[Mlir-commits] [clang] [llvm] [mlir] Add f8E4M3 IEEE 754 type (PR #97118)
Alexander Pivovarov
llvmlistbot at llvm.org
Fri Jun 28 21:59:14 PDT 2024
https://github.com/apivovarov updated https://github.com/llvm/llvm-project/pull/97118
>From e5fb4d85889472fd9880e5262d3a9cf00e1b1f89 Mon Sep 17 00:00:00 2001
From: Alexander Pivovarov <pivovaa at amazon.com>
Date: Fri, 28 Jun 2024 19:33:21 +0000
Subject: [PATCH 1/2] Rename f8E4M3 to f8E4M3FN in mlir.extras.types py package
---
llvm/lib/Support/APFloat.cpp | 4 ++--
llvm/unittests/ADT/APFloatTest.cpp | 12 ++++++------
mlir/python/mlir/extras/types.py | 2 +-
3 files changed, 9 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Support/APFloat.cpp b/llvm/lib/Support/APFloat.cpp
index 47618bc325951..3017a9b976658 100644
--- a/llvm/lib/Support/APFloat.cpp
+++ b/llvm/lib/Support/APFloat.cpp
@@ -83,8 +83,8 @@ enum class fltNanEncoding {
// exponent is all 1s and the significand is non-zero.
IEEE,
- // Represents the behavior in the Float8E4M3 floating point type where NaN is
- // represented by having the exponent and mantissa set to all 1s.
+ // Represents the behavior in the Float8E4M3FN floating point type where NaN
+ // is represented by having the exponent and mantissa set to all 1s.
// This behavior matches the FP8 E4M3 type described in
// https://arxiv.org/abs/2209.05433. We treat both signed and unsigned NaNs
// as non-signalling, although the paper does not state whether the NaN
diff --git a/llvm/unittests/ADT/APFloatTest.cpp b/llvm/unittests/ADT/APFloatTest.cpp
index cf6bbd313c6c6..86a25f4394e19 100644
--- a/llvm/unittests/ADT/APFloatTest.cpp
+++ b/llvm/unittests/ADT/APFloatTest.cpp
@@ -5508,8 +5508,8 @@ TEST(APFloatTest, ConvertE4M3FNToE5M2) {
EXPECT_TRUE(losesInfo);
EXPECT_EQ(status, APFloat::opInexact);
- // Convert E4M3 denormal to E5M2 normal. Should not be truncated, despite the
- // destination format having one fewer significand bit
+ // Convert E4M3FN denormal to E5M2 normal. Should not be truncated, despite
+ // the destination format having one fewer significand bit
test = APFloat(APFloat::Float8E4M3FN(), "0x1.Cp-7");
status = test.convert(APFloat::Float8E5M2(), APFloat::rmNearestTiesToEven,
&losesInfo);
@@ -5647,8 +5647,8 @@ TEST(APFloatTest, Float8E4M3FNAdd) {
int category;
APFloat::roundingMode roundingMode = APFloat::rmNearestTiesToEven;
} AdditionTests[] = {
- // Test addition operations involving NaN, overflow, and the max E4M3
- // value (448) because E4M3 differs from IEEE-754 types in these regards
+ // Test addition operations involving NaN, overflow, and the max E4M3FN
+ // value (448) because E4M3FN differs from IEEE-754 types in these regards
{FromStr("448"), FromStr("16"), "448", APFloat::opInexact,
APFloat::fcNormal},
{FromStr("448"), FromStr("18"), "NaN",
@@ -6278,8 +6278,8 @@ TEST(APFloatTest, ConvertE4M3FNUZToE5M2FNUZ) {
EXPECT_TRUE(losesInfo);
EXPECT_EQ(status, APFloat::opInexact);
- // Convert E4M3 denormal to E5M2 normal. Should not be truncated, despite the
- // destination format having one fewer significand bit
+ // Convert E4M3FNUZ denormal to E5M2 normal. Should not be truncated, despite
+ // the destination format having one fewer significand bit
losesInfo = true;
test = APFloat(APFloat::Float8E4M3FNUZ(), "0x1.Cp-8");
status = test.convert(APFloat::Float8E5M2FNUZ(), APFloat::rmNearestTiesToEven,
diff --git a/mlir/python/mlir/extras/types.py b/mlir/python/mlir/extras/types.py
index db9e8229fb288..b93c46b172a9a 100644
--- a/mlir/python/mlir/extras/types.py
+++ b/mlir/python/mlir/extras/types.py
@@ -68,7 +68,7 @@ def ui(width):
bf16 = lambda: BF16Type.get()
f8E5M2 = lambda: Float8E5M2Type.get()
-f8E4M3 = lambda: Float8E4M3FNType.get()
+f8E4M3FN = lambda: Float8E4M3FNType.get()
f8E4M3B11FNUZ = lambda: Float8E4M3B11FNUZType.get()
none = lambda: NoneType.get()
>From 79567a1d3bd3e68946e4a80edbb6690e499e07b7 Mon Sep 17 00:00:00 2001
From: Alexander Pivovarov <pivovaa at amazon.com>
Date: Fri, 28 Jun 2024 21:09:33 +0000
Subject: [PATCH 2/2] Add f8E4M3 IEEE 754 type
---
clang/include/clang/AST/Stmt.h | 6 +--
clang/lib/AST/MicrosoftMangle.cpp | 1 +
llvm/include/llvm/ADT/APFloat.h | 6 +++
llvm/lib/Support/APFloat.cpp | 20 ++++++++++
llvm/unittests/ADT/APFloatTest.cpp | 38 +++++++++++++++++++
mlir/include/mlir-c/BuiltinTypes.h | 10 +++++
