[Mlir-commits] [mlir] [mlir][spirv] Add bfloat16 support (PR #141458)
Darren Wihandi
llvmlistbot at llvm.org
Mon May 26 10:35:52 PDT 2025
https://github.com/fairywreath updated https://github.com/llvm/llvm-project/pull/141458
>From 45349e623c10722a7576ff81d1dc889cd13889d1 Mon Sep 17 00:00:00 2001
From: fairywreath <nerradfour at gmail.com>
Date: Mon, 26 May 2025 02:33:38 -0400
Subject: [PATCH 1/2] [mlir][spirv] Add bfloat16 support
---
.../Dialect/SPIRV/IR/SPIRVArithmeticOps.td | 6 +-
.../mlir/Dialect/SPIRV/IR/SPIRVBase.td | 40 +++++++++++--
.../mlir/Dialect/SPIRV/IR/SPIRVCastOps.td | 12 ++--
mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp | 5 +-
mlir/lib/Dialect/SPIRV/IR/SPIRVTypes.cpp | 2 +-
.../Target/SPIRV/Serialization/Serializer.cpp | 3 +
.../FuncToSPIRV/types-to-spirv.mlir | 12 ----
.../test/Dialect/SPIRV/IR/arithmetic-ops.mlir | 9 ++-
mlir/test/Dialect/SPIRV/IR/cast-ops.mlir | 56 +++++++++++++++++++
.../SPIRV/IR/khr-cooperative-matrix-ops.mlir | 29 ++++++++++
mlir/test/Dialect/SPIRV/IR/types.mlir | 5 --
11 files changed, 142 insertions(+), 37 deletions(-)
diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td
index 22d5afcd77381..daa1b2b328115 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td
@@ -445,12 +445,12 @@ def SPIRV_DotOp : SPIRV_Op<"Dot",
}];
let arguments = (ins
- SPIRV_VectorOf<SPIRV_Float>:$vector1,
- SPIRV_VectorOf<SPIRV_Float>:$vector2
+ SPIRV_VectorOf<SPIRV_FloatOrBFloat16>:$vector1,
+ SPIRV_VectorOf<SPIRV_FloatOrBFloat16>:$vector2
);
let results = (outs
- SPIRV_Float:$result
+ SPIRV_FloatOrBFloat16:$result
);
let assemblyFormat = "operands attr-dict `:` type($vector1) `->` type($result)";
diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td
index 8fd533db83d9a..5d4469954e5b7 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td
@@ -344,6 +344,7 @@ def SPV_KHR_subgroup_rotate : I32EnumAttrCase<"SPV_KHR_subgroup
def SPV_KHR_non_semantic_info : I32EnumAttrCase<"SPV_KHR_non_semantic_info", 29>;
def SPV_KHR_terminate_invocation : I32EnumAttrCase<"SPV_KHR_terminate_invocation", 30>;
def SPV_KHR_cooperative_matrix : I32EnumAttrCase<"SPV_KHR_cooperative_matrix", 31>;
+def SPV_KHR_bfloat16 : I32EnumAttrCase<"SPV_KHR_bfloat16", 32>;
def SPV_EXT_demote_to_helper_invocation : I32EnumAttrCase<"SPV_EXT_demote_to_helper_invocation", 1000>;
def SPV_EXT_descriptor_indexing : I32EnumAttrCase<"SPV_EXT_descriptor_indexing", 1001>;
@@ -436,7 +437,7 @@ def SPIRV_ExtensionAttr :
SPV_KHR_fragment_shader_barycentric, SPV_KHR_ray_cull_mask,
SPV_KHR_uniform_group_instructions, SPV_KHR_subgroup_rotate,
SPV_KHR_non_semantic_info, SPV_KHR_terminate_invocation,
- SPV_KHR_cooperative_matrix,
+ SPV_KHR_cooperative_matrix, SPV_KHR_bfloat16,
SPV_EXT_demote_to_helper_invocation, SPV_EXT_descriptor_indexing,
SPV_EXT_fragment_fully_covered, SPV_EXT_fragment_invocation_density,
SPV_EXT_fragment_shader_interlock, SPV_EXT_physical_storage_buffer,
@@ -1412,6 +1413,23 @@ def SPIRV_C_ShaderStereoViewNV : I32EnumAttrCase<"Shade
Extension<[SPV_NV_stereo_view_rendering]>
];
}
+def SPIRV_C_BFloat16TypeKHR : I32EnumAttrCase<"BFloat16TypeKHR", 5116> {
+ list<Availability> availability = [
+ Extension<[SPV_KHR_bfloat16]>
+ ];
+}
+def SPIRV_C_BFloat16DotProductKHR : I32EnumAttrCase<"BFloat16DotProductKHR", 5117> {
+ list<I32EnumAttrCase> implies = [SPIRV_C_BFloat16TypeKHR];
+ list<Availability> availability = [
+ Extension<[SPV_KHR_bfloat16]>
+ ];
+}
+def SPIRV_C_BFloat16CooperativeMatrixKHR : I32EnumAttrCase<"BFloat16CooperativeMatrixKHR", 5118> {
