Skip to content

Commit 2c58063

Browse files
sergey-kozubsergey-kozub
authored
[MLIR] Add f4E2M1FN type (#108877)
This PR adds `f4E2M1FN` type to mlir. `f4E2M1FN` type is proposed in [OpenCompute MX Specification](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf). It defines a 4-bit floating point number with bit layout S1E2M1. Unlike IEEE-754 types, there are no infinity or NaN values. ```c f4E2M1FN - 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.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 ``` Related PRs: - [PR-95392](#95392) [APFloat] Add APFloat support for FP4 data type - [PR-105573](#105573) [MLIR] Add f6E3M2FN type - was used as a template for this PR - [PR-107999](#107999) [MLIR] Add f6E2M3FN type
1 parent 63b2595 commit 2c58063

File tree

24 files changed

+133
-7
lines changed

24 files changed

+133
-7
lines changed

‎mlir/include/mlir-c/BuiltinTypes.h

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -79,6 +79,16 @@ MLIR_CAPI_EXPORTED bool mlirTypeIsAFloat(MlirType type);
7979
/// Returns the bitwidth of a floating-point type.
8080
MLIR_CAPI_EXPORTED unsigned mlirFloatTypeGetWidth(MlirType type);
8181

82+
/// Returns the typeID of an Float4E2M1FN type.
83+
MLIR_CAPI_EXPORTED MlirTypeID mlirFloat4E2M1FNTypeGetTypeID(void);
84+
85+
/// Checks whether the given type is an f4E2M1FN type.
86+
MLIR_CAPI_EXPORTED bool mlirTypeIsAFloat4E2M1FN(MlirType type);
87+
88+
/// Creates an f4E2M1FN type in the given context. The type is owned by the
89+
/// context.
90+
MLIR_CAPI_EXPORTED MlirType mlirFloat4E2M1FNTypeGet(MlirContext ctx);
91+
8292
/// Returns the typeID of an Float6E2M3FN type.
8393
MLIR_CAPI_EXPORTED MlirTypeID mlirFloat6E2M3FNTypeGetTypeID(void);
8494

‎mlir/include/mlir/IR/Builders.h

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -60,6 +60,7 @@ class Builder {
6060
Attribute metadata = Attribute());
6161

6262
// Types.
63+
FloatType getFloat4E2M1FNType();
6364
FloatType getFloat6E2M3FNType();
6465
FloatType getFloat6E3M2FNType();
6566
FloatType getFloat8E5M2Type();

‎mlir/include/mlir/IR/BuiltinTypes.h

Lines changed: 10 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -67,6 +67,7 @@ class FloatType : public Type {
6767
static FloatType getFloat8E4M3FNUZ(MLIRContext *ctx);
6868
static FloatType getFloat8E4M3B11FNUZ(MLIRContext *ctx);
6969
static FloatType getFloat8E3M4(MLIRContext *ctx);
70+
static FloatType getFloat4E2M1FN(MLIRContext *ctx);
7071
static FloatType getFloat6E2M3FN(MLIRContext *ctx);
7172
static FloatType getFloat6E3M2FN(MLIRContext *ctx);
7273

@@ -415,11 +416,15 @@ inline bool BaseMemRefType::isValidElementType(Type type) {
415416
}
416417

417418
inline bool FloatType::classof(Type type) {
418-
return llvm::isa<Float6E2M3FNType, Float6E3M2FNType, Float8E5M2Type,
419-
Float8E4M3Type, Float8E4M3FNType, Float8E5M2FNUZType,
420-
Float8E4M3FNUZType, Float8E4M3B11FNUZType, Float8E3M4Type,
421-
BFloat16Type, Float16Type, FloatTF32Type, Float32Type,
422-
Float64Type, Float80Type, Float128Type>(type);
419+
return llvm::isa<
420+
Float4E2M1FNType, Float6E2M3FNType, Float6E3M2FNType, Float8E5M2Type,
421+
Float8E4M3Type, Float8E4M3FNType, Float8E5M2FNUZType, Float8E4M3FNUZType,
422+
Float8E4M3B11FNUZType, Float8E3M4Type, BFloat16Type, Float16Type,
423+
FloatTF32Type, Float32Type, Float64Type, Float80Type, Float128Type>(type);
424+
}
425+
426+
inline FloatType FloatType::getFloat4E2M1FN(MLIRContext *ctx) {
427+
return Float4E2M1FNType::get(ctx);
423428
}
424429

