diff --git a/src/calng/AgipdCorrection.py b/src/calng/AgipdCorrection.py
index 144295641c6805ec2d6a320bdc6331d4ee263f16..c295f6ac6fa0c676786e47bb526e49fd9c9e56d7 100644
--- a/src/calng/AgipdCorrection.py
+++ b/src/calng/AgipdCorrection.py
@@ -12,7 +12,6 @@ from karabo.bound import (
)
from karabo.common.states import State
-from . import utils
from ._version import version as deviceVersion
from .agipd_gpu import AgipdGainMode, AgipdGpuRunner, BadPixelValues, CorrectionFlags
from .base_correction import BaseCorrection, add_correction_step_schema
@@ -164,17 +163,6 @@ class AgipdCorrection(BaseCorrection):
self.get("dataFormat.pixelsY"),
)
- @property
- def output_data_shape(self):
- return utils.shape_after_transpose(
- (
- self.get("dataFormat.memoryCells"),
- self.get("dataFormat.pixelsX"),
- self.get("dataFormat.pixelsY"),
- ),
- self._output_transpose,
- )
-
def __init__(self, config):
super().__init__(config)
# TODO: different gpu runner for fixed gain mode
@@ -188,11 +176,6 @@ class AgipdCorrection(BaseCorrection):
"g_gain_value": config.get("corrections.relGainXray.gGainValue"),
}
- self._output_transpose = {
- "pixels-fast": None,
- "memorycells-fast": (2, 1, 0),
- "no-reshape": None,
- }[config.get("dataFormat.outputAxisOrder")]
self._update_shapes()
# configurability: overriding md_additional_offset
@@ -268,7 +251,10 @@ class AgipdCorrection(BaseCorrection):
buffer_handle, buffer_array = self._shmem_buffer.next_slot()
self.gpu_runner.load_cell_table(cell_table)
self.gpu_runner.correct(self._correction_flag_enabled)
- self.gpu_runner.reshape(out=buffer_array)
+ self.gpu_runner.reshape(
+ output_order=self._schema_cache["dataFormat.outputAxisOrder"],
+ out=buffer_array,
+ )
# after reshape, data for dataOutput is now safe in its own buffer
if do_generate_preview:
if self._correction_flag_enabled != self._correction_flag_preview:
diff --git a/src/calng/DsscCorrection.py b/src/calng/DsscCorrection.py
index 349e5414f6d13e11b720fdcbedbec8a4fe1428a2..04916db116346e31999f0f4515d5b5fe1afeb34b 100644
--- a/src/calng/DsscCorrection.py
+++ b/src/calng/DsscCorrection.py
@@ -4,7 +4,6 @@ import numpy as np
from karabo.bound import KARABO_CLASSINFO, VECTOR_STRING_ELEMENT
from karabo.common.states import State
-from . import utils
from ._version import version as deviceVersion
from .base_correction import BaseCorrection, add_correction_step_schema
from .calcat_utils import DsscCalcatFriend, DsscConstants
@@ -48,20 +47,8 @@ class DsscCorrection(BaseCorrection):
self.get("dataFormat.pixelsX"),
)
- @property
- def output_data_shape(self):
- return utils.shape_after_transpose(
- self.input_data_shape, self._output_transpose
- )
-
def __init__(self, config):
super().__init__(config)
- self._output_transpose = {
- "pixels-fast": (0, 2, 1),
- "memorycells-fast": (2, 1, 0),
- "no-reshape": None,
- }[config.get("dataFormat.outputAxisOrder")]
- self._update_shapes()
self.updateState(State.ON)
def process_input(self, data, metadata):
@@ -124,7 +111,10 @@ class DsscCorrection(BaseCorrection):
buffer_handle, buffer_array = self._shmem_buffer.next_slot()
self.gpu_runner.load_cell_table(cell_table)
self.gpu_runner.correct(self._correction_flag_enabled)
- self.gpu_runner.reshape(out=buffer_array)
+ self.gpu_runner.reshape(
+ output_order=self._schema_cache["dataFormat.outputAxisOrder"],
+ out=buffer_array,
+ )
if do_generate_preview:
if self._correction_flag_enabled != self._correction_flag_preview:
self.gpu_runner.correct(self._correction_flag_preview)
diff --git a/src/calng/agipd_gpu.py b/src/calng/agipd_gpu.py
index fd935ebade2ff0ea9794bc2018271f320f4e0a37..3b3802ab633f9cd89cb76934477bcb38b3e83101 100644
--- a/src/calng/agipd_gpu.py
+++ b/src/calng/agipd_gpu.py
@@ -26,6 +26,7 @@ class AgipdGainMode(enum.IntEnum):
class AgipdGpuRunner(base_gpu.BaseGpuRunner):
_kernel_source_filename = "agipd_gpu_kernels.cpp"
+ _corrected_axis_order = "cxy"
def __init__(
self,
@@ -33,7 +34,6 @@ class AgipdGpuRunner(base_gpu.BaseGpuRunner):
pixels_y,
memory_cells,
constant_memory_cells,
- output_transpose=(1, 2, 0), # default: memorycells-fast
input_data_dtype=cupy.uint16,
