[mlir][gpu] Modifies `gpu.launch_func` to allow lowering it after gpu-to-llvm.

[fmorac]

Summary

For an explanation of these patches see D154153.

Commit message:
In order to lower gpu.launch_func after running gpu-to-llvm it must be
able to handle lowered types -eg. index -> i64. This patch also allows the op
to refer to GPU binaries and not only GPU modules.

Depends on D154132.

[fmorac]

Correct dependencies.

[fmorac]

Update dependencies.

[fmorac]

Rebasing.

[krzysz00]

Comments wrt assembly format

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

1081

This is just AllTypesMatch in the tablegen

1092

I don't think this'll be needed with an AllTypesMatch

[fmorac]

Added the AllTypesMatch trait.

[mehdi_amini]

Please add tests here

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

328–329

Update the comment?

1001–1039

Since we're touching the build methods, can you update these to use properties instead?

should look like:

Properties &prop = result.getOrAddProperties<Properties>();
prop.operand_segment_sizes = builder.getDenseI32ArrayAttr(segmentSizes);

[fmorac]

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

328–329

I'll update it.

1001–1039

Yes.

[mehdi_amini]

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

1001–1039

Actually prop.odsOperandSegmentSizes, I just changed it ;)

[fmorac]

Change builders to use properties.

[mehdi_amini]

Can you add tests for the new syntax and new types?

[fmorac]

In D154137#4556082, @mehdi_amini wrote:

Can you add tests for the new syntax and new types?

Sure, I can add syntax & type tests here, however tests for verifying the compatibility with gpu.binary have to be added after the introduction of #gpu.select_object.

[fmorac]

Added tests for the syntax.

[mehdi_amini]

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

1039

Sorry, I removed the ods prefix...

[fmorac]

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

1039

Is that change in trunk? Because I just did a build and it's still using odsOperandSegmentSizes.

[mehdi_amini]

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

1039

I forgot to push... now it is there :)

[fmorac]

Removing Ods prefix.

mlir/include/mlir/Dialect/GPU/IR/GPUOps.td

let extraClassDeclaration = [{
     /// Verifies the body of the function.
     LogicalResult verifyBody();
   }];
   let hasCustomAssemblyFormat = 1;
 
   let hasVerifier = 1;
 }
 
-def GPU_LaunchFuncOp : GPU_Op<"launch_func",
-                              [GPU_AsyncOpInterface, AttrSizedOperandSegments]>,
+def LaunchIndx : AnyTypeOf<[Index, I32, I64]>;
+
+def GPU_LaunchFuncOp :GPU_Op<"launch_func", [
+      GPU_AsyncOpInterface, AttrSizedOperandSegments,
+      AllTypesMatch<["gridSizeX", "gridSizeY", "gridSizeZ", "blockSizeX",
+                     "blockSizeY", "blockSizeZ"]>]>,
     Arguments<(ins Variadic<GPU_AsyncToken>:$asyncDependencies,
                SymbolRefAttr:$kernel,
-               Index:$gridSizeX, Index:$gridSizeY, Index:$gridSizeZ,
-               Index:$blockSizeX, Index:$blockSizeY, Index:$blockSizeZ,
+               LaunchIndx:$gridSizeX, LaunchIndx:$gridSizeY, LaunchIndx:$gridSizeZ,
+               LaunchIndx:$blockSizeX, LaunchIndx:$blockSizeY, LaunchIndx:$blockSizeZ,
                Optional<I32>:$dynamicSharedMemorySize,
-               Variadic<AnyType>:$kernelOperands)>,
+               Variadic<AnyType>:$kernelOperands,
+               Optional<AnyType>:$asyncObject)>,
     Results<(outs Optional<GPU_AsyncToken>:$asyncToken)> {
   let summary = "Launches a function as a GPU kernel";
 
   let description = [{
     Launch a kernel function on the specified grid of thread blocks.
     `gpu.launch` operations are lowered to `gpu.launch_func` operations by
     outlining the kernel body into a function in a dedicated module, which
     reflects the separate compilation process. The kernel function is required
def GPU_LaunchFuncOp :GPU_Op<"launch_func", [
 
   let skipDefaultBuilders = 1;
 
   let builders = [
     OpBuilder<(ins "GPUFuncOp":$kernelFunc, "KernelDim3":$gridSize,
       "KernelDim3":$blockSize, "Value":$dynamicSharedMemorySize,
       "ValueRange":$kernelOperands,
       CArg<"Type", "nullptr">:$asyncTokenType,
-      CArg<"ValueRange", "{}">:$asyncDependencies)>
+      CArg<"ValueRange", "{}">:$asyncDependencies)>,
+    OpBuilder<(ins "SymbolRefAttr":$kernel, "KernelDim3":$gridSize,
+      "KernelDim3":$blockSize, "Value":$dynamicSharedMemorySize,
+      "ValueRange":$kernelOperands,
+      CArg<"Value", "nullptr">:$asyncObject)>
   ];
 
   let extraClassDeclaration = [{
     /// The name of the kernel's containing module.
     StringAttr getKernelModuleName();
 
