Skip to content

Commit

Permalink
[Clang][RISCV] Introduce tuple types for RVV bfloat16 (#72216)
Browse files Browse the repository at this point in the history 
The first commit extends the capacity from the compiler infrastructure,
and the second commit continues the effort in #71140 to introduce tuple
types for bfloat16.
  • Loading branch information
@eopXD
eopXD committed Nov 15, 2023
1 parent 3f743fd commit 8434b0b
Show file tree
Hide file tree
Showing 6 changed files with 121 additions and 4 deletions.
56 changes: 56 additions & 0 deletions clang/include/clang/Basic/RISCVVTypes.def
View file
Original file line number Diff line number Diff line change
Expand Up @@ -452,6 +452,62 @@ RVV_VECTOR_TYPE_FLOAT("__rvv_float64m2x4_t", RvvFloat64m2x4, RvvFloat64m2x4Ty, 2

RVV_VECTOR_TYPE_FLOAT("__rvv_float64m4x2_t", RvvFloat64m4x2, RvvFloat64m4x2Ty, 4, 64, 2)

//===- BFloat16 tuple types -------------------------------------------------===//
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x2_t", RvvBFloat16mf4x2, RvvBFloat16mf4x2Ty,
1, 16, 2)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x3_t", RvvBFloat16mf4x3, RvvBFloat16mf4x3Ty,
1, 16, 3)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x4_t", RvvBFloat16mf4x4, RvvBFloat16mf4x4Ty,
1, 16, 4)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x5_t", RvvBFloat16mf4x5, RvvBFloat16mf4x5Ty,
1, 16, 5)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x6_t", RvvBFloat16mf4x6, RvvBFloat16mf4x6Ty,
1, 16, 6)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x7_t", RvvBFloat16mf4x7, RvvBFloat16mf4x7Ty,
1, 16, 7)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf4x8_t", RvvBFloat16mf4x8, RvvBFloat16mf4x8Ty,
1, 16, 8)

RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x2_t", RvvBFloat16mf2x2, RvvBFloat16mf2x2Ty,
2, 16, 2)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x3_t", RvvBFloat16mf2x3, RvvBFloat16mf2x3Ty,
2, 16, 3)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x4_t", RvvBFloat16mf2x4, RvvBFloat16mf2x4Ty,
2, 16, 4)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x5_t", RvvBFloat16mf2x5, RvvBFloat16mf2x5Ty,
2, 16, 5)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x6_t", RvvBFloat16mf2x6, RvvBFloat16mf2x6Ty,
2, 16, 6)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x7_t", RvvBFloat16mf2x7, RvvBFloat16mf2x7Ty,
2, 16, 7)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16mf2x8_t", RvvBFloat16mf2x8, RvvBFloat16mf2x8Ty,
2, 16, 8)

RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x2_t", RvvBFloat16m1x2, RvvBFloat16m1x2Ty,
4, 16, 2)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x3_t", RvvBFloat16m1x3, RvvBFloat16m1x3Ty,
4, 16, 3)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x4_t", RvvBFloat16m1x4, RvvBFloat16m1x4Ty,
4, 16, 4)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x5_t", RvvBFloat16m1x5, RvvBFloat16m1x5Ty,
4, 16, 5)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x6_t", RvvBFloat16m1x6, RvvBFloat16m1x6Ty,
4, 16, 6)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x7_t", RvvBFloat16m1x7, RvvBFloat16m1x7Ty,
4, 16, 7)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m1x8_t", RvvBFloat16m1x8, RvvBFloat16m1x8Ty,
4, 16, 8)

RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m2x2_t", RvvBFloat16m2x2, RvvBFloat16m2x2Ty,
8, 16, 2)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m2x3_t", RvvBFloat16m2x3, RvvBFloat16m2x3Ty,
8, 16, 3)
RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m2x4_t", RvvBFloat16m2x4, RvvBFloat16m2x4Ty,
8, 16, 4)

RVV_VECTOR_TYPE_BFLOAT("__rvv_bfloat16m4x2_t", RvvBFloat16m4x2, RvvBFloat16m4x2Ty,
16, 16, 2)

#undef RVV_VECTOR_TYPE_BFLOAT
#undef RVV_VECTOR_TYPE_FLOAT
#undef RVV_VECTOR_TYPE_INT
Expand Down
2 changes: 1 addition & 1 deletion clang/include/clang/Serialization/ASTBitCodes.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1101,7 +1101,7 @@ enum PredefinedTypeIDs {
///
/// Type IDs for non-predefined types will start at
/// NUM_PREDEF_TYPE_IDs.
const unsigned NUM_PREDEF_TYPE_IDS = 500;
const unsigned NUM_PREDEF_TYPE_IDS = 600;

