[MLIR] Add f6E2M3FN type

This PR adds `f6E2M3FN` type to mlir.

`f6E2M3FN` type is proposed in [OpenCompute MX Specification](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf).
It defines a 6-bit floating point number with bit layout S1E2M3. Unlike IEEE-754 types, there are no infinity or NaN values.

```c
f6E2M3FN
- Exponent bias: 1
- Maximum stored exponent value: 3 (binary 11)
- Maximum unbiased exponent value: 3 - 1 = 2
- Minimum stored exponent value: 1 (binary 01)
- Minimum unbiased exponent value: 1 − 1 = 0
- Has Positive and Negative zero
- Doesn't have infinity
- Doesn't have NaNs

Additional details:
- Zeros (+/-): S.00.000
- Max normal number: S.11.111 = ±2^(2) x (1 + 0.875) = ±7.5
- Min normal number: S.01.000 = ±2^(0) = ±1.0
- Max subnormal number: S.00.111 = ±2^(0) x 0.875 = ±0.875
- Min subnormal number: S.00.001 = ±2^(0) x 0.125 = ±0.125
```

Related PRs:
- [PR-94735](#94735) [APFloat] Add APFloat support for FP6 data types
- [PR-105573](#105573) [MLIR] Add f6E3M2FN type - was used as a template for this PR

mlir/include/mlir-c/BuiltinTypes.h

@@ -79,6 +79,16 @@ MLIR_CAPI_EXPORTED bool mlirTypeIsAFloat(MlirType type);
 /// Returns the bitwidth of a floating-point type.
 MLIR_CAPI_EXPORTED unsigned mlirFloatTypeGetWidth(MlirType type);
 
+/// Returns the typeID of an Float6E2M3FN type.
+MLIR_CAPI_EXPORTED MlirTypeID mlirFloat6E2M3FNTypeGetTypeID(void);
+
+/// Checks whether the given type is an f6E2M3FN type.
+MLIR_CAPI_EXPORTED bool mlirTypeIsAFloat6E2M3FN(MlirType type);
+
+/// Creates an f6E2M3FN type in the given context. The type is owned by the
+/// context.
+MLIR_CAPI_EXPORTED MlirType mlirFloat6E2M3FNTypeGet(MlirContext ctx);
+
 /// Returns the typeID of an Float6E3M2FN type.
 MLIR_CAPI_EXPORTED MlirTypeID mlirFloat6E3M2FNTypeGetTypeID(void);
 

mlir/include/mlir/IR/Builders.h

@@ -60,6 +60,7 @@ class Builder {
                        Attribute metadata = Attribute());
 
   // Types.
+  FloatType getFloat6E2M3FNType();
   FloatType getFloat6E3M2FNType();
   FloatType getFloat8E5M2Type();
   FloatType getFloat8E4M3Type();

mlir/include/mlir/IR/BuiltinTypes.h

@@ -67,6 +67,7 @@ class FloatType : public Type {
   static FloatType getFloat8E4M3FNUZ(MLIRContext *ctx);
   static FloatType getFloat8E4M3B11FNUZ(MLIRContext *ctx);
   static FloatType getFloat8E3M4(MLIRContext *ctx);
+  static FloatType getFloat6E2M3FN(MLIRContext *ctx);
   static FloatType getFloat6E3M2FN(MLIRContext *ctx);
 
   /// Methods for support type inquiry through isa, cast, and dyn_cast.
@@ -414,11 +415,15 @@ inline bool BaseMemRefType::isValidElementType(Type type) {
 }
 
 inline bool FloatType::classof(Type type) {
-  return llvm::isa<Float6E3M2FNType, Float8E5M2Type, Float8E4M3Type,
-                   Float8E4M3FNType, Float8E5M2FNUZType, Float8E4M3FNUZType,
-                   Float8E4M3B11FNUZType, Float8E3M4Type, BFloat16Type,
-                   Float16Type, FloatTF32Type, Float32Type, Float64Type,
-                   Float80Type, Float128Type>(type);
+  return llvm::isa<Float6E2M3FNType, Float6E3M2FNType, Float8E5M2Type,
+                   Float8E4M3Type, Float8E4M3FNType, Float8E5M2FNUZType,
+                   Float8E4M3FNUZType, Float8E4M3B11FNUZType, Float8E3M4Type,
+                   BFloat16Type, Float16Type, FloatTF32Type, Float32Type,
+                   Float64Type, Float80Type, Float128Type>(type);
+}
+
+inline FloatType FloatType::getFloat6E2M3FN(MLIRContext *ctx) {
+  return Float6E2M3FNType::get(ctx);
 }
 
