[mlir][AMDGPU] Set uniform-work-group-size=true by default

GPU kernels generated via typical MLIR mechanisms make the assumption
that all workgroups are of uniform size, and so, as in OpenMP, it is
appropriate to set the "uniform-work-group-size"="true" attribute on
these functions by default. This commit makes that choiec.

In the event it is needed,t his commit adds
`rocdl.uniform_wrok_group_size` as an attribute to be set on LLVM
functions that can be used to override the default.

Author: krzysz00

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

@@ -37,6 +37,12 @@ def ROCDL_Dialect : Dialect {
     static constexpr ::llvm::StringLiteral getReqdWorkGroupSizeAttrName() {
       return ::llvm::StringLiteral("rocdl.reqd_work_group_size");
     }
+    /// MLIR's gpu-related infrastructure effectively assume uniform workgroup
+    /// sizes, so this attribute defaults to "true" on `rocdl.kernel` functions.
+    /// It is provided here to allow overriding this assumption.
+    static constexpr ::llvm::StringLiteral getUniformWorkGroupSizeAttrName() {
+      return ::llvm::StringLiteral("rocdl.uniform_work_group_size");
+    }
 
     /// The address space value that represents global memory.
     static constexpr unsigned kGlobalMemoryAddressSpace = 1;

mlir/lib/Target/LLVMIR/Dialect/ROCDL/ROCDLToLLVMIRTranslation.cpp

@@ -99,6 +99,12 @@ class ROCDLDialectLLVMIRTranslationInterface
       if (!llvmFunc->hasFnAttribute("amdgpu-flat-work-group-size")) {
         llvmFunc->addFnAttr("amdgpu-flat-work-group-size", "1,256");
       }
+      // MLIR's GPU kernel APIs all assume and produce uniformly-sized
+      // workgroups, so the lowering of the `rocdl.kernel` marker encodes this
+      // assumption. This assumption may be overridden by setting
+      // `rocdl.uniform_work_group_size` on a given function.
+      if (!llvmFunc->hasFnAttribute("uniform-work-group-size"))
+        llvmFunc->addFnAttr("uniform-work-group-size", "true");
     }
     // Override flat-work-group-size
     // TODO: update clients to rocdl.flat_work_group_size instead,
@@ -133,7 +139,19 @@ class ROCDLDialectLLVMIRTranslationInterface
       llvmAttrValue.append(value.getValue());
       llvmFunc->addFnAttr("amdgpu-flat-work-group-size", llvmAttrValue);
     }
-
+    if (ROCDL::ROCDLDialect::getUniformWorkGroupSizeAttrName() ==
+        attribute.getName()) {
+      auto func = dyn_cast<LLVM::LLVMFuncOp>(op);
+      if (!func)
+        return failure();
+      auto value = dyn_cast<BoolAttr>(attribute.getValue());
+      if (!value)
+        return failure();
+      llvm::Function *llvmFunc =
+          moduleTranslation.lookupFunction(func.getName());
+      llvmFunc->addFnAttr("uniform-work-group-size",
+                          value.getValue() ? "true" : "false");
+    }
     // Set reqd_work_group_size metadata
     if (ROCDL::ROCDLDialect::getReqdWorkGroupSizeAttrName() ==
         attribute.getName()) {

mlir/test/Target/LLVMIR/rocdl.mlir

@@ -56,6 +56,12 @@ llvm.func @known_block_sizes()
   llvm.return
 }
 
+llvm.func @kernel_func_no_uniform_work_groups() attributes {rocdl.kernel, rocdl.uniform_work_group_size = false} {
+  // CHECK-LABEL: amdgpu_kernel void @kernel_func_no_uniform_work_groups()
+  // CHECK: #[[$KERNEL_NO_UNIFORM_WORK_GROUPS_ATTRS:[0-9]+]]
+  llvm.return
+}
+
 llvm.func @rocdl.lane_id() -> i32 {
   // CHECK: [[mbcntlo:%.+]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
   // CHECK-NEXT: call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 [[mbcntlo]])
@@ -489,8 +495,9 @@ llvm.func @rocdl_8bit_floats(%source: i32, %stoch: i32) -> i32 {
   llvm.return %source5 : i32
 }
 
-// CHECK-DAG: attributes #[[$KERNEL_ATTRS]] = { "amdgpu-flat-work-group-size"="1,256" }
+// CHECK-DAG: attributes #[[$KERNEL_ATTRS]] = { "amdgpu-flat-work-group-size"="1,256" "uniform-work-group-size"="true" }
 // CHECK-DAG: attributes #[[$KERNEL_WORKGROUP_ATTRS]] = { "amdgpu-flat-work-group-size"="1,1024"
 // CHECK-DAG: attributes #[[$KNOWN_BLOCK_SIZE_ATTRS]] = { "amdgpu-flat-work-group-size"="128,128"
+// CHECK-DAG: attributes #[[$KERNEL_NO_UNIFORM_WORK_GROUPS_ATTRS]] = { "amdgpu-flat-work-group-size"="1,256" "uniform-work-group-size"="false" }
 // CHECK-DAG: ![[$RANGE]] = !{i32 0, i32 64}
 // CHECK-DAG: ![[$REQD_WORK_GROUP_SIZE]] = !{i32 16, i32 4, i32 2}