CUTLASS 3.3.0 (#1167)

* Release 3.3.0

Adds support for mixed precision GEMMs On Hopper and Ampere
Adds support for < 16B aligned GEMMs on Hopper
Enhancements to EVT
Enhancements to Python interface
Enhancements to Sub-byte type handling in CuTe
Several other bug-fixes and performance improvements.

* minor doc update

tools/library/src/handle.cu

@@ -1183,7 +1183,7 @@ Operation const* find_gemm_operation_for_parallel_reduction(Operation const *ope
     return nullptr;
   }
 
-  // return matching gemm operation (same tile shape, stages, warp count, and instruction)
+  // return matching gemm opertion (same tile shape, stages, warp count, and instruction)
   for (auto op : it->second) {
     if (op->description().tile_description == operation->description().tile_description) {
       return op;

tools/library/src/library_internal.h

@@ -287,6 +287,10 @@ template <> struct OpcodeClassMap<arch::OpClassTensorOp> {
   static OpcodeClassID const kId = OpcodeClassID::kTensorOp;
 };
 
+template <> struct OpcodeClassMap<arch::OpClassSparseTensorOp> {
+  static OpcodeClassID const kId = OpcodeClassID::kSparseTensorOp;
+};
+
 template <> struct OpcodeClassMap<arch::OpClassWmmaTensorOp> {
   static OpcodeClassID const kId = OpcodeClassID::kWmmaTensorOp;
 };

tools/library/src/operation_table.cu

@@ -47,9 +47,7 @@ void OperationTable::append(Manifest const &manifest) {
 
   // Insert operations into appropriate data structure
   for (auto const & operation : manifest) {
-
     OperationDescription const &desc = operation->description();
-
     // insert all gemm operation into operation table
     if (desc.kind == OperationKind::kGemm) {
       GemmDescription const &gemm_desc = static_cast<GemmDescription const &>(desc);

tools/library/src/reduction/init_reduction_operations.cu

@@ -42,6 +42,7 @@ namespace library {
 ///////////////////////////////////////////////////////////////////////////////////////////////
 //                             CUTLASS Reduction Instances                                   //
 ///////////////////////////////////////////////////////////////////////////////////////////////
+
 void initialize_reduce_add_linear_combination_f16_f16_f16(Manifest &manifest);
 void initialize_reduce_add_linear_combination_f32_f32_f16(Manifest &manifest);
 void initialize_reduce_add_linear_combination_f32_f32_bf16(Manifest &manifest);

tools/library/src/reduction/reduction_device.cu

@@ -146,7 +146,6 @@ void initialize_reduce_add_linear_combination_f32_f32_bf16(Manifest &manifest) {
   ));
 }
 
-
 void initialize_reduce_add_linear_combination_f32_f32_f32(Manifest &manifest) {
 
   using ElementWorkspace = float;

tools/library/src/reference/gemm_fp_mixed_input.cu

@@ -78,7 +78,7 @@ void initialize_gemm_reference_operations_fp_mixed_input(Manifest &manifest) {
     float,
     float
   >(manifest);
-  
+
   make_gemm_real_canonical_layouts<
     half_t,
     uint8_t,
@@ -151,4 +151,3 @@ void initialize_gemm_reference_operations_fp_mixed_input(Manifest &manifest) {
 } // namespace cutlass
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
-

tools/library/src/reference/initialize_reference_operations.cu

@@ -84,6 +84,7 @@ void initialize_reference_operations(Manifest &manifest) {
   initialize_gemm_reference_operations_fp32out(manifest);
   initialize_gemm_reference_operations_fp_other(manifest);
   initialize_gemm_reference_operations_fp_mixed_input(manifest);
+
 }
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////

tools/library/src/util.cu

@@ -333,7 +333,7 @@ static struct {
 }
 OperationKind_enumerants[] = {
   {"eq_gemm", "EqGemm", OperationKind::kEqGemm}, 
-  {"gemm", "Gemm", OperationKind::kGemm},               
+  {"gemm", "Gemm", OperationKind::kGemm},
   {"rank_k", "RankK", OperationKind::kRankK},
   {"rank_2k", "Rank2K", OperationKind::kRank2K},
   {"trmm", "Trmm", OperationKind::kTrmm},