Merge pull request #321 from effigies/enh/mem_estimate

ENH: Dynamically choose number of resampling threads to adapt to memory constraints

sdcflows/workflows/apply/correction.py

@@ -26,7 +26,7 @@
 from niworkflows.engine.workflows import LiterateWorkflow as Workflow
 
 
-def init_unwarp_wf(omp_nthreads=1, debug=False, name="unwarp_wf"):
+def init_unwarp_wf(*, free_mem=None, omp_nthreads=1, debug=False, name="unwarp_wf"):
     r"""
     Set up a workflow that unwarps the input :abbr:`EPI (echo-planar imaging)` dataset.
 
@@ -103,9 +103,24 @@ def init_unwarp_wf(omp_nthreads=1, debug=False, name="unwarp_wf"):
 
     rotime = pe.Node(GetReadoutTime(), name="rotime")
     rotime.interface._always_run = debug
-    resample = pe.Node(ApplyCoeffsField(
-        num_threads=omp_nthreads if not debug else 1
-    ), name="resample")
+
+    # resample is memory-hungry; choose a smaller number of threads
+    # if we know how much memory we have to work with
+    mem_per_thread = 5  # True for a 128x128x84 image; should generalize
+    if debug:
+        num_threads = 1
+    elif free_mem is not None:
+        mem_gb = min(0.9 * free_mem, mem_per_thread * omp_nthreads)
+        num_threads = max(int(mem_gb // mem_per_thread), 1)
+    else:
+        num_threads = omp_nthreads
+
+    resample = pe.Node(
+        ApplyCoeffsField(num_threads=num_threads),
+        mem_gb=mem_per_thread * num_threads,
+        name="resample",
+    )
+
     merge = pe.Node(MergeSeries(), name="merge")
     average = pe.Node(RobustAverage(mc_method=None), name="average")