output_data_dtype=cupy.float32,
bad_pixel_mask_value=cupy.nan,
@@ -52,7 +52,6 @@ class AgipdGpuRunner(base_gpu.BaseGpuRunner):
pixels_y,
memory_cells,
constant_memory_cells,
- output_transpose,
input_data_dtype,
output_data_dtype,
)
@@ -247,9 +246,12 @@ class AgipdGpuRunner(base_gpu.BaseGpuRunner):
),
)
- def get_gain_map(self, out=None):
+ def get_gain_map(self, output_order, out=None):
return cupy.ascontiguousarray(
- cupy.transpose(self.gain_map_gpu, self.output_transpose)
+ cupy.transpose(
+ self.gain_map_gpu,
+ utils.transpose_order(self._corrected_axis_order, output_order),
+ )
).get(out=out)
def _init_kernels(self):
diff --git a/src/calng/base_correction.py b/src/calng/base_correction.py
index 82022c15acd2a2c40f72b65417490603a1896b5b..fd3fbdc0cbe514de13ef8aa7cd0a4d21cc669e5b 100644
--- a/src/calng/base_correction.py
+++ b/src/calng/base_correction.py
@@ -41,11 +41,12 @@ PROCESSING_STATE_TIMEOUT = 10
@KARABO_CLASSINFO("BaseCorrection", deviceVersion)
class BaseCorrection(PythonDevice):
- _correction_flag_class = None # subclass must override this with some enum class
+ _correction_flag_class = None # subclass must set to some enum class
_gpu_runner_class = None # subclass must set this
_gpu_runner_init_args = {} # subclass can set this (TODO: remove, design better)
_managed_keys = [
"outputShmemBufferSize",
+ "dataFormat.outputAxisOrder",
"dataFormat.outputImageDtype",
"preview.enable",
"preview.pulse",
@@ -59,12 +60,13 @@ class BaseCorrection(PythonDevice):
"constantParameters.memoryCells",
"dataFormat.pixelsX",
"dataFormat.pixelsY",
+ "dataFormat.outputAxisOrder",
"preview.enable",
"preview.pulse",
"preview.trainIdModulo",
"processingStateTimeout",
"state",
- }
+ } # subclass should be aware of cache, but does not need to extend
def _load_constant_to_gpu(constant_name, constant_data):
raise NotImplementedError()
@@ -75,7 +77,15 @@ class BaseCorrection(PythonDevice):
@property
def output_data_shape(self):
- raise NotImplementedError()
+ axis_lengths = {
+ "x": self._schema_cache["dataFormat.pixelsX"],
+ "y": self._schema_cache["dataFormat.pixelsY"],
+ "c": self._schema_cache["dataFormat.memoryCells"],
+ }
+ return tuple(
+ axis_lengths[axis]
+ for axis in self._schema_cache["dataFormat.outputAxisOrder"]
+ )
@staticmethod
def expectedParameters(expected):
@@ -175,13 +185,13 @@ class BaseCorrection(PythonDevice):
.key("dataFormat.outputAxisOrder")
.displayedName("Output axis order")
.description(
- "Axes of main data output can be reordered after correction. Choose "
- "between 'pixels-fast' (memory_cell, x, y), 'memorycells-fast' "
- "(x, y, memory_cell), and 'no-reshape'"
+ "Axes of main data output can be reordered after correction. Axis order "
+ "is specified as string consisting of 'x', 'y', and 'c', with the "
+ "latter indicating the memory cell axis. The default value of 'cxy' "
+ "puts pixels along the fast axes."
)
- .options("pixels-fast,memorycells-fast,no-reshape")
.assignmentOptional()
- .defaultValue("pixels-fast")
+ .defaultValue("cxy")
.commit(),
VECTOR_UINT32_ELEMENT(expected)
.key("dataFormat.inputDataShape")
@@ -337,6 +347,11 @@ class BaseCorrection(PythonDevice):
}
super().__init__(config)
+ if not sorted(config.get("dataFormat.outputAxisOrder")) == ["c", "x", "y"]:
+ # TODO: figure out how to get this information to operator
+ self.log_status_error("Invalid output axis order string")
+ return
+
self.KARABO_ON_DATA("dataInput", self.process_input)
self.KARABO_ON_EOS("dataInput", self.handle_eos)
@@ -574,7 +589,6 @@ class BaseCorrection(PythonDevice):
self.get("dataFormat.pixelsY"),
self.get("dataFormat.memoryCells"),
int(self.get("constantParameters.memoryCells")),
- output_transpose=self._output_transpose,
input_data_dtype=self.input_data_dtype,
output_data_dtype=self.output_data_dtype,
**self._gpu_runner_init_args,
diff --git a/src/calng/base_gpu.py b/src/calng/base_gpu.py
index 868d42e0866e2be0866f6e2db45a9f92c93dabc2..e51bc99b0f67a24a2fc91cfbf39710af78be7e7f 100644
--- a/src/calng/base_gpu.py
+++ b/src/calng/base_gpu.py
@@ -29,13 +29,16 @@ class BaseGpuRunner:
correction steps are appropriate given the constants loaded so far.
