@@ -336,14 +342,14 @@ class AnalyticalExecution:
336
342
  device=data.device, dtype=self.dtype)
337
343
  if self.beam == "parallel":
338
344
  geometry = (geometry[0], geometry[1] - centre, geometry[2])
339
- return parallel_weighted_backproject(data, *geometry, self.pitch[0], *shape, self.spacing[0])
345
+ return parallel_weighted_backproject(data, *geometry, self.pitch[0], *shape, self.voxel_spacing)
340
346
  geometry = (geometry[0] - centre, geometry[1] - centre, *geometry[2:])
341
347
  isocenter = self.isocenter.to(data.device) - centre
342
348
  if self.beam == "fan":
343
349
  return fan_weighted_backproject(data, *geometry, self.pitch[0], *shape,
344
- self.spacing[0], isocenter=isocenter)
350
+ self.voxel_spacing, isocenter=isocenter)
345
351
  return cone_weighted_backproject(data, *geometry, *shape, *self.pitch,
346
- self.spacing[0], isocenter=isocenter)
352
+ self.voxel_spacing, isocenter=isocenter)
347
353
 
348
354
  def _tile(self, spatial):
349
355
  shape = tuple(part.stop - part.start for part in spatial)