mlir/include/mlir/IR/Builders.h | 1 +
mlir/include/mlir/IR/BuiltinTypes.h | 13 +++++--
mlir/include/mlir/IR/BuiltinTypes.td | 19 ++++++++++
mlir/include/mlir/IR/CommonTypeConstraints.td | 2 +
mlir/include/mlir/IR/Types.h | 1 +
mlir/lib/AsmParser/TokenKinds.def | 1 +
mlir/lib/AsmParser/TypeParser.cpp | 4 ++
mlir/lib/Bindings/Python/IRTypes.cpp | 23 ++++++++++-
mlir/lib/CAPI/IR/BuiltinTypes.cpp | 12 ++++++
.../Conversion/LLVMCommon/TypeConverter.cpp | 5 ++-
.../Transforms/EmulateUnsupportedFloats.cpp | 1 +
mlir/lib/IR/AsmPrinter.cpp | 1 +
mlir/lib/IR/Builders.cpp | 4 ++
mlir/lib/IR/BuiltinTypes.cpp | 7 +++-
mlir/lib/IR/MLIRContext.cpp | 5 +++
mlir/lib/IR/Types.cpp | 1 +
mlir/python/mlir/_mlir_libs/_mlir/ir.pyi | 14 +++++++
mlir/python/mlir/extras/types.py | 2 +
mlir/test/IR/attribute.mlir | 4 ++
mlir/test/Target/LLVMIR/llvmir.mlir | 3 ++
mlir/test/python/ir/builtin_types.py | 9 +++++
mlir/utils/lldb-scripts/mlirDataFormatters.py | 1 +
mlir/utils/tree-sitter-mlir/grammar.js | 3 +-
29 files changed, 204 insertions(+), 13 deletions(-)
diff --git a/clang/include/clang/AST/Stmt.h b/clang/include/clang/AST/Stmt.h
index 9cd7a364cd3f1..4edb04502f83e 100644
--- a/clang/include/clang/AST/Stmt.h
+++ b/clang/include/clang/AST/Stmt.h
@@ -460,10 +460,10 @@ class alignas(void *) Stmt {
unsigned : NumExprBits;
static_assert(
- llvm::APFloat::S_MaxSemantics < 16,
- "Too many Semantics enum values to fit in bitfield of size 4");
+ llvm::APFloat::S_MaxSemantics < 32,
+ "Too many Semantics enum values to fit in bitfield of size 5");
LLVM_PREFERRED_TYPE(llvm::APFloat::Semantics)
- unsigned Semantics : 4; // Provides semantics for APFloat construction
+ unsigned Semantics : 5; // Provides semantics for APFloat construction
LLVM_PREFERRED_TYPE(bool)
unsigned IsExact : 1;
};
diff --git a/clang/lib/AST/MicrosoftMangle.cpp b/clang/lib/AST/MicrosoftMangle.cpp
index 7f1e9ab02ec26..a2e270df276cb 100644
--- a/clang/lib/AST/MicrosoftMangle.cpp
+++ b/clang/lib/AST/MicrosoftMangle.cpp
@@ -946,6 +946,7 @@ void MicrosoftCXXNameMangler::mangleFloat(llvm::APFloat Number) {
case APFloat::S_IEEEquad: Out << 'Y'; break;
case APFloat::S_PPCDoubleDouble: Out << 'Z'; break;
case APFloat::S_Float8E5M2:
+ case APFloat::S_Float8E4M3:
case APFloat::S_Float8E4M3FN:
case APFloat::S_Float8E5M2FNUZ:
case APFloat::S_Float8E4M3FNUZ:
diff --git a/llvm/include/llvm/ADT/APFloat.h b/llvm/include/llvm/ADT/APFloat.h
index db2fa480655c6..bff8e6490d1de 100644
--- a/llvm/include/llvm/ADT/APFloat.h
+++ b/llvm/include/llvm/ADT/APFloat.h
@@ -166,6 +166,9 @@ struct APFloatBase {
// This format's exponent bias is 16, instead of the 15 (2 ** (5 - 1) - 1)
// that IEEE precedent would imply.
S_Float8E5M2FNUZ,
+ // 8-bit floating point number following IEEE-754 conventions with bit
+ // layout S1E4M3.
+ S_Float8E4M3,
// 8-bit floating point number mostly following IEEE-754 conventions with
// bit layout S1E4M3 as described in https://arxiv.org/abs/2209.05433.
// Unlike IEEE-754 types, there are no infinity values, and NaN is
@@ -217,6 +220,7 @@ struct APFloatBase {
static const fltSemantics &PPCDoubleDouble() LLVM_READNONE;
static const fltSemantics &Float8E5M2() LLVM_READNONE;
static const fltSemantics &Float8E5M2FNUZ() LLVM_READNONE;
+ static const fltSemantics &Float8E4M3() LLVM_READNONE;
static const fltSemantics &Float8E4M3FN() LLVM_READNONE;
static const fltSemantics &Float8E4M3FNUZ() LLVM_READNONE;
static const fltSemantics &Float8E4M3B11FNUZ() LLVM_READNONE;
@@ -638,6 +642,7 @@ class IEEEFloat final : public APFloatBase {
APInt convertPPCDoubleDoubleAPFloatToAPInt() const;
APInt convertFloat8E5M2APFloatToAPInt() const;
APInt convertFloat8E5M2FNUZAPFloatToAPInt() const;
+ APInt convertFloat8E4M3APFloatToAPInt() const;
APInt convertFloat8E4M3FNAPFloatToAPInt() const;
APInt convertFloat8E4M3FNUZAPFloatToAPInt() const;
APInt convertFloat8E4M3B11FNUZAPFloatToAPInt() const;
@@ -656,6 +661,7 @@ class IEEEFloat final : public APFloatBase {
void initFromPPCDoubleDoubleAPInt(const APInt &api);
void initFromFloat8E5M2APInt(const APInt &api);
void initFromFloat8E5M2FNUZAPInt(const APInt &api);
+ void initFromFloat8E4M3APInt(const APInt &api);
void initFromFloat8E4M3FNAPInt(const APInt &api);
void initFromFloat8E4M3FNUZAPInt(const APInt &api);
void initFromFloat8E4M3B11FNUZAPInt(const APInt &api);
