v4.2 release. (#2587)

* Fix default cluster callback values to 1 to avoid profiler failure when these values are not set in command line.

* v4.2 release.

tools/util/include/cutlass/util/command_line.h

@@ -41,6 +41,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
+#include <unordered_map>
 
 #include <cuda_runtime.h>
 
@@ -89,6 +90,16 @@ struct CommandLine {
     }
   }
 
+  /**
+   * Constructor to represent a command line from a map of [argument] -> [value]
+   */
+  CommandLine(std::unordered_map<std::string, std::string>& arg_map) {
+    for (const auto& [key, value] : arg_map) {
+      keys.push_back(key);
+      values.push_back(value);
+    }
+  }
+
   /**
    * Checks whether a flag "--<flag>" is present in the commandline
    */

tools/util/include/cutlass/util/cublas_wrappers.hpp

@@ -51,7 +51,9 @@
 // User could potentially define ComplexFloat/ComplexDouble instead of std::
 #ifndef BLAM_COMPLEX_TYPES
 #define BLAM_COMPLEX_TYPES 1
-#include <cuda/std/complex>
+#include "cutlass/cutlass.h"
+#include CUDA_STD_HEADER(complex)
+
 namespace blam {
 template <typename T>
 using Complex       = cuda::std::complex<T>;