[MLIR][SPIRV] Add intel joint matrix ops

[nirvedhmeshram]

Summary

Adding these ops according to the spec provided here
https://github.com/intel/llvm/blob/c2f8602c843d2ba87dc1c9057f1422f23a77ed66/sycl/doc/design/spirv-extensions/SPV_INTEL_joint_matrix.asciidoc

[antiagainst]

Cool, thanks for adding this! A few comments inlined.

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td

446

Nit: match the order of definition in the above.

1395

This should not imply Shader, which is the ~root capability for Vulkan API? I didn't find the spec saying it implies Shader.

1474

Nit: match the order of definitions above

4058

Nit: Alignment is a bit off here.

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVJointMatrixOps.td

1

Nit: alignment is off here.

46

Now we have v1.6. Should make this as v1.6 here. Also applies for following ops.

67

This seems to be the doc for the type, not for JointMatrixLoadINTEL?

86

This does not match with the assemblyFormat anymore. With assemblyFormat we don't need to document this; so just remove it.

120

Please define the layout as an attribute like SPV_ScopeAttr, instead of using the magic SPV_Integer here. You can follow how SPV_ScopeAttr is defined and copy that for Matrix Layout (https://github.com/intel/llvm/blob/c2f8602c843d2ba87dc1c9057f1422f23a77ed66/sycl/doc/design/spirv-extensions/SPV_INTEL_joint_matrix.asciidoc#matrix-layout). Please make sure to put it outside of the autogen region in SPIRVBase.td. Defining it that way will give us better IR parsing/printing and allows wiring up capability checks.

It's a bit tricker given this SPV_MatrixLayoutAttr would need special treatment like SPV_Scope. Search where SPV_ScopeAttr are listed in SPIRVUtilsGen.cpp and make sure you add SPV_MatrixLayoutAttr there.

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTypes.h

441

Also getLayout?

mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp

351

It should be rows 'x' columns 'x' element-type ',' layout ',' scope?

352

Missing layout here? As said before we should define SPV_MatrixLayoutAttr like SPV_ScopeAttr. Then you can add the parsing rules for matrix layout like scope in the following. So the full type should look like !spv.jointmatrix<8x8xf32, RowMajor, Subgroup>.

mlir/lib/Dialect/SPIRV/IR/SPIRVOps.cpp

3924

Hmm, StorageBuffer and PhysicalStorageBuffer are actually in the Shader capability hierarchy and for Vulkan. Are we sure we want those here? I'd assume this extension is more for OpenCL/oneAPI, which I'd think we need Workgroup , CrossWorkGroup, and some other ones fitting OpenCL/oneAPI. (The spec is unclear about this particular part, which could be a good question to ask.)

mlir/lib/Target/SPIRV/Serialization/Serializer.cpp

612

Missing layout here.

mlir/test/Dialect/SPIRV/IR/joint-matrix-ops.mlir

1

This is the test for IR correctness. Please also add a test file under Target/SPIRV/ to test serialization and deserialization.

[antiagainst]

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVAttributes.td

67

Please also put a comment in the above to mention the URL of the spec.

[antiagainst]

mlir/test/Dialect/SPIRV/IR/joint-matrix-ops.mlir

53

Actually these element wise ops are not explicitly said to be allowed. I suspect they should. But good to double check with Intel folks if you can.

[nirvedhmeshram]

Addressing reviewer comments

[nirvedhmeshram]

mlir/lib/Dialect/SPIRV/IR/SPIRVOps.cpp

3924

I did not get a response from intel, https://github.com/intel/llvm/issues/6568 at the moment. I am only allowing Workgroup and CrossWorkGroup for now

mlir/test/Dialect/SPIRV/IR/joint-matrix-ops.mlir

53

Let me know if you want me to remove this support.

[antiagainst]

Awesome, just some final nits!

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td

3836

Nit: Like others, use SPV_ML_* here. (ML is the short for MatrixLayout.)

3836

Could you define these new enum attribute after L3969? Here, this is the section for autogenerated enum attributes; it will be removed once we run the script to refresh definitions.

mlir/lib/Dialect/SPIRV/IR/SPIRVOps.cpp

3924

Okay SG!

mlir/test/Dialect/SPIRV/IR/joint-matrix-ops.mlir

53

Let's wait for Intel's reply. Fine to keep it right now.

[antiagainst]

Awesome! Do you have write access to llvm repo? Do you need me to land it?

[nirvedhmeshram]

In D131586#3724518, @antiagainst wrote:

Awesome! Do you have write access to llvm repo? Do you need me to land it?

yes I do, will land it, thanks.

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td

class SPV_ArithmeticBinaryOp<string mnemonic, Type type,
                        list<Trait> traits = []> :
       // Operands type same as result type.
       SPV_BinaryOp<mnemonic, type, type,
                    !listconcat(traits,
                                [NoSideEffect, SameOperandsAndResultType])> {
   // In addition to normal types arithmetic instructions can support cooperative
   // matrix.
   let arguments = (ins
-    SPV_ScalarOrVectorOrCoopMatrixOf<type>:$operand1,
-    SPV_ScalarOrVectorOrCoopMatrixOf<type>:$operand2
+    SPV_ScalarOrVectorOrCoopMatrixOfOrJointMatrixOf<type>:$operand1,
+    SPV_ScalarOrVectorOrCoopMatrixOfOrJointMatrixOf<type>:$operand2
   );
 
   let results = (outs
-    SPV_ScalarOrVectorOrCoopMatrixOf<type>:$result
+    SPV_ScalarOrVectorOrCoopMatrixOfOrJointMatrixOf<type>:$result
   );
   let assemblyFormat = "operands attr-dict `:` type($result)";
 }
 
 class SPV_ArithmeticUnaryOp<string mnemonic, Type type,
                             list<Trait> traits = []> :
       // Operand type same as result type.
       SPV_UnaryOp<mnemonic, type, type,

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVAttributes.td

def SPV_CooperativeMatrixPropertiesNVAttr :
   );
   let assemblyFormat = "`<` struct(params) `>`";
 }
 
 def SPV_CooperativeMatrixPropertiesNVArrayAttr :
     TypedArrayAttrBase<SPV_CooperativeMatrixPropertiesNVAttr,
                        "CooperativeMatrixPropertiesNV array attribute">;
 
+def SPV_JointMatrixPropertiesINTELAttr :
+    SPV_Attr<"JointMatrixPropertiesINTEL", "joint_matrix_props"> {
+  let parameters = (ins
+    "int":$m_size,
+    "int":$n_size,
+    "int":$k_size,
+    "mlir::Type":$a_type,
+    "mlir::Type":$b_type,
+    "mlir::Type":$c_type,
+    "mlir::Type":$result_type,
+    "mlir::spirv::ScopeAttr":$scope
+  );
+  let assemblyFormat = "`<` struct(params) `>`";
+}
+
+def SPV_JointMatrixPropertiesINTELArrayAttr :
+    TypedArrayAttrBase<SPV_JointMatrixPropertiesINTELAttr,
+                       "JointMatrixPropertiesINTEL array attribute">;
+
 // This attribute specifies the limits for various resources on the target
 // architecture.
 //
 // See https://www.khronos.org/registry/vulkan/specs/1.2-extensions/html/vkspec.html#limits
 // for the complete list of limits and their explanation for the Vulkan API.
 // The following ones are those affecting SPIR-V CodeGen. Their default value
 // are the from Vulkan limit requirements:
 // https://www.khronos.org/registry/vulkan/specs/1.2-extensions/html/vkspec.html#limits-minmax

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVBase.td

 def SPV_INTEL_blocking_pipes                     : I32EnumAttrCase<"SPV_INTEL_blocking_pipes", 4022>;
 def SPV_INTEL_fpga_reg                           : I32EnumAttrCase<"SPV_INTEL_fpga_reg", 4023>;
 def SPV_INTEL_long_constant_composite            : I32EnumAttrCase<"SPV_INTEL_long_constant_composite", 4024>;
 def SPV_INTEL_optnone                            : I32EnumAttrCase<"SPV_INTEL_optnone", 4025>;
 def SPV_INTEL_debug_module                       : I32EnumAttrCase<"SPV_INTEL_debug_module", 4026>;
 def SPV_INTEL_fp_fast_math_mode                  : I32EnumAttrCase<"SPV_INTEL_fp_fast_math_mode", 4027>;
 def SPV_INTEL_memory_access_aliasing             : I32EnumAttrCase<"SPV_INTEL_memory_access_aliasing", 4028>;
 def SPV_INTEL_split_barrier                      : I32EnumAttrCase<"SPV_INTEL_split_barrier", 4029>;
+def SPV_INTEL_joint_matrix                       : I32EnumAttrCase<"SPV_INTEL_joint_matrix", 4030>;
 
