[mlir][spirv] Add missing group non-uniform bitwise and logical ops (#73475)

This covers the following ops:
`spirv.GroupNonUniform` x {`Bitwise`, `Logical`} x {`And`, `Or`, `Xor`}

We need these to efficiently lower from the `gpu.subgroup_reduce` op.

Author: kuhar

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

@@ -4462,6 +4462,12 @@ def SPIRV_OC_OpGroupNonUniformFMin        : I32EnumAttrCase<"OpGroupNonUniformFM
 def SPIRV_OC_OpGroupNonUniformSMax        : I32EnumAttrCase<"OpGroupNonUniformSMax", 356>;
 def SPIRV_OC_OpGroupNonUniformUMax        : I32EnumAttrCase<"OpGroupNonUniformUMax", 357>;
 def SPIRV_OC_OpGroupNonUniformFMax        : I32EnumAttrCase<"OpGroupNonUniformFMax", 358>;
+def SPIRV_OC_OpGroupNonUniformBitwiseAnd  : I32EnumAttrCase<"OpGroupNonUniformBitwiseAnd", 359>;
+def SPIRV_OC_OpGroupNonUniformBitwiseOr   : I32EnumAttrCase<"OpGroupNonUniformBitwiseOr", 360>;
+def SPIRV_OC_OpGroupNonUniformBitwiseXor  : I32EnumAttrCase<"OpGroupNonUniformBitwiseXor", 361>;
+def SPIRV_OC_OpGroupNonUniformLogicalAnd  : I32EnumAttrCase<"OpGroupNonUniformLogicalAnd", 362>;
+def SPIRV_OC_OpGroupNonUniformLogicalOr   : I32EnumAttrCase<"OpGroupNonUniformLogicalOr", 363>;
+def SPIRV_OC_OpGroupNonUniformLogicalXor  : I32EnumAttrCase<"OpGroupNonUniformLogicalXor", 364>;
 def SPIRV_OC_OpSubgroupBallotKHR          : I32EnumAttrCase<"OpSubgroupBallotKHR", 4421>;
 def SPIRV_OC_OpSDot                       : I32EnumAttrCase<"OpSDot", 4450>;
 def SPIRV_OC_OpUDot                       : I32EnumAttrCase<"OpUDot", 4451>;
@@ -4570,6 +4576,9 @@ def SPIRV_OpcodeAttr :
       SPIRV_OC_OpGroupNonUniformSMin, SPIRV_OC_OpGroupNonUniformUMin,
       SPIRV_OC_OpGroupNonUniformFMin, SPIRV_OC_OpGroupNonUniformSMax,
       SPIRV_OC_OpGroupNonUniformUMax, SPIRV_OC_OpGroupNonUniformFMax,
+      SPIRV_OC_OpGroupNonUniformBitwiseAnd, SPIRV_OC_OpGroupNonUniformBitwiseOr,
+      SPIRV_OC_OpGroupNonUniformBitwiseXor, SPIRV_OC_OpGroupNonUniformLogicalAnd,
+      SPIRV_OC_OpGroupNonUniformLogicalOr, SPIRV_OC_OpGroupNonUniformLogicalXor,
       SPIRV_OC_OpSubgroupBallotKHR, SPIRV_OC_OpSDot, SPIRV_OC_OpUDot,
       SPIRV_OC_OpSUDot, SPIRV_OC_OpSDotAccSat, SPIRV_OC_OpUDotAccSat,
       SPIRV_OC_OpSUDotAccSat,

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

@@ -997,4 +997,304 @@ def SPIRV_GroupNonUniformUMinOp : SPIRV_GroupNonUniformArithmeticOp<"GroupNonUni
 
