[mlir][gpu][spirv] Gpu min/max reductions lowering to SPIR-V

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp

const OpHandler handlers[] = {
       {ReduceType::ADD,
        &createGroupReduceOpImpl<spv::GroupIAddOp, spv::GroupNonUniformIAddOp>,
        &createGroupReduceOpImpl<spv::GroupFAddOp, spv::GroupNonUniformFAddOp>},
       {ReduceType::MUL,
        &createGroupReduceOpImpl<spv::GroupIMulKHROp,
                                 spv::GroupNonUniformIMulOp>,
        &createGroupReduceOpImpl<spv::GroupFMulKHROp,
                                 spv::GroupNonUniformFMulOp>},
+      {ReduceType::MIN,
+       &createGroupReduceOpImpl<spv::GroupSMinOp, spv::GroupNonUniformSMinOp>,
+       &createGroupReduceOpImpl<spv::GroupFMinOp, spv::GroupNonUniformFMinOp>},
+      {ReduceType::MAX,
+       &createGroupReduceOpImpl<spv::GroupSMaxOp, spv::GroupNonUniformSMaxOp>,
+       &createGroupReduceOpImpl<spv::GroupFMaxOp, spv::GroupNonUniformFMaxOp>},
   };
 
   for (auto &handler : handlers)
     if (handler.type == opType)
       return (handler.*handlerPtr)(builder, loc, arg, isGroup, isUniform);
 
   return std::nullopt;
 }

mlir/test/Conversion/GPUToSPIRV/reductions.mlir

gpu.func @test(%arg : i32) kernel
     attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
     // CHECK: %{{.*}} = spirv.GroupNonUniformIMul "Subgroup" "Reduce" %[[ARG]] : i32
     %reduced = gpu.subgroup_reduce mul %arg : (i32) -> (i32)
     gpu.return
   }
 }
 
 }
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupFMin <Workgroup> <Reduce> %[[ARG]] : f32
+    %reduced = gpu.all_reduce min %arg uniform {} : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformFMin "Workgroup" "Reduce" %[[ARG]] : f32
+    %reduced = gpu.all_reduce min %arg {} : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupSMin <Workgroup> <Reduce> %[[ARG]] : i32
+    %reduced = gpu.all_reduce min %arg uniform {} : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformSMin "Workgroup" "Reduce" %[[ARG]] : i32
+    %reduced = gpu.all_reduce min %arg {} : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupFMin <Subgroup> <Reduce> %[[ARG]] : f32
+    %reduced = gpu.subgroup_reduce min %arg uniform : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformFMin "Subgroup" "Reduce" %[[ARG]] : f32
+    %reduced = gpu.subgroup_reduce min %arg : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupSMin <Subgroup> <Reduce> %[[ARG]] : i32
+    %reduced = gpu.subgroup_reduce min %arg uniform : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformSMin "Subgroup" "Reduce" %[[ARG]] : i32
+    %reduced = gpu.subgroup_reduce min %arg : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupFMax <Workgroup> <Reduce> %[[ARG]] : f32
+    %reduced = gpu.all_reduce max %arg uniform {} : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformFMax "Workgroup" "Reduce" %[[ARG]] : f32
+    %reduced = gpu.all_reduce max %arg {} : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupSMax <Workgroup> <Reduce> %[[ARG]] : i32
+    %reduced = gpu.all_reduce max %arg uniform {} : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformSMax "Workgroup" "Reduce" %[[ARG]] : i32
+    %reduced = gpu.all_reduce max %arg {} : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupFMax <Subgroup> <Reduce> %[[ARG]] : f32
+    %reduced = gpu.subgroup_reduce max %arg uniform : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  gpu.func @test(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformFMax "Subgroup" "Reduce" %[[ARG]] : f32
+    %reduced = gpu.subgroup_reduce max %arg : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupSMax <Subgroup> <Reduce> %[[ARG]] : i32
+    %reduced = gpu.subgroup_reduce max %arg uniform : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupNonUniformArithmetic, GroupUniformArithmeticKHR], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: i32)
+  gpu.func @test(%arg : i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformSMax "Subgroup" "Reduce" %[[ARG]] : i32
+    %reduced = gpu.subgroup_reduce max %arg : (i32) -> (i32)
+    gpu.return
+  }
+}
+
+}