diff --git a/llvm/lib/Support/APFloat.cpp b/llvm/lib/Support/APFloat.cpp
index 3017a9b976658..79e0094b243b2 100644
--- a/llvm/lib/Support/APFloat.cpp
+++ b/llvm/lib/Support/APFloat.cpp
@@ -136,6 +136,7 @@ static constexpr fltSemantics semIEEEquad = {16383, -16382, 113, 128};
static constexpr fltSemantics semFloat8E5M2 = {15, -14, 3, 8};
static constexpr fltSemantics semFloat8E5M2FNUZ = {
15, -15, 3, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::NegativeZero};
+static constexpr fltSemantics semFloat8E4M3 = {7, -6, 4, 8};
static constexpr fltSemantics semFloat8E4M3FN = {
8, -6, 4, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::AllOnes};
static constexpr fltSemantics semFloat8E4M3FNUZ = {
@@ -208,6 +209,8 @@ const llvm::fltSemantics &APFloatBase::EnumToSemantics(Semantics S) {
return Float8E5M2();
case S_Float8E5M2FNUZ:
return Float8E5M2FNUZ();
+ case S_Float8E4M3:
+ return Float8E4M3();
case S_Float8E4M3FN:
return Float8E4M3FN();
case S_Float8E4M3FNUZ:
@@ -246,6 +249,8 @@ APFloatBase::SemanticsToEnum(const llvm::fltSemantics &Sem) {
return S_Float8E5M2;
else if (&Sem == &llvm::APFloat::Float8E5M2FNUZ())
return S_Float8E5M2FNUZ;
+ else if (&Sem == &llvm::APFloat::Float8E4M3())
+ return S_Float8E4M3;
else if (&Sem == &llvm::APFloat::Float8E4M3FN())
return S_Float8E4M3FN;
else if (&Sem == &llvm::APFloat::Float8E4M3FNUZ())
@@ -276,6 +281,7 @@ const fltSemantics &APFloatBase::PPCDoubleDouble() {
}
const fltSemantics &APFloatBase::Float8E5M2() { return semFloat8E5M2; }
const fltSemantics &APFloatBase::Float8E5M2FNUZ() { return semFloat8E5M2FNUZ; }
+const fltSemantics &APFloatBase::Float8E4M3() { return semFloat8E4M3; }
const fltSemantics &APFloatBase::Float8E4M3FN() { return semFloat8E4M3FN; }
const fltSemantics &APFloatBase::Float8E4M3FNUZ() { return semFloat8E4M3FNUZ; }
const fltSemantics &APFloatBase::Float8E4M3B11FNUZ() {
@@ -3617,6 +3623,11 @@ APInt IEEEFloat::convertFloat8E5M2FNUZAPFloatToAPInt() const {
return convertIEEEFloatToAPInt<semFloat8E5M2FNUZ>();
}
+APInt IEEEFloat::convertFloat8E4M3APFloatToAPInt() const {
+ assert(partCount() == 1);
+ return convertIEEEFloatToAPInt<semFloat8E4M3>();
+}
+
APInt IEEEFloat::convertFloat8E4M3FNAPFloatToAPInt() const {
assert(partCount() == 1);
return convertIEEEFloatToAPInt<semFloat8E4M3FN>();
@@ -3681,6 +3692,9 @@ APInt IEEEFloat::bitcastToAPInt() const {
if (semantics == (const llvm::fltSemantics *)&semFloat8E5M2FNUZ)
return convertFloat8E5M2FNUZAPFloatToAPInt();
+ if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3)
+ return convertFloat8E4M3APFloatToAPInt();
+
if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3FN)
return convertFloat8E4M3FNAPFloatToAPInt();
@@ -3902,6 +3916,10 @@ void IEEEFloat::initFromFloat8E5M2FNUZAPInt(const APInt &api) {
initFromIEEEAPInt<semFloat8E5M2FNUZ>(api);
}
+void IEEEFloat::initFromFloat8E4M3APInt(const APInt &api) {
+ initFromIEEEAPInt<semFloat8E4M3>(api);
+}
+
void IEEEFloat::initFromFloat8E4M3FNAPInt(const APInt &api) {
initFromIEEEAPInt<semFloat8E4M3FN>(api);
}
@@ -3951,6 +3969,8 @@ void IEEEFloat::initFromAPInt(const fltSemantics *Sem, const APInt &api) {
return initFromFloat8E5M2APInt(api);
if (Sem == &semFloat8E5M2FNUZ)
return initFromFloat8E5M2FNUZAPInt(api);
+ if (Sem == &semFloat8E4M3)
+ return initFromFloat8E4M3APInt(api);
if (Sem == &semFloat8E4M3FN)
return initFromFloat8E4M3FNAPInt(api);
if (Sem == &semFloat8E4M3FNUZ)
diff --git a/llvm/unittests/ADT/APFloatTest.cpp b/llvm/unittests/ADT/APFloatTest.cpp
index 86a25f4394e19..132e95bd45f77 100644
--- a/llvm/unittests/ADT/APFloatTest.cpp
+++ b/llvm/unittests/ADT/APFloatTest.cpp
@@ -2133,6 +2133,8 @@ TEST(APFloatTest, getZero) {
{&APFloat::Float8E5M2(), true, true, {0x80ULL, 0}, 1},
{&APFloat::Float8E5M2FNUZ(), false, false, {0, 0}, 1},
{&APFloat::Float8E5M2FNUZ(), true, false, {0, 0}, 1},
+ {&APFloat::Float8E4M3(), false, true, {0, 0}, 1},
+ {&APFloat::Float8E4M3(), true, true, {0x80ULL, 0}, 1},
{&APFloat::Float8E4M3FN(), false, true, {0, 0}, 1},
{&APFloat::Float8E4M3FN(), true, true, {0x80ULL, 0}, 1},
{&APFloat::Float8E4M3FNUZ(), false, false, {0, 0}, 1},
@@ -6846,6 +6848,42 @@ TEST(APFloatTest, Float8E5M2ToFloat) {
EXPECT_TRUE(std::isnan(QNaN.convertToFloat()));
}
+TEST(APFloatTest, Float8E4M3ToFloat) {
+ APFloat PosZero = APFloat::getZero(APFloat::Float8E4M3());
+ APFloat PosZeroToFloat(PosZero.convertToFloat());