 def SPV_NV_compute_shader_derivatives    : I32EnumAttrCase<"SPV_NV_compute_shader_derivatives", 5000>;
 def SPV_NV_cooperative_matrix            : I32EnumAttrCase<"SPV_NV_cooperative_matrix", 5001>;
 def SPV_NV_fragment_shader_barycentric   : I32EnumAttrCase<"SPV_NV_fragment_shader_barycentric", 5002>;
 def SPV_NV_geometry_shader_passthrough   : I32EnumAttrCase<"SPV_NV_geometry_shader_passthrough", 5003>;
 def SPV_NV_mesh_shader                   : I32EnumAttrCase<"SPV_NV_mesh_shader", 5004>;
 def SPV_NV_ray_tracing                   : I32EnumAttrCase<"SPV_NV_ray_tracing", 5005>;
 def SPV_NV_sample_mask_override_coverage : I32EnumAttrCase<"SPV_NV_sample_mask_override_coverage", 5006>;
SPV_I32EnumAttr<"Extension", "supported SPIR-V extensions", "ext", [
       SPV_INTEL_variable_length_array, SPV_INTEL_fpga_memory_attributes,
       SPV_INTEL_unstructured_loop_controls, SPV_INTEL_fpga_loop_controls,
       SPV_INTEL_arbitrary_precision_integers, SPV_INTEL_arbitrary_precision_floating_point,
       SPV_INTEL_kernel_attributes, SPV_INTEL_fpga_memory_accesses,
       SPV_INTEL_fpga_cluster_attributes, SPV_INTEL_loop_fuse,
       SPV_INTEL_fpga_buffer_location, SPV_INTEL_arbitrary_precision_fixed_point,
       SPV_INTEL_usm_storage_classes, SPV_INTEL_io_pipes, SPV_INTEL_blocking_pipes,
       SPV_INTEL_fpga_reg, SPV_INTEL_long_constant_composite, SPV_INTEL_optnone,
-      SPV_INTEL_debug_module, SPV_INTEL_fp_fast_math_mode,
+      SPV_INTEL_debug_module, SPV_INTEL_fp_fast_math_mode, SPV_INTEL_joint_matrix,
       SPV_INTEL_memory_access_aliasing, SPV_INTEL_split_barrier,
       SPV_NV_compute_shader_derivatives, SPV_NV_cooperative_matrix,
       SPV_NV_fragment_shader_barycentric, SPV_NV_geometry_shader_passthrough,
       SPV_NV_mesh_shader, SPV_NV_ray_tracing, SPV_NV_sample_mask_override_coverage,
       SPV_NV_shader_image_footprint, SPV_NV_shader_sm_builtins,
       SPV_NV_shader_subgroup_partitioned, SPV_NV_shading_rate,
       SPV_NV_stereo_view_rendering, SPV_NV_viewport_array2, SPV_NV_bindless_texture,
       SPV_NV_ray_tracing_motion_blur, SPV_NVX_multiview_per_view_attributes

 }
 def SPV_C_ShaderStereoViewNV                          : I32EnumAttrCase<"ShaderStereoViewNV", 5259> {
   list<I32EnumAttrCase> implies = [SPV_C_ShaderViewportMaskNV];
   list<Availability> availability = [
     Extension<[SPV_NV_stereo_view_rendering]>
   ];
 }
 
+def SPV_C_JointMatrixINTEL                         : I32EnumAttrCase<"JointMatrixINTEL", 6118> {
+  list<I32EnumAttrCase> implies = [SPV_C_Shader];
+  list<Availability> availability = [
+    Extension<[SPV_INTEL_joint_matrix]>
+  ];
+}
+
 def SPV_CapabilityAttr :
     SPV_I32EnumAttr<"Capability", "valid SPIR-V Capability", "capability", [
       SPV_C_Matrix, SPV_C_Addresses, SPV_C_Linkage, SPV_C_Kernel, SPV_C_Float16,
       SPV_C_Float64, SPV_C_Int64, SPV_C_Groups, SPV_C_Int16, SPV_C_Int8,
       SPV_C_Sampled1D, SPV_C_SampledBuffer, SPV_C_GroupNonUniform, SPV_C_ShaderLayer,
       SPV_C_ShaderViewportIndex, SPV_C_UniformDecoration, SPV_C_SubgroupBallotKHR,
       SPV_C_SubgroupVoteKHR, SPV_C_StorageBuffer16BitAccess,
       SPV_C_StoragePushConstant16, SPV_C_StorageInputOutput16, SPV_C_DeviceGroup,
SPV_I32EnumAttr<"Capability", "valid SPIR-V Capability", "capability", [
       SPV_C_RayQueryKHR, SPV_C_RayTracingKHR, SPV_C_Float16ImageAMD,
       SPV_C_ImageGatherBiasLodAMD, SPV_C_FragmentMaskAMD, SPV_C_StencilExportEXT,
       SPV_C_ImageReadWriteLodAMD, SPV_C_Int64ImageEXT, SPV_C_ShaderClockKHR,
       SPV_C_FragmentFullyCoveredEXT, SPV_C_MeshShadingNV, SPV_C_FragmentDensityEXT,
       SPV_C_ShaderNonUniform, SPV_C_RuntimeDescriptorArray,
       SPV_C_StorageTexelBufferArrayDynamicIndexing, SPV_C_RayTracingNV,
       SPV_C_RayTracingMotionBlurNV, SPV_C_PhysicalStorageBufferAddresses,
       SPV_C_RayTracingProvisionalKHR, SPV_C_CooperativeMatrixNV,
+      SPV_C_JointMatrixINTEL,
       SPV_C_FragmentShaderSampleInterlockEXT,
       SPV_C_FragmentShaderShadingRateInterlockEXT, SPV_C_ShaderSMBuiltinsNV,
       SPV_C_FragmentShaderPixelInterlockEXT, SPV_C_DemoteToHelperInvocation,
       SPV_C_IntegerFunctions2INTEL, SPV_C_TessellationPointSize,
       SPV_C_GeometryPointSize, SPV_C_ImageCubeArray, SPV_C_ImageRect,
       SPV_C_GeometryStreams, SPV_C_MultiViewport,
       SPV_C_WorkgroupMemoryExplicitLayout8BitAccessKHR, SPV_C_VariablePointers,
       SPV_C_RayTraversalPrimitiveCullingKHR, SPV_C_SampleMaskOverrideCoverageNV,

 
 def SPV_ScopeAttr :
     SPV_I32EnumAttr<"Scope", "valid SPIR-V Scope", "scope", [
       SPV_S_CrossDevice, SPV_S_Device, SPV_S_Workgroup, SPV_S_Subgroup,
       SPV_S_Invocation, SPV_S_QueueFamily, SPV_S_ShaderCallKHR
     ]>;
 
 def SPV_SC_None        : I32BitEnumAttrCaseNone<"None">;
 def SPV_SC_Flatten     : I32BitEnumAttrCaseBit<"Flatten", 0>;
 def SPV_SC_DontFlatten : I32BitEnumAttrCaseBit<"DontFlatten", 1>;
 
 def SPV_SelectionControlAttr :
     SPV_BitEnumAttr<"SelectionControl", "valid SPIR-V SelectionControl", "selection_control", [
       SPV_SC_None, SPV_SC_Flatten, SPV_SC_DontFlatten
     ]>;
 
 def SPV_SC_UniformConstant         : I32EnumAttrCase<"UniformConstant", 0>;

 class SignlessOrUnsignedIntOfWidths<list<int> widths> :
     AnyTypeOf<!foreach(w, widths, IOrUI<w>),
               !interleave(widths, "/") # "-bit signless/unsigned integer">;
 
 def SPV_IsArrayType : CPred<"$_self.isa<::mlir::spirv::ArrayType>()">;
 def SPV_IsCooperativeMatrixType :
   CPred<"$_self.isa<::mlir::spirv::CooperativeMatrixNVType>()">;
 def SPV_IsImageType : CPred<"$_self.isa<::mlir::spirv::ImageType>()">;
+def SPV_IsJointMatrixType :
+  CPred<"$_self.isa<::mlir::spirv::JointMatrixINTELType>()">;
 def SPV_IsMatrixType : CPred<"$_self.isa<::mlir::spirv::MatrixType>()">;
 def SPV_IsPtrType : CPred<"$_self.isa<::mlir::spirv::PointerType>()">;
 def SPV_IsRTArrayType : CPred<"$_self.isa<::mlir::spirv::RuntimeArrayType>()">;
 def SPV_IsSampledImageType : CPred<"$_self.isa<::mlir::spirv::SampledImageType>()">;
 def SPV_IsStructType : CPred<"$_self.isa<::mlir::spirv::StructType>()">;
 
 
 // See https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#_types
def SPV_AnyPtr : DialectType<SPIRV_Dialect, SPV_IsPtrType,
                              "any SPIR-V pointer type">;
 def SPV_AnyArray : DialectType<SPIRV_Dialect, SPV_IsArrayType,
                                "any SPIR-V array type">;
 def SPV_AnyCooperativeMatrix : DialectType<SPIRV_Dialect,
                                SPV_IsCooperativeMatrixType,
                                "any SPIR-V cooperative matrix type">;
 def SPV_AnyImage : DialectType<SPIRV_Dialect, SPV_IsImageType,
                                 "any SPIR-V image type">;
+def SPV_AnyJointMatrix : DialectType<SPIRV_Dialect,
+                               SPV_IsJointMatrixType,
+                               "any SPIR-V joint matrix type">;
 def SPV_AnyMatrix : DialectType<SPIRV_Dialect, SPV_IsMatrixType,
                                 "any SPIR-V matrix type">;
 def SPV_AnyRTArray : DialectType<SPIRV_Dialect, SPV_IsRTArrayType,
                                  "any SPIR-V runtime array type">;
 def SPV_AnyStruct : DialectType<SPIRV_Dialect, SPV_IsStructType,
                                 "any SPIR-V struct type">;
 def SPV_AnySampledImage : DialectType<SPIRV_Dialect, SPV_IsSampledImageType,
                                 "any SPIR-V sampled image type">;
 
 def SPV_Numerical : AnyTypeOf<[SPV_Integer, SPV_Float]>;
 def SPV_Scalar : AnyTypeOf<[SPV_Numerical, SPV_Bool]>;
 def SPV_Aggregate : AnyTypeOf<[SPV_AnyArray, SPV_AnyRTArray, SPV_AnyStruct]>;
 def SPV_Composite :
     AnyTypeOf<[SPV_Vector, SPV_AnyArray, SPV_AnyRTArray, SPV_AnyStruct,
-               SPV_AnyCooperativeMatrix, SPV_AnyMatrix]>;
+               SPV_AnyCooperativeMatrix, SPV_AnyJointMatrix, SPV_AnyMatrix]>;
 def SPV_Type : AnyTypeOf<[
     SPV_Void, SPV_Bool, SPV_Integer, SPV_Float, SPV_Vector,
     SPV_AnyPtr, SPV_AnyArray, SPV_AnyRTArray, SPV_AnyStruct,
-    SPV_AnyCooperativeMatrix, SPV_AnyMatrix, SPV_AnySampledImage
+    SPV_AnyCooperativeMatrix, SPV_AnyJointMatrix, SPV_AnyMatrix,
+    SPV_AnySampledImage
   ]>;
 
 def SPV_SignedInt : SignedIntOfWidths<[8, 16, 32, 64]>;
 def SPV_SignlessOrUnsignedInt : SignlessOrUnsignedIntOfWidths<[8, 16, 32, 64]>;
 
 class SPV_CoopMatrixOfType<list<Type> allowedTypes> :
   ContainerType<AnyTypeOf<allowedTypes>, SPV_IsCooperativeMatrixType,
     "$_self.cast<::mlir::spirv::CooperativeMatrixNVType>().getElementType()",
     "Cooperative Matrix">;
 