 inline FloatType FloatType::getFloat6E3M2FN(MLIRContext *ctx) {

mlir/include/mlir/IR/BuiltinTypes.td

@@ -233,6 +233,27 @@ def Builtin_Float8E3M4 : Builtin_FloatType<"Float8E3M4", "f8E3M4"> {
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// Float6E2M3FNType
+
+def Builtin_Float6E2M3FN : Builtin_FloatType<"Float6E2M3FN", "f6E2M3FN"> {
+  let summary = "6-bit floating point with 3 bits exponent and 2 bit mantissa";
+  let description = [{
+    An 6-bit floating point type with 1 sign bit, 2 bits exponent and 3 bits
+    mantissa. This is not a standard type as defined by IEEE-754, but it
+    follows similar conventions with the following characteristics:
+
+      * bit encoding: S1E2M3
+      * exponent bias: 1
+      * infinities: Not supported
+      * NaNs: Not supported
+      * denormals when exponent is 0
+
+    Open Compute Project (OCP) microscaling formats (MX) specification:
+    https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // Float6E3M2FNType
 

mlir/include/mlir/IR/CommonTypeConstraints.td

@@ -344,6 +344,8 @@ def F8E5M2FNUZ : Type<CPred<"$_self.isFloat8E5M2FNUZ()">, "f8E5M2FNUZ type">,
                  BuildableType<"$_builder.getFloat8E5M2FNUZType()">;
 def F8E3M4 : Type<CPred<"$_self.isFloat8E3M4()">, "f8E3M4 type">,
              BuildableType<"$_builder.getFloat8E3M4Type()">;
+def F6E2M3FN : Type<CPred<"$_self.isFloat6E2M3FN()">, "f6E2M3FN type">,
+               BuildableType<"$_builder.getFloat6E2M3FNType()">;
 def F6E3M2FN : Type<CPred<"$_self.isFloat6E3M2FN()">, "f6E3M2FN type">,
                BuildableType<"$_builder.getFloat6E3M2FNType()">;
 

mlir/include/mlir/IR/Types.h

@@ -125,6 +125,7 @@ class Type {
   // Convenience predicates.  This is only for floating point types,
   // derived types should use isa/dyn_cast.
   bool isIndex() const;
+  bool isFloat6E2M3FN() const;
   bool isFloat6E3M2FN() const;
   bool isFloat8E5M2() const;
   bool isFloat8E4M3() const;

mlir/lib/AsmParser/TokenKinds.def

@@ -101,6 +101,7 @@ TOK_KEYWORD(f8E5M2FNUZ)
 TOK_KEYWORD(f8E4M3FNUZ)
 TOK_KEYWORD(f8E4M3B11FNUZ)
 TOK_KEYWORD(f8E3M4)
+TOK_KEYWORD(f6E2M3FN)
 TOK_KEYWORD(f6E3M2FN)
 TOK_KEYWORD(f128)
 TOK_KEYWORD(false)

mlir/lib/AsmParser/TypeParser.cpp

@@ -39,6 +39,7 @@ OptionalParseResult Parser::parseOptionalType(Type &type) {
   case Token::kw_tuple:
   case Token::kw_vector:
   case Token::inttype:
+  case Token::kw_f6E2M3FN:
   case Token::kw_f6E3M2FN:
   case Token::kw_f8E5M2:
   case Token::kw_f8E4M3:
@@ -304,6 +305,9 @@ Type Parser::parseNonFunctionType() {
   }
 
   // float-type
+  case Token::kw_f6E2M3FN:
+    consumeToken(Token::kw_f6E2M3FN);
+    return builder.getFloat6E2M3FNType();
   case Token::kw_f6E3M2FN:
     consumeToken(Token::kw_f6E3M2FN);
     return builder.getFloat6E3M2FNType();

mlir/lib/Bindings/Python/IRTypes.cpp

@@ -124,6 +124,27 @@ class PyFloatType : public PyConcreteType<PyFloatType> {
   }
 };
 