+ list<I32EnumAttrCase> implies = [SPIRV_C_BFloat16TypeKHR, SPIRV_C_CooperativeMatrixKHR];
+ list<Availability> availability = [
+ Extension<[SPV_KHR_bfloat16]>
+ ];
+}
def SPIRV_C_Bfloat16ConversionINTEL : I32EnumAttrCase<"Bfloat16ConversionINTEL", 6115> {
list<Availability> availability = [
@@ -1518,7 +1536,8 @@ def SPIRV_CapabilityAttr :
SPIRV_C_StorageTexelBufferArrayNonUniformIndexing,
SPIRV_C_ShaderViewportIndexLayerEXT, SPIRV_C_ShaderViewportMaskNV,
SPIRV_C_ShaderStereoViewNV, SPIRV_C_Bfloat16ConversionINTEL,
- SPIRV_C_CacheControlsINTEL
+ SPIRV_C_CacheControlsINTEL, SPIRV_C_BFloat16TypeKHR,
+ SPIRV_C_BFloat16DotProductKHR, SPIRV_C_BFloat16CooperativeMatrixKHR,
]>;
def SPIRV_AM_Logical : I32EnumAttrCase<"Logical", 0>;
@@ -3217,6 +3236,16 @@ def SPIRV_ExecutionModelAttr :
SPIRV_EM_TaskEXT, SPIRV_EM_MeshEXT
]>;
+def SPIRV_FPE_BFloat16KHR : I32EnumAttrCase<"BFloat16KHR", 0> {
+ list<Availability> availability = [
+ Capability<[SPIRV_C_BFloat16TypeKHR]>
+ ];
+}
+def SPIRV_FPEncodingAttr :
+ SPIRV_I32EnumAttr<"FPEncoding", "valid SPIR-V FPEncoding", "f_p_encoding", [
+ SPIRV_FPE_BFloat16KHR
+ ]>;
+
def SPIRV_FC_None : I32BitEnumAttrCaseNone<"None">;
def SPIRV_FC_Inline : I32BitEnumAttrCaseBit<"Inline", 0>;
def SPIRV_FC_DontInline : I32BitEnumAttrCaseBit<"DontInline", 1>;
@@ -4163,8 +4192,9 @@ def SPIRV_Int32 : TypeAlias<I32, "Int32">;
def SPIRV_Float32 : TypeAlias<F32, "Float32">;
def SPIRV_Float : FloatOfWidths<[16, 32, 64]>;
def SPIRV_Float16or32 : FloatOfWidths<[16, 32]>;
+def SPIRV_FloatOrBFloat16 : AnyTypeOf<[SPIRV_Float, BF16]>;
def SPIRV_Vector : VectorOfLengthAndType<[2, 3, 4, 8, 16],
- [SPIRV_Bool, SPIRV_Integer, SPIRV_Float]>;
+ [SPIRV_Bool, SPIRV_Integer, SPIRV_FloatOrBFloat16]>;
// Component type check is done in the type parser for the following SPIR-V
// dialect-specific types so we use "Any" here.
def SPIRV_AnyPtr : DialectType<SPIRV_Dialect, SPIRV_IsPtrType,
@@ -4194,9 +4224,9 @@ def SPIRV_Composite :
AnyTypeOf<[SPIRV_Vector, SPIRV_AnyArray, SPIRV_AnyRTArray, SPIRV_AnyStruct,
SPIRV_AnyCooperativeMatrix, SPIRV_AnyMatrix]>;
def SPIRV_Type : AnyTypeOf<[
- SPIRV_Void, SPIRV_Bool, SPIRV_Integer, SPIRV_Float, SPIRV_Vector,
+ SPIRV_Void, SPIRV_Bool, SPIRV_Integer, SPIRV_FloatOrBFloat16, SPIRV_Vector,
SPIRV_AnyPtr, SPIRV_AnyArray, SPIRV_AnyRTArray, SPIRV_AnyStruct,
- SPIRV_AnyCooperativeMatrix, SPIRV_AnyMatrix, SPIRV_AnySampledImage
+ SPIRV_AnyCooperativeMatrix, SPIRV_AnyMatrix, SPIRV_AnySampledImage,
]>;
def SPIRV_SignedInt : SignedIntOfWidths<[8, 16, 32, 64]>;
diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVCastOps.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVCastOps.td
index b05ee0251df5b..29571cf138ebf 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVCastOps.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVCastOps.td
@@ -86,7 +86,7 @@ def SPIRV_BitcastOp : SPIRV_Op<"Bitcast", [Pure]> {
// -----
-def SPIRV_ConvertFToSOp : SPIRV_CastOp<"ConvertFToS", SPIRV_Integer, SPIRV_Float, []> {
+def SPIRV_ConvertFToSOp : SPIRV_CastOp<"ConvertFToS", SPIRV_Integer, SPIRV_FloatOrBFloat16, []> {
let summary = [{
Convert value numerically from floating point to signed integer, with
round toward 0.0.
@@ -111,7 +111,7 @@ def SPIRV_ConvertFToSOp : SPIRV_CastOp<"ConvertFToS", SPIRV_Integer, SPIRV_Float
// -----
-def SPIRV_ConvertFToUOp : SPIRV_CastOp<"ConvertFToU", SPIRV_Integer, SPIRV_Float, []> {
+def SPIRV_ConvertFToUOp : SPIRV_CastOp<"ConvertFToU", SPIRV_Integer, SPIRV_FloatOrBFloat16, []> {
let summary = [{
Convert value numerically from floating point to unsigned integer, with
round toward 0.0.
