Merge pull request rapidsai#34 from sklam/fix/iss32

sklam · web-flow · commit f1412560f7ba · 2017-08-23T10:59:06.000-05:00
Use better gridsize calc to avoid cuda launch error
diff --git a/src/binaryops.cu b/src/binaryops.cu
@@ -1,3 +1,5 @@
+#include <algorithm>
+
 #include <gdf/gdf.h>
 #include <gdf/utils.h>
 #include <gdf/errorutils.h>
@@ -41,7 +43,8 @@ struct BinaryOp {
                                                gpu_binary_op<T, Tout, F>)
         );
         // find needed gridsize
-        int gridsize = (lhs->size + blocksize - 1) / blocksize;
+        int neededgridsize = (lhs->size + blocksize - 1) / blocksize;
+        int gridsize = std::min(mingridsize, neededgridsize);
 
         F functor;
         gpu_binary_op<<<gridsize, blocksize>>>(
diff --git a/src/unaryops.cu b/src/unaryops.cu
@@ -1,4 +1,5 @@
 #include <cmath>
+#include <algorithm>
 
 #include <gdf/gdf.h>
 #include <gdf/utils.h>
@@ -43,7 +44,8 @@ struct UnaryOp {
                                                gpu_unary_op<T, Tout, F>)
         );
         // find needed gridsize
-        int gridsize = (input->size + blocksize - 1) / blocksize;
+        int neededgridsize = (input->size + blocksize - 1) / blocksize;
+        int gridsize = std::min(neededgridsize, mingridsize);
 
         F functor;
         gpu_unary_op<<<gridsize, blocksize>>>(