+ EXPECT_TRUE(PosZeroToFloat.isPosZero());
+ APFloat NegZero = APFloat::getZero(APFloat::Float8E4M3(), true);
+ APFloat NegZeroToFloat(NegZero.convertToFloat());
+ EXPECT_TRUE(NegZeroToFloat.isNegZero());
+
+ APFloat One(APFloat::Float8E4M3(), "1.0");
+ EXPECT_EQ(1.0F, One.convertToFloat());
+ APFloat Two(APFloat::Float8E4M3(), "2.0");
+ EXPECT_EQ(2.0F, Two.convertToFloat());
+
+ APFloat PosLargest = APFloat::getLargest(APFloat::Float8E4M3(), false);
+ EXPECT_EQ(240.0F, PosLargest.convertToFloat());
+ APFloat NegLargest = APFloat::getLargest(APFloat::Float8E4M3(), true);
+ EXPECT_EQ(-240.0F, NegLargest.convertToFloat());
+ APFloat PosSmallest =
+ APFloat::getSmallestNormalized(APFloat::Float8E4M3(), false);
+ EXPECT_EQ(0x1.p-6, PosSmallest.convertToFloat());
+ APFloat NegSmallest =
+ APFloat::getSmallestNormalized(APFloat::Float8E4M3(), true);
+ EXPECT_EQ(-0x1.p-6, NegSmallest.convertToFloat());
+
+ APFloat SmallestDenorm = APFloat::getSmallest(APFloat::Float8E4M3(), false);
+ EXPECT_TRUE(SmallestDenorm.isDenormal());
+ EXPECT_EQ(0x1.p-9, SmallestDenorm.convertToFloat());
+
+ APFloat PosInf = APFloat::getInf(APFloat::Float8E4M3());
+ EXPECT_EQ(std::numeric_limits<float>::infinity(), PosInf.convertToFloat());
+ APFloat NegInf = APFloat::getInf(APFloat::Float8E4M3(), true);
+ EXPECT_EQ(-std::numeric_limits<float>::infinity(), NegInf.convertToFloat());
+ APFloat QNaN = APFloat::getQNaN(APFloat::Float8E4M3());
+ EXPECT_TRUE(std::isnan(QNaN.convertToFloat()));
+}
+
TEST(APFloatTest, Float8E4M3FNToFloat) {
APFloat PosZero = APFloat::getZero(APFloat::Float8E4M3FN());
APFloat PosZeroToFloat(PosZero.convertToFloat());
diff --git a/mlir/include/mlir-c/BuiltinTypes.h b/mlir/include/mlir-c/BuiltinTypes.h
index 99c5e3f46b04c..2212087b9898f 100644
--- a/mlir/include/mlir-c/BuiltinTypes.h
+++ b/mlir/include/mlir-c/BuiltinTypes.h
@@ -89,6 +89,16 @@ MLIR_CAPI_EXPORTED bool mlirTypeIsAFloat8E5M2(MlirType type);
/// context.
MLIR_CAPI_EXPORTED MlirType mlirFloat8E5M2TypeGet(MlirContext ctx);
+/// Returns the typeID of an Float8E4M3 type.
+MLIR_CAPI_EXPORTED MlirTypeID mlirFloat8E4M3TypeGetTypeID(void);
+
+/// Checks whether the given type is an f8E4M3 type.
+MLIR_CAPI_EXPORTED bool mlirTypeIsAFloat8E4M3(MlirType type);
+
+/// Creates an f8E4M3 type in the given context. The type is owned by the
+/// context.
+MLIR_CAPI_EXPORTED MlirType mlirFloat8E4M3TypeGet(MlirContext ctx);
+
/// Returns the typeID of an Float8E4M3FN type.
MLIR_CAPI_EXPORTED MlirTypeID mlirFloat8E4M3FNTypeGetTypeID(void);
diff --git a/mlir/include/mlir/IR/Builders.h b/mlir/include/mlir/IR/Builders.h
index 0d5fa719d0dee..1c4d329fbf0d8 100644
--- a/mlir/include/mlir/IR/Builders.h
+++ b/mlir/include/mlir/IR/Builders.h
@@ -61,6 +61,7 @@ class Builder {
// Types.
FloatType getFloat8E5M2Type();
+ FloatType getFloat8E4M3Type();
FloatType getFloat8E4M3FNType();
FloatType getFloat8E5M2FNUZType();
FloatType getFloat8E4M3FNUZType();
diff --git a/mlir/include/mlir/IR/BuiltinTypes.h b/mlir/include/mlir/IR/BuiltinTypes.h
index 5579b138668d2..f4c05cafa1fd9 100644
--- a/mlir/include/mlir/IR/BuiltinTypes.h
+++ b/mlir/include/mlir/IR/BuiltinTypes.h
@@ -56,6 +56,7 @@ class FloatType : public Type {
static FloatType getF80(MLIRContext *ctx);
static FloatType getF128(MLIRContext *ctx);
static FloatType getFloat8E5M2(MLIRContext *ctx);
+ static FloatType getFloat8E4M3(MLIRContext *ctx);
static FloatType getFloat8E4M3FN(MLIRContext *ctx);
static FloatType getFloat8E5M2FNUZ(MLIRContext *ctx);
static FloatType getFloat8E4M3FNUZ(MLIRContext *ctx);
@@ -405,16 +406,20 @@ inline bool BaseMemRefType::isValidElementType(Type type) {
}
inline bool FloatType::classof(Type type) {
- return llvm::isa<Float8E5M2Type, Float8E4M3FNType, Float8E5M2FNUZType,
- Float8E4M3FNUZType, Float8E4M3B11FNUZType, BFloat16Type,
- Float16Type, FloatTF32Type, Float32Type, Float64Type,
- Float80Type, Float128Type>(type);
+ return llvm::isa<
+ Float8E5M2Type, Float8E4M3Type, Float8E4M3FNType, Float8E5M2FNUZType,
+ Float8E4M3FNUZType, Float8E4M3B11FNUZType, BFloat16Type, Float16Type,
+ FloatTF32Type, Float32Type, Float64Type, Float80Type, Float128Type>(type);
}
inline FloatType FloatType::getFloat8E5M2(MLIRContext *ctx) {
return Float8E5M2Type::get(ctx);
}
+inline FloatType FloatType::getFloat8E4M3(MLIRContext *ctx) {