@@ -138,7 +138,7 @@ def SPIRV_ConvertFToUOp : SPIRV_CastOp<"ConvertFToU", SPIRV_Integer, SPIRV_Float
// -----
def SPIRV_ConvertSToFOp : SPIRV_CastOp<"ConvertSToF",
- SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
SPIRV_Integer,
[SignedOp]> {
let summary = [{
@@ -165,7 +165,7 @@ def SPIRV_ConvertSToFOp : SPIRV_CastOp<"ConvertSToF",
// -----
def SPIRV_ConvertUToFOp : SPIRV_CastOp<"ConvertUToF",
- SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
SPIRV_Integer,
[UnsignedOp]> {
let summary = [{
@@ -192,8 +192,8 @@ def SPIRV_ConvertUToFOp : SPIRV_CastOp<"ConvertUToF",
// -----
def SPIRV_FConvertOp : SPIRV_CastOp<"FConvert",
- SPIRV_Float,
- SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
+ SPIRV_FloatOrBFloat16,
[UsableInSpecConstantOp]> {
let summary = [{
Convert value numerically from one floating-point width to another
diff --git a/mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp b/mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp
index 0cf5f0823be63..a21acef1c4b43 100644
--- a/mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp
+++ b/mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp
@@ -175,10 +175,7 @@ static Type parseAndVerifyType(SPIRVDialect const &dialect,
// Check other allowed types
if (auto t = llvm::dyn_cast<FloatType>(type)) {
- if (type.isBF16()) {
- parser.emitError(typeLoc, "cannot use 'bf16' to compose SPIR-V types");
- return Type();
- }
+ // TODO: All float types are allowed for now, but this should be fixed.
} else if (auto t = llvm::dyn_cast<IntegerType>(type)) {
if (!ScalarType::isValid(t)) {
parser.emitError(typeLoc,
diff --git a/mlir/lib/Dialect/SPIRV/IR/SPIRVTypes.cpp b/mlir/lib/Dialect/SPIRV/IR/SPIRVTypes.cpp
index 337df3a5a65f0..5da3164ad4d14 100644
--- a/mlir/lib/Dialect/SPIRV/IR/SPIRVTypes.cpp
+++ b/mlir/lib/Dialect/SPIRV/IR/SPIRVTypes.cpp
@@ -505,7 +505,7 @@ bool ScalarType::classof(Type type) {
}
bool ScalarType::isValid(FloatType type) {
- return llvm::is_contained({16u, 32u, 64u}, type.getWidth()) && !type.isBF16();
+ return llvm::is_contained({16u, 32u, 64u}, type.getWidth());
}
bool ScalarType::isValid(IntegerType type) {
diff --git a/mlir/lib/Target/SPIRV/Serialization/Serializer.cpp b/mlir/lib/Target/SPIRV/Serialization/Serializer.cpp
index 15e06616f4492..b43f22db55a2e 100644
--- a/mlir/lib/Target/SPIRV/Serialization/Serializer.cpp
+++ b/mlir/lib/Target/SPIRV/Serialization/Serializer.cpp
@@ -523,6 +523,9 @@ LogicalResult Serializer::prepareBasicType(
if (auto floatType = dyn_cast<FloatType>(type)) {
typeEnum = spirv::Opcode::OpTypeFloat;
operands.push_back(floatType.getWidth());
+ if (floatType.isBF16()) {
+ operands.push_back(static_cast<uint32_t>(spirv::FPEncoding::BFloat16KHR));
+ }
return success();
}
diff --git a/mlir/test/Conversion/FuncToSPIRV/types-to-spirv.mlir b/mlir/test/Conversion/FuncToSPIRV/types-to-spirv.mlir
index 82d750755ffe2..2e34c9ff54012 100644
--- a/mlir/test/Conversion/FuncToSPIRV/types-to-spirv.mlir
+++ b/mlir/test/Conversion/FuncToSPIRV/types-to-spirv.mlir
@@ -206,18 +206,6 @@ func.func @float64(%arg0: f64) { return }
// -----
-// Check that bf16 is not supported.
-module attributes {
- spirv.target_env = #spirv.target_env<#spirv.vce<v1.0, [], []>, #spirv.resource_limits<>>
-} {
-
-// CHECK-NOT: spirv.func @bf16_type
-func.func @bf16_type(%arg0: bf16) { return }
-
-} // end module
-
-// -----
-
//===----------------------------------------------------------------------===//
// Complex types
//===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/SPIRV/IR/arithmetic-ops.mlir b/mlir/test/Dialect/SPIRV/IR/arithmetic-ops.mlir
index 2d0c86e08de5a..301a5bab9ab1a 100644
--- a/mlir/test/Dialect/SPIRV/IR/arithmetic-ops.mlir
+++ b/mlir/test/Dialect/SPIRV/IR/arithmetic-ops.mlir
@@ -265,6 +265,13 @@ func.func @dot(%arg0: vector<4xf32>, %arg1: vector<4xf32>) -> f32 {
// -----
+func.func @dot_bf16(%arg0: vector<4xbf16>, %arg1: vector<4xbf16>) -> bf16 {
+ %0 = spirv.Dot %arg0, %arg1 : vector<4xbf16> -> bf16
+ return %0 : bf16
+}
+
+// -----
+
// expected-note @+1 {{prior use here}}
func.func @dot(%arg0: vector<4xf32>, %arg1: vector<3xf32>) -> f32 {
// expected-error @+1 {{use of value '%arg1' expects different type than prior uses}}
@@ -283,7 +290,7 @@ func.func @dot(%arg0: vector<4xf32>, %arg1: vector<4xf32>) -> f16 {
// -----
func.func @dot(%arg0: vector<4xi32>, %arg1: vector<4xi32>) -> i32 {
- // expected-error @+1 {{'spirv.Dot' op operand #0 must be vector of 16/32/64-bit float values of length 2/3/4/8/16}}