+class SPV_JointMatrixOfType<list<Type> allowedTypes> :
+  ContainerType<AnyTypeOf<allowedTypes>, SPV_IsJointMatrixType,
+    "$_self.cast<::mlir::spirv::JointMatrixINTELType>().getElementType()",
+    "Joint Matrix">;
+
 class SPV_ScalarOrVectorOf<Type type> :
     AnyTypeOf<[type, VectorOfLengthAndType<[2, 3, 4, 8, 16], [type]>]>;
 
 class SPV_ScalarOrVectorOrCoopMatrixOf<Type type> :
     AnyTypeOf<[type, VectorOfLengthAndType<[2, 3, 4, 8, 16], [type]>,
                SPV_CoopMatrixOfType<[type]>]>;
 
+class SPV_ScalarOrVectorOrJointMatrixOf<Type type> :
+    AnyTypeOf<[type, VectorOfLengthAndType<[2, 3, 4, 8, 16], [type]>,
+               SPV_JointMatrixOfType<[type]>]>;
+
+class SPV_ScalarOrVectorOrCoopMatrixOfOrJointMatrixOf<Type type> :
+    AnyTypeOf<[type, VectorOfLengthAndType<[2, 3, 4, 8, 16], [type]>,
+               SPV_CoopMatrixOfType<[type]>, SPV_JointMatrixOfType<[type]> ]>;
+
 def SPV_ScalarOrVector : AnyTypeOf<[SPV_Scalar, SPV_Vector]>;
 def SPV_ScalarOrVectorOrPtr : AnyTypeOf<[SPV_ScalarOrVector, SPV_AnyPtr]>;
 
 class SPV_Vec4<Type type> : VectorOfLengthAndType<[4], [type]>;
 def SPV_IntVec4 : SPV_Vec4<SPV_Integer>;
 def SPV_IOrUIVec4 : SPV_Vec4<SPV_SignlessOrUnsignedInt>;
 def SPV_Int32Vec4 : SPV_Vec4<AnyI32>;
 

 def SPV_OC_OpCooperativeMatrixLoadNV   : I32EnumAttrCase<"OpCooperativeMatrixLoadNV", 5359>;
 def SPV_OC_OpCooperativeMatrixStoreNV  : I32EnumAttrCase<"OpCooperativeMatrixStoreNV", 5360>;
 def SPV_OC_OpCooperativeMatrixMulAddNV : I32EnumAttrCase<"OpCooperativeMatrixMulAddNV", 5361>;
 def SPV_OC_OpCooperativeMatrixLengthNV : I32EnumAttrCase<"OpCooperativeMatrixLengthNV", 5362>;
 def SPV_OC_OpSubgroupBlockReadINTEL    : I32EnumAttrCase<"OpSubgroupBlockReadINTEL", 5575>;
 def SPV_OC_OpSubgroupBlockWriteINTEL   : I32EnumAttrCase<"OpSubgroupBlockWriteINTEL", 5576>;
 def SPV_OC_OpAssumeTrueKHR             : I32EnumAttrCase<"OpAssumeTrueKHR", 5630>;
 def SPV_OC_OpAtomicFAddEXT             : I32EnumAttrCase<"OpAtomicFAddEXT", 6035>;
+def SPV_OC_OpTypeJointMatrixINTEL      : I32EnumAttrCase<"OpTypeJointMatrixINTEL", 6119>;
+def SPV_OC_OpJointMatrixLoadINTEL      : I32EnumAttrCase<"OpJointMatrixLoadINTEL", 6120>;
+def SPV_OC_OpJointMatrixStoreINTEL     : I32EnumAttrCase<"OpJointMatrixStoreINTEL", 6121>;
+def SPV_OC_OpJointMatrixMadINTEL       : I32EnumAttrCase<"OpJointMatrixMadINTEL", 6122>;
+def SPV_OC_OpTypejointMatrixWorkItemLengthINTEL : I32EnumAttrCase<"OpJointMatrixWorkItemLengthINTEL", 6410>;
 
 def SPV_OpcodeAttr :
     SPV_I32EnumAttr<"Opcode", "valid SPIR-V instructions", "opcode", [
       SPV_OC_OpNop, SPV_OC_OpUndef, SPV_OC_OpSourceContinued, SPV_OC_OpSource,
       SPV_OC_OpSourceExtension, SPV_OC_OpName, SPV_OC_OpMemberName, SPV_OC_OpString,
       SPV_OC_OpLine, SPV_OC_OpExtension, SPV_OC_OpExtInstImport, SPV_OC_OpExtInst,
       SPV_OC_OpMemoryModel, SPV_OC_OpEntryPoint, SPV_OC_OpExecutionMode,
       SPV_OC_OpCapability, SPV_OC_OpTypeVoid, SPV_OC_OpTypeBool, SPV_OC_OpTypeInt,
SPV_I32EnumAttr<"Opcode", "valid SPIR-V instructions", "opcode", [
       SPV_OC_OpGroupNonUniformIMul, SPV_OC_OpGroupNonUniformFMul,
       SPV_OC_OpGroupNonUniformSMin, SPV_OC_OpGroupNonUniformUMin,
       SPV_OC_OpGroupNonUniformFMin, SPV_OC_OpGroupNonUniformSMax,
       SPV_OC_OpGroupNonUniformUMax, SPV_OC_OpGroupNonUniformFMax,
       SPV_OC_OpSubgroupBallotKHR, SPV_OC_OpTypeCooperativeMatrixNV,
       SPV_OC_OpCooperativeMatrixLoadNV, SPV_OC_OpCooperativeMatrixStoreNV,
       SPV_OC_OpCooperativeMatrixMulAddNV, SPV_OC_OpCooperativeMatrixLengthNV,
       SPV_OC_OpSubgroupBlockReadINTEL, SPV_OC_OpSubgroupBlockWriteINTEL,
-      SPV_OC_OpAssumeTrueKHR, SPV_OC_OpAtomicFAddEXT
+      SPV_OC_OpAssumeTrueKHR, SPV_OC_OpAtomicFAddEXT,
+      SPV_OC_OpTypeJointMatrixINTEL, SPV_OC_OpJointMatrixLoadINTEL,
+      SPV_OC_OpJointMatrixStoreINTEL, SPV_OC_OpJointMatrixMadINTEL,
+      SPV_OC_OpTypejointMatrixWorkItemLengthINTEL
     ]>;
 
 // End opcode section. Generated from SPIR-V spec; DO NOT MODIFY!
 
 //===----------------------------------------------------------------------===//
 // SPIR-V op definitions
 //===----------------------------------------------------------------------===//
 

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVCastOps.td

 include "mlir/Interfaces/SideEffectInterfaces.td"
 
 class SPV_CastOp<string mnemonic, Type resultType, Type operandType,
                  list<Trait> traits = []> :
       SPV_Op<mnemonic,
              !listconcat(traits,
                          [NoSideEffect, SameOperandsAndResultShape])> {
   let arguments = (ins
-    SPV_ScalarOrVectorOrCoopMatrixOf<operandType>:$operand
+    SPV_ScalarOrVectorOrCoopMatrixOfOrJointMatrixOf<operandType>:$operand
   );
 
   let results = (outs
-    SPV_ScalarOrVectorOrCoopMatrixOf<resultType>:$result
+    SPV_ScalarOrVectorOrCoopMatrixOfOrJointMatrixOf<resultType>:$result
   );
   let assemblyFormat = [{
     $operand attr-dict `:` type($operand) `to` type($result)
   }];
 }
 