+ return Float8E4M3Type::get(ctx);
+}
+
inline FloatType FloatType::getFloat8E4M3FN(MLIRContext *ctx) {
return Float8E4M3FNType::get(ctx);
}
diff --git a/mlir/include/mlir/IR/BuiltinTypes.td b/mlir/include/mlir/IR/BuiltinTypes.td
index 4cade83dd3c32..2eed105e81f05 100644
--- a/mlir/include/mlir/IR/BuiltinTypes.td
+++ b/mlir/include/mlir/IR/BuiltinTypes.td
@@ -97,6 +97,25 @@ def Builtin_Float8E5M2 : Builtin_FloatType<"Float8E5M2", "f8E5M2"> {
}];
}
+//===----------------------------------------------------------------------===//
+// Float8E4M3Type
+
+def Builtin_Float8E4M3 : Builtin_FloatType<"Float8E4M3", "f8E4M3"> {
+ let summary = "8-bit floating point with 3 bit mantissa";
+ let description = [{
+ An 8-bit floating point type with 1 sign bit, 4 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: S1E4M3
+ * exponent bias: 7
+ * infinities: supported with exponent set to all 1s and mantissa 0s
+ * NaNs: supported with exponent bits set to all 1s and mantissa of
+ (001, 010, 011, 100, 101, 110, 111)
+ * denormals when exponent is 0
+ }];
+}
+
//===----------------------------------------------------------------------===//
// Float8E4M3FNType
diff --git a/mlir/include/mlir/IR/CommonTypeConstraints.td b/mlir/include/mlir/IR/CommonTypeConstraints.td
index af4f13dc09360..cac37c8fec4d3 100644
--- a/mlir/include/mlir/IR/CommonTypeConstraints.td
+++ b/mlir/include/mlir/IR/CommonTypeConstraints.td
@@ -331,6 +331,8 @@ def F8E4M3FN : Type<CPred<"$_self.isFloat8E4M3FN()">, "f8E4M3FN type">,
BuildableType<"$_builder.getFloat8E4M3FNType()">;
def F8E5M2 : Type<CPred<"$_self.isFloat8E5M2()">, "f8E5M2 type">,
BuildableType<"$_builder.getFloat8E5M2Type()">;
+def F8E4M3 : Type<CPred<"$_self.isFloat8E4M3()">, "f8E4M3 type">,
+ BuildableType<"$_builder.getFloat8E4M3Type()">;
def F8E4M3FNUZ : Type<CPred<"$_self.isFloat8E4M3FNUZ()">, "f8E4M3FNUZ type">,
BuildableType<"$_builder.getFloat8E4M3FNUZType()">;
def F8E4M3B11FNUZ : Type<CPred<"$_self.isFloat8E4M3B11FNUZ()">, "f8E4M3B11FNUZ type">,
diff --git a/mlir/include/mlir/IR/Types.h b/mlir/include/mlir/IR/Types.h
index 65824531fdc90..a32de33114e40 100644
--- a/mlir/include/mlir/IR/Types.h
+++ b/mlir/include/mlir/IR/Types.h
@@ -126,6 +126,7 @@ class Type {
// derived types should use isa/dyn_cast.
bool isIndex() const;
bool isFloat8E5M2() const;
+ bool isFloat8E4M3() const;
bool isFloat8E4M3FN() const;
bool isFloat8E5M2FNUZ() const;
bool isFloat8E4M3FNUZ() const;
diff --git a/mlir/lib/AsmParser/TokenKinds.def b/mlir/lib/AsmParser/TokenKinds.def
index 297e074594530..eb3154c6da42e 100644
--- a/mlir/lib/AsmParser/TokenKinds.def
+++ b/mlir/lib/AsmParser/TokenKinds.def
@@ -95,6 +95,7 @@ TOK_KEYWORD(f32)
TOK_KEYWORD(f64)
TOK_KEYWORD(f80)
TOK_KEYWORD(f8E5M2)
+TOK_KEYWORD(f8E4M3)
TOK_KEYWORD(f8E4M3FN)
TOK_KEYWORD(f8E5M2FNUZ)
TOK_KEYWORD(f8E4M3FNUZ)
diff --git a/mlir/lib/AsmParser/TypeParser.cpp b/mlir/lib/AsmParser/TypeParser.cpp
index 5da931b77b3be..90ba56af670f6 100644
--- a/mlir/lib/AsmParser/TypeParser.cpp
+++ b/mlir/lib/AsmParser/TypeParser.cpp
@@ -41,6 +41,7 @@ OptionalParseResult Parser::parseOptionalType(Type &type) {
case Token::kw_vector:
case Token::inttype:
case Token::kw_f8E5M2:
+ case Token::kw_f8E4M3:
case Token::kw_f8E4M3FN:
case Token::kw_f8E5M2FNUZ:
case Token::kw_f8E4M3FNUZ:
@@ -305,6 +306,9 @@ Type Parser::parseNonFunctionType() {
case Token::kw_f8E5M2:
consumeToken(Token::kw_f8E5M2);
return builder.getFloat8E5M2Type();
+ case Token::kw_f8E4M3:
+ consumeToken(Token::kw_f8E4M3);
+ return builder.getFloat8E4M3Type();
case Token::kw_f8E4M3FN:
consumeToken(Token::kw_f8E4M3FN);
return builder.getFloat8E4M3FNType();
diff --git a/mlir/lib/Bindings/Python/IRTypes.cpp b/mlir/lib/Bindings/Python/IRTypes.cpp
index e1e4eb999b3aa..5e0aebc03e2c1 100644
--- a/mlir/lib/Bindings/Python/IRTypes.cpp
+++ b/mlir/lib/Bindings/Python/IRTypes.cpp
@@ -143,7 +143,7 @@ class PyFloat8E4M3FNType
}
};
-/// Floating Point Type subclass - Float8M5E2Type.
+/// Floating Point Type subclass - Float8E5M2Type.
class PyFloat8E5M2Type : public PyConcreteType<PyFloat8E5M2Type, PyFloatType> {
public:
static constexpr IsAFunctionTy isaFunction = mlirTypeIsAFloat8E5M2;
@@ -163,6 +163,26 @@ class PyFloat8E5M2Type : public PyConcreteType<PyFloat8E5M2Type, PyFloatType> {
}
};
+/// Floating Point Type subclass - Float8E4M3Type.