+ // expected-error @+1 {{'spirv.Dot' op operand #0 must be vector of 16/32/64-bit float or bfloat16 type values of length 2/3/4/8/16}}
%0 = spirv.Dot %arg0, %arg1 : vector<4xi32> -> i32
return %0 : i32
}
diff --git a/mlir/test/Dialect/SPIRV/IR/cast-ops.mlir b/mlir/test/Dialect/SPIRV/IR/cast-ops.mlir
index 34d0109e6bb44..4480a1f3720f2 100644
--- a/mlir/test/Dialect/SPIRV/IR/cast-ops.mlir
+++ b/mlir/test/Dialect/SPIRV/IR/cast-ops.mlir
@@ -110,6 +110,14 @@ func.func @convert_f_to_s_vector(%arg0 : vector<3xf32>) -> vector<3xi32> {
// -----
+func.func @convert_bf16_to_s32_scalar(%arg0 : bf16) -> i32 {
+ // CHECK: {{%.*}} = spirv.ConvertFToS {{%.*}} : bf16 to i32
+ %0 = spirv.ConvertFToS %arg0 : bf16 to i32
+ spirv.ReturnValue %0 : i32
+}
+
+// -----
+
//===----------------------------------------------------------------------===//
// spirv.ConvertFToU
//===----------------------------------------------------------------------===//
@@ -146,6 +154,14 @@ func.func @convert_f_to_u.coopmatrix(%arg0 : !spirv.coopmatrix<8x16xf32, Subgrou
// -----
+func.func @convert_bf16_to_u32_scalar(%arg0 : bf16) -> i32 {
+ // CHECK: {{%.*}} = spirv.ConvertFToU {{%.*}} : bf16 to i32
+ %0 = spirv.ConvertFToU %arg0 : bf16 to i32
+ spirv.ReturnValue %0 : i32
+}
+
+// -----
+
//===----------------------------------------------------------------------===//
// spirv.ConvertSToF
//===----------------------------------------------------------------------===//
@@ -174,6 +190,14 @@ func.func @convert_s_to_f_vector(%arg0 : vector<3xi32>) -> vector<3xf32> {
// -----
+func.func @convert_s32_to_bf16_scalar(%arg0 : i32) -> bf16 {
+ // CHECK: {{%.*}} = spirv.ConvertSToF {{%.*}} : i32 to bf16
+ %0 = spirv.ConvertSToF %arg0 : i32 to bf16
+ spirv.ReturnValue %0 : bf16
+}
+
+// -----
+
//===----------------------------------------------------------------------===//
// spirv.ConvertUToF
//===----------------------------------------------------------------------===//
@@ -202,6 +226,14 @@ func.func @convert_u_to_f_vector(%arg0 : vector<3xi32>) -> vector<3xf32> {
// -----
+func.func @convert_u32_to_bf16_scalar(%arg0 : i32) -> bf16 {
+ // CHECK: {{%.*}} = spirv.ConvertUToF {{%.*}} : i32 to bf16
+ %0 = spirv.ConvertUToF %arg0 : i32 to bf16
+ spirv.ReturnValue %0 : bf16
+}
+
+// -----
+
//===----------------------------------------------------------------------===//
// spirv.FConvert
//===----------------------------------------------------------------------===//
@@ -238,6 +270,30 @@ func.func @f_convert_vector(%arg0 : f32) -> f32 {
// -----
+func.func @f_convert_bf16_to_f32_scalar(%arg0 : bf16) -> f32 {
+ // CHECK: {{%.*}} = spirv.FConvert {{%.*}} : bf16 to f32
+ %0 = spirv.FConvert %arg0 : bf16 to f32
+ spirv.ReturnValue %0 : f32
+}
+
+// -----
+
+func.func @f_convert_f32_to_bf16_vector(%arg0 : vector<3xf32>) -> vector<3xbf16> {
+ // CHECK: {{%.*}} = spirv.FConvert {{%.*}} : vector<3xf32> to vector<3xbf16>
+ %0 = spirv.FConvert %arg0 : vector<3xf32> to vector<3xbf16>
+ spirv.ReturnValue %0 : vector<3xbf16>
+}
+
+// -----
+
+func.func @f_convert_f32_to_bf16_coop_matrix(%arg0 : !spirv.coopmatrix<8x16xf32, Subgroup, MatrixA>) -> !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA> {
+ // CHECK: {{%.*}} = spirv.FConvert {{%.*}} : !spirv.coopmatrix<8x16xf32, Subgroup, MatrixA> to !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>
+ %0 = spirv.FConvert %arg0 : !spirv.coopmatrix<8x16xf32, Subgroup, MatrixA> to !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>
+ spirv.ReturnValue %0 : !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>
+}
+
+// -----
+
//===----------------------------------------------------------------------===//
// spirv.SConvert
//===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir b/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir
index d3e1dbc229ef9..8929e63639c97 100644
--- a/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir
+++ b/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir
@@ -31,6 +31,15 @@ spirv.func @cooperative_matrix_load_memoperand(%ptr : !spirv.ptr<i32, StorageBuf
spirv.Return
}
+// CHECK-LABEL: @cooperative_matrix_load_bf16
+spirv.func @cooperative_matrix_load_bf16(%ptr : !spirv.ptr<bf16, StorageBuffer>, %stride : i32) "None" {
+ // CHECK: {{%.+}} = spirv.KHR.CooperativeMatrixLoad {{%.*}}, {{%.*}}, <RowMajor>
+ // CHECK-SAME: : !spirv.ptr<bf16, StorageBuffer>, i32 -> !spirv.coopmatrix<16x8xbf16, Workgroup, MatrixA>
+ %0 = spirv.KHR.CooperativeMatrixLoad %ptr, %stride, <RowMajor> :