 // -----
 
+def SPIRV_GroupNonUniformBitwiseAndOp :
+  SPIRV_GroupNonUniformArithmeticOp<"GroupNonUniformBitwiseAnd",
+    SPIRV_Integer, []> {
+  let summary = [{
+    A bitwise `and` group operation of all Value operands contributed by active
+    invocations in the group.
+  }];
+
+  let description = [{
+    Result Type must be a scalar or vector of integer type.
+
+    Execution is a Scope. It must be either Workgroup or Subgroup.
+
+    The identity I for Operation is ~0. If Operation is ClusteredReduce,
+    ClusterSize must be present.
+
+    The type of Value must be the same as Result Type.
+
+    ClusterSize is the size of cluster to use. ClusterSize must be a scalar
+    of integer type, whose Signedness operand is 0. ClusterSize must come
+    from a constant instruction. ClusterSize must be at least 1, and must be
+    a power of 2. If ClusterSize is greater than the declared SubGroupSize,
+    executing this instruction results in undefined behavior.
+
+    <!-- End of AutoGen section -->
+
+    #### Example:
+
+    ```mlir
+    %four = spirv.Constant 4 : i32
+    %scalar = ... : i32
+    %vector = ... : vector<4xi32>
+    %0 = spirv.GroupNonUniformBitwiseAnd "Workgroup" "Reduce" %scalar : i32
+    %1 = spirv.GroupNonUniformBitwiseAnd "Subgroup" "ClusteredReduce"
+           %vector cluster_size(%four) : vector<4xi32>
+    ```
+  }];
+
+  let availability = [
+    MinVersion<SPIRV_V_1_3>,
+    MaxVersion<SPIRV_V_1_6>,
+    Extension<[]>,
+    Capability<[SPIRV_C_GroupNonUniformArithmetic,
+                SPIRV_C_GroupNonUniformClustered,
+                SPIRV_C_GroupNonUniformPartitionedNV]>
+  ];
+}
+
+// -----
+
+def SPIRV_GroupNonUniformBitwiseOrOp :
+  SPIRV_GroupNonUniformArithmeticOp<"GroupNonUniformBitwiseOr",
+    SPIRV_Integer, []> {
+  let summary = [{
+    A bitwise `or` group operation of all Value operands contributed by active
+    invocations in the group.
+  }];
+
+  let description = [{
+    Result Type must be a scalar or vector of integer type.
+
+    Execution is a Scope. It must be either Workgroup or Subgroup.
+
+    The identity I for Operation is 0. If Operation is ClusteredReduce,
+    ClusterSize must be present.
+
+    The type of Value must be the same as Result Type.
+
+    ClusterSize is the size of cluster to use. ClusterSize must be a scalar
+    of integer type, whose Signedness operand is 0. ClusterSize must come
+    from a constant instruction. ClusterSize must be at least 1, and must be
+    a power of 2. If ClusterSize is greater than the declared SubGroupSize,
+    executing this instruction results in undefined behavior.
+
+    <!-- End of AutoGen section -->
+
+    #### Example:
+
+    ```mlir
+    %four = spirv.Constant 4 : i32
+    %scalar = ... : i32
+    %vector = ... : vector<4xi32>
+    %0 = spirv.GroupNonUniformBitwiseOr "Workgroup" "Reduce" %scalar : i32
+    %1 = spirv.GroupNonUniformBitwiseOr "Subgroup" "ClusteredReduce"
+           %vector cluster_size(%four) : vector<4xi32>
+    ```
+  }];
+
+  let availability = [
+    MinVersion<SPIRV_V_1_3>,
+    MaxVersion<SPIRV_V_1_6>,
+    Extension<[]>,
+    Capability<[SPIRV_C_GroupNonUniformArithmetic,
+                SPIRV_C_GroupNonUniformClustered,
+                SPIRV_C_GroupNonUniformPartitionedNV]>
+  ];
+}
+
+// -----
+
+def SPIRV_GroupNonUniformBitwiseXorOp :
+  SPIRV_GroupNonUniformArithmeticOp<"GroupNonUniformBitwiseXor",
+    SPIRV_Integer, []> {
+  let summary = [{
+    A bitwise `xor` group operation of all Value operands contributed by active
+    invocations in the group.
+  }];
+
+  let description = [{
+    Result Type must be a scalar or vector of integer type.
+
+    Execution is a Scope. It must be either Workgroup or Subgroup.
+
+    The identity I for Operation is 0. If Operation is ClusteredReduce,
+    ClusterSize must be present.
+
+    The type of Value must be the same as Result Type.
+
+    ClusterSize is the size of cluster to use. ClusterSize must be a scalar
+    of integer type, whose Signedness operand is 0. ClusterSize must come
+    from a constant instruction. ClusterSize must be at least 1, and must be
+    a power of 2. If ClusterSize is greater than the declared SubGroupSize,
+    executing this instruction results in undefined behavior.
+
+    <!-- End of AutoGen section -->
+
+    #### Example:
+
+    ```mlir
+    %four = spirv.Constant 4 : i32
+    %scalar = ... : i32
+    %vector = ... : vector<4xi32>
+    %0 = spirv.GroupNonUniformBitwiseXor "Workgroup" "Reduce" %scalar : i32
+    %1 = spirv.GroupNonUniformBitwiseXor "Subgroup" "ClusteredReduce"
+           %vector cluster_size(%four) : vector<4xi32>
+    ```
+  }];
+
+  let availability = [
+    MinVersion<SPIRV_V_1_3>,
+    MaxVersion<SPIRV_V_1_6>,
+    Extension<[]>,
+    Capability<[SPIRV_C_GroupNonUniformArithmetic,
+                SPIRV_C_GroupNonUniformClustered,
+                SPIRV_C_GroupNonUniformPartitionedNV]>
+  ];
+}
+
+// -----
+
