Add a pass that inserts getters for all cubins found via nvvm.cubin

annotations.

Getters are required as there are currently no global constants in MLIR and this
is an easy way to unblock CUDA execution while waiting for those.

PiperOrigin-RevId: 255169002

diff --git a/mlir/include/mlir/Conversion/GPUToCUDA/GPUToCUDAPass.h b/mlir/include/mlir/Conversion/GPUToCUDA/GPUToCUDAPass.h
index dac703a..7457a07 100644 (file)
--- a/mlir/include/mlir/Conversion/GPUToCUDA/GPUToCUDAPass.h
+++ b/mlir/include/mlir/Conversion/GPUToCUDA/GPUToCUDAPass.h
@@ -51,6 +51,9 @@ createConvertGPUKernelToCubinPass(CubinGenerator cubinGenerator);
 /// ontop of CUDA.
 ModulePassBase *createConvertGpuLaunchFuncToCudaCallsPass();
 
+/// Creates a pass to augment a module with getter functions for all contained
+/// cubins as encoded via the 'nvvm.cubin' attribute.
+ModulePassBase *createGenerateCubinAccessorPass();
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_GPUTOCUDA_GPUTOCUDAPASS_H_

diff --git a/mlir/lib/Conversion/GPUToCUDA/CMakeLists.txt b/mlir/lib/Conversion/GPUToCUDA/CMakeLists.txt
index 4eddb78..fbaf36c 100644 (file)
--- a/mlir/lib/Conversion/GPUToCUDA/CMakeLists.txt
+++ b/mlir/lib/Conversion/GPUToCUDA/CMakeLists.txt
@@ -4,6 +4,7 @@ if(MLIR_CUDA_CONVERSIONS_ENABLED)
   add_llvm_library(MLIRGPUtoCUDATransforms
     ConvertKernelFuncToCubin.cpp
     ConvertLaunchFuncToCudaCalls.cpp
+    GenerateCubinAccessors.cpp
   )
   target_link_libraries(MLIRGPUtoCUDATransforms
     MLIRGPU

diff --git a/mlir/lib/Conversion/GPUToCUDA/GenerateCubinAccessors.cpp b/mlir/lib/Conversion/GPUToCUDA/GenerateCubinAccessors.cpp
new file mode 100644 (file)
index 0000000..c1d4af3
--- /dev/null
+++ b/mlir/lib/Conversion/GPUToCUDA/GenerateCubinAccessors.cpp
@@ -0,0 +1,147 @@
+//===- GenerateCubinAccessors.cpp - MLIR GPU lowering passes --------------===//
+//
+// Copyright 2019 The MLIR Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+// =============================================================================
+//
+// This file implements a pass to generate LLVMIR functions that return the
+// data stored in nvvm.cubin char* blob.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/GPU/GPUDialect.h"
+#include "mlir/IR/Attributes.h"
+#include "mlir/IR/Builders.h"
+#include "mlir/IR/Function.h"
+#include "mlir/IR/Identifier.h"
+#include "mlir/IR/Module.h"
+#include "mlir/IR/StandardTypes.h"
+#include "mlir/LLVMIR/LLVMDialect.h"
+#include "mlir/Pass/Pass.h"
+#include "mlir/Pass/PassRegistry.h"
+
+#include "llvm/ADT/STLExtras.h"
+
+namespace mlir {
+namespace {
+
+// TODO(herhut): Move to shared location.
+constexpr const char *kCubinAnnotation = "nvvm.cubin";
+constexpr const char *kCubinGetterAnnotation = "nvvm.cubingetter";
+constexpr const char *kCubinGetterSuffix = "_cubin";
+constexpr const char *kMallocHelperName = "mcuMalloc";
+
+/// A pass generating getter functions for all cubin blobs annotated on
+/// functions via the nvvm.cubin attribute.
+///
+/// The functions allocate memory using a mcuMalloc helper function with the
+/// signature void *mcuMalloc(int32_t size). This function has to be provided
+/// by the actual runner that executes the generated code.
+///
+/// This is a stop-gap measure until MLIR supports global constants.
+class GpuGenerateCubinAccessorsPass
+    : public ModulePass<GpuGenerateCubinAccessorsPass> {
+private:
+  Function *getMallocHelper(Location loc, Builder &builder) {
+    Function *result = getModule().getNamedFunction(kMallocHelperName);
+    if (!result) {
+      result = new Function(
+          loc, kMallocHelperName,
+          builder.getFunctionType(
+              ArrayRef<Type>{LLVM::LLVMType::getInt32Ty(llvmDialect)},
+              LLVM::LLVMType::getInt8PtrTy(llvmDialect)));
+      getModule().getFunctions().push_back(result);
+    }
+    return result;
+  }
+
+  // Generates a function that returns a char array at runtime that contains the
+  // data from blob. As there are currently no global constants, this uses a
+  // sequence of store operations.
+  // TODO(herhut): Use global constants instead.
+  Function *generateCubinAccessor(Builder &builder, Function &orig,
+                                  StringAttr blob) {
+    Location loc = orig.getLoc();
+    SmallString<128> nameBuffer(orig.getName());
+    nameBuffer.append(kCubinGetterSuffix);
+    // Generate a function that returns void*.
+    Function *result = new Function(
+        loc, mlir::Identifier::get(nameBuffer, &getContext()),
+        builder.getFunctionType(ArrayRef<Type>{},
+                                LLVM::LLVMType::getInt8PtrTy(llvmDialect)));
+    // Insert a body block that just returns the constant.
+    OpBuilder ob(result->getBody());
+    ob.createBlock();
+    auto sizeConstant = ob.create<LLVM::ConstantOp>(
+        loc, LLVM::LLVMType::getInt32Ty(llvmDialect),
+        builder.getI32IntegerAttr(blob.getValue().size()));
+    auto memory =
+        ob.create<LLVM::CallOp>(
+              loc, ArrayRef<Type>{LLVM::LLVMType::getInt8PtrTy(llvmDialect)},
+              builder.getFunctionAttr(getMallocHelper(loc, builder)),
+              ArrayRef<Value *>{sizeConstant})
+            .getResult(0);
+    for (auto byte : llvm::enumerate(blob.getValue().bytes())) {
+      auto index = ob.create<LLVM::ConstantOp>(
+          loc, LLVM::LLVMType::getInt32Ty(llvmDialect),
+          builder.getI32IntegerAttr(byte.index()));
+      auto gep =
+          ob.create<LLVM::GEPOp>(loc, LLVM::LLVMType::getInt8PtrTy(llvmDialect),
+                                 memory, ArrayRef<Value *>{index});
+      auto value = ob.create<LLVM::ConstantOp>(
+          loc, LLVM::LLVMType::getInt8Ty(llvmDialect),
+          builder.getIntegerAttr(builder.getIntegerType(8), byte.value()));
+      ob.create<LLVM::StoreOp>(loc, value, gep);
+    }
+    ob.create<LLVM::ReturnOp>(loc, ArrayRef<Value *>{memory});
+    // Store the name of the getter on the function for easier lookup.
+    orig.setAttr(kCubinGetterAnnotation, builder.getFunctionAttr(result));
+    return result;
+  }
+
+public:
+  // Run the dialect converter on the module.
+  void runOnModule() override {
+    llvmDialect =
+        getModule().getContext()->getRegisteredDialect<LLVM::LLVMDialect>();
+    Builder builder(getModule().getContext());
+
+    auto &functions = getModule().getFunctions();
+    for (auto it = functions.begin(); it != functions.end();) {
+      // Move iterator to after the current function so that potential insertion
+      // of the accessor is after the kernel with cubin iself.
+      Function &orig = *it++;
+      StringAttr cubinBlob = orig.getAttrOfType<StringAttr>(kCubinAnnotation);
+      if (!cubinBlob)
+        continue;
+      it =
+          functions.insert(it, generateCubinAccessor(builder, orig, cubinBlob));
+    }
+  }
+
+private:
+  LLVM::LLVMDialect *llvmDialect;
+};
+
+} // anonymous namespace
+
+ModulePassBase *createGenerateCubinAccessorPass() {
+  return new GpuGenerateCubinAccessorsPass();
+}
+
+static PassRegistration<GpuGenerateCubinAccessorsPass>
+    pass("generate-cubin-accessors",
+         "Generate LLVMIR functions that give access to cubin data");
+
+} // namespace mlir