+ !spirv.ptr<bf16, StorageBuffer>, i32 -> !spirv.coopmatrix<16x8xbf16, Workgroup, MatrixA>
+ spirv.Return
+}
+
// CHECK-LABEL: @cooperative_matrix_load_vector_ptr_type
spirv.func @cooperative_matrix_load_vector_ptr_type(%ptr : !spirv.ptr<vector<4xi32>, StorageBuffer>, %stride : i32) "None" {
// CHECK: {{%.*}} = spirv.KHR.CooperativeMatrixLoad {{%.*}}, {{%.*}}, <RowMajor>, <Volatile> :
@@ -225,6 +234,26 @@ spirv.func @cooperative_matrix_muladd_f32(%a : !spirv.coopmatrix<4x4xf32, Subgro
spirv.Return
}
+spirv.func @cooperative_matrix_muladd_bf16_bf16(%a : !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>,
+ %b : !spirv.coopmatrix<16x4xbf16, Subgroup, MatrixB>,
+ %c : !spirv.coopmatrix<8x4xbf16, Subgroup, MatrixAcc>) "None" {
+ %r = spirv.KHR.CooperativeMatrixMulAdd %a, %b, %c :
+ !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>,
+ !spirv.coopmatrix<16x4xbf16, Subgroup, MatrixB> ->
+ !spirv.coopmatrix<8x4xbf16, Subgroup, MatrixAcc>
+ spirv.Return
+}
+
+spirv.func @cooperative_matrix_muladd_bf16_f32(%a : !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>,
+ %b : !spirv.coopmatrix<16x4xbf16, Subgroup, MatrixB>,
+ %c : !spirv.coopmatrix<8x4xf32, Subgroup, MatrixAcc>) "None" {
+ %r = spirv.KHR.CooperativeMatrixMulAdd %a, %b, %c :
+ !spirv.coopmatrix<8x16xbf16, Subgroup, MatrixA>,
+ !spirv.coopmatrix<16x4xbf16, Subgroup, MatrixB> ->
+ !spirv.coopmatrix<8x4xf32, Subgroup, MatrixAcc>
+ spirv.Return
+}
+
spirv.func @cooperative_matrix_muladd_i8_i32(%a : !spirv.coopmatrix<8x16xi8, Subgroup, MatrixA>,
%b : !spirv.coopmatrix<16x4xi8, Subgroup, MatrixB>,
%c : !spirv.coopmatrix<8x4xi32, Subgroup, MatrixAcc>) "None" {
diff --git a/mlir/test/Dialect/SPIRV/IR/types.mlir b/mlir/test/Dialect/SPIRV/IR/types.mlir
index b63a08d96e6af..a81fe72a8362e 100644
--- a/mlir/test/Dialect/SPIRV/IR/types.mlir
+++ b/mlir/test/Dialect/SPIRV/IR/types.mlir
@@ -57,11 +57,6 @@ func.func private @tensor_type(!spirv.array<4xtensor<4xf32>>) -> ()
// -----
-// expected-error @+1 {{cannot use 'bf16' to compose SPIR-V types}}
-func.func private @bf16_type(!spirv.array<4xbf16>) -> ()
-
-// -----
-
// expected-error @+1 {{only 1/8/16/32/64-bit integer type allowed but found 'i256'}}
func.func private @i256_type(!spirv.array<4xi256>) -> ()
>From d9815d4addcfa9ed4c5efeb6c572d0b6c1576863 Mon Sep 17 00:00:00 2001
From: fairywreath <nerradfour at gmail.com>
Date: Mon, 26 May 2025 11:35:10 -0600
Subject: [PATCH 2/2] Properly implement arithmetic coop matrix ops with bf16
---
.../Dialect/SPIRV/IR/SPIRVArithmeticOps.td | 96 ++++++++++++++-----
.../mlir/Dialect/SPIRV/IR/SPIRVBase.td | 12 ++-
.../SPIRV/IR/khr-cooperative-matrix-ops.mlir | 68 +++++++++++--
3 files changed, 138 insertions(+), 38 deletions(-)
diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td
index daa1b2b328115..850e0d165f4cf 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td
@@ -24,8 +24,8 @@ class SPIRV_ArithmeticBinaryOp<string mnemonic, Type type,
SPIRV_BinaryOp<mnemonic, type, type,
!listconcat(traits,
[Pure, SameOperandsAndResultType])> {
- // In addition to normal types arithmetic instructions can support cooperative
- // matrix.
+ // TODO: Arithmetic operations that use this definition do not support cooperative matrices,
+ // these need to be fixed.
let arguments = (ins
SPIRV_ScalarOrVectorOrCoopMatrixOf<type>:$operand1,
SPIRV_ScalarOrVectorOrCoopMatrixOf<type>:$operand2
@@ -37,20 +37,43 @@ class SPIRV_ArithmeticBinaryOp<string mnemonic, Type type,
let assemblyFormat = "operands attr-dict `:` type($result)";
}
-class SPIRV_ArithmeticUnaryOp<string mnemonic, Type type,
+class SPIRV_ArithmeticWithCoopMatrixBinaryOp<string mnemonic,
+ Type scalarVectorType,
+ Type coopMatrixType,
+ list<Trait> traits = []> :
+ // Operands type same as result type.
+ SPIRV_BinaryOp<mnemonic, coopMatrixType, coopMatrixType,
+ !listconcat(traits,
+ [Pure, SameOperandsAndResultType])> {
+ // In addition to normal types these arithmetic instructions can support
+ // cooperative matrix.
+ let arguments = (ins
+ SPIRV_ScalarOrVectorOfOrCoopMatrixOf<scalarVectorType, coopMatrixType>:$operand1,
+ SPIRV_ScalarOrVectorOfOrCoopMatrixOf<scalarVectorType, coopMatrixType>:$operand2
+ );
+
+ let results = (outs
+ SPIRV_ScalarOrVectorOfOrCoopMatrixOf<scalarVectorType, coopMatrixType>:$result
+ );
+ let assemblyFormat = "operands attr-dict `:` type($result)";
+}
+
+class SPIRV_ArithmeticUnaryOp<string mnemonic,
+ Type scalarVectorType,
+ Type coopMatrixType,
list<Trait> traits = []> :
// Operand type same as result type.