 // -----
 

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVJointMatrixOps.td

+//===- SPIRVJointMatrixOps.td - joint matmul ---*- tablegen -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This is the op definition spec of joint matrix multiply extension ops.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_DIALECT_SPIRV_IR_JOINT_MATRIX_OPS
+#define MLIR_DIALECT_SPIRV_IR_JOINT_MATRIX_OPS
+
+// -----
+
+def SPV_JointMatrixWorkItemLengthINTELOp : SPV_Op<"JointMatrixWorkItemLengthINTEL",
+  [NoSideEffect]> {
+  let summary = "See extension SPV_INTEL_joint_matrix";
+
+  let description = [{
+    Return number of components owned by the current work-item in 
+    a joint matrix.
+
+    Result Type must be an 32-bit unsigned integer type scalar.
+
+    Type is a joint matrix type.
+
+    ``` {.ebnf}
+    joint-matrix-length-op ::= ssa-id `=` `spv.JointMatrixWorkItemLengthINTEL
+                                    ` : ` joint-matrix-type
+    ```
+
+    For example:
+
+    ```
+    %0 = spv.JointMatrixWorkItemLengthINTEL : !spv.jointmatrix<Subgroup, i32, 8, 16>
+    ```
+  }];
+
+  let assemblyFormat = "attr-dict `:` $type";
+
+  let availability = [
+    MinVersion<SPV_V_1_0>,
+    MaxVersion<SPV_V_1_5>,
+    Extension<[SPV_INTEL_joint_matrix]>,
+    Capability<[SPV_C_JointMatrixINTEL]>
+  ];
+
+  let arguments = (ins
+    TypeAttr:$type
+  );
+
+  let results = (outs
+    SPV_Int32:$result
+  );
+  let hasVerifier = 0;
+}
+
+// -----
+
+def SPV_JointMatrixLoadINTELOp : SPV_Op<"JointMatrixLoadINTEL", []> {
+  let summary = "See extension SPV_INTEL_joint_matrix";
+
+  let description = [{
+    Declare a matrix type.
+
+    Component Type is the type of each component in the resulting type. 
+    It must be a scalar numerical type.
+
+    Row Count is the number of rows in the matrix type. It must be a 
+    constant unsigned 32-bit integer. Behavior is undefined when Row 
+    Count is 0 or OpConstantNull.
+
+    Column Count is the number of columns in the matrix type. It must 
+    be a constant unsigned 32-bit integer. Behavior is undefined when 
+    Column Count is 0 or OpConstantNull.
+
+    Layout indicates how the values are arranged internally in the 
+    matrix type. It must be the result of a constant instruction.
+
+    Scope is memory scope for operations on the matrix. It must be the
+    result of a constant instruction with scalar integer type.
+
+    ### Custom assembly form
+
+    ``` {.ebnf}
+    joint-matrixload-op ::= ssa-id `=` `spv.JointMatrixLoadINTEL`
+                              ssa-use `,` ssa-use `,` ssa-use `,
+                              ` ssa-use
+                              (`[` memory-access `]`)? ` : `
+                              pointer-type `as`
+                              joint-matrix-type
+    ```
+
+    For example:
+
+    ```
+    %0 = spv.JointMatrixLoadINTEL %ptr, %stride, %Layout, %Scope
+         : !spv.ptr<i32, StorageBuffer> as !spv.jointmatrix<i32, Workgroup, 16, 8>
+    ```
+  }];
+
+  let assemblyFormat = [{
+    $scope operands attr-dict `:` `(` type(operands) `)` `->` type($result)
+  }];
+
+  let availability = [
+    MinVersion<SPV_V_1_0>,
+    MaxVersion<SPV_V_1_5>,
+    Extension<[SPV_INTEL_joint_matrix]>,
+    Capability<[SPV_C_JointMatrixINTEL]>
+  ];
+
+  let arguments = (ins
+    SPV_ScopeAttr:$scope,
+    SPV_AnyPtr:$pointer,
+    SPV_Integer:$stride,
+    SPV_Integer:$layout,
+    OptionalAttr<SPV_MemoryAccessAttr>:$memory_access
+  );
+
+  let results = (outs
+    SPV_AnyJointMatrix:$result
+  );
+}
+
+// -----
+
+def SPV_JointMatrixMadINTELOp : SPV_Op<"JointMatrixMadINTEL",
+  [NoSideEffect, AllTypesMatch<["c", "result"]>]> {
+  let summary = "See extension SPV_INTEL_joint_matrix";
+
+  let description = [{
+    Multiply matrix A by matrix B and add matrix C to the result 
+    of the multiplication: A*B+C. Here A is a M x K matrix, B is 
+    a K x N matrix and C is a M x N matrix.
+
+    Behavior is undefined if sizes of operands do not meet the 
+    conditions above. All operands and the Result Type must be 
+    OpTypeJointMatrixINTEL.
+
+    A must be a OpTypeJointMatrixINTEL whose Component Type is a 
+    signed numerical type, Row Count equals to M and Column Count 
+    equals to K
+
+    B must be a OpTypeJointMatrixINTEL whose Component Type is a 
+    signed numerical type, Row Count equals to K and Column Count 
+    equals to N
+
+    C and Result Type must be a OpTypeJointMatrixINTEL with Row 
+    Count equals to M and Column Count equals to N
+
+    Scope is syncronization scope for operation on the matrix. 
+    It must be the result of a constant instruction with scalar 
+    integer type.
+
+    ``` {.ebnf}
+    joint-matrixmuladd-op ::= ssa-id `=` `spv.JointMatrixMulAddINTEL`
+                              ssa-use `,` ssa-use `,` ssa-use `
+                              ,` ssa_use` : `
+                              a-joint-matrix-type,
+                              b-joint-matrix-type ->
+                              result-joint-matrix-type
+    ```
+    For example:
+
+    ```
+    %0 = spv.JointMatrixMulAddINTEL %arg0, %arg1, %arg2, %scope  :
+      !spv.jointmatrix<Subgroup, i32, 8, 16>
+    ```
+  }];
+
+  let assemblyFormat = [{
+    $scope operands attr-dict`:` type($a) `,` type($b) `->` type($c)
+  }];
+
+  let availability = [
+    MinVersion<SPV_V_1_0>,
+    MaxVersion<SPV_V_1_5>,
+    Extension<[SPV_INTEL_joint_matrix]>,
+    Capability<[SPV_C_JointMatrixINTEL]>
+  ];
+
+  let arguments = (ins
+    SPV_ScopeAttr:$scope,
+    SPV_AnyJointMatrix:$a,
+    SPV_AnyJointMatrix:$b,
+    SPV_AnyJointMatrix:$c
+  );
+
+  let results = (outs
+    SPV_AnyJointMatrix:$result
+  );
+}
+
+// -----
+
+def SPV_JointMatrixStoreINTELOp : SPV_Op<"JointMatrixStoreINTEL", []> {
+  let summary = "See extension SPV_INTEL_joint_matrix";
+
+  let description = [{
+    Store a matrix through a pointer.
+
+    Pointer is the pointer to store through. It specifies 
+    start of memory region where elements of the matrix must 
+    be stored and arranged according to Layout.
+
+    Object is the matrix to store. It must be 
+    OpTypeJointMatrixINTEL.
+
+    Stride is the number of elements in memory between beginnings 
+    of successive rows, columns (or words) of the Object. It must 
+    be a scalar integer type.
+
+    Layout indicates how the values stored to memory are arranged. 
+    It must be the result of a constant instruction.
+
+    Scope is syncronization scope for operation on the matrix. 
+    It must be the result of a constant instruction with scalar 
+    integer type.
+
+    If present, any Memory Operands must begin with a memory operand 
+    literal. If not present, it is the same as specifying the memory 
+    operand None.
+
+    ``` {.ebnf}
+    joint-matrix-store-op ::= `spv.JointMatrixStoreINTEL `
+                              ssa-use `, ` ssa-use `, `
+                              ssa-use `, ` ssa-use `, `
+                              ssa-use `,  `(`[` memory-access `]`)? `:`
+                              pointer-type `,` spirv-element-type
+    ```
+
+    For example:
+
+    ```
+      spv.JointMatrixStoreINTEL %arg0, %arg2, %arg1, %arg3, %arg4 :
+        !spv.ptr<i32, StorageBuffer>, !spv.jointmatrix<Workgroup, i32, 16, 8>
+    ```
+  }];
+
+   let assemblyFormat = [{
+    $scope operands attr-dict `:` `(` type(operands) `)`
+  }];
+
+  let availability = [
+    MinVersion<SPV_V_1_0>,
+    MaxVersion<SPV_V_1_5>,
+    Extension<[SPV_INTEL_joint_matrix]>,
+    Capability<[SPV_C_JointMatrixINTEL]>
+  ];
+
+  let arguments = (ins
+    SPV_ScopeAttr:$scope,
+    SPV_AnyPtr:$pointer,
+    SPV_AnyJointMatrix:$object,
+    SPV_Integer:$stride,
+    SPV_Integer:$layout,
+    OptionalAttr<SPV_MemoryAccessAttr>:$memory_access
+  );
+
+  let results = (outs);
+}
+
+// -----
+
+#endif // MLIR_DIALECT_SPIRV_IR_JOINT_MATRIX_OPS

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVOps.td

 include "mlir/Dialect/SPIRV/IR/SPIRVArithmeticOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVAtomicOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVBarrierOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVBitOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVCastOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVCompositeOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVControlFlowOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVCooperativeMatrixOps.td"
+include "mlir/Dialect/SPIRV/IR/SPIRVJointMatrixOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVGLOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVGroupOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVImageOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVLogicalOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVMatrixOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVMemoryOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVMiscOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVNonUniformOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVCLOps.td"
 include "mlir/Dialect/SPIRV/IR/SPIRVStructureOps.td"
 include "mlir/Interfaces/SideEffectInterfaces.td"
 
 #endif // MLIR_DIALECT_SPIRV_IR_OPS

mlir/include/mlir/Dialect/SPIRV/IR/SPIRVTypes.h

 
 namespace mlir {
 namespace spirv {
 
 namespace detail {
 struct ArrayTypeStorage;
 struct CooperativeMatrixTypeStorage;
 struct ImageTypeStorage;
+struct JointMatrixTypeStorage;
 struct MatrixTypeStorage;
 struct PointerTypeStorage;
 struct RuntimeArrayTypeStorage;
 struct SampledImageTypeStorage;
 struct StructTypeStorage;
 