+def SPIRV_GroupNonUniformLogicalAndOp :
+  SPIRV_GroupNonUniformArithmeticOp<"GroupNonUniformLogicalAnd",
+    SPIRV_Bool, []> {
+  let summary = [{
+    A logical `and` group operation of all Value operands contributed by active
+    invocations in the group.
+  }];
+
+  let description = [{
+    Result Type must be a scalar or vector of Boolean type.
+
+    Execution is a Scope. It must be either Workgroup or Subgroup.
+
+    The identity I for Operation is ~0. If Operation is ClusteredReduce,
+    ClusterSize must be present.
+
+    The type of Value must be the same as Result Type.
+
+    ClusterSize is the size of cluster to use. ClusterSize must be a scalar
+    of integer type, whose Signedness operand is 0. ClusterSize must come
+    from a constant instruction. ClusterSize must be at least 1, and must be
+    a power of 2. If ClusterSize is greater than the declared SubGroupSize,
+    executing this instruction results in undefined behavior.
+
+    <!-- End of AutoGen section -->
+
+    #### Example:
+
+    ```mlir
+    %four = spirv.Constant 4 : i32
+    %scalar = ... : i1
+    %vector = ... : vector<4xi1>
+    %0 = spirv.GroupNonUniformLogicalAnd "Workgroup" "Reduce" %scalar : i1
+    %1 = spirv.GroupNonUniformLogicalAnd "Subgroup" "ClusteredReduce"
+           %vector cluster_size(%four) : vector<4xi1>
+    ```
+  }];
+
+  let availability = [
+    MinVersion<SPIRV_V_1_3>,
+    MaxVersion<SPIRV_V_1_6>,
+    Extension<[]>,
+    Capability<[SPIRV_C_GroupNonUniformArithmetic,
+                SPIRV_C_GroupNonUniformClustered,
+                SPIRV_C_GroupNonUniformPartitionedNV]>
+  ];
+}
+
+// -----
+
+def SPIRV_GroupNonUniformLogicalOrOp :
+  SPIRV_GroupNonUniformArithmeticOp<"GroupNonUniformLogicalOr",
+    SPIRV_Bool, []> {
+  let summary = [{
+    A logical `or` group operation of all Value operands contributed by active
+    invocations in the group.
+  }];
+
+  let description = [{
+    Result Type must be a scalar or vector of Boolean type.
+
+    Execution is a Scope. It must be either Workgroup or Subgroup.
+
+    The identity I for Operation is 0. If Operation is ClusteredReduce,
+    ClusterSize must be present.
+
+    The type of Value must be the same as Result Type.
+
+    ClusterSize is the size of cluster to use. ClusterSize must be a scalar
+    of integer type, whose Signedness operand is 0. ClusterSize must come
+    from a constant instruction. ClusterSize must be at least 1, and must be
+    a power of 2. If ClusterSize is greater than the declared SubGroupSize,
+    executing this instruction results in undefined behavior.
+
+    <!-- End of AutoGen section -->
+
+    #### Example:
+
+    ```mlir
+    %four = spirv.Constant 4 : i32
+    %scalar = ... : i1
+    %vector = ... : vector<4xi1>
+    %0 = spirv.GroupNonUniformLogicalOr "Workgroup" "Reduce" %scalar : i1
+    %1 = spirv.GroupNonUniformLogicalOr "Subgroup" "ClusteredReduce"
+           %vector cluster_size(%four) : vector<4xi1>
+    ```
+  }];
+
+  let availability = [
+    MinVersion<SPIRV_V_1_3>,
+    MaxVersion<SPIRV_V_1_6>,
+    Extension<[]>,
+    Capability<[SPIRV_C_GroupNonUniformArithmetic,
+                SPIRV_C_GroupNonUniformClustered,
+                SPIRV_C_GroupNonUniformPartitionedNV]>
+  ];
+}
+
+// -----
+
+def SPIRV_GroupNonUniformLogicalXorOp :
+  SPIRV_GroupNonUniformArithmeticOp<"GroupNonUniformLogicalXor",
+    SPIRV_Bool, []> {
+  let summary = [{
+    A logical `xor` group operation of all Value operands contributed by active
+    invocations in the group.
+  }];
+
+  let description = [{
+    Result Type must be a scalar or vector of Boolean type.
+
+    Execution is a Scope. It must be either Workgroup or Subgroup.
+
+    The identity I for Operation is 0. If Operation is ClusteredReduce,
+    ClusterSize must be present.
+
+    The type of Value must be the same as Result Type.
+
+    ClusterSize is the size of cluster to use. ClusterSize must be a scalar
+    of integer type, whose Signedness operand is 0. ClusterSize must come
+    from a constant instruction. ClusterSize must be at least 1, and must be
+    a power of 2. If ClusterSize is greater than the declared SubGroupSize,
+    executing this instruction results in undefined behavior.
+
+    <!-- End of AutoGen section -->
+
+    #### Example:
+
+    ```mlir
+    %four = spirv.Constant 4 : i32
+    %scalar = ... : i1
+    %vector = ... : vector<4xi1>
+    %0 = spirv.GroupNonUniformLogicalXor "Workgroup" "Reduce" %scalar : i1
+    %1 = spirv.GroupNonUniformLogicalXor "Subgroup" "ClusteredReduce"
+           %vector cluster_size(%four) : vector<4xi>
+    ```
+  }];
+
+  let availability = [
+    MinVersion<SPIRV_V_1_3>,
+    MaxVersion<SPIRV_V_1_6>,
+    Extension<[]>,
+    Capability<[SPIRV_C_GroupNonUniformArithmetic,
+                SPIRV_C_GroupNonUniformClustered,
+                SPIRV_C_GroupNonUniformPartitionedNV]>
+  ];
+}
+
+// -----
+
 #endif // MLIR_DIALECT_SPIRV_IR_NON_UNIFORM_OPS