+class PyFloat8E4M3Type : public PyConcreteType<PyFloat8E4M3Type, PyFloatType> {
+public:
+ static constexpr IsAFunctionTy isaFunction = mlirTypeIsAFloat8E4M3;
+ static constexpr GetTypeIDFunctionTy getTypeIdFunction =
+ mlirFloat8E4M3TypeGetTypeID;
+ static constexpr const char *pyClassName = "Float8E4M3Type";
+ using PyConcreteType::PyConcreteType;
+
+ static void bindDerived(ClassTy &c) {
+ c.def_static(
+ "get",
+ [](DefaultingPyMlirContext context) {
+ MlirType t = mlirFloat8E4M3TypeGet(context->get());
+ return PyFloat8E4M3Type(context->getRef(), t);
+ },
+ py::arg("context") = py::none(), "Create a float8_e4m3 type.");
+ }
+};
+
/// Floating Point Type subclass - Float8E4M3FNUZ.
class PyFloat8E4M3FNUZType
: public PyConcreteType<PyFloat8E4M3FNUZType, PyFloatType> {
@@ -840,6 +860,7 @@ void mlir::python::populateIRTypes(py::module &m) {
PyIndexType::bind(m);
PyFloat8E4M3FNType::bind(m);
PyFloat8E5M2Type::bind(m);
+ PyFloat8E4M3Type::bind(m);
PyFloat8E4M3FNUZType::bind(m);
PyFloat8E4M3B11FNUZType::bind(m);
PyFloat8E5M2FNUZType::bind(m);
diff --git a/mlir/lib/CAPI/IR/BuiltinTypes.cpp b/mlir/lib/CAPI/IR/BuiltinTypes.cpp
index c94c070144a7e..6e22caca78ff1 100644
--- a/mlir/lib/CAPI/IR/BuiltinTypes.cpp
+++ b/mlir/lib/CAPI/IR/BuiltinTypes.cpp
@@ -98,6 +98,18 @@ MlirType mlirFloat8E5M2TypeGet(MlirContext ctx) {
return wrap(FloatType::getFloat8E5M2(unwrap(ctx)));
}
+MlirTypeID mlirFloat8E4M3TypeGetTypeID() {
+ return wrap(Float8E4M3Type::getTypeID());
+}
+
+bool mlirTypeIsAFloat8E4M3(MlirType type) {
+ return unwrap(type).isFloat8E4M3();
+}
+
+MlirType mlirFloat8E4M3TypeGet(MlirContext ctx) {
+ return wrap(FloatType::getFloat8E4M3(unwrap(ctx)));
+}
+
MlirTypeID mlirFloat8E4M3FNTypeGetTypeID() {
return wrap(Float8E4M3FNType::getTypeID());
}
diff --git a/mlir/lib/Conversion/LLVMCommon/TypeConverter.cpp b/mlir/lib/Conversion/LLVMCommon/TypeConverter.cpp
index f5620a6a7cd91..0861e3b192d91 100644
--- a/mlir/lib/Conversion/LLVMCommon/TypeConverter.cpp
+++ b/mlir/lib/Conversion/LLVMCommon/TypeConverter.cpp
@@ -233,8 +233,9 @@ Type LLVMTypeConverter::convertIntegerType(IntegerType type) const {
}
Type LLVMTypeConverter::convertFloatType(FloatType type) const {
- if (type.isFloat8E5M2() || type.isFloat8E4M3FN() || type.isFloat8E5M2FNUZ() ||
- type.isFloat8E4M3FNUZ() || type.isFloat8E4M3B11FNUZ())
+ if (type.isFloat8E5M2() || type.isFloat8E4M3() || type.isFloat8E4M3FN() ||
+ type.isFloat8E5M2FNUZ() || type.isFloat8E4M3FNUZ() ||
+ type.isFloat8E4M3B11FNUZ())
return IntegerType::get(&getContext(), type.getWidth());
return type;
}
diff --git a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
index 8e1cb474feee7..a362c8500aa5b 100644
--- a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
+++ b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
@@ -56,6 +56,7 @@ static std::optional<FloatType> parseFloatType(MLIRContext *ctx,
Builder b(ctx);
return llvm::StringSwitch<std::optional<FloatType>>(name)
.Case("f8E5M2", b.getFloat8E5M2Type())
+ .Case("f8E4M3", b.getFloat8E4M3Type())
.Case("f8E4M3FN", b.getFloat8E4M3FNType())
.Case("f8E5M2FNUZ", b.getFloat8E5M2FNUZType())
.Case("f8E4M3FNUZ", b.getFloat8E4M3FNUZType())
diff --git a/mlir/lib/IR/AsmPrinter.cpp b/mlir/lib/IR/AsmPrinter.cpp
index 13eb18036eeec..e5b1291afce2b 100644
--- a/mlir/lib/IR/AsmPrinter.cpp
+++ b/mlir/lib/IR/AsmPrinter.cpp
@@ -2576,6 +2576,7 @@ void AsmPrinter::Impl::printTypeImpl(Type type) {
})
.Case<IndexType>([&](Type) { os << "index"; })
.Case<Float8E5M2Type>([&](Type) { os << "f8E5M2"; })
+ .Case<Float8E4M3Type>([&](Type) { os << "f8E4M3"; })
.Case<Float8E4M3FNType>([&](Type) { os << "f8E4M3FN"; })
.Case<Float8E5M2FNUZType>([&](Type) { os << "f8E5M2FNUZ"; })
.Case<Float8E4M3FNUZType>([&](Type) { os << "f8E4M3FNUZ"; })
diff --git a/mlir/lib/IR/Builders.cpp b/mlir/lib/IR/Builders.cpp
index d49f69a7b7ae6..d0eb2d8fbae9d 100644
--- a/mlir/lib/IR/Builders.cpp
+++ b/mlir/lib/IR/Builders.cpp
@@ -38,6 +38,10 @@ FloatType Builder::getFloat8E5M2Type() {
return FloatType::getFloat8E5M2(context);
}
+FloatType Builder::getFloat8E4M3Type() {
+ return FloatType::getFloat8E4M3(context);
+}
+
FloatType Builder::getFloat8E4M3FNType() {
return FloatType::getFloat8E4M3FN(context);
}
diff --git a/mlir/lib/IR/BuiltinTypes.cpp b/mlir/lib/IR/BuiltinTypes.cpp
index 179797cb943a1..faa944937e007 100644
--- a/mlir/lib/IR/BuiltinTypes.cpp
+++ b/mlir/lib/IR/BuiltinTypes.cpp
@@ -87,8 +87,9 @@ IntegerType IntegerType::scaleElementBitwidth(unsigned scale) {
//===----------------------------------------------------------------------===//
unsigned FloatType::getWidth() {
- if (llvm::isa<Float8E5M2Type, Float8E4M3FNType, Float8E5M2FNUZType,
- Float8E4M3FNUZType, Float8E4M3B11FNUZType>(*this))
+ if (llvm::isa<Float8E5M2Type, Float8E4M3Type, Float8E4M3FNType,
+ Float8E5M2FNUZType, Float8E4M3FNUZType, Float8E4M3B11FNUZType>(
+ *this))
return 8;
if (llvm::isa<Float16Type, BFloat16Type>(*this))
return 16;
@@ -107,6 +108,8 @@ unsigned FloatType::getWidth() {
const llvm::fltSemantics &FloatType::getFloatSemantics() {
if (llvm::isa<Float8E5M2Type>(*this))
return APFloat::Float8E5M2();
+ if (llvm::isa<Float8E4M3Type>(*this))
+ return APFloat::Float8E4M3();
if (llvm::isa<Float8E4M3FNType>(*this))
return APFloat::Float8E4M3FN();
if (llvm::isa<Float8E5M2FNUZType>(*this))
diff --git a/mlir/lib/IR/MLIRContext.cpp b/mlir/lib/IR/MLIRContext.cpp
index 214b354c5347e..12336701c9ca0 100644
--- a/mlir/lib/IR/MLIRContext.cpp
+++ b/mlir/lib/IR/MLIRContext.cpp
@@ -222,6 +222,7 @@ class MLIRContextImpl {
/// Cached Type Instances.
