perf: avoid extra dtype call

cpcloud · cpcloud · commit 1c9e7aa44dc6 · 2025-10-23T11:59:59.000-04:00
diff --git a/numba_cuda/numba/cuda/api.py b/numba_cuda/numba/cuda/api.py
@@ -39,10 +39,9 @@ def from_cuda_array_interface(desc, owner=None, sync=True):
 
     shape = desc["shape"]
     strides = desc.get("strides")
-    dtype = np.dtype(desc["typestr"])
 
     shape, strides, dtype = prepare_shape_strides_dtype(
-        shape, strides, dtype, order="C"
+        shape, strides, desc["typestr"], order="C"
     )
     size = driver.memory_size_from_info(shape, strides, dtype.itemsize)
 

Original file line number	Diff line number	Diff line change
`@@ -39,10 +39,9 @@ def from_cuda_array_interface(desc, owner=None, sync=True):`
`39`	`39`
`40`	`40`	`shape = desc["shape"]`
`41`	`41`	`strides = desc.get("strides")`
`42`		`- dtype = np.dtype(desc["typestr"])`
`43`	`42`
`44`	`43`	`shape, strides, dtype = prepare_shape_strides_dtype(`
`45`		`- shape, strides, dtype, order="C"`
	`44`	`+ shape, strides, desc["typestr"], order="C"`
`46`	`45`	`)`
`47`	`46`	`size = driver.memory_size_from_info(shape, strides, dtype.itemsize)`
`48`	`47`