- SPIRV_UnaryOp<mnemonic, type, type,
+ SPIRV_UnaryOp<mnemonic, coopMatrixType, coopMatrixType,
!listconcat(traits,
[Pure, SameOperandsAndResultType])> {
// In addition to normal types arithmetic instructions can support cooperative
// matrix.
let arguments = (ins
- SPIRV_ScalarOrVectorOrCoopMatrixOf<type>:$operand
+ SPIRV_ScalarOrVectorOfOrCoopMatrixOf<scalarVectorType, coopMatrixType>:$operand1
);
let results = (outs
- SPIRV_ScalarOrVectorOrCoopMatrixOf<type>:$result
+ SPIRV_ScalarOrVectorOfOrCoopMatrixOf<scalarVectorType, coopMatrixType>:$result
);
let assemblyFormat = "operands attr-dict `:` type($result)";
}
@@ -82,7 +105,10 @@ class SPIRV_ArithmeticExtendedBinaryOp<string mnemonic,
// -----
-def SPIRV_FAddOp : SPIRV_ArithmeticBinaryOp<"FAdd", SPIRV_Float, [Commutative]> {
+def SPIRV_FAddOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"FAdd",
+ SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
+ [Commutative]> {
let summary = "Floating-point addition of Operand 1 and Operand 2.";
let description = [{
@@ -104,7 +130,10 @@ def SPIRV_FAddOp : SPIRV_ArithmeticBinaryOp<"FAdd", SPIRV_Float, [Commutative]>
// -----
-def SPIRV_FDivOp : SPIRV_ArithmeticBinaryOp<"FDiv", SPIRV_Float, []> {
+def SPIRV_FDivOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"FDiv",
+ SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
+ []> {
let summary = "Floating-point division of Operand 1 divided by Operand 2.";
let description = [{
@@ -154,7 +183,10 @@ def SPIRV_FModOp : SPIRV_ArithmeticBinaryOp<"FMod", SPIRV_Float, []> {
// -----
-def SPIRV_FMulOp : SPIRV_ArithmeticBinaryOp<"FMul", SPIRV_Float, [Commutative]> {
+def SPIRV_FMulOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"FMul",
+ SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
+ [Commutative]> {
let summary = "Floating-point multiplication of Operand 1 and Operand 2.";
let description = [{
@@ -176,7 +208,10 @@ def SPIRV_FMulOp : SPIRV_ArithmeticBinaryOp<"FMul", SPIRV_Float, [Commutative]>
// -----
-def SPIRV_FNegateOp : SPIRV_ArithmeticUnaryOp<"FNegate", SPIRV_Float, []> {
+def SPIRV_FNegateOp : SPIRV_ArithmeticUnaryOp<"FNegate",
+ SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
+ []> {
let summary = [{
Inverts the sign bit of Operand. (Note, however, that OpFNegate is still
considered a floating-point instruction, and so is subject to the
@@ -229,7 +264,10 @@ def SPIRV_FRemOp : SPIRV_ArithmeticBinaryOp<"FRem", SPIRV_Float, []> {
// -----
-def SPIRV_FSubOp : SPIRV_ArithmeticBinaryOp<"FSub", SPIRV_Float, []> {
+def SPIRV_FSubOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"FSub",
+ SPIRV_Float,
+ SPIRV_FloatOrBFloat16,
+ []> {
let summary = "Floating-point subtraction of Operand 2 from Operand 1.";
let description = [{
@@ -251,9 +289,10 @@ def SPIRV_FSubOp : SPIRV_ArithmeticBinaryOp<"FSub", SPIRV_Float, []> {
// -----
-def SPIRV_IAddOp : SPIRV_ArithmeticBinaryOp<"IAdd",
- SPIRV_Integer,
- [Commutative, UsableInSpecConstantOp]> {
+def SPIRV_IAddOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"IAdd",
+ SPIRV_Integer,
+ SPIRV_Integer,
+ [Commutative, UsableInSpecConstantOp]> {
let summary = "Integer addition of Operand 1 and Operand 2.";
let description = [{
@@ -322,9 +361,10 @@ def SPIRV_IAddCarryOp : SPIRV_ArithmeticExtendedBinaryOp<"IAddCarry",
// -----
-def SPIRV_IMulOp : SPIRV_ArithmeticBinaryOp<"IMul",
- SPIRV_Integer,
- [Commutative, UsableInSpecConstantOp]> {
+def SPIRV_IMulOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"IMul",
+ SPIRV_Integer,
+ SPIRV_Integer,