// Ensure we do not overrun the predefined types we reserved
// in the enum PredefinedTypeIDs above.
Expand Down
5 changes: 3 additions & 2 deletions clang/include/clang/Support/RISCVVIntrinsicUtils.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -97,13 +97,14 @@ enum class TypeModifier : uint8_t {
UnsignedInteger = 1 << 3,
SignedInteger = 1 << 4,
Float = 1 << 5,
BFloat = 1 << 6,
// LMUL1 should be kind of VectorTypeModifier, but that might come with
// Widening2XVector for widening reduction.
// However that might require VectorTypeModifier become bitmask rather than
// simple enum, so we decide keek LMUL1 in TypeModifier for code size
// optimization of clang binary size.
LMUL1 = 1 << 6,
MaxOffset = 6,
LMUL1 = 1 << 7,
MaxOffset = 7,
LLVM_MARK_AS_BITMASK_ENUM(LMUL1),
};

Expand Down
3 changes: 3 additions & 0 deletions clang/lib/Support/RISCVVIntrinsicUtils.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -857,6 +857,9 @@ void RVVType::applyModifier(const PrototypeDescriptor &Transformer) {
case TypeModifier::Float:
ScalarType = ScalarTypeKind::Float;
break;
case TypeModifier::BFloat:
ScalarType = ScalarTypeKind::BFloat;
break;
case TypeModifier::LMUL1:
LMUL = LMULType(0);
// Update ElementBitwidth need to update Scale too.
Expand Down
55 changes: 55 additions & 0 deletions clang/test/CodeGen/RISCV/rvv-intrinsics-handcrafted/rvv-intrinsic-datatypes.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -305,6 +305,31 @@
// CHECK-NEXT: [[F64M2X3:%.*]] = alloca { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }, align 8
// CHECK-NEXT: [[F64M2X4:%.*]] = alloca { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> }, align 8
// CHECK-NEXT: [[F64M4X2:%.*]] = alloca { <vscale x 4 x double>, <vscale x 4 x double> }, align 8
// CHECK-NEXT: [[BF16MF4X2:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF4X3:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF4X4:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF4X5:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF4X6:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF4X7:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF4X8:%.*]] = alloca { <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat>, <vscale x 1 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X2:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X3:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X4:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X5:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X6:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X7:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16MF2X8:%.*]] = alloca { <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat>, <vscale x 2 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X2:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X3:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X4:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X5:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X6:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X7:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M1X8:%.*]] = alloca { <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat>, <vscale x 4 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M2X2:%.*]] = alloca { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M2X3:%.*]] = alloca { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M2X4:%.*]] = alloca { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> }, align 2
// CHECK-NEXT: [[BF16M4X2:%.*]] = alloca { <vscale x 16 x bfloat>, <vscale x 16 x bfloat> }, align 2
// CHECK-NEXT: ret void
//
void foo () {
Expand Down Expand Up @@ -664,4 +689,34 @@ void foo () {
vfloat64m2x4_t f64m2x4;

vfloat64m4x2_t f64m4x2;
// bf16
vbfloat16mf4x2_t bf16mf4x2;
vbfloat16mf4x3_t bf16mf4x3;
vbfloat16mf4x4_t bf16mf4x4;
vbfloat16mf4x5_t bf16mf4x5;
vbfloat16mf4x6_t bf16mf4x6;
vbfloat16mf4x7_t bf16mf4x7;
vbfloat16mf4x8_t bf16mf4x8;

vbfloat16mf2x2_t bf16mf2x2;
vbfloat16mf2x3_t bf16mf2x3;
vbfloat16mf2x4_t bf16mf2x4;
vbfloat16mf2x5_t bf16mf2x5;
vbfloat16mf2x6_t bf16mf2x6;
vbfloat16mf2x7_t bf16mf2x7;
vbfloat16mf2x8_t bf16mf2x8;

vbfloat16m1x2_t bf16m1x2;
vbfloat16m1x3_t bf16m1x3;
vbfloat16m1x4_t bf16m1x4;
vbfloat16m1x5_t bf16m1x5;
vbfloat16m1x6_t bf16m1x6;
vbfloat16m1x7_t bf16m1x7;
vbfloat16m1x8_t bf16m1x8;

vbfloat16m2x2_t bf16m2x2;
vbfloat16m2x3_t bf16m2x3;
vbfloat16m2x4_t bf16m2x4;

vbfloat16m4x2_t bf16m4x2;
}
4 changes: 3 additions & 1 deletion clang/utils/TableGen/RISCVVEmitter.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -401,7 +401,9 @@ void RVVEmitter::createHeader(raw_ostream &OS) {
auto TupleT = TypeCache.computeType(
BT, Log2LMUL,
PrototypeDescriptor(BaseTypeModifier::Vector, getTupleVTM(NF),
TypeModifier::Float));
(BT == BasicType::BFloat16
? TypeModifier::BFloat
: TypeModifier::Float)));
if (TupleT)
printType(*TupleT);
}
Expand Down

0 comments on commit 8434b0b

Please sign in to comment.