Cuda 12 sm86

ocelot/include/ocelot/executive/CooperativeThreadArray.h

@@ -475,6 +475,7 @@ namespace executive {
 		void eval_Ex2(CTAContext &context, const ir::PTXInstruction &instr);
 		void eval_Exit(CTAContext &context, const ir::PTXInstruction &instr);
 		void eval_Fma(CTAContext &context, const ir::PTXInstruction &instr);
+		void eval_Mma(CTAContext &context, const ir::PTXInstruction &instr);
 		void eval_Isspacep(CTAContext &context,
 			const ir::PTXInstruction &instr);
 		void eval_Ld(CTAContext &context, const ir::PTXInstruction &instr);
@@ -562,4 +563,3 @@ namespace executive {
 }
 
 #endif
-

ocelot/include/ocelot/ir/PTXInstruction.h

@@ -54,6 +54,7 @@ namespace ir {
 			Mad24,
 			Mad,
 			MadC,
+			Mma,
 			Max,
 			Membar,
 			Min,
@@ -143,6 +144,12 @@ namespace ir {
 			None = 0,
 			CC = 1
 		};
+
+		enum MmaShape {
+			MmaM16N8K8,
+			MmaM16N8K16,
+			MmaShape_Invalid
+		};
 
 		enum Volatility {
 			Nonvolatile = 0,
@@ -426,6 +433,9 @@ namespace ir {
 		/*! indicates data type of instruction */
 		PTXOperand::DataType type;
 
+		/*! Shape for MMA instructions */
+		MmaShape mmaShape;
+
 		/*! Flag containing one or more floating-point modifiers */
 		unsigned int modifier;
 
@@ -602,4 +612,3 @@ namespace ir {
 }
 
 #endif
-

ocelot/include/ocelot/ir/PTXOperand.h

@@ -25,6 +25,7 @@ namespace ir {
 	typedef int32_t PTXS32;
 	typedef int64_t PTXS64;
 
+	typedef _Float16 PTXF16;
 	typedef float PTXF32;
 	typedef double PTXF64;
 
@@ -62,6 +63,8 @@ namespace ir {
 			u64,
 			f16,
 			f32,
+			bf16,
+			tf32,
 			f64,
 			b8,
 			b16,
@@ -285,4 +288,3 @@ namespace std {
 }
 
 #endif
-

ocelot/include/ocelot/parser/PTXParser.h

@@ -137,6 +137,7 @@ namespace parser
 
 				private:
 					void _setImmediateTypes();
+					void _setMovVectorImmediateTypes();
 					std::string _nameInContext( const std::string& name );
 
 					OperandWrapper* _getOperand( const std::string& name );
@@ -236,6 +237,7 @@ namespace parser
 					void constantOperand( double value );
 					void indexedOperand( const std::string& name, 
 						YYLTYPE& location, long long int value );
+					void vectorOperand( unsigned int elements );
 					void addressableOperand( const std::string& name, 
 						long long int value, YYLTYPE& location, 
 						bool invert );
@@ -267,6 +269,7 @@ namespace parser
 					void instruction();
 					void instruction( const std::string& opcode, int dataType );
 					void instruction( const std::string& opcode );
+					void mma( int shape, int accumulatorType, int aType, int bType, int cType );
 					void tex( int dataType );
 					void tld4( int dataType );
 					void callPrototypeName( const std::string& identifier );
@@ -366,4 +369,3 @@ namespace parser
 }
 
 #endif
-