 } // namespace detail
 
public:
   unsigned getColumns() const;
 
   void getExtensions(SPIRVType::ExtensionArrayRefVector &extensions,
                      Optional<StorageClass> storage = llvm::None);
   void getCapabilities(SPIRVType::CapabilityArrayRefVector &capabilities,
                        Optional<StorageClass> storage = llvm::None);
 };
 
+// SPIR-V joint matrix type
+class JointMatrixINTELType
+    : public Type::TypeBase<JointMatrixINTELType, CompositeType,
+                            detail::JointMatrixTypeStorage> {
+public:
+  using Base::Base;
+
+  static JointMatrixINTELType get(Type elementType, Scope scope, unsigned rows,
+                                  unsigned columns);
+  Type getElementType() const;
+
+  /// Return the scope of the joint matrix.
+  Scope getScope() const;
+  /// return the number of rows of the matrix.
+  unsigned getRows() const;
+  /// return the number of columns of the matrix.
+  unsigned getColumns() const;
+
+  void getExtensions(SPIRVType::ExtensionArrayRefVector &extensions,
+                     Optional<StorageClass> storage = llvm::None);
+  void getCapabilities(SPIRVType::CapabilityArrayRefVector &capabilities,
+                       Optional<StorageClass> storage = llvm::None);
+};
+
 // SPIR-V matrix type
 class MatrixType : public Type::TypeBase<MatrixType, CompositeType,
                                          detail::MatrixTypeStorage> {
 public:
   using Base::Base;
 
   static MatrixType get(Type columnType, uint32_t columnCount);
 

mlir/lib/Dialect/SPIRV/IR/SPIRVDialect.cpp

static Type parseCooperativeMatrixType(SPIRVDialect const &dialect,
   if (parser.parseComma() || parseEnumKeywordAttr(scope, parser, "scope <id>"))
     return Type();
 
   if (parser.parseGreater())
     return Type();
   return CooperativeMatrixNVType::get(elementTy, scope, dims[0], dims[1]);
 }
 
+// joint-matrix-type ::= `!spv.jointmatrix` `<` element-type ',' scope ','
+//                                                   rows ',' columns>`
+static Type parseJointMatrixType(SPIRVDialect const &dialect,
+                                 DialectAsmParser &parser) {
+  if (parser.parseLess())
+    return Type();
+
+  SmallVector<int64_t, 2> dims;
+  SMLoc countLoc = parser.getCurrentLocation();
+  if (parser.parseDimensionList(dims, /*allowDynamic=*/false))
+    return Type();
+
+  if (dims.size() != 2) {
+    parser.emitError(countLoc, "expected rows and columns size");
+    return Type();
+  }
+
+  auto elementTy = parseAndVerifyType(dialect, parser);
+  if (!elementTy)
+    return Type();
+
+  Scope scope;
+  if (parser.parseComma() || parseEnumKeywordAttr(scope, parser, "scope <id>"))
+    return Type();
+
+  if (parser.parseGreater())
+    return Type();
+  return JointMatrixINTELType::get(elementTy, scope, dims[0], dims[1]);
+}
+
 // TODO: Reorder methods to be utilities first and parse*Type
 // methods in alphabetical order
 //
 // storage-class ::= `UniformConstant`
 //                 | `Uniform`
 //                 | `Workgroup`
 //                 | <and other storage classes...>
 //
Type SPIRVDialect::parseType(DialectAsmParser &parser) const {
   StringRef keyword;
   if (parser.parseKeyword(&keyword))
     return Type();
 
   if (keyword == "array")
     return parseArrayType(*this, parser);
   if (keyword == "coopmatrix")
     return parseCooperativeMatrixType(*this, parser);
+  if (keyword == "jointmatrix")
+    return parseJointMatrixType(*this, parser);
   if (keyword == "image")
     return parseImageType(*this, parser);
   if (keyword == "ptr")
     return parsePointerType(*this, parser);
   if (keyword == "rtarray")
     return parseRuntimeArrayType(*this, parser);
   if (keyword == "sampled_image")
     return parseSampledImageType(*this, parser);

 }
 
 static void print(CooperativeMatrixNVType type, DialectAsmPrinter &os) {
   os << "coopmatrix<" << type.getRows() << "x" << type.getColumns() << "x";
   os << type.getElementType() << ", " << stringifyScope(type.getScope());
   os << ">";
 }
 
+static void print(JointMatrixINTELType type, DialectAsmPrinter &os) {
+  os << "jointmatrix<" << type.getRows() << "x" << type.getColumns() << "x";
+  os << type.getElementType() << ", " << stringifyScope(type.getScope());
+  os << ">";
+}
+
 static void print(MatrixType type, DialectAsmPrinter &os) {
   os << "matrix<" << type.getNumColumns() << " x " << type.getColumnType();
   os << ">";
 }
 
 void SPIRVDialect::printType(Type type, DialectAsmPrinter &os) const {
   TypeSwitch<Type>(type)
-      .Case<ArrayType, CooperativeMatrixNVType, PointerType, RuntimeArrayType,
-            ImageType, SampledImageType, StructType, MatrixType>(
-          [&](auto type) { print(type, os); })
+      .Case<ArrayType, CooperativeMatrixNVType, JointMatrixINTELType,
+            PointerType, RuntimeArrayType, ImageType, SampledImageType,
+            StructType, MatrixType>([&](auto type) { print(type, os); })
       .Default([](Type) { llvm_unreachable("unhandled SPIR-V type"); });
 }
 
 //===----------------------------------------------------------------------===//
 // Constant
 //===----------------------------------------------------------------------===//
 
 Operation *SPIRVDialect::materializeConstant(OpBuilder &builder,

mlir/lib/Dialect/SPIRV/IR/SPIRVOps.cpp

static LogicalResult verifyCastOp(Operation *op,
 
   if (auto coopMatrixType =
           operandType.dyn_cast<spirv::CooperativeMatrixNVType>()) {
     operandType = coopMatrixType.getElementType();
     resultType =
         resultType.cast<spirv::CooperativeMatrixNVType>().getElementType();
   }
 
+  if (auto jointMatrixType =
+          operandType.dyn_cast<spirv::JointMatrixINTELType>()) {
+    operandType = jointMatrixType.getElementType();
+    resultType =
+        resultType.cast<spirv::JointMatrixINTELType>().getElementType();
+  }
+
   auto operandTypeBitWidth = operandType.getIntOrFloatBitWidth();
   auto resultTypeBitWidth = resultType.getIntOrFloatBitWidth();
   auto isSameBitWidth = operandTypeBitWidth == resultTypeBitWidth;
 
   if (requireSameBitWidth) {
     if (!isSameBitWidth) {
       return op->emitOpError(
                  "expected the same bit widths for operand type and result "
if (constituents.size() != 1)
              << "1, but provided " << constituents.size();
     if (coopType.getElementType() != constituents.front().getType())
       return emitOpError("operand type mismatch: expected operand type ")
              << coopType.getElementType() << ", but provided "
              << constituents.front().getType();
     return success();
   }
 
+  if (auto jointType = cType.dyn_cast<spirv::JointMatrixINTELType>()) {
+    if (constituents.size() != 1)
+      return emitOpError("has incorrect number of operands: expected ")
+             << "1, but provided " << constituents.size();
+    if (jointType.getElementType() != constituents.front().getType())
+      return emitOpError("operand type mismatch: expected operand type ")
+             << jointType.getElementType() << ", but provided "
+             << constituents.front().getType();
+    return success();
+  }
+
   if (constituents.size() == cType.getNumElements()) {
     for (auto index : llvm::seq<uint32_t>(0, constituents.size())) {
       if (constituents[index].getType() != cType.getElementType(index)) {
         return emitOpError("operand type mismatch: expected operand type ")
                << cType.getElementType(index) << ", but provided "
                << constituents[index].getType();
       }
     }
if (typeA.getElementType() != typeB.getElementType() ||
     return op.emitOpError("matrix element type must match");
   return success();
 }
 
 LogicalResult spirv::CooperativeMatrixMulAddNVOp::verify() {
   return verifyCoopMatrixMulAdd(*this);
 }
 
+static LogicalResult
+verifyPointerAndJointMatrixType(Operation *op, Type pointer, Type jointMatrix) {
+  Type pointeeType = pointer.cast<spirv::PointerType>().getPointeeType();
+  if (!pointeeType.isa<spirv::ScalarType>() && !pointeeType.isa<VectorType>())
+    return op->emitError(
+               "Pointer must point to a scalar or vector type but provided ")
+           << pointeeType;
+  spirv::StorageClass storage =
+      pointer.cast<spirv::PointerType>().getStorageClass();
+  if (storage != spirv::StorageClass::Workgroup &&
+      storage != spirv::StorageClass::StorageBuffer &&
+      storage != spirv::StorageClass::PhysicalStorageBuffer)
+    return op->emitError(
+               "Pointer storage class must be Workgroup, StorageBuffer or "
+               "PhysicalStorageBufferEXT but provided ")
+           << stringifyStorageClass(storage);
+  return success();
+}
+
+//===----------------------------------------------------------------------===//
+// spv.JointMatrixLoadINTEL
+//===----------------------------------------------------------------------===//
+
+LogicalResult spirv::JointMatrixLoadINTELOp::verify() {
+  return verifyPointerAndJointMatrixType(*this, pointer().getType(),
+                                         result().getType());
+}
+
+//===----------------------------------------------------------------------===//
+// spv.JointMatrixStoreINTEL
+//===----------------------------------------------------------------------===//
+
+LogicalResult spirv::JointMatrixStoreINTELOp::verify() {
+  return verifyPointerAndJointMatrixType(*this, pointer().getType(),
+                                         object().getType());
+}
+
+//===----------------------------------------------------------------------===//
+// spv.JointMatrixMadINTEL
+//===----------------------------------------------------------------------===//
+
+static LogicalResult verifyJointMatrixMad(spirv::JointMatrixMadINTELOp op) {
+  if (op.c().getType() != op.result().getType())
+    return op.emitOpError("result and third operand must have the same type");
+  auto typeA = op.a().getType().cast<spirv::JointMatrixINTELType>();
+  auto typeB = op.b().getType().cast<spirv::JointMatrixINTELType>();
+  auto typeC = op.c().getType().cast<spirv::JointMatrixINTELType>();
+  auto typeR = op.result().getType().cast<spirv::JointMatrixINTELType>();
+  if (typeA.getRows() != typeR.getRows() ||
+      typeA.getColumns() != typeB.getRows() ||
+      typeB.getColumns() != typeR.getColumns())
+    return op.emitOpError("matrix size must match");
+  if (typeR.getScope() != typeA.getScope() ||
+      typeR.getScope() != typeB.getScope() ||
+      typeR.getScope() != typeC.getScope())
+    return op.emitOpError("matrix scope must match");
+  if (typeA.getElementType() != typeB.getElementType() ||
+      typeR.getElementType() != typeC.getElementType())
+    return op.emitOpError("matrix element type must match");
+  return success();
+}
+
+LogicalResult spirv::JointMatrixMadINTELOp::verify() {
+  return verifyJointMatrixMad(*this);
+}
+
 //===----------------------------------------------------------------------===//
 // spv.MatrixTimesScalar
 //===----------------------------------------------------------------------===//
 