     /// The name of the kernel.
     StringAttr getKernelName();
let extraClassDeclaration = [{
     // This needs to quietly verify if attributes with names defined below are
     // present since it is run before the verifier of this op.
     friend LogicalResult GPUDialect::verifyOperationAttribute(Operation *,
                                                               NamedAttribute);
   }];
 
   let assemblyFormat = [{
       custom<AsyncDependencies>(type($asyncToken), $asyncDependencies)
+      (`<` $asyncObject^ `:` type($asyncObject) `>`)?
       $kernel
-      `blocks` `in` ` ` `(`$gridSizeX`,` $gridSizeY`,` $gridSizeZ`)`
-      `threads` `in` ` ` `(`$blockSizeX`,` $blockSizeY`,` $blockSizeZ`)`
+      `blocks` `in` ` ` `(` $gridSizeX `,` $gridSizeY `,` $gridSizeZ `)`
+      `threads` `in` ` ` `(` $blockSizeX `,` $blockSizeY `,` $blockSizeZ `)`
+      custom<LaunchDimType>(type($gridSizeX))
       (`dynamic_shared_memory_size` $dynamicSharedMemorySize^)?
       custom<LaunchFuncOperands>($kernelOperands, type($kernelOperands)) attr-dict
   }];
   let hasVerifier = 1;
 }
 
 def GPU_LaunchOp : GPU_Op<"launch", [
       AutomaticAllocationScope, AttrSizedOperandSegments, GPU_AsyncOpInterface,

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

if (!launchOp->getParentOp() ||
         launchOp->getParentOp()->getParentOp() != module)
       return success();
 
     // Ignore launch ops with missing attributes here. The errors will be
     // reported by the verifiers of those ops.
     if (!launchOp->getAttrOfType<SymbolRefAttr>(
             LaunchFuncOp::getKernelAttrName(launchOp->getName())))
       return success();
 
-    // Check that `launch_func` refers to a well-formed GPU kernel module.
-    StringAttr kernelModuleName = launchOp.getKernelModuleName();
-    auto kernelModule = module.lookupSymbol<GPUModuleOp>(kernelModuleName);
+    // Check that `launch_func` refers to a well-formed GPU kernel container.
+    StringAttr kernelContainerName = launchOp.getKernelModuleName();
+    Operation *kernelContainer = module.lookupSymbol(kernelContainerName);
+    if (!kernelContainer)
+      return launchOp.emitOpError()
+             << "kernel container '" << kernelContainerName.getValue()
+             << "' is undefined";
+
+    // If the container is a GPU binary op return success.
+    if (isa<BinaryOp>(kernelContainer))
+      return success();
+
+    auto kernelModule = dyn_cast<GPUModuleOp>(kernelContainer);
     if (!kernelModule)
       return launchOp.emitOpError()
-             << "kernel module '" << kernelModuleName.getValue()
+             << "kernel module '" << kernelContainerName.getValue()
              << "' is undefined";
 