+ [Commutative, UsableInSpecConstantOp]> {
let summary = "Integer multiplication of Operand 1 and Operand 2.";
let description = [{
@@ -354,9 +394,10 @@ def SPIRV_IMulOp : SPIRV_ArithmeticBinaryOp<"IMul",
// -----
-def SPIRV_ISubOp : SPIRV_ArithmeticBinaryOp<"ISub",
- SPIRV_Integer,
- [UsableInSpecConstantOp]> {
+def SPIRV_ISubOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"ISub",
+ SPIRV_Integer,
+ SPIRV_Integer,
+ [UsableInSpecConstantOp]> {
let summary = "Integer subtraction of Operand 2 from Operand 1.";
let description = [{
@@ -460,9 +501,10 @@ def SPIRV_DotOp : SPIRV_Op<"Dot",
// -----
-def SPIRV_SDivOp : SPIRV_ArithmeticBinaryOp<"SDiv",
- SPIRV_Integer,
- [UsableInSpecConstantOp]> {
+def SPIRV_SDivOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"SDiv",
+ SPIRV_Integer,
+ SPIRV_Integer,
+ [UsableInSpecConstantOp]> {
let summary = "Signed-integer division of Operand 1 divided by Operand 2.";
let description = [{
@@ -560,6 +602,7 @@ def SPIRV_SMulExtendedOp : SPIRV_ArithmeticExtendedBinaryOp<"SMulExtended",
// -----
def SPIRV_SNegateOp : SPIRV_ArithmeticUnaryOp<"SNegate",
+ SPIRV_Integer,
SPIRV_Integer,
[UsableInSpecConstantOp]> {
let summary = "Signed-integer subtract of Operand from zero.";
@@ -622,9 +665,10 @@ def SPIRV_SRemOp : SPIRV_ArithmeticBinaryOp<"SRem",
// -----
-def SPIRV_UDivOp : SPIRV_ArithmeticBinaryOp<"UDiv",
- SPIRV_Integer,
- [UnsignedOp, UsableInSpecConstantOp]> {
+def SPIRV_UDivOp : SPIRV_ArithmeticWithCoopMatrixBinaryOp<"UDiv",
+ SPIRV_Integer,
+ SPIRV_Integer,
+ [UnsignedOp, UsableInSpecConstantOp]> {
let summary = "Unsigned-integer division of Operand 1 divided by Operand 2.";
let description = [{
diff --git a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td
index 5d4469954e5b7..ccc6bd76ca1d0 100644
--- a/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td
+++ b/mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td
@@ -4245,16 +4245,24 @@ class SPIRV_MatrixOfType<list<Type> allowedTypes> :
class SPIRV_VectorOf<Type type> :
VectorOfLengthAndType<[2, 3, 4, 8, 16], [type]>;
+class SPIRV_CoopMatrixOf<Type type> :
+ SPIRV_CoopMatrixOfType<[type]>;
+
class SPIRV_ScalarOrVectorOf<Type type> :
AnyTypeOf<[type, SPIRV_VectorOf<type>]>;
class SPIRV_ScalarOrVectorOrCoopMatrixOf<Type type> :
AnyTypeOf<[type, SPIRV_VectorOf<type>,
- SPIRV_CoopMatrixOfType<[type]>]>;
+ SPIRV_CoopMatrixOf<type>]>;
+
+class SPIRV_ScalarOrVectorOfOrCoopMatrixOf<Type scalarVectorType,
+ Type coopMatrixType> :
+ AnyTypeOf<[scalarVectorType, SPIRV_VectorOf<scalarVectorType>,
+ SPIRV_CoopMatrixOf<coopMatrixType>]>;
class SPIRV_MatrixOrCoopMatrixOf<Type type> :
AnyTypeOf<[SPIRV_AnyMatrix,
- SPIRV_CoopMatrixOfType<[type]>]>;
+ SPIRV_CoopMatrixOf<type>]>;
class SPIRV_MatrixOf<Type type> :
SPIRV_MatrixOfType<[type]>;
diff --git a/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir b/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir
index 8929e63639c97..4d161a3193505 100644
--- a/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir
+++ b/mlir/test/Dialect/SPIRV/IR/khr-cooperative-matrix-ops.mlir
@@ -437,6 +437,9 @@ spirv.func @cooperative_matrix_muladd_matrix_operands(%a : !spirv.coopmatrix<8x1
!matA_f32 = !spirv.coopmatrix<2x2xf32, Subgroup, MatrixA>
!matB_f32 = !spirv.coopmatrix<2x2xf32, Subgroup, MatrixB>
+!matA_bf16 = !spirv.coopmatrix<2x2xbf16, Subgroup, MatrixA>
+!matB_bf16 = !spirv.coopmatrix<2x2xbf16, Subgroup, MatrixB>
+
// These tests are kept in the same order as the list of compatible ops in the
// SPV_KHR_cooperative_matrix extension spec.