 LogicalResult spirv::MatrixTimesScalarOp::verify() {
   // We already checked that result and matrix are both of matrix type in the
   // auto-generated verify method.
 
LogicalResult spirv::SpecConstantCompositeOp::verify() {
   auto constituents = this->constituents().getValue();
 
   if (!cType)
     return emitError("result type must be a composite type, but provided ")
            << type();
 
   if (cType.isa<spirv::CooperativeMatrixNVType>())
     return emitError("unsupported composite type  ") << cType;
+  if (cType.isa<spirv::JointMatrixINTELType>())
+    return emitError("unsupported composite type  ") << cType;
   if (constituents.size() != cType.getNumElements())
     return emitError("has incorrect number of operands: expected ")
            << cType.getNumElements() << ", but provided "
            << constituents.size();
 
   for (auto index : llvm::seq<uint32_t>(0, constituents.size())) {
     auto constituent = constituents[index].cast<FlatSymbolRefAttr>();
 

mlir/lib/Dialect/SPIRV/IR/SPIRVTypes.cpp

 
 //===----------------------------------------------------------------------===//
 // CompositeType
 //===----------------------------------------------------------------------===//
 
 bool CompositeType::classof(Type type) {
   if (auto vectorType = type.dyn_cast<VectorType>())
     return isValid(vectorType);
-  return type
-      .isa<spirv::ArrayType, spirv::CooperativeMatrixNVType, spirv::MatrixType,
+  return type.isa<spirv::ArrayType, spirv::CooperativeMatrixNVType,
+                  spirv::JointMatrixINTELType, spirv::MatrixType,
                   spirv::RuntimeArrayType, spirv::StructType>();
 }
 
 bool CompositeType::isValid(VectorType type) {
   switch (type.getNumElements()) {
   case 2:
   case 3:
   case 4:
   case 8:
   case 16:
     break;
   default:
     return false;
   }
   return type.getRank() == 1 && type.getElementType().isa<ScalarType>();
 }
 
 Type CompositeType::getElementType(unsigned index) const {
   return TypeSwitch<Type, Type>(*this)
-      .Case<ArrayType, CooperativeMatrixNVType, RuntimeArrayType, VectorType>(
+      .Case<ArrayType, CooperativeMatrixNVType, JointMatrixINTELType,
+            RuntimeArrayType, VectorType>(
           [](auto type) { return type.getElementType(); })
       .Case<MatrixType>([](MatrixType type) { return type.getColumnType(); })
       .Case<StructType>(
           [index](StructType type) { return type.getElementType(index); })
       .Default(
           [](Type) -> Type { llvm_unreachable("invalid composite type"); });
 }
 
 unsigned CompositeType::getNumElements() const {
   if (auto arrayType = dyn_cast<ArrayType>())
     return arrayType.getNumElements();
   if (auto matrixType = dyn_cast<MatrixType>())
     return matrixType.getNumColumns();
   if (auto structType = dyn_cast<StructType>())
     return structType.getNumElements();
   if (auto vectorType = dyn_cast<VectorType>())
     return vectorType.getNumElements();
   if (isa<CooperativeMatrixNVType>()) {
     llvm_unreachable(
         "invalid to query number of elements of spirv::CooperativeMatrix type");
   }
+  if (isa<JointMatrixINTELType>()) {
+    llvm_unreachable(
+        "invalid to query number of elements of spirv::JointMatrix type");
+  }
   if (isa<RuntimeArrayType>()) {
     llvm_unreachable(
         "invalid to query number of elements of spirv::RuntimeArray type");
   }
   llvm_unreachable("invalid composite type");
 }
 
 bool CompositeType::hasCompileTimeKnownNumElements() const {
-  return !isa<CooperativeMatrixNVType, RuntimeArrayType>();
+  return !isa<CooperativeMatrixNVType, JointMatrixINTELType,
+              RuntimeArrayType>();
 }
 
 void CompositeType::getExtensions(
     SPIRVType::ExtensionArrayRefVector &extensions,
     Optional<StorageClass> storage) {
   TypeSwitch<Type>(*this)
-      .Case<ArrayType, CooperativeMatrixNVType, MatrixType, RuntimeArrayType,
-            StructType>(
+      .Case<ArrayType, CooperativeMatrixNVType, JointMatrixINTELType,
+            MatrixType, RuntimeArrayType, StructType>(
           [&](auto type) { type.getExtensions(extensions, storage); })
       .Case<VectorType>([&](VectorType type) {
         return type.getElementType().cast<ScalarType>().getExtensions(
             extensions, storage);
       })
       .Default([](Type) { llvm_unreachable("invalid composite type"); });
 }
 
 void CompositeType::getCapabilities(
     SPIRVType::CapabilityArrayRefVector &capabilities,
     Optional<StorageClass> storage) {
   TypeSwitch<Type>(*this)
-      .Case<ArrayType, CooperativeMatrixNVType, MatrixType, RuntimeArrayType,
-            StructType>(
+      .Case<ArrayType, CooperativeMatrixNVType, JointMatrixINTELType,
+            MatrixType, RuntimeArrayType, StructType>(
           [&](auto type) { type.getCapabilities(capabilities, storage); })
       .Case<VectorType>([&](VectorType type) {
         auto vecSize = getNumElements();
         if (vecSize == 8 || vecSize == 16) {
           static const Capability caps[] = {Capability::Vector16};
           ArrayRef<Capability> ref(caps, llvm::array_lengthof(caps));
           capabilities.push_back(ref);
         }
void CooperativeMatrixNVType::getCapabilities(
     Optional<StorageClass> storage) {
   getElementType().cast<SPIRVType>().getCapabilities(capabilities, storage);
   static const Capability caps[] = {Capability::CooperativeMatrixNV};
   ArrayRef<Capability> ref(caps, llvm::array_lengthof(caps));
   capabilities.push_back(ref);
 }
 
 //===----------------------------------------------------------------------===//
+// JointMatrixType
+//===----------------------------------------------------------------------===//
+
+struct spirv::detail::JointMatrixTypeStorage : public TypeStorage {
+  using KeyTy = std::tuple<Type, Scope, unsigned, unsigned>;
+
+  static JointMatrixTypeStorage *construct(TypeStorageAllocator &allocator,
+                                           const KeyTy &key) {
+    return new (allocator.allocate<JointMatrixTypeStorage>())
+        JointMatrixTypeStorage(key);
+  }
+
+  bool operator==(const KeyTy &key) const {
+    return key == KeyTy(elementType, scope, rows, columns);
+  }
+
+  JointMatrixTypeStorage(const KeyTy &key)
+      : elementType(std::get<0>(key)), rows(std::get<2>(key)),
+        columns(std::get<3>(key)), scope(std::get<1>(key)) {}
+
+  Type elementType;
+  unsigned rows;
+  unsigned columns;
+  Scope scope;
+};
+
+JointMatrixINTELType JointMatrixINTELType::get(Type elementType, Scope scope,
+                                               unsigned rows,
+                                               unsigned columns) {
+  return Base::get(elementType.getContext(), elementType, scope, rows, columns);
+}
+
+Type JointMatrixINTELType::getElementType() const {
+  return getImpl()->elementType;
+}
+
+Scope JointMatrixINTELType::getScope() const { return getImpl()->scope; }
+
+unsigned JointMatrixINTELType::getRows() const { return getImpl()->rows; }
+
+unsigned JointMatrixINTELType::getColumns() const { return getImpl()->columns; }
+
+void JointMatrixINTELType::getExtensions(
+    SPIRVType::ExtensionArrayRefVector &extensions,
+    Optional<StorageClass> storage) {
+  getElementType().cast<SPIRVType>().getExtensions(extensions, storage);
+  static const Extension exts[] = {Extension::SPV_INTEL_joint_matrix};
+  ArrayRef<Extension> ref(exts, llvm::array_lengthof(exts));
+  extensions.push_back(ref);
+}
+
+void JointMatrixINTELType::getCapabilities(
+    SPIRVType::CapabilityArrayRefVector &capabilities,
+    Optional<StorageClass> storage) {
+  getElementType().cast<SPIRVType>().getCapabilities(capabilities, storage);
+  static const Capability caps[] = {Capability::JointMatrixINTEL};
+  ArrayRef<Capability> ref(caps, llvm::array_lengthof(caps));
+  capabilities.push_back(ref);
+}
+
+//===----------------------------------------------------------------------===//
 // ImageType
 //===----------------------------------------------------------------------===//
 
 template <typename T>
 static constexpr unsigned getNumBits() {
   return 0;
 }
 template <>
void MatrixType::getCapabilities(
   getColumnType().cast<SPIRVType>().getCapabilities(capabilities, storage);
 }
 