diff --git a/mlir/test/Conversion/GPUToCUDA/insert-cubin-getter.mlir b/mlir/test/Conversion/GPUToCUDA/insert-cubin-getter.mlir
new file mode 100644 (file)
index 0000000..19090ce
--- /dev/null
+++ b/mlir/test/Conversion/GPUToCUDA/insert-cubin-getter.mlir
@@ -0,0 +1,31 @@
+// RUN: mlir-opt %s --generate-cubin-accessors | FileCheck %s
+
+func @kernel(!llvm.float, !llvm<"float*">)
+// CHECK: attributes  {gpu.kernel, nvvm.cubin = "CUBIN", nvvm.cubingetter = @kernel_cubin}
+  attributes  {gpu.kernel, nvvm.cubin = "CUBIN"}
+
+// CHECK: func @mcuMalloc(!llvm.i32) -> !llvm<"i8*">
+// CHECK: func @kernel_cubin() -> !llvm<"i8*">
+// CHECK-NEXT: %0 = llvm.constant(5 : i32) : !llvm.i32
+// CHECK-NEXT: %1 = llvm.call @mcuMalloc(%0) : (!llvm.i32) -> !llvm<"i8*">
+// CHECK-NEXT: %2 = llvm.constant(0 : i32) : !llvm.i32
+// CHECK-NEXT: %3 = llvm.getelementptr %1[%2] : (!llvm<"i8*">, !llvm.i32) -> !llvm<"i8*">
+// CHECK-NEXT: %4 = llvm.constant(67 : i8) : !llvm.i8
+// CHECK-NEXT: llvm.store %4, %3 : !llvm<"i8*">
+// CHECK-NEXT: %5 = llvm.constant(1 : i32) : !llvm.i32
+// CHECK-NEXT: %6 = llvm.getelementptr %1[%5] : (!llvm<"i8*">, !llvm.i32) -> !llvm<"i8*">
+// CHECK-NEXT: %7 = llvm.constant(85 : i8) : !llvm.i8
+// CHECK-NEXT: llvm.store %7, %6 : !llvm<"i8*">
+// CHECK-NEXT: %8 = llvm.constant(2 : i32) : !llvm.i32
+// CHECK-NEXT: %9 = llvm.getelementptr %1[%8] : (!llvm<"i8*">, !llvm.i32) -> !llvm<"i