Fix edge case in offset padding for empty cells

zatkins-dev · zatkins-dev · commit 76c9e6994d93 · 2025-04-17T21:20:52.000-07:00
diff --git a/backends/cuda-ref/ceed-cuda-ref-restriction.c b/backends/cuda-ref/ceed-cuda-ref-restriction.c
@@ -550,16 +550,18 @@ int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode,
     CeedCallBackend(CeedMalloc(num_elem, &points_per_elem));
     for (CeedInt i = 0; i < num_elem; i++) {
       CeedInt num_points = offsets[i + 1] - offsets[i];
+      CeedInt last_point = offsets[offsets[i]] * num_comp;
 
       points_per_elem[i] = num_points;
       at_points_size += num_points;
       // -- Copy all points in element
       for (CeedInt j = 0; j < num_points; j++) {
         offsets_padded[i * max_points + j] = offsets[offsets[i] + j] * num_comp;
+        last_point                         = offsets_padded[i * max_points + j];
       }
       // -- Replicate out last point in element
       for (CeedInt j = num_points; j < max_points; j++) {
-        offsets_padded[i * max_points + j] = offsets[offsets[i] + num_points - 1] * num_comp;
+        offsets_padded[i * max_points + j] = last_point;
       }
     }
     CeedCallBackend(CeedSetHostCeedIntArray(offsets, copy_mode, at_points_size, &impl->h_offsets_at_points_owned, &impl->h_offsets_at_points_borrowed,
diff --git a/backends/hip-ref/ceed-hip-ref-restriction.c b/backends/hip-ref/ceed-hip-ref-restriction.c
@@ -551,16 +551,18 @@ int CeedElemRestrictionCreate_Hip(CeedMemType mem_type, CeedCopyMode copy_mode,
     CeedCallBackend(CeedMalloc(num_elem, &points_per_elem));
     for (CeedInt i = 0; i < num_elem; i++) {
       CeedInt num_points = offsets[i + 1] - offsets[i];
+      CeedInt last_point = offsets[offsets[i]] * num_comp;
 
       points_per_elem[i] = num_points;
       at_points_size += num_points;
       // -- Copy all points in element
       for (CeedInt j = 0; j < num_points; j++) {
         offsets_padded[i * max_points + j] = offsets[offsets[i] + j] * num_comp;
+        last_point                         = offsets_padded[i * max_points + j];
       }
       // -- Replicate out last point in element
       for (CeedInt j = num_points; j < max_points; j++) {
-        offsets_padded[i * max_points + j] = offsets[offsets[i] + num_points - 1] * num_comp;
+        offsets_padded[i * max_points + j] = last_point;
       }
     }
     CeedCallBackend(CeedSetHostCeedIntArray(offsets, copy_mode, at_points_size, &impl->h_offsets_at_points_owned, &impl->h_offsets_at_points_borrowed,

Original file line number	Diff line number	Diff line change
`@@ -550,16 +550,18 @@ int CeedElemRestrictionCreate_Cuda(CeedMemType mem_type, CeedCopyMode copy_mode,`
`550`	`550`	`CeedCallBackend(CeedMalloc(num_elem, &points_per_elem));`
`551`	`551`	`for (CeedInt i = 0; i < num_elem; i++) {`
`552`	`552`	`CeedInt num_points = offsets[i + 1] - offsets[i];`
	`553`	`+ CeedInt last_point = offsets[offsets[i]] * num_comp;`
`553`	`554`
`554`	`555`	`points_per_elem[i] = num_points;`
`555`	`556`	`at_points_size += num_points;`
`556`	`557`	`// -- Copy all points in element`
`557`	`558`	`for (CeedInt j = 0; j < num_points; j++) {`
`558`	`559`	`offsets_padded[i * max_points + j] = offsets[offsets[i] + j] * num_comp;`
	`560`	`+ last_point = offsets_padded[i * max_points + j];`
`559`	`561`	`}`
`560`	`562`	`// -- Replicate out last point in element`
`561`	`563`	`for (CeedInt j = num_points; j < max_points; j++) {`
`562`		`- offsets_padded[i * max_points + j] = offsets[offsets[i] + num_points - 1] * num_comp;`
	`564`	`+ offsets_padded[i * max_points + j] = last_point;`
`563`	`565`	`}`
`564`	`566`	`}`
`565`	`567`	`CeedCallBackend(CeedSetHostCeedIntArray(offsets, copy_mode, at_points_size, &impl->h_offsets_at_points_owned, &impl->h_offsets_at_points_borrowed,`
Original file line number	Diff line number	Diff line change
`@@ -551,16 +551,18 @@ int CeedElemRestrictionCreate_Hip(CeedMemType mem_type, CeedCopyMode copy_mode,`
`551`	`551`	`CeedCallBackend(CeedMalloc(num_elem, &points_per_elem));`
`552`	`552`	`for (CeedInt i = 0; i < num_elem; i++) {`
`553`	`553`	`CeedInt num_points = offsets[i + 1] - offsets[i];`
	`554`	`+ CeedInt last_point = offsets[offsets[i]] * num_comp;`
`554`	`555`
`555`	`556`	`points_per_elem[i] = num_points;`
`556`	`557`	`at_points_size += num_points;`
`557`	`558`	`// -- Copy all points in element`
`558`	`559`	`for (CeedInt j = 0; j < num_points; j++) {`
`559`	`560`	`offsets_padded[i * max_points + j] = offsets[offsets[i] + j] * num_comp;`
	`561`	`+ last_point = offsets_padded[i * max_points + j];`
`560`	`562`	`}`
`561`	`563`	`// -- Replicate out last point in element`
`562`	`564`	`for (CeedInt j = num_points; j < max_points; j++) {`
`563`		`- offsets_padded[i * max_points + j] = offsets[offsets[i] + num_points - 1] * num_comp;`
	`565`	`+ offsets_padded[i * max_points + j] = last_point;`
`564`	`566`	`}`
`565`	`567`	`}`
`566`	`568`	`CeedCallBackend(CeedSetHostCeedIntArray(offsets, copy_mode, at_points_size, &impl->h_offsets_at_points_owned, &impl->h_offsets_at_points_borrowed,`