@@ -449,8 +452,8 @@ spirv.func @snegate(%a: !matA_i32, %b: !matB_i32) "None" {
spirv.Return
}
-// CHECK-LABEL: @fnegate
-spirv.func @fnegate(%a: !matA_f32, %b: !matB_f32) "None" {
+// CHECK-LABEL: @fnegate_f32
+spirv.func @fnegate_f32(%a: !matA_f32, %b: !matB_f32) "None" {
// CHECK: spirv.FNegate {{%.*}} : !spirv.coopmatrix
// CHECK-NEXT: spirv.FNegate {{%.*}} : !spirv.coopmatrix
%p = spirv.FNegate %a : !matA_f32
@@ -458,6 +461,15 @@ spirv.func @fnegate(%a: !matA_f32, %b: !matB_f32) "None" {
spirv.Return
}
+// CHECK-LABEL: @fnegate_bf16
+spirv.func @fnegate_bf16(%a: !matA_bf16, %b: !matB_bf16) "None" {
+ // CHECK: spirv.FNegate {{%.*}} : !spirv.coopmatrix
+ // CHECK-NEXT: spirv.FNegate {{%.*}} : !spirv.coopmatrix
+ %p = spirv.FNegate %a : !matA_bf16
+ %q = spirv.FNegate %b : !matB_bf16
+ spirv.Return
+}
+
// CHECK-LABEL: @iadd
spirv.func @iadd(%a: !matA_i32, %b: !matB_i32) "None" {
// CHECK: spirv.IAdd {{%.*}}, {{%.*}} : !spirv.coopmatrix
@@ -467,8 +479,8 @@ spirv.func @iadd(%a: !matA_i32, %b: !matB_i32) "None" {
spirv.Return
}
-// CHECK-LABEL: @fadd
-spirv.func @fadd(%a: !matA_f32, %b: !matB_f32) "None" {
+// CHECK-LABEL: @fadd_f32
+spirv.func @fadd_f32(%a: !matA_f32, %b: !matB_f32) "None" {
// CHECK: spirv.FAdd {{%.*}}, {{%.*}} : !spirv.coopmatrix
// CHECK-NEXT: spirv.FAdd {{%.*}}, {{%.*}} : !spirv.coopmatrix
%p = spirv.FAdd %a, %a : !matA_f32
@@ -476,6 +488,15 @@ spirv.func @fadd(%a: !matA_f32, %b: !matB_f32) "None" {
spirv.Return
}
+// CHECK-LABEL: @fadd_bf16
+spirv.func @fadd_bf16(%a: !matA_bf16, %b: !matB_bf16) "None" {
+ // CHECK: spirv.FAdd {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ // CHECK-NEXT: spirv.FAdd {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ %p = spirv.FAdd %a, %a : !matA_bf16
+ %q = spirv.FAdd %b, %b : !matB_bf16
+ spirv.Return
+}
+
// CHECK-LABEL: @isub
spirv.func @isub(%a: !matA_i32, %b: !matB_i32) "None" {
// CHECK: spirv.ISub {{%.*}}, {{%.*}} : !spirv.coopmatrix
@@ -485,8 +506,8 @@ spirv.func @isub(%a: !matA_i32, %b: !matB_i32) "None" {
spirv.Return
}
-// CHECK-LABEL: @fsub
-spirv.func @fsub(%a: !matA_f32, %b: !matB_f32) "None" {
+// CHECK-LABEL: @fsub_f32
+spirv.func @fsub_f32(%a: !matA_f32, %b: !matB_f32) "None" {
// CHECK: spirv.FSub {{%.*}}, {{%.*}} : !spirv.coopmatrix
// CHECK-NEXT: spirv.FSub {{%.*}}, {{%.*}} : !spirv.coopmatrix
%p = spirv.FSub %a, %a : !matA_f32
@@ -494,8 +515,17 @@ spirv.func @fsub(%a: !matA_f32, %b: !matB_f32) "None" {
spirv.Return
}
-// CHECK-LABEL: @fmul
-spirv.func @fmul(%a: !matA_f32, %b: !matB_f32) "None" {
+// CHECK-LABEL: @fsub_bf16
+spirv.func @fsub_bf16(%a: !matA_bf16, %b: !matB_bf16) "None" {
+ // CHECK: spirv.FSub {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ // CHECK-NEXT: spirv.FSub {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ %p = spirv.FSub %a, %a : !matA_bf16
+ %q = spirv.FSub %b, %b : !matB_bf16
+ spirv.Return
+}
+
+// CHECK-LABEL: @fmul_f32
+spirv.func @fmul_f32(%a: !matA_f32, %b: !matB_f32) "None" {
// CHECK: spirv.FMul {{%.*}}, {{%.*}} : !spirv.coopmatrix
// CHECK-NEXT: spirv.FMul {{%.*}}, {{%.*}} : !spirv.coopmatrix
%p = spirv.FMul %a, %a : !matA_f32
@@ -503,6 +533,15 @@ spirv.func @fmul(%a: !matA_f32, %b: !matB_f32) "None" {
spirv.Return
}
+// CHECK-LABEL: @fmul_bf16
+spirv.func @fmul_bf16(%a: !matA_bf16, %b: !matB_bf16) "None" {
+ // CHECK: spirv.FMul {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ // CHECK-NEXT: spirv.FMul {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ %p = spirv.FMul %a, %a : !matA_bf16
+ %q = spirv.FMul %b, %b : !matB_bf16
+ spirv.Return
+}
+
// CHECK-LABEL: @imul
spirv.func @imul(%a: !matA_i32, %b: !matB_i32) "None" {
// CHECK: spirv.IMul {{%.*}}, {{%.*}} : !spirv.coopmatrix
@@ -512,8 +551,8 @@ spirv.func @imul(%a: !matA_i32, %b: !matB_i32) "None" {
spirv.Return
}
-// CHECK-LABEL: @fdiv
-spirv.func @fdiv(%a: !matA_f32, %b: !matB_f32) "None" {
+// CHECK-LABEL: @fdiv_f32
+spirv.func @fdiv_f32(%a: !matA_f32, %b: !matB_f32) "None" {
// CHECK: spirv.FDiv {{%.*}}, {{%.*}} : !spirv.coopmatrix
// CHECK-NEXT: spirv.FDiv {{%.*}}, {{%.*}} : !spirv.coopmatrix
%p = spirv.FDiv %a, %a : !matA_f32
@@ -521,6 +560,15 @@ spirv.func @fdiv(%a: !matA_f32, %b: !matB_f32) "None" {
spirv.Return
}
+// CHECK-LABEL: @fdiv_bf16
+spirv.func @fdiv_bf16(%a: !matA_bf16, %b: !matB_bf16) "None" {
+ // CHECK: spirv.FDiv {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ // CHECK-NEXT: spirv.FDiv {{%.*}}, {{%.*}} : !spirv.coopmatrix
+ %p = spirv.FDiv %a, %a : !matA_bf16
+ %q = spirv.FDiv %b, %b : !matB_bf16
+ spirv.Return
+}
+
// CHECK-LABEL: @sdiv
spirv.func @sdiv(%a: !matA_i32, %b: !matB_i32) "None" {
// CHECK: spirv.SDiv {{%.*}}, {{%.*}} : !spirv.coopmatrix
More information about the Mlir-commits
mailing list