"""
+ # These must be set by subclass
+ _kernel_source_filename = None
+ _corrected_axis_order = None
+
def __init__(
self,
pixels_x,
pixels_y,
memory_cells,
constant_memory_cells,
- output_transpose=(2, 1, 0), # default: memorycells-fast
input_data_dtype=np.uint16,
output_data_dtype=np.float32,
):
@@ -47,15 +50,11 @@ class BaseGpuRunner:
self.pixels_x = pixels_x
self.pixels_y = pixels_y
self.memory_cells = memory_cells
- self.output_transpose = output_transpose
if constant_memory_cells == 0:
# if not set, guess same as input; may save one recompilation
self.constant_memory_cells = memory_cells
else:
self.constant_memory_cells = constant_memory_cells
- self.output_shape = utils.shape_after_transpose(
- self.processed_shape, self.output_transpose
- )
# preview will only be single memory cell
self.preview_shape = (self.pixels_x, self.pixels_y)
self.input_data_dtype = input_data_dtype
@@ -69,7 +68,7 @@ class BaseGpuRunner:
self.processed_data_gpu = cupy.empty(
self.processed_shape, dtype=output_data_dtype
)
- self.reshaped_data_gpu = cupy.empty(self.output_shape, dtype=output_data_dtype)
+ self.reshaped_data_gpu = None # currently not reusing buffer
self.preview_raw = np.empty(self.preview_shape, dtype=np.float32)
self.preview_corrected = np.empty(self.preview_shape, dtype=np.float32)
@@ -105,22 +104,22 @@ class BaseGpuRunner:
"""
raise NotImplementedError()
- def reshape(self, out=None):
- """Move axes to desired output order
+ def reshape(self, output_order, out=None):
+ """Move axes to desired output order and copy to host memory
The out parameter is passed directly to the get function of GPU array: if
None, then a new ndarray (in host memory) is returned. If not None, then data
will be loaded into the provided array, which must match shape / dtype.
"""
# TODO: avoid copy
- if self.output_transpose is None:
- self.reshaped_data_gpu = cupy.ascontiguousarray(
- cupy.squeeze(self.processed_data_gpu)
- )
+ # TODO: check if necessary to make contiguous
+ if output_order == self._corrected_axis_order:
+ self.reshaped_data_gpu = cupy.ascontiguousarray(self.processed_data_gpu)
else:
self.reshaped_data_gpu = cupy.ascontiguousarray(
cupy.transpose(
- cupy.squeeze(self.processed_data_gpu), self.output_transpose
+ self.processed_data_gpu,
+ utils.transpose_order(self._corrected_axis_order, output_order),
)
)
return self.reshaped_data_gpu.get(out=out)
diff --git a/src/calng/dssc_gpu.py b/src/calng/dssc_gpu.py
index 8c1995e39733f462c36f7a663da3b1ec2dbe0d16..2b4bac53f025f8bcbe3dde08355a35294319713d 100644
--- a/src/calng/dssc_gpu.py
+++ b/src/calng/dssc_gpu.py
@@ -13,6 +13,7 @@ class CorrectionFlags(enum.IntFlag):
class DsscGpuRunner(base_gpu.BaseGpuRunner):
_kernel_source_filename = "dssc_gpu_kernels.cpp"
+ _corrected_axis_order = "cyx"
def __init__(
self,
@@ -20,7 +21,6 @@ class DsscGpuRunner(base_gpu.BaseGpuRunner):
pixels_y,
memory_cells,
constant_memory_cells,
- output_transpose=(2, 1, 0), # default: memorycells-fast
input_data_dtype=np.uint16,
output_data_dtype=np.float32,
):
@@ -31,7 +31,6 @@ class DsscGpuRunner(base_gpu.BaseGpuRunner):
pixels_y,
memory_cells,
constant_memory_cells,
- output_transpose,
input_data_dtype,
output_data_dtype,
)
diff --git a/src/calng/utils.py b/src/calng/utils.py
index 7c4c78e5f47e1b338f9da54afe172632f72064dc..0acf8cfd401a8bb358dd6d33090f6861a0a671be 100644
--- a/src/calng/utils.py
+++ b/src/calng/utils.py
@@ -39,6 +39,19 @@ def threadsafe_cache(fun):
return aux
+@functools.lru_cache
+def transpose_order(axes_in, axes_out):
+ """Computes the order of axes_out relative to axes_in for transposition purposes
+
+ Both axes_in and axes_out are assumed to be strings in which each letter represents
+ an axis (duck typing accepts: any iterable of hashable elements). They should
+ probably be of the same length and have no repetitions, but this is not enforced.
+ Off-label use voids warranty.
+ """
+ axis_order = {axis: index for index, axis in enumerate(axes_in)}
+ return tuple(axis_order[axis] for axis in axes_out)
+
+
_np_typechar_to_c_typestring = {
"?": "bool",
"B": "unsigned char",