425430
inline FloatType FloatType::getFloat6E2M3FN(MLIRContext *ctx) {

‎mlir/include/mlir/IR/BuiltinTypes.td

Lines changed: 21 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -233,6 +233,27 @@ def Builtin_Float8E3M4 : Builtin_FloatType<"Float8E3M4", "f8E3M4"> {
233233
}];
234234
}
235235

236+
//===----------------------------------------------------------------------===//
237+
// Float4E2M1FNType
238+
239+
def Builtin_Float4E2M1FN : Builtin_FloatType<"Float4E2M1FN", "f4E2M1FN"> {
240+
let summary = "4-bit floating point with 2-bit exponent and 1-bit mantissa";
241+
let description = [{
242+
An 4-bit floating point type with 1 sign bit, 2 bits exponent and 1 bit
243+
mantissa. This is not a standard type as defined by IEEE-754, but it
244+
follows similar conventions with the following characteristics:
245+
246+
* bit encoding: S1E2M1
247+
* exponent bias: 1
248+
* infinities: Not supported
249+
* NaNs: Not supported
250+
* denormals when exponent is 0
251+
252+
Open Compute Project (OCP) microscaling formats (MX) specification:
253+
https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
254+
}];
255+
}
256+
236257
//===----------------------------------------------------------------------===//
237258
// Float6E2M3FNType
238259

‎mlir/include/mlir/IR/CommonTypeConstraints.td

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -347,6 +347,8 @@ def F8E5M2FNUZ : Type<CPred<"$_self.isFloat8E5M2FNUZ()">, "f8E5M2FNUZ type">,
347347
BuildableType<"$_builder.getFloat8E5M2FNUZType()">;
348348
def F8E3M4 : Type<CPred<"$_self.isFloat8E3M4()">, "f8E3M4 type">,
349349
BuildableType<"$_builder.getFloat8E3M4Type()">;
350+
def F4E2M1FN : Type<CPred<"$_self.isFloat4E2M1FN()">, "f4E2M1FN type">,
351+
BuildableType<"$_builder.getFloat4E2M1FNType()">;
350352
def F6E2M3FN : Type<CPred<"$_self.isFloat6E2M3FN()">, "f6E2M3FN type">,
351353
BuildableType<"$_builder.getFloat6E2M3FNType()">;
352354
def F6E3M2FN : Type<CPred<"$_self.isFloat6E3M2FN()">, "f6E3M2FN type">,

‎mlir/include/mlir/IR/Types.h

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -125,6 +125,7 @@ class Type {
125125
// Convenience predicates. This is only for floating point types,
126126
// derived types should use isa/dyn_cast.
127127
bool isIndex() const;
128+
bool isFloat4E2M1FN() const;
128129
bool isFloat6E2M3FN() const;
129130
bool isFloat6E3M2FN() const;
130131
bool isFloat8E5M2() const;

‎mlir/lib/AsmParser/TokenKinds.def

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -101,6 +101,7 @@ TOK_KEYWORD(f8E5M2FNUZ)
101101
TOK_KEYWORD(f8E4M3FNUZ)
102102
TOK_KEYWORD(f8E4M3B11FNUZ)
103103
TOK_KEYWORD(f8E3M4)
104+
TOK_KEYWORD(f4E2M1FN)
104105
TOK_KEYWORD(f6E2M3FN)
105106
TOK_KEYWORD(f6E3M2FN)
106107
TOK_KEYWORD(f128)