     // Check that `launch_func` refers to a well-formed kernel function.
     Operation *kernelFunc = module.lookupSymbol(launchOp.getKernelAttr());
     if (!kernelFunc)
       return launchOp.emitOpError("kernel function '")
              << launchOp.getKernel() << "' is undefined";
     auto kernelConvertedFunction = dyn_cast<FunctionOpInterface>(kernelFunc);
result.addOperands({gridSize.x, gridSize.y, gridSize.z, getBlockSize.x,
                       getBlockSize.y, getBlockSize.z});
   if (dynamicSharedMemorySize)
     result.addOperands(dynamicSharedMemorySize);
   result.addOperands(kernelOperands);
   auto kernelModule = kernelFunc->getParentOfType<GPUModuleOp>();
   auto kernelSymbol =
       SymbolRefAttr::get(kernelModule.getNameAttr(),
                          {SymbolRefAttr::get(kernelFunc.getNameAttr())});
-  result.addAttribute(getKernelAttrName(result.name), kernelSymbol);
-  SmallVector<int32_t, 9> segmentSizes(9, 1);
-  segmentSizes.front() = asyncDependencies.size();
-  segmentSizes[segmentSizes.size() - 2] = dynamicSharedMemorySize ? 1 : 0;
-  segmentSizes.back() = static_cast<int32_t>(kernelOperands.size());
-  result.addAttribute(getOperandSegmentSizeAttr(),
-                      builder.getDenseI32ArrayAttr(segmentSizes));
+
+  Properties &prop = result.getOrAddProperties<Properties>();
+  prop.kernel = kernelSymbol;
+  size_t segmentSizesLen = std::size(prop.operandSegmentSizes);
+  // Initialize the segment sizes to 1.
+  for (auto &sz : prop.operandSegmentSizes)
+    sz = 1;
+  prop.operandSegmentSizes[0] = asyncDependencies.size();
+  prop.operandSegmentSizes[segmentSizesLen - 3] =
+      dynamicSharedMemorySize ? 1 : 0;
+  prop.operandSegmentSizes[segmentSizesLen - 2] =
+      static_cast<int32_t>(kernelOperands.size());
+  prop.operandSegmentSizes[segmentSizesLen - 1] = 0;
+}
+
+void LaunchFuncOp::build(OpBuilder &builder, OperationState &result,
+                         SymbolRefAttr kernel, KernelDim3 gridSize,
+                         KernelDim3 getBlockSize, Value dynamicSharedMemorySize,
+                         ValueRange kernelOperands, Value asyncObject) {
+  // Add grid and block sizes as op operands, followed by the data operands.
+  result.addOperands({gridSize.x, gridSize.y, gridSize.z, getBlockSize.x,
+                      getBlockSize.y, getBlockSize.z});
+  if (dynamicSharedMemorySize)
+    result.addOperands(dynamicSharedMemorySize);
+  result.addOperands(kernelOperands);
+  if (asyncObject)
+    result.addOperands(asyncObject);
+  Properties &prop = result.getOrAddProperties<Properties>();
+  prop.kernel = kernel;
+  size_t segmentSizesLen = std::size(prop.operandSegmentSizes);
+  // Initialize the segment sizes to 1.
+  for (auto &sz : prop.operandSegmentSizes)
+    sz = 1;
+  prop.operandSegmentSizes[0] = 0;
+  prop.operandSegmentSizes[segmentSizesLen - 3] =
+      dynamicSharedMemorySize ? 1 : 0;
+  prop.operandSegmentSizes[segmentSizesLen - 2] =
+      static_cast<int32_t>(kernelOperands.size());
+  prop.operandSegmentSizes[segmentSizesLen - 1] = asyncObject ? 1 : 0;
 }
 
 StringAttr LaunchFuncOp::getKernelModuleName() {
   return getKernel().getRootReference();
 }
 
 StringAttr LaunchFuncOp::getKernelName() {
   return getKernel().getLeafReference();
LogicalResult LaunchFuncOp::verify() {
   if (!module->getAttrOfType<UnitAttr>(
           GPUDialect::getContainerModuleAttrName()))
     return emitOpError("expected the closest surrounding module to have the '" +
                        GPUDialect::getContainerModuleAttrName() +
                        "' attribute");
 
   return success();
 }
 
+static ParseResult parseLaunchDimType(OpAsmParser &parser, Type &dimTy) {
+  if (succeeded(parser.parseOptionalColon())) {
+    if (parser.parseType(dimTy))
+      return failure();
+  } else {
+    dimTy = IndexType::get(parser.getContext());
+  }
+  return success();
+}
+
+static void printLaunchDimType(OpAsmPrinter &printer, Operation *op,
+                               Type dimTy) {
+  if (!dimTy.isIndex())
+    printer << ": " << dimTy;
+}
+
 static ParseResult parseLaunchFuncOperands(
     OpAsmParser &parser,
     SmallVectorImpl<OpAsmParser::UnresolvedOperand> &argNames,
     SmallVectorImpl<Type> &argTypes) {
   if (parser.parseOptionalKeyword("args"))
     return success();
 
   auto parseElement = [&]() -> ParseResult {

mlir/test/Dialect/GPU/invalid.mlir

 }
 
 // -----
 
 module attributes {gpu.container_module} {
   func.func @launch_func_missing_callee_attribute(%sz : index) {
     // expected-error@+1 {{'gpu.launch_func' op requires attribute 'kernel'}}
     "gpu.launch_func"(%sz, %sz, %sz, %sz, %sz, %sz)
-        {operandSegmentSizes = array<i32: 0, 1, 1, 1, 1, 1, 1, 0, 0>}
+        {operandSegmentSizes = array<i32: 0, 1, 1, 1, 1, 1, 1, 0, 0, 0>}
         : (index, index, index, index, index, index) -> ()
     return
   }
 }
 
