[mlir][gpu] Add an offloading handler attribute to gpu.module (#78047)

This patch adds an optional offloading handler attribute to the`gpu.module` op. This attribute will be used during `gpu-module-to-binary` pass to override the offloading handler used in the `gpu.binary` op.
ROCm-Developer-Tools · Jan 15, 2024 · 5b4f2b9 · 5b4f2b9
1 parent 8b6b882
commit 5b4f2b9
Show file tree

Hide file tree

Showing 6 changed files with 74 additions and 12 deletions.
diff --git a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
@@ -1194,7 +1194,9 @@ def GPU_BarrierOp : GPU_Op<"barrier"> {
 def GPU_GPUModuleOp : GPU_Op<"module", [
       DataLayoutOpInterface, HasDefaultDLTIDataLayout, IsolatedFromAbove,
       SymbolTable, Symbol, SingleBlockImplicitTerminator<"ModuleEndOp">
-    ]>, Arguments<(ins OptionalAttr<GPUNonEmptyTargetArrayAttr>:$targets)> {
+    ]>, Arguments<(ins
+          OptionalAttr<GPUNonEmptyTargetArrayAttr>:$targets,
+          OptionalAttr<OffloadingTranslationAttr>:$offloadingHandler)> {
   let summary = "A top level compilation unit containing code to be run on a GPU.";
   let description = [{
     GPU module contains code that is intended to be run on a GPU. A host device
@@ -1215,22 +1217,33 @@ def GPU_GPUModuleOp : GPU_Op<"module", [
     how to transform modules into binary strings and are used by the
     `gpu-module-to-binary` pass to transform modules into GPU binaries.
 