+/// Floating Point Type subclass - Float6E2M3FNType.
+class PyFloat6E2M3FNType
+    : public PyConcreteType<PyFloat6E2M3FNType, PyFloatType> {
+public:
+  static constexpr IsAFunctionTy isaFunction = mlirTypeIsAFloat6E2M3FN;
+  static constexpr GetTypeIDFunctionTy getTypeIdFunction =
+      mlirFloat6E2M3FNTypeGetTypeID;
+  static constexpr const char *pyClassName = "Float6E2M3FNType";
+  using PyConcreteType::PyConcreteType;
+
+  static void bindDerived(ClassTy &c) {
+    c.def_static(
+        "get",
+        [](DefaultingPyMlirContext context) {
+          MlirType t = mlirFloat6E2M3FNTypeGet(context->get());
+          return PyFloat6E2M3FNType(context->getRef(), t);
+        },
+        py::arg("context") = py::none(), "Create a float6_e2m3fn type.");
+  }
+};
+
 /// Floating Point Type subclass - Float6E3M2FNType.
 class PyFloat6E3M2FNType
     : public PyConcreteType<PyFloat6E3M2FNType, PyFloatType> {
@@ -901,6 +922,7 @@ void mlir::python::populateIRTypes(py::module &m) {
   PyIntegerType::bind(m);
   PyFloatType::bind(m);
   PyIndexType::bind(m);
+  PyFloat6E2M3FNType::bind(m);
   PyFloat6E3M2FNType::bind(m);
   PyFloat8E4M3FNType::bind(m);
   PyFloat8E5M2Type::bind(m);

mlir/lib/CAPI/IR/BuiltinTypes.cpp

@@ -85,6 +85,18 @@ unsigned mlirFloatTypeGetWidth(MlirType type) {
   return llvm::cast<FloatType>(unwrap(type)).getWidth();
 }
 
+MlirTypeID mlirFloat6E2M3FNTypeGetTypeID() {
+  return wrap(Float6E2M3FNType::getTypeID());
+}
+
+bool mlirTypeIsAFloat6E2M3FN(MlirType type) {
+  return unwrap(type).isFloat6E2M3FN();
+}
+
+MlirType mlirFloat6E2M3FNTypeGet(MlirContext ctx) {
+  return wrap(FloatType::getFloat6E2M3FN(unwrap(ctx)));
+}
+
 MlirTypeID mlirFloat6E3M2FNTypeGetTypeID() {
   return wrap(Float6E3M2FNType::getTypeID());
 }

mlir/lib/Conversion/LLVMCommon/TypeConverter.cpp

@@ -250,7 +250,7 @@ Type LLVMTypeConverter::convertFloatType(FloatType type) const {
   if (type.isFloat8E5M2() || type.isFloat8E4M3() || type.isFloat8E4M3FN() ||
       type.isFloat8E5M2FNUZ() || type.isFloat8E4M3FNUZ() ||
       type.isFloat8E4M3B11FNUZ() || type.isFloat8E3M4() ||
-      type.isFloat6E3M2FN())
+      type.isFloat6E2M3FN() || type.isFloat6E3M2FN())
     return IntegerType::get(&getContext(), type.getWidth());
   return type;
 }

mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp

@@ -55,6 +55,7 @@ static std::optional<FloatType> parseFloatType(MLIRContext *ctx,
                                                StringRef name) {
   Builder b(ctx);
   return llvm::StringSwitch<std::optional<FloatType>>(name)
+      .Case("f6E2M3FN", b.getFloat6E2M3FNType())
       .Case("f6E3M2FN", b.getFloat6E3M2FNType())
       .Case("f8E5M2", b.getFloat8E5M2Type())
       .Case("f8E4M3", b.getFloat8E4M3Type())

mlir/lib/IR/AsmPrinter.cpp

@@ -2575,6 +2575,7 @@ void AsmPrinter::Impl::printTypeImpl(Type type) {
                            opaqueTy.getTypeData());
       })
       .Case<IndexType>([&](Type) { os << "index"; })
+      .Case<Float6E2M3FNType>([&](Type) { os << "f6E2M3FN"; })
       .Case<Float6E3M2FNType>([&](Type) { os << "f6E3M2FN"; })
       .Case<Float8E5M2Type>([&](Type) { os << "f8E5M2"; })
       .Case<Float8E4M3Type>([&](Type) { os << "f8E4M3"; })

mlir/lib/IR/Builders.cpp

@@ -34,6 +34,10 @@ Location Builder::getFusedLoc(ArrayRef<Location> locs, Attribute metadata) {
 // Types.
 //===----------------------------------------------------------------------===//
 
+FloatType Builder::getFloat6E2M3FNType() {
+  return FloatType::getFloat6E2M3FN(context);
+}
+
 FloatType Builder::getFloat6E3M2FNType() {
   return FloatType::getFloat6E3M2FN(context);
 }

mlir/lib/IR/BuiltinTypes.cpp

@@ -101,6 +101,8 @@ unsigned FloatType::getWidth() {
 
 /// Returns the floating semantics for the given type.
 const llvm::fltSemantics &FloatType::getFloatSemantics() {
+  if (llvm::isa<Float6E2M3FNType>(*this))
+    return APFloat::Float6E2M3FN();
   if (llvm::isa<Float6E3M2FNType>(*this))
     return APFloat::Float6E3M2FN();
   if (llvm::isa<Float8E5M2Type>(*this))

mlir/lib/IR/MLIRContext.cpp

@@ -221,6 +221,7 @@ class MLIRContextImpl {
   llvm::DenseMap<StringRef, AbstractType *> nameToType;
 
