CUTLASS 3.5.1 (#1623)

* CUTLASS 3.5.1

* updates, optimizations, fixes

tools/profiler/include/cutlass/profiler/gemm_operation_profiler.h

@@ -84,6 +84,8 @@ public:
     int batch_count{1};
 
     cutlass::library::RasterOrder raster_order{cutlass::library::RasterOrder::kHeuristic};
+    int swizzle_size{1};
+
     // gemm with parallel interleaved reduction
     // gemm epilogue (alpha, beta) = (1.0, 0.0)
     // reduction epilogue (alpha, beta) = (GemmProblem::alpha, GemmProblem::beta)

tools/profiler/include/cutlass/profiler/options.h

@@ -72,7 +72,7 @@ public:
     // Methods
     //
 
-    Library(CommandLine const &cmdline);
+    explicit Library(CommandLine const &cmdline);
 
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -94,7 +94,7 @@ public:
     // Methods
     //
 
-    Device(CommandLine const &cmdline);
+    explicit Device(CommandLine const &cmdline);
 
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -128,7 +128,7 @@ public:
     // Methods
     //
 
-    Initialization(CommandLine const &cmdline);
+    explicit Initialization(CommandLine const &cmdline);
     
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -170,7 +170,7 @@ public:
     // Methods
     //
 
-    Verification(CommandLine const &cmdline);
+    explicit Verification(CommandLine const &cmdline);
   
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -186,22 +186,26 @@ public:
   struct Profiling {
 
     /// Number of workspaces to rotate through to avoid cache-resident working sets
-    int workspace_count;
+    int workspace_count{0};
 
     /// Number of iterations to warmup each kernel prior to profiling
-    int warmup_iterations;
+    int warmup_iterations{10};
 
     /// Number of iterations to profile each kernel - if 0, kernels are launched up to the profiling duration
-    int iterations;
+    int iterations{100};
 
     /// Number of ms to sleep between profiling periods (ms)
-    int sleep_duration;
+    int sleep_duration{50};
 
     /// If true, profiling is actually conducted.
-    bool enabled;
+    bool enabled{true};
 
     /// If true, profiling returns an error code if no kernels are found to match the filters.
-    bool error_on_no_match = false;
+    bool error_on_no_match{false};
+
+    /// If true, profiling returns an error code if no kernel are profiled
+    // Sometimes the kernel matches but failed to profile (e.g. can_implement() error)
+    bool error_if_nothing_is_profiled{false};
 
     /// List of providers of each functionality to be profiled
     ProviderVector providers;
@@ -210,7 +214,7 @@ public:
     // Methods
     //
 
-    Profiling(CommandLine const &cmdline);
+    explicit Profiling(CommandLine const &cmdline);
 
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -255,7 +259,7 @@ public:
     // Methods
     //
 
-    Report(CommandLine const &cmdline);
+    explicit Report(CommandLine const &cmdline);
     
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -277,7 +281,7 @@ public:
     // Methods
     //
 
-    About(CommandLine const &cmdline);
+    explicit About(CommandLine const &cmdline);
     
     void print_usage(std::ostream &out) const;
     void print_options(std::ostream &out, int indent = 0) const;
@@ -320,7 +324,7 @@ public:
 
 public:
 
-  Options(CommandLine const &cmdline);
+  explicit Options(CommandLine const &cmdline);
 
   void print_usage(std::ostream &out) const;
   void print_options(std::ostream &out) const;

tools/profiler/include/cutlass/profiler/problem_space.h

@@ -256,7 +256,7 @@ struct ScalarArgument : public KernelArgument {
     virtual std::ostream &print(std::ostream &out) const;
   };
 
-  using ValueCollection = std::vector<std::string>;
+  using ValueCollection = std::vector<decltype(ScalarValue::value)>;
 
   /// Abstract base class to iterate over values within arguments
   struct ScalarValueIterator : public KernelArgument::ValueIterator {
@@ -271,7 +271,7 @@ struct ScalarArgument : public KernelArgument {
     // Methods
     //
 
-    ScalarValueIterator(ScalarArgument const *argument = nullptr);
+    explicit ScalarValueIterator(ScalarArgument const *argument = nullptr);
 
     virtual void operator++();
     virtual bool operator==(ValueIterator const &it) const;
@@ -292,7 +292,7 @@ struct ScalarArgument : public KernelArgument {
   //
 
   /// Default ctor
-  ScalarArgument(
+  explicit ScalarArgument(
     ArgumentDescription const *description
   ): 
     KernelArgument(description) { }
@@ -632,7 +632,7 @@ struct TensorArgument : public KernelArgument {
     // Methods
     //
 
-    TensorValueIterator(TensorArgument const *argument_);
+    explicit TensorValueIterator(TensorArgument const *argument_);
 
     virtual void operator++();
     virtual bool operator==(ValueIterator const &it) const;
@@ -649,7 +649,7 @@ struct TensorArgument : public KernelArgument {
   //
 
   /// Default ctor
-  TensorArgument(
+  explicit TensorArgument(
     ArgumentDescription const *description
   ): 
     KernelArgument(description) { }
@@ -690,7 +690,7 @@ struct EnumeratedTypeArgument : public KernelArgument {
     virtual std::ostream &print(std::ostream &out) const;
   };
 
-  using ValueCollection = std::vector<std::string>;
+  using ValueCollection = std::vector<decltype(EnumeratedTypeValue::element)>;
 
   /// Abstract base class to iterate over values within arguments
   struct EnumeratedTypeValueIterator : public KernelArgument::ValueIterator {
@@ -705,7 +705,7 @@ struct EnumeratedTypeArgument : public KernelArgument {
     // Methods
     //
 
-    EnumeratedTypeValueIterator(EnumeratedTypeArgument const *argument_ = nullptr);
+    explicit EnumeratedTypeValueIterator(EnumeratedTypeArgument const *argument_ = nullptr);
 
     virtual void operator++();
     virtual bool operator==(ValueIterator const &it) const;
@@ -725,7 +725,7 @@ struct EnumeratedTypeArgument : public KernelArgument {
   //
 
   /// Default ctor
-  EnumeratedTypeArgument(ArgumentDescription const *description):
+  explicit EnumeratedTypeArgument(ArgumentDescription const *description):
     KernelArgument(description) {}
 
   virtual bool not_null() const {
@@ -819,7 +819,7 @@ public:
   //
 
   /// Default ctor
-  ProblemSpace() {}
+  ProblemSpace() = default;
 
   /// Constructs a problem space from a vector of arguments. This vector must outlive
   /// the ProblemSpace object, which stores pointers to objects within the