‎mlir/lib/AsmParser/TypeParser.cpp

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -39,6 +39,7 @@ OptionalParseResult Parser::parseOptionalType(Type &type) {
3939
case Token::kw_tuple:
4040
case Token::kw_vector:
4141
case Token::inttype:
42+
case Token::kw_f4E2M1FN:
4243
case Token::kw_f6E2M3FN:
4344
case Token::kw_f6E3M2FN:
4445
case Token::kw_f8E5M2:
@@ -305,6 +306,9 @@ Type Parser::parseNonFunctionType() {
305306
}
306307

307308
// float-type
309+
case Token::kw_f4E2M1FN:
310+
consumeToken(Token::kw_f4E2M1FN);
311+
return builder.getFloat4E2M1FNType();
308312
case Token::kw_f6E2M3FN:
309313
consumeToken(Token::kw_f6E2M3FN);
310314
return builder.getFloat6E2M3FNType();

‎mlir/lib/Bindings/Python/IRTypes.cpp

Lines changed: 22 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -124,6 +124,27 @@ class PyFloatType : public PyConcreteType<PyFloatType> {
124124
}
125125
};
126126

127+
/// Floating Point Type subclass - Float4E2M1FNType.
128+
class PyFloat4E2M1FNType
129+
: public PyConcreteType<PyFloat4E2M1FNType, PyFloatType> {
130+
public:
131+
static constexpr IsAFunctionTy isaFunction = mlirTypeIsAFloat4E2M1FN;
132+
static constexpr GetTypeIDFunctionTy getTypeIdFunction =
133+
mlirFloat4E2M1FNTypeGetTypeID;
134+
static constexpr const char *pyClassName = "Float4E2M1FNType";
135+
using PyConcreteType::PyConcreteType;
136+
137+
static void bindDerived(ClassTy &c) {
138+
c.def_static(
139+
"get",
140+
[](DefaultingPyMlirContext context) {
141+
MlirType t = mlirFloat4E2M1FNTypeGet(context->get());
142+
return PyFloat4E2M1FNType(context->getRef(), t);
143+
},
144+
py::arg("context") = py::none(), "Create a float4_e2m1fn type.");
145+
}
146+
};
147+
127148
/// Floating Point Type subclass - Float6E2M3FNType.
128149
class PyFloat6E2M3FNType
129150
: public PyConcreteType<PyFloat6E2M3FNType, PyFloatType> {
@@ -922,6 +943,7 @@ void mlir::python::populateIRTypes(py::module &m) {
922943
PyIntegerType::bind(m);
923944
PyFloatType::bind(m);
924945
PyIndexType::bind(m);
946+
PyFloat4E2M1FNType::bind(m);
925947
PyFloat6E2M3FNType::bind(m);
926948
PyFloat6E3M2FNType::bind(m);
927949
PyFloat8E4M3FNType::bind(m);

‎mlir/lib/CAPI/IR/BuiltinTypes.cpp

Lines changed: 12 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -85,6 +85,18 @@ unsigned mlirFloatTypeGetWidth(MlirType type) {
8585
return llvm::cast<FloatType>(unwrap(type)).getWidth();
8686
}
8787

88+
MlirTypeID mlirFloat4E2M1FNTypeGetTypeID() {
89+
return wrap(Float4E2M1FNType::getTypeID());
90+
}
91+
92+
bool mlirTypeIsAFloat4E2M1FN(MlirType type) {
93+
return unwrap(type).isFloat4E2M1FN();
94+
}
95+
96+
MlirType mlirFloat4E2M1FNTypeGet(MlirContext ctx) {
97+
return wrap(FloatType::getFloat4E2M1FN(unwrap(ctx)));
98+
}
99+
88100
MlirTypeID mlirFloat6E2M3FNTypeGetTypeID() {
89101
return wrap(Float6E2M3FNType::getTypeID());
90102
}

0 commit comments

Comments
 (0)