 // -----
 
 module attributes {gpu.container_module} {
   func.func @launch_func_no_function_attribute(%sz : index) {
     // expected-error@+1 {{custom op 'gpu.launch_func' invalid kind of attribute specified}}
     gpu.launch_func "foo" blocks in (%sz, %sz, %sz) threads in (%sz, %sz, %sz)
     return
   }
 }
 
 // -----
 
 module attributes {gpu.container_module} {
   func.func @launch_func_undefined_module(%sz : index) {
-    // expected-error@+1 {{kernel module 'kernels' is undefined}}
+    // expected-error@+1 {{kernel container 'kernels' is undefined}}
     gpu.launch_func @kernels::@kernel_1 blocks in (%sz, %sz, %sz) threads in (%sz, %sz, %sz)
     return
   }
 }
 
 // -----
 
 module attributes {gpu.container_module} {

mlir/test/Dialect/GPU/ops.mlir

module attributes {gpu.container_module} {
 
   func.func private @two_value_generator() -> (f32, memref<?xf32, 1>)
 
   func.func @foo() {
     %0 = "op"() : () -> (f32)
     %1 = "op"() : () -> (memref<?xf32, 1>)
     // CHECK: %{{.*}} = arith.constant 8
     %cst = arith.constant 8 : index
+    %cstI64 = arith.constant 8 : i64
     %c0 = arith.constant 0 : i32
     %t0 = gpu.wait async
+    %lowStream = llvm.mlir.null : !llvm.ptr
 
     // CHECK: gpu.launch_func @kernels::@kernel_1 blocks in (%{{.*}}, %{{.*}}, %{{.*}}) threads in (%{{.*}}, %{{.*}}, %{{.*}}) args(%{{.*}} : f32, %{{.*}} : memref<?xf32, 1>)
     gpu.launch_func @kernels::@kernel_1 blocks in (%cst, %cst, %cst) threads in (%cst, %cst, %cst) args(%0 : f32, %1 : memref<?xf32, 1>)
 
     gpu.launch_func @kernels::@kernel_1 blocks in (%cst, %cst, %cst) threads in (%cst, %cst, %cst) dynamic_shared_memory_size %c0 args(%0 : f32, %1 : memref<?xf32, 1>)
 
     // CHECK: gpu.launch_func @kernels::@kernel_2 blocks in (%{{.*}}, %{{.*}}, %{{.*}}) threads in (%{{.*}}, %{{.*}}, %{{.*}})
     gpu.launch_func @kernels::@kernel_2 blocks in (%cst, %cst, %cst) threads in (%cst, %cst, %cst)
 
     // CHECK: %{{.*}} = gpu.launch_func async [%{{.*}}] @kernels::@kernel_2 blocks in (%{{.*}}, %{{.*}}, %{{.*}}) threads in (%{{.*}}, %{{.*}}, %{{.*}})
     %t1 = gpu.launch_func async [%t0] @kernels::@kernel_2  blocks in (%cst, %cst, %cst) threads in (%cst, %cst, %cst)
 
+    // CHECK: gpu.launch_func <%{{.*}} : !llvm.ptr> @kernels::@kernel_1 blocks in (%{{.*}}, %{{.*}}, %{{.*}}) threads in (%{{.*}}, %{{.*}}, %{{.*}}) : i64 args(%{{.*}} : f32, %{{.*}} : memref<?xf32, 1>)
+    gpu.launch_func <%lowStream : !llvm.ptr> @kernels::@kernel_1 blocks in (%cstI64, %cstI64, %cstI64) threads in (%cstI64, %cstI64, %cstI64) : i64 args(%0 : f32, %1 : memref<?xf32, 1>)
+
+    // CHECK: gpu.launch_func @kernels::@kernel_1 blocks in (%{{.*}}, %{{.*}}, %{{.*}}) threads in (%{{.*}}, %{{.*}}, %{{.*}}) : i32 args(%{{.*}} : f32, %{{.*}} : memref<?xf32, 1>)
+    gpu.launch_func @kernels::@kernel_1 blocks in (%c0, %c0, %c0) threads in (%c0, %c0, %c0) : i32 args(%0 : f32, %1 : memref<?xf32, 1>)
+
     // CHECK: %[[VALUES:.*]]:2 = call
     %values:2 = func.call @two_value_generator() : () -> (f32, memref<?xf32, 1>)
     // CHECK: gpu.launch_func @kernels::@kernel_1 {{.*}} args(%[[VALUES]]#0 : f32, %[[VALUES]]#1 : memref<?xf32, 1>)
     gpu.launch_func @kernels::@kernel_1 blocks in (%cst, %cst, %cst) threads in (%cst, %cst, %cst) args(%values#0 : f32, %values#1 : memref<?xf32, 1>)
 
     return
   }