   /// Cached Type Instances.
+  Float6E2M3FNType f6E2M3FNTy;
   Float6E3M2FNType f6E3M2FNTy;
   Float8E5M2Type f8E5M2Ty;
   Float8E4M3Type f8E4M3Ty;
@@ -314,6 +315,7 @@ MLIRContext::MLIRContext(const DialectRegistry &registry, Threading setting)
 
   //// Types.
   /// Floating-point Types.
+  impl->f6E2M3FNTy = TypeUniquer::get<Float6E2M3FNType>(this);
   impl->f6E3M2FNTy = TypeUniquer::get<Float6E3M2FNType>(this);
   impl->f8E5M2Ty = TypeUniquer::get<Float8E5M2Type>(this);
   impl->f8E4M3Ty = TypeUniquer::get<Float8E4M3Type>(this);
@@ -1015,6 +1017,9 @@ AbstractType::lookup(StringRef name, MLIRContext *context) {
 /// This should not be used directly.
 StorageUniquer &MLIRContext::getTypeUniquer() { return getImpl().typeUniquer; }
 
+Float6E2M3FNType Float6E2M3FNType::get(MLIRContext *context) {
+  return context->getImpl().f6E2M3FNTy;
+}
 Float6E3M2FNType Float6E3M2FNType::get(MLIRContext *context) {
   return context->getImpl().f6E3M2FNTy;
 }

mlir/lib/IR/Types.cpp

@@ -34,6 +34,7 @@ Type AbstractType::replaceImmediateSubElements(Type type,
 
 MLIRContext *Type::getContext() const { return getDialect().getContext(); }
 
+bool Type::isFloat6E2M3FN() const { return llvm::isa<Float6E2M3FNType>(*this); }
 bool Type::isFloat6E3M2FN() const { return llvm::isa<Float6E3M2FNType>(*this); }
 bool Type::isFloat8E5M2() const { return llvm::isa<Float8E5M2Type>(*this); }
 bool Type::isFloat8E4M3() const { return llvm::isa<Float8E4M3Type>(*this); }

mlir/python/mlir/_mlir_libs/_mlir/ir.pyi

@@ -120,6 +120,7 @@ __all__ = [
     "F32Type",
     "F64Type",
     "FlatSymbolRefAttr",
+    "Float6E2M3FNType",
     "Float6E3M2FNType",
     "Float8E3M4Type",
     "Float8E4M3B11FNUZType",
@@ -1540,6 +1541,19 @@ class FlatSymbolRefAttr(Attribute):
         Returns the value of the FlatSymbolRef attribute as a string
         """
 
+class Float6E2M3FNType(FloatType):
+    static_typeid: ClassVar[TypeID]
+    @staticmethod
+    def get(context: Optional[Context] = None) -> Float6E2M3FNType:
+        """
+        Create a float6_e2m3fn type.
+        """
+    @staticmethod
+    def isinstance(other: Type) -> bool: ...
+    def __init__(self, cast_from_type: Type) -> None: ...
+    @property
+    def typeid(self) -> TypeID: ...
+
 class Float6E3M2FNType(FloatType):
     static_typeid: ClassVar[TypeID]
     @staticmethod

mlir/python/mlir/extras/types.py

@@ -12,6 +12,7 @@
     F16Type,
     F32Type,
     F64Type,
+    Float6E2M3FNType,
     Float6E3M2FNType,
     Float8E3M4Type,
     Float8E4M3B11FNUZType,
@@ -75,6 +76,7 @@ def ui(width):
 f8E4M3FN = lambda: Float8E4M3FNType.get()
 f8E4M3B11FNUZ = lambda: Float8E4M3B11FNUZType.get()
 f8E3M4 = lambda: Float8E3M4Type.get()
+f6E2M3FN = lambda: Float6E2M3FNType.get()
 f6E3M2FN = lambda: Float6E3M2FNType.get()
 
 none = lambda: NoneType.get()

mlir/test/IR/attribute.mlir

@@ -36,6 +36,10 @@ func.func @any_attr_of_fail() {
 //===----------------------------------------------------------------------===//
 
 func.func @float_attrs_pass() {
+  "test.float_attrs"() {
+    // CHECK: float_attr = 2.000000e+00 : f6E2M3FN
+    float_attr = 2. : f6E2M3FN
+  } : () -> ()
   "test.float_attrs"() {
     // CHECK: float_attr = 2.000000e+00 : f6E3M2FN
     float_attr = 2. : f6E3M2FN

mlir/test/Target/LLVMIR/llvmir.mlir

@@ -42,6 +42,9 @@ llvm.mlir.global internal @int_global_undef() : i64
 // CHECK: @externally_initialized_global = internal externally_initialized global i32 0
 llvm.mlir.global internal @externally_initialized_global(0 : i32) {externally_initialized} : i32
 
