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.

include/cute/numeric/arithmetic_tuple.hpp

@@ -469,7 +469,9 @@ CUTE_HOST_DEVICE void print(ArithmeticTupleIterator<ArithTuple> const& iter)
 template <class T, int... Ns>
 CUTE_HOST_DEVICE void print(ScaledBasis<T,Ns...> const& e)
 {
-  print(e.value()); (void(printf("@%d", Ns)), ...);
+  print(e.value());
+  // Param pack trick to print in reverse
+  [[maybe_unused]] int dummy; (dummy = ... = (void(printf("@%d", Ns)), 0));
 }
 
 #if !defined(__CUDACC_RTC__)
@@ -482,7 +484,9 @@ CUTE_HOST std::ostream& operator<<(std::ostream& os, ArithmeticTupleIterator<Ari
 template <class T, int... Ns>
 CUTE_HOST std::ostream& operator<<(std::ostream& os, ScaledBasis<T,Ns...> const& e)
 {
-  os << e.value(); (void(os << "@" << Ns), ...);
+  os << e.value();
+  // Param pack trick to print in reverse
+  [[maybe_unused]] int dummy; (dummy = ... = (void(os << "@" << Ns),0));
   return os;
 }
 #endif

include/cute/numeric/int.hpp

@@ -29,9 +29,9 @@
  *
  **************************************************************************************************/
 #pragma once
-
+#include "cutlass/cutlass.h"
 #if defined(__CUDACC_RTC__)
-#include <cuda/std/cstdint>
+#include CUDA_STD_HEADER(cstdint)
 #else
 #include <cstdint>
 #endif
@@ -85,6 +85,8 @@ using CUTE_STL_NAMESPACE::uint16_t;
 using CUTE_STL_NAMESPACE::uint32_t;
 using CUTE_STL_NAMESPACE::uint64_t;
 using cutlass::uint128_t;
+using cutlass::uint256_t;
+
 template <int N> struct uint_bit;
 template <> struct uint_bit<  1> { using type = uint1_t; };
 template <> struct uint_bit<  2> { using type = uint2_t; };
@@ -95,6 +97,8 @@ template <> struct uint_bit< 16> { using type = uint16_t; };
 template <> struct uint_bit< 32> { using type = uint32_t; };
 template <> struct uint_bit< 64> { using type = uint64_t; };
 template <> struct uint_bit<128> { using type = cutlass::uint128_t; };
+template <> struct uint_bit<256> { using type = cutlass::uint256_t; };
+
 template <int N>
 using uint_bit_t = typename uint_bit<N>::type;
 

include/cute/numeric/integral_ratio.hpp

@@ -225,6 +225,27 @@ operator==(C<c>, R<a,b>) {
   return {};
 }
 
+template <auto a, auto b, auto x, auto y>
+CUTE_HOST_DEVICE constexpr
+bool_constant<R<a,b>::num * R<x,y>::den < R<x,y>::num * R<a,b>::den>
+operator<(R<a,b>, R<x,y>) {
+  return {};
+}
+
+template <auto a, auto b, auto c>
+CUTE_HOST_DEVICE constexpr
+bool_constant<R<a,b>::num < c * R<a,b>::den>
+operator<(R<a,b>, C<c>) {
+  return {};
+}
+
+template <auto c, auto x, auto y>
+CUTE_HOST_DEVICE constexpr
+bool_constant<c * R<x,y>::den < R<x,y>::num>
+operator<(C<c>, R<x,y>) {
+  return {};
+}
+
 ///////////////////////
 // Special functions //
 ///////////////////////