 //===----------------------------------------------------------------------===//
 // SPIR-V Dialect
 //===----------------------------------------------------------------------===//
 
 void SPIRVDialect::registerTypes() {
-  addTypes<ArrayType, CooperativeMatrixNVType, ImageType, MatrixType,
-           PointerType, RuntimeArrayType, SampledImageType, StructType>();
+  addTypes<ArrayType, CooperativeMatrixNVType, ImageType, JointMatrixINTELType,
+           MatrixType, PointerType, RuntimeArrayType, SampledImageType,
+           StructType>();
 }

mlir/lib/Target/SPIRV/Deserialization/DeserializeOps.cpp

LogicalResult spirv::Deserializer::processInstruction(
   case spirv::Opcode::OpTypeSampledImage:
   case spirv::Opcode::OpTypeRuntimeArray:
   case spirv::Opcode::OpTypeStruct:
   case spirv::Opcode::OpTypePointer:
   case spirv::Opcode::OpTypeCooperativeMatrixNV:
     return processType(opcode, operands);
   case spirv::Opcode::OpTypeForwardPointer:
     return processTypeForwardPointer(operands);
+  case spirv::Opcode::OpTypeJointMatrixINTEL:
+    return processType(opcode, operands);
   case spirv::Opcode::OpConstant:
     return processConstant(operands, /*isSpec=*/false);
   case spirv::Opcode::OpSpecConstant:
     return processConstant(operands, /*isSpec=*/true);
   case spirv::Opcode::OpConstantComposite:
     return processConstantComposite(operands);
   case spirv::Opcode::OpSpecConstantComposite:
     return processSpecConstantComposite(operands);

mlir/lib/Target/SPIRV/Deserialization/Deserializer.h

private:
   LogicalResult processOpTypePointer(ArrayRef<uint32_t> operands);
 
   LogicalResult processArrayType(ArrayRef<uint32_t> operands);
 
   LogicalResult processCooperativeMatrixType(ArrayRef<uint32_t> operands);
 
   LogicalResult processFunctionType(ArrayRef<uint32_t> operands);
 
+  LogicalResult processJointMatrixType(ArrayRef<uint32_t> operands);
+
   LogicalResult processImageType(ArrayRef<uint32_t> operands);
 
   LogicalResult processSampledImageType(ArrayRef<uint32_t> operands);
 
   LogicalResult processRuntimeArrayType(ArrayRef<uint32_t> operands);
 
   LogicalResult processStructType(ArrayRef<uint32_t> operands);
 

mlir/lib/Target/SPIRV/Deserialization/Deserializer.cpp

case spirv::Opcode::OpTypePointer: {
     return processOpTypePointer(operands);
   } break;
   case spirv::Opcode::OpTypeArray:
     return processArrayType(operands);
   case spirv::Opcode::OpTypeCooperativeMatrixNV:
     return processCooperativeMatrixType(operands);
   case spirv::Opcode::OpTypeFunction:
     return processFunctionType(operands);
+  case spirv::Opcode::OpTypeJointMatrixINTEL:
+    return processJointMatrixType(operands);
   case spirv::Opcode::OpTypeImage:
     return processImageType(operands);
   case spirv::Opcode::OpTypeSampledImage:
     return processSampledImageType(operands);
   case spirv::Opcode::OpTypeRuntimeArray:
     return processRuntimeArrayType(operands);
   case spirv::Opcode::OpTypeStruct:
     return processStructType(operands);
spirv::Deserializer::processCooperativeMatrixType(ArrayRef<uint32_t> operands) {
   unsigned columns = getConstantInt(operands[4]).getInt();
 
   typeMap[operands[0]] = spirv::CooperativeMatrixNVType::get(
       elementTy, scope.value(), rows, columns);
   return success();
 }
 
 LogicalResult
+spirv::Deserializer::processJointMatrixType(ArrayRef<uint32_t> operands) {
+  if (operands.size() != 5) {
+    return emitError(unknownLoc, "OpTypeJointMatrix must have element "
+                                 "type and row x column parameters");
+  }
+
+  Type elementTy = getType(operands[1]);
+  if (!elementTy) {
+    return emitError(unknownLoc, "OpTypeJointMatrix references undefined <id> ")
+           << operands[1];
+  }
+
+  auto scope = spirv::symbolizeScope(getConstantInt(operands[2]).getInt());
+  if (!scope) {
+    return emitError(unknownLoc,
+                     "OpTypeCooperativeMatrix references undefined scope <id> ")
+           << operands[2];
+  }
+
+  unsigned rows = getConstantInt(operands[3]).getInt();
+  unsigned columns = getConstantInt(operands[4]).getInt();
+
+  typeMap[operands[0]] =
+      spirv::JointMatrixINTELType::get(elementTy, scope.value(), rows, columns);
+  return success();
+}
+
+LogicalResult
 spirv::Deserializer::processRuntimeArrayType(ArrayRef<uint32_t> operands) {
   if (operands.size() != 2) {
     return emitError(unknownLoc, "OpTypeRuntimeArray must have two operands");
   }
   Type memberType = getType(operands[1]);
   if (!memberType) {
     return emitError(unknownLoc,
                      "OpTypeRuntimeArray references undefined <id> ")

mlir/lib/Target/SPIRV/Serialization/Serializer.cpp

if (auto cooperativeMatrixType =
     operands.push_back(elementTypeID);
     operands.push_back(
         getConstantOp(static_cast<uint32_t>(cooperativeMatrixType.getScope())));
     operands.push_back(getConstantOp(cooperativeMatrixType.getRows()));
     operands.push_back(getConstantOp(cooperativeMatrixType.getColumns()));
     return success();
   }
 
+  if (auto jointMatrixType = type.dyn_cast<spirv::JointMatrixINTELType>()) {
+    uint32_t elementTypeID = 0;
+    if (failed(processTypeImpl(loc, jointMatrixType.getElementType(),
+                               elementTypeID, serializationCtx))) {
+      return failure();
+    }
+    typeEnum = spirv::Opcode::OpTypeJointMatrixINTEL;
+    auto getConstantOp = [&](uint32_t id) {
+      auto attr = IntegerAttr::get(IntegerType::get(type.getContext(), 32), id);
+      return prepareConstantInt(loc, attr);
+    };
+    operands.push_back(elementTypeID);
+    operands.push_back(
+        getConstantOp(static_cast<uint32_t>(jointMatrixType.getScope())));
+    operands.push_back(getConstantOp(jointMatrixType.getRows()));
+    operands.push_back(getConstantOp(jointMatrixType.getColumns()));
+    return success();
+  }
+
   if (auto matrixType = type.dyn_cast<spirv::MatrixType>()) {
     uint32_t elementTypeID = 0;
     if (failed(processTypeImpl(loc, matrixType.getColumnType(), elementTypeID,
                                serializationCtx))) {
       return failure();
     }
     typeEnum = spirv::Opcode::OpTypeMatrix;
     operands.push_back(elementTypeID);