Float8E5M2Type f8E5M2Ty;
+ Float8E4M3Type f8E4M3Ty;
Float8E4M3FNType f8E4M3FNTy;
Float8E5M2FNUZType f8E5M2FNUZTy;
Float8E4M3FNUZType f8E4M3FNUZTy;
@@ -312,6 +313,7 @@ MLIRContext::MLIRContext(const DialectRegistry ®istry, Threading setting)
//// Types.
/// Floating-point Types.
impl->f8E5M2Ty = TypeUniquer::get<Float8E5M2Type>(this);
+ impl->f8E4M3Ty = TypeUniquer::get<Float8E4M3Type>(this);
impl->f8E4M3FNTy = TypeUniquer::get<Float8E4M3FNType>(this);
impl->f8E5M2FNUZTy = TypeUniquer::get<Float8E5M2FNUZType>(this);
impl->f8E4M3FNUZTy = TypeUniquer::get<Float8E4M3FNUZType>(this);
@@ -1012,6 +1014,9 @@ StorageUniquer &MLIRContext::getTypeUniquer() { return getImpl().typeUniquer; }
Float8E5M2Type Float8E5M2Type::get(MLIRContext *context) {
return context->getImpl().f8E5M2Ty;
}
+Float8E4M3Type Float8E4M3Type::get(MLIRContext *context) {
+ return context->getImpl().f8E4M3Ty;
+}
Float8E4M3FNType Float8E4M3FNType::get(MLIRContext *context) {
return context->getImpl().f8E4M3FNTy;
}
diff --git a/mlir/lib/IR/Types.cpp b/mlir/lib/IR/Types.cpp
index 1d1ba6df4db2f..e8cd28bf9e85d 100644
--- a/mlir/lib/IR/Types.cpp
+++ b/mlir/lib/IR/Types.cpp
@@ -35,6 +35,7 @@ Type AbstractType::replaceImmediateSubElements(Type type,
MLIRContext *Type::getContext() const { return getDialect().getContext(); }
bool Type::isFloat8E5M2() const { return llvm::isa<Float8E5M2Type>(*this); }
+bool Type::isFloat8E4M3() const { return llvm::isa<Float8E4M3Type>(*this); }
bool Type::isFloat8E4M3FN() const { return llvm::isa<Float8E4M3FNType>(*this); }
bool Type::isFloat8E5M2FNUZ() const {
return llvm::isa<Float8E5M2FNUZType>(*this);
diff --git a/mlir/python/mlir/_mlir_libs/_mlir/ir.pyi b/mlir/python/mlir/_mlir_libs/_mlir/ir.pyi
index 317e688076304..224e77a3f46be 100644
--- a/mlir/python/mlir/_mlir_libs/_mlir/ir.pyi
+++ b/mlir/python/mlir/_mlir_libs/_mlir/ir.pyi
@@ -123,6 +123,7 @@ __all__ = [
"Float8E4M3B11FNUZType",
"Float8E4M3FNType",
"Float8E4M3FNUZType",
+ "Float8E4M3Type",
"Float8E5M2FNUZType",
"Float8E5M2Type",
"FloatAttr",
@@ -1575,6 +1576,19 @@ class Float8E4M3FNUZType(FloatType):
@property
def typeid(self) -> TypeID: ...
+class Float8E4M3Type(FloatType):
+ static_typeid: ClassVar[TypeID]
+ @staticmethod
+ def get(context: Optional[Context] = None) -> Float8E4M3Type:
+ """
+ Create a float8_e4m3 type.
+ """
+ @staticmethod
+ def isinstance(other: Type) -> bool: ...
+ def __init__(self, cast_from_type: Type) -> None: ...
+ @property
+ def typeid(self) -> TypeID: ...