+    Modules can contain an optional `OffloadingTranslationAttr` attribute. This
+    attribute will be used during the `gpu-module-to-binary` pass to specify the
+    `OffloadingTranslationAttr` used when creating the `gpu.binary` operation.
+
     ```
     gpu.module @symbol_name {
       gpu.func {}
         ...
       gpu.module_end
     }
-    gpu.module @symbol_name2 [#nvvm.target, #rocdl.target<chip = "gfx90a">] {
+    // Module with offloading handler and target attributes.
+    gpu.module @symbol_name2 <#gpu.select_object<1>> [
+        #nvvm.target,
+        #rocdl.target<chip = "gfx90a">] {
       gpu.func {}
         ...
       gpu.module_end
     }
     ```
   }];
   let builders = [
-    OpBuilder<(ins "StringRef":$name, CArg<"ArrayAttr", "{}">:$targets)>,
-    OpBuilder<(ins "StringRef":$name, "ArrayRef<Attribute>":$targets)>
+    OpBuilder<(ins "StringRef":$name,
+                   CArg<"ArrayAttr", "{}">:$targets,
+                   CArg<"Attribute", "{}">:$handler)>,
+    OpBuilder<(ins "StringRef":$name,
+                   "ArrayRef<Attribute>":$targets,
+                   CArg<"Attribute", "{}">:$handler)>
   ];
   let regions = (region SizedRegion<1>:$bodyRegion);
   let hasCustomAssemblyFormat = 1;

diff --git a/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp b/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp
@@ -1724,19 +1724,24 @@ LogicalResult gpu::ReturnOp::verify() {
 //===----------------------------------------------------------------------===//
 
 void GPUModuleOp::build(OpBuilder &builder, OperationState &result,
-                        StringRef name, ArrayAttr targets) {
+                        StringRef name, ArrayAttr targets,
+                        Attribute offloadingHandler) {
   ensureTerminator(*result.addRegion(), builder, result.location);
   result.attributes.push_back(builder.getNamedAttr(
       ::mlir::SymbolTable::getSymbolAttrName(), builder.getStringAttr(name)));
 
+  Properties &props = result.getOrAddProperties<Properties>();
   if (targets)
-    result.getOrAddProperties<Properties>().targets = targets;
+    props.targets = targets;
+  props.offloadingHandler = offloadingHandler;
 }
 
 void GPUModuleOp::build(OpBuilder &builder, OperationState &result,
-                        StringRef name, ArrayRef<Attribute> targets) {
+                        StringRef name, ArrayRef<Attribute> targets,
+                        Attribute offloadingHandler) {
   build(builder, result, name,
-        targets.empty() ? ArrayAttr() : builder.getArrayAttr(targets));
+        targets.empty() ? ArrayAttr() : builder.getArrayAttr(targets),
+        offloadingHandler);
 }
 
 ParseResult GPUModuleOp::parse(OpAsmParser &parser, OperationState &result) {
@@ -1747,14 +1752,24 @@ ParseResult GPUModuleOp::parse(OpAsmParser &parser, OperationState &result) {
                              result.attributes))
     return failure();
 
+  Properties &props = result.getOrAddProperties<Properties>();
+
+  // Parse the optional offloadingHandler
+  if (succeeded(parser.parseOptionalLess())) {
+    if (parser.parseAttribute(props.offloadingHandler))
+      return failure();
+    if (parser.parseGreater())
+      return failure();
+  }
+
   // Parse the optional array of target attributes.
   OptionalParseResult targetsAttrResult =
       parser.parseOptionalAttribute(targetsAttr, Type{});
   if (targetsAttrResult.has_value()) {
     if (failed(*targetsAttrResult)) {
       return failure();
     }
-    result.getOrAddProperties<Properties>().targets = targetsAttr;
+    props.targets = targetsAttr;
   }
 
   // If module attributes are present, parse them.
@@ -1775,15 +1790,22 @@ void GPUModuleOp::print(OpAsmPrinter &p) {
   p << ' ';
   p.printSymbolName(getName());
 
+  if (Attribute attr = getOffloadingHandlerAttr()) {
+    p << " <";
+    p.printAttribute(attr);
+    p << ">";
+  }
+
   if (Attribute attr = getTargetsAttr()) {
     p << ' ';
     p.printAttribute(attr);
     p << ' ';
   }
 
-  p.printOptionalAttrDictWithKeyword(
-      (*this)->getAttrs(),
-      {mlir::SymbolTable::getSymbolAttrName(), getTargetsAttrName()});
+  p.printOptionalAttrDictWithKeyword((*this)->getAttrs(),
+                                     {mlir::SymbolTable::getSymbolAttrName(),
+                                      getTargetsAttrName(),
+                                      getOffloadingHandlerAttrName()});
   p << ' ';
   p.printRegion(getRegion(), /*printEntryBlockArgs=*/false,
                 /*printBlockTerminators=*/false);

diff --git a/mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp b/mlir/lib/Dialect/GPU/Transforms/ModuleToBinary.cpp
@@ -124,6 +124,11 @@ LogicalResult moduleSerializer(GPUModuleOp op,
     }
     objects.push_back(object);
   }
+  if (auto moduleHandler =
+          dyn_cast_or_null<OffloadingLLVMTranslationAttrInterface>(
+              op.getOffloadingHandlerAttr());
+      !handler && moduleHandler)
+    handler = moduleHandler;
   builder.setInsertionPointAfter(op);
   builder.create<gpu::BinaryOp>(op.getLoc(), op.getName(), handler,
                                 builder.getArrayAttr(objects));

diff --git a/mlir/test/Dialect/GPU/invalid.mlir b/mlir/test/Dialect/GPU/invalid.mlir
@@ -818,3 +818,10 @@ func.func @main(%arg0 : index) {
   return
 }
 
+// -----
+
+module attributes {gpu.container_module} {
+  // expected-error@+1 {{expected attribute value}}
+  gpu.module @kernel <> {
+  }
+}
diff --git a/mlir/test/Dialect/GPU/module-to-binary-nvvm.mlir b/mlir/test/Dialect/GPU/module-to-binary-nvvm.mlir
@@ -22,4 +22,16 @@ module attributes {gpu.container_module} {
       llvm.return
     }
   }
+
+  // CHECK-LABEL:gpu.binary @kernel_module3 <#gpu.select_object<1 : i64>>
+  // CHECK:[#gpu.object<#nvvm.target<chip = "sm_70">, offload = "{{.*}}">, #gpu.object<#nvvm.target<chip = "sm_80">, offload = "{{.*}}">]
+  gpu.module @kernel_module3 <#gpu.select_object<1>> [
+      #nvvm.target<chip = "sm_70">,
+      #nvvm.target<chip = "sm_80">] {
+    llvm.func @kernel(%arg0: i32, %arg1: !llvm.ptr,
+        %arg2: !llvm.ptr, %arg3: i64, %arg4: i64,
+        %arg5: i64) attributes {gpu.kernel} {
+      llvm.return
+    }
+  }
 }
diff --git a/mlir/test/Dialect/GPU/ops.mlir b/mlir/test/Dialect/GPU/ops.mlir
@@ -423,3 +423,6 @@ gpu.module @module_with_two_target [#nvvm.target, #rocdl.target<chip = "gfx90a">
     gpu.return
   }
 }
+
+gpu.module @module_with_offload_handler <#gpu.select_object<0>> [#nvvm.target] {
+}