+// CHECK: @f6E2M3FN_global_as_i6 = internal global i6 12
+llvm.mlir.global internal @f6E2M3FN_global_as_i6(1.5 : f6E2M3FN) : i6
+
 // CHECK: @f6E3M2FN_global_as_i6 = internal global i6 14
 llvm.mlir.global internal @f6E3M2FN_global_as_i6(1.5 : f6E3M2FN) : i6
 

mlir/test/python/ir/builtin_types.py

@@ -113,6 +113,8 @@ def testTypeIsInstance():
 def testFloatTypeSubclasses():
     ctx = Context()
     # CHECK: True
+    print(isinstance(Type.parse("f6E2M3FN", ctx), FloatType))
+    # CHECK: True
     print(isinstance(Type.parse("f6E3M2FN", ctx), FloatType))
     # CHECK: True
     print(isinstance(Type.parse("f8E3M4", ctx), FloatType))
@@ -235,6 +237,8 @@ def testIndexType():
 @run
 def testFloatType():
     with Context():
+        # CHECK: float: f6E2M3FN
+        print("float:", Float6E2M3FNType.get())
         # CHECK: float: f6E3M2FN
         print("float:", Float6E3M2FNType.get())
         # CHECK: float: f8E3M4
@@ -613,6 +617,7 @@ def testTypeIDs():
         types = [
             (IntegerType, IntegerType.get_signless(16)),
             (IndexType, IndexType.get()),
+            (Float6E2M3FNType, Float6E2M3FNType.get()),
             (Float6E3M2FNType, Float6E3M2FNType.get()),
             (Float8E3M4Type, Float8E3M4Type.get()),
             (Float8E4M3Type, Float8E4M3Type.get()),
@@ -639,6 +644,7 @@ def testTypeIDs():
 
         # CHECK: IntegerType(i16)
         # CHECK: IndexType(index)
+        # CHECK: Float6E2M3FNType(f6E2M3FN)
         # CHECK: Float6E3M2FNType(f6E3M2FN)
         # CHECK: Float8E3M4Type(f8E3M4)
         # CHECK: Float8E4M3Type(f8E4M3)
@@ -719,6 +725,9 @@ def print_downcasted(typ):
         # CHECK: F64Type
         # CHECK: F64Type(f64)
         print_downcasted(F64Type.get())
+        # CHECK: Float6E2M3FNType
+        # CHECK: Float6E2M3FNType(f6E2M3FN)
+        print_downcasted(Float6E2M3FNType.get())
         # CHECK: Float6E3M2FNType
         # CHECK: Float6E3M2FNType(f6E3M2FN)
         print_downcasted(Float6E3M2FNType.get())

mlir/utils/lldb-scripts/mlirDataFormatters.py

@@ -50,6 +50,7 @@ def build_ptr_str_from_addr(addrValue: lldb.SBValue, type: lldb.SBType):
     "mlir::CallSiteLoc": '"loc(callsite(...))"',
     "mlir::FusedLoc": '"loc(fused<...>[...])"',
     "mlir::UnknownLoc": '"loc(unknown)"',
+    "mlir::Float6E2M3FNType": '"f6E2M3FN"',
     "mlir::Float6E3M2FNType": '"f6E3M2FN"',
     "mlir::Float8E5M2Type": '"f8E5M2"',
     "mlir::Float8E4M3Type": '"f8E4M3"',

mlir/utils/tree-sitter-mlir/grammar.js

@@ -231,7 +231,7 @@ const common = {
       token(seq(choice('si', 'ui', 'i'), /[1-9]/, repeat(/[0-9]/))),
   float_type : $ => token(
       choice('f16', 'f32', 'f64', 'f80', 'f128', 'bf16', 'f8E3M4', 'f8E4M3FN',
-             'f8E4M3', 'f8E5M2', 'f6E3M2FN')),
+             'f8E4M3', 'f8E5M2', 'f6E2M3FN', 'f6E3M2FN')),
   index_type : $ => token('index'),
   none_type : $ => token('none'),
   complex_type : $ => seq(token('complex'), '<', $._prim_type, '>'),