mlir/test/Dialect/SPIRV/IR/joint-matrix-ops.mlir

+// RUN: mlir-opt -allow-unregistered-dialect -split-input-file -verify-diagnostics %s | FileCheck %s
+
+// CHECK-LABEL: @joint_matrix_load
+spv.func @joint_matrix_load(%ptr : !spv.ptr<i32, StorageBuffer>, %stride : i32, %layout : i32) "None" {
+  // CHECK: {{%.*}} = spv.JointMatrixLoadINTEL <Subgroup> {{%.*}}, {{%.*}}, {{%.*}} : (!spv.ptr<i32, StorageBuffer>, i32, i32) -> !spv.jointmatrix<16x8xi32, Workgroup>
+  %0 = spv.JointMatrixLoadINTEL <Subgroup> %ptr, %stride, %layout : (!spv.ptr<i32, StorageBuffer>, i32, i32) -> !spv.jointmatrix<16x8xi32, Workgroup>
+  spv.Return
+}
+
+// -----
+// CHECK-LABEL: @joint_matrix_load_memaccess
+spv.func @joint_matrix_load_memaccess(%ptr : !spv.ptr<i32, StorageBuffer>, %stride : i32, %layout : i32) "None" {
+  // CHECK: {{%.*}} = spv.JointMatrixLoadINTEL <Subgroup> {{%.*}}, {{%.*}}, {{%.*}} {Volatile} : (!spv.ptr<i32, StorageBuffer>, i32, i32) -> !spv.jointmatrix<8x16xi32, Subgroup>
+  %0 = spv.JointMatrixLoadINTEL <Subgroup> %ptr, %stride, %layout {Volatile} : (!spv.ptr<i32, StorageBuffer>, i32, i32) -> !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_load_diff_ptr_type
+spv.func @joint_matrix_load_diff_ptr_type(%ptr : !spv.ptr<vector<4xi32>, StorageBuffer>, %stride : i32, %layout : i32) "None" {
+  // CHECK: {{%.*}} = spv.JointMatrixLoadINTEL <Subgroup> {{%.*}}, {{%.*}}, {{%.*}} {Volatile} : (!spv.ptr<vector<4xi32>, StorageBuffer>, i32, i32) -> !spv.jointmatrix<8x16xi32, Subgroup>
+  %0 = spv.JointMatrixLoadINTEL <Subgroup> %ptr, %stride, %layout {Volatile} : (!spv.ptr<vector<4xi32>, StorageBuffer>, i32, i32) -> !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_store
+spv.func @joint_matrix_store(%ptr : !spv.ptr<i32, StorageBuffer>, %stride : i32, %m : !spv.jointmatrix<8x16xi32, Workgroup>,  %layout : i32) "None" {
+  // CHECK: spv.JointMatrixStoreINTEL <Subgroup> {{%.*}}, {{%.*}}, {{%.*}}, {{%.*}} : (!spv.ptr<i32, StorageBuffer>, !spv.jointmatrix<8x16xi32, Workgroup>, i32, i32)
+  spv.JointMatrixStoreINTEL <Subgroup> %ptr, %m, %stride, %layout : (!spv.ptr<i32, StorageBuffer>, !spv.jointmatrix<8x16xi32, Workgroup> , i32, i32)
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_store_memaccess
+spv.func @joint_matrix_store_memaccess(%ptr : !spv.ptr<i32, StorageBuffer>, %m : !spv.jointmatrix<8x16xi32, Subgroup>, %stride : i32, %layout : i32) "None" {
+  // CHECK: spv.JointMatrixStoreINTEL <Subgroup> {{%.*}}, {{%.*}}, {{%.*}}, {{%.*}} {Volatile} : (!spv.ptr<i32, StorageBuffer>, !spv.jointmatrix<8x16xi32, Subgroup>, i32, i32)
+  spv.JointMatrixStoreINTEL <Subgroup> %ptr, %m, %stride, %layout {Volatile} : (!spv.ptr<i32, StorageBuffer>, !spv.jointmatrix<8x16xi32, Subgroup>, i32, i32)
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_length
+spv.func @joint_matrix_length() -> i32 "None" {
+  // CHECK: {{%.*}} = spv.JointMatrixWorkItemLengthINTEL : !spv.jointmatrix<8x16xi32, Subgroup>
+  %0 = spv.JointMatrixWorkItemLengthINTEL : !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.ReturnValue %0 : i32
+}
+
+// CHECK-LABEL: @joint_matrix_muladd
+spv.func @joint_matrix_muladd(%a : !spv.jointmatrix<8x32xi8, Subgroup>, %b : !spv.jointmatrix<32x8xi8, Subgroup>, %c : !spv.jointmatrix<8x8xi32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.JointMatrixMadINTEL <Subgroup> {{%.*}}, {{%.*}}, {{%.*}}  : !spv.jointmatrix<8x32xi8, Subgroup>, !spv.jointmatrix<32x8xi8, Subgroup> -> !spv.jointmatrix<8x8xi32, Subgroup>
+  %r = spv.JointMatrixMadINTEL <Subgroup> %a, %b, %c : !spv.jointmatrix<8x32xi8, Subgroup>, !spv.jointmatrix<32x8xi8, Subgroup> -> !spv.jointmatrix<8x8xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_add
+spv.func @joint_matrix_add(%a : !spv.jointmatrix<8x16xi32, Subgroup>, %b : !spv.jointmatrix<8x16xi32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.IAdd {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xi32, Subgroup>
+  %r = spv.IAdd %a, %b : !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_sub
+spv.func @joint_matrix_sub(%a : !spv.jointmatrix<8x16xi32, Subgroup>, %b : !spv.jointmatrix<8x16xi32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.ISub {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xi32, Subgroup>
+  %r = spv.ISub %a, %b : !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_sdiv
+spv.func @joint_matrix_sdiv(%a : !spv.jointmatrix<8x16xi32, Subgroup>, %b : !spv.jointmatrix<8x16xi32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.SDiv {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xi32, Subgroup>
+  %r = spv.SDiv %a, %b : !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_udiv
+spv.func @joint_matrix_udiv(%a : !spv.jointmatrix<8x16xi32, Subgroup>, %b : !spv.jointmatrix<8x16xi32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.UDiv {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xi32, Subgroup>
+  %r = spv.UDiv %a, %b : !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_fadd
+spv.func @joint_matrix_fadd(%a : !spv.jointmatrix<8x16xf32, Subgroup>, %b : !spv.jointmatrix<8x16xf32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.FAdd {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xf32, Subgroup>
+  %r = spv.FAdd %a, %b : !spv.jointmatrix<8x16xf32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_fsub
+spv.func @joint_matrix_fsub(%a : !spv.jointmatrix<8x16xf32, Subgroup>, %b : !spv.jointmatrix<8x16xf32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.FSub {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xf32, Subgroup>
+  %r = spv.FSub %a, %b : !spv.jointmatrix<8x16xf32, Subgroup>
+  spv.Return
+}
+
+// CHECK-LABEL: @joint_matrix_fdiv
+spv.func @joint_matrix_fdiv(%a : !spv.jointmatrix<8x16xf32, Subgroup>, %b : !spv.jointmatrix<8x16xf32, Subgroup>) "None" {
+  // CHECK: {{%.*}} = spv.FDiv {{%.*}}, {{%.*}} : !spv.jointmatrix<8x16xf32, Subgroup>
+  %r = spv.FDiv %a, %b : !spv.jointmatrix<8x16xf32, Subgroup>
+  spv.Return
+}
+
+// -----
+
+// CHECK-LABEL: @joint_matrix_access_chain
+spv.func @joint_matrix_access_chain(%a : !spv.ptr<!spv.jointmatrix<8x16xf32, Subgroup>, Function>) -> !spv.ptr<f32, Function> "None" {
+  %0 = spv.Constant 0: i32
+  // CHECK: {{%.*}} = spv.AccessChain {{%.*}}[{{%.*}}] : !spv.ptr<!spv.jointmatrix<8x16xf32, Subgroup>, Function>, i32
+  %1 = spv.AccessChain %a[%0] : !spv.ptr<!spv.jointmatrix<8x16xf32, Subgroup>, Function>, i32
+  spv.ReturnValue %1 : !spv.ptr<f32, Function>
+}
+
+// -----
+
+spv.func @joint_matrix_muladd(%a : !spv.jointmatrix<16x16xi32, Subgroup>, %b : !spv.jointmatrix<16x8xi32, Subgroup>, %c : !spv.jointmatrix<8x8xi32, Subgroup>) "None" {
+  // expected-error @+1 {{'spv.JointMatrixMadINTEL' op matrix size must match}}
+  %r = spv.JointMatrixMadINTEL <Subgroup> %a, %b, %c : !spv.jointmatrix<16x16xi32, Subgroup>, !spv.jointmatrix<16x8xi32, Subgroup> -> !spv.jointmatrix<8x8xi32, Subgroup>
+  spv.Return
+}
+
+// -----
+
+spv.func @joint_matrix_muladd(%a : !spv.jointmatrix<8x16xi32, Subgroup>, %b : !spv.jointmatrix<8x8xi32, Subgroup>, %c : !spv.jointmatrix<8x8xi32, Subgroup>) "None" {
+  // expected-error @+1 {{'spv.JointMatrixMadINTEL' op matrix size must match}}
+  %r = spv.JointMatrixMadINTEL <Subgroup> %a, %b, %c : !spv.jointmatrix<8x16xi32, Subgroup>, !spv.jointmatrix<8x8xi32, Subgroup> -> !spv.jointmatrix<8x8xi32, Subgroup>
+  spv.Return
+}
+
+// -----
+
+spv.func @joint_matrix_muladd(%a : !spv.jointmatrix<8x16xi32, Subgroup>, %b : !spv.jointmatrix<16x8xi32, Workgroup>, %c : !spv.jointmatrix<8x8xi32, Subgroup>) "None" {
+  // expected-error @+1 {{'spv.JointMatrixMadINTEL' op matrix scope must match}}
+  %r = spv.JointMatrixMadINTEL <Subgroup> %a, %b, %c : !spv.jointmatrix<8x16xi32, Subgroup>, !spv.jointmatrix<16x8xi32, Workgroup> -> !spv.jointmatrix<8x8xi32, Subgroup>
+  spv.Return
+}
+
+// -----
+
+spv.func @joint_matrix_muladd(%a : !spv.jointmatrix<8x16xf32, Subgroup>, %b : !spv.jointmatrix<16x8xi32, Subgroup>, %c : !spv.jointmatrix<8x8xi32, Subgroup>) "None" {
+  // expected-error @+1 {{matrix element type must match}}
+  %r = spv.JointMatrixMadINTEL <Subgroup> %a, %b, %c : !spv.jointmatrix<8x16xf32, Subgroup>, !spv.jointmatrix<16x8xi32, Subgroup> -> !spv.jointmatrix<8x8xi32, Subgroup>
+  spv.Return
+}
+
+// -----
+
+spv.func @joint_matrix_load_memaccess(%ptr : !spv.ptr<!spv.struct<(f32 [0])>, StorageBuffer>, %stride : i32, %layout : i32) "None" {
+  // expected-error @+1 {{Pointer must point to a scalar or vector type}}
+  %0 = spv.JointMatrixLoadINTEL <Subgroup> %ptr, %stride, %layout : (!spv.ptr<!spv.struct<(f32 [0])>, StorageBuffer>, i32, i32 )-> !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}
+
+// -----
+
+spv.func @joint_matrix_load_memaccess(%ptr : !spv.ptr<i32, Function>, %stride : i32, %layout : i32) "None" {
+  // expected-error @+1 {{Pointer storage class must be Workgroup, StorageBuffer or PhysicalStorageBufferEXT}}
+  %0 = spv.JointMatrixLoadINTEL <Subgroup> %ptr, %stride, %layout : (!spv.ptr<i32, Function>, i32, i32) -> !spv.jointmatrix<8x16xi32, Subgroup>
+  spv.Return
+}