+
class Float8E5M2FNUZType(FloatType):
static_typeid: ClassVar[TypeID]
@staticmethod
diff --git a/mlir/python/mlir/extras/types.py b/mlir/python/mlir/extras/types.py
index b93c46b172a9a..fde9909a8f9d6 100644
--- a/mlir/python/mlir/extras/types.py
+++ b/mlir/python/mlir/extras/types.py
@@ -14,6 +14,7 @@
F64Type,
Float8E4M3B11FNUZType,
Float8E4M3FNType,
+ Float8E4M3Type,
Float8E5M2Type,
FunctionType,
IndexType,
@@ -68,6 +69,7 @@ def ui(width):
bf16 = lambda: BF16Type.get()
f8E5M2 = lambda: Float8E5M2Type.get()
+f8E4M3 = lambda: Float8E4M3Type.get()
f8E4M3FN = lambda: Float8E4M3FNType.get()
f8E4M3B11FNUZ = lambda: Float8E4M3B11FNUZType.get()
diff --git a/mlir/test/IR/attribute.mlir b/mlir/test/IR/attribute.mlir
index 291d5832fce79..362e98134ee4a 100644
--- a/mlir/test/IR/attribute.mlir
+++ b/mlir/test/IR/attribute.mlir
@@ -40,6 +40,10 @@ func.func @float_attrs_pass() {
// CHECK: float_attr = 2.000000e+00 : f8E5M2
float_attr = 2. : f8E5M2
} : () -> ()
+ "test.float_attrs"() {
+ // CHECK: float_attr = 2.000000e+00 : f8E4M3
+ float_attr = 2. : f8E4M3
+ } : () -> ()
"test.float_attrs"() {
// CHECK: float_attr = 2.000000e+00 : f8E4M3FN
float_attr = 2. : f8E4M3FN
diff --git a/mlir/test/Target/LLVMIR/llvmir.mlir b/mlir/test/Target/LLVMIR/llvmir.mlir
index 7116f3b64d7f5..61e849c813756 100644
--- a/mlir/test/Target/LLVMIR/llvmir.mlir
+++ b/mlir/test/Target/LLVMIR/llvmir.mlir
@@ -39,6 +39,9 @@ llvm.mlir.global internal constant @string_const("foobar") : !llvm.array<6 x i8>
// CHECK: @int_global_undef = internal global i64 undef
llvm.mlir.global internal @int_global_undef() : i64
+// CHECK: @f8E4M3_global_as_i8 = internal global i8 60
+llvm.mlir.global internal @f8E4M3_global_as_i8(1.5 : f8E4M3) : i8
+
// CHECK: @f8E4M3FN_global_as_i8 = internal global i8 60
llvm.mlir.global internal @f8E4M3FN_global_as_i8(1.5 : f8E4M3FN) : i8
diff --git a/mlir/test/python/ir/builtin_types.py b/mlir/test/python/ir/builtin_types.py
index cfe377c703717..3178f58cf2e74 100644
--- a/mlir/test/python/ir/builtin_types.py
+++ b/mlir/test/python/ir/builtin_types.py
@@ -113,6 +113,8 @@ def testTypeIsInstance():
def testFloatTypeSubclasses():
ctx = Context()
# CHECK: True
+ print(isinstance(Type.parse("f8E4M3", ctx), FloatType))
+ # CHECK: True
print(isinstance(Type.parse("f8E4M3FN", ctx), FloatType))
# CHECK: True
print(isinstance(Type.parse("f8E5M2", ctx), FloatType))
@@ -229,6 +231,8 @@ def testIndexType():
@run
def testFloatType():
with Context():
+ # CHECK: float: f8E4M3
+ print("float:", Float8E4M3Type.get())
# CHECK: float: f8E4M3FN
print("float:", Float8E4M3FNType.get())
# CHECK: float: f8E5M2
@@ -601,6 +605,7 @@ def testTypeIDs():
types = [
(IntegerType, IntegerType.get_signless(16)),
(IndexType, IndexType.get()),
+ (Float8E4M3Type, Float8E4M3Type.get()),
(Float8E4M3FNType, Float8E4M3FNType.get()),
(Float8E5M2Type, Float8E5M2Type.get()),
(Float8E4M3FNUZType, Float8E4M3FNUZType.get()),
@@ -624,6 +629,7 @@ def testTypeIDs():
# CHECK: IntegerType(i16)
# CHECK: IndexType(index)
+ # CHECK: Float8E4M3Type(f8E4M3)
# CHECK: Float8E4M3FNType(f8E4M3FN)
# CHECK: Float8E5M2Type(f8E5M2)
# CHECK: Float8E4M3FNUZType(f8E4M3FNUZ)
@@ -704,6 +710,9 @@ def print_downcasted(typ):
# CHECK: Float8E4M3B11FNUZType
# CHECK: Float8E4M3B11FNUZType(f8E4M3B11FNUZ)
print_downcasted(Float8E4M3B11FNUZType.get())
+ # CHECK: Float8E4M3Type
+ # CHECK: Float8E4M3Type(f8E4M3)
+ print_downcasted(Float8E4M3Type.get())
# CHECK: Float8E4M3FNType
# CHECK: Float8E4M3FNType(f8E4M3FN)
print_downcasted(Float8E4M3FNType.get())
diff --git a/mlir/utils/lldb-scripts/mlirDataFormatters.py b/mlir/utils/lldb-scripts/mlirDataFormatters.py
index 3282b4a32d503..ed0ee431fd7d8 100644
--- a/mlir/utils/lldb-scripts/mlirDataFormatters.py
+++ b/mlir/utils/lldb-scripts/mlirDataFormatters.py
@@ -51,6 +51,7 @@ def build_ptr_str_from_addr(addrValue: lldb.SBValue, type: lldb.SBType):
"mlir::FusedLoc": '"loc(fused<...>[...])"',
"mlir::UnknownLoc": '"loc(unknown)"',
"mlir::Float8E5M2Type": '"f8E5M2"',
+ "mlir::Float8E4M3Type": '"f8E4M3"',
"mlir::Float8E4M3FNType": '"f8E4M3FN"',
"mlir::Float8E5M2FNUZType": '"f8E5M2FNUZ"',
"mlir::Float8E4M3FNUZType": '"f8E4M3FNUZ"',
diff --git a/mlir/utils/tree-sitter-mlir/grammar.js b/mlir/utils/tree-sitter-mlir/grammar.js
index 0e7075b124b55..a657874f894b7 100644
--- a/mlir/utils/tree-sitter-mlir/grammar.js
+++ b/mlir/utils/tree-sitter-mlir/grammar.js
@@ -230,7 +230,8 @@ const common = {
integer_type : $ =>
token(seq(choice('si', 'ui', 'i'), /[1-9]/, repeat(/[0-9]/))),
float_type : $ => token(
- choice('f16', 'f32', 'f64', 'f80', 'f128', 'bf16', 'f8E4M3FN', 'f8E5M2')),
+ choice('f16', 'f32', 'f64', 'f80', 'f128', 'bf16', 'f8E4M3FN', 'f8E4M3',
+ 'f8E5M2')),
index_type : $ => token('index'),
none_type : $ => token('none'),
complex_type : $ => seq(token('complex'), '<', $._prim_type, '>'),
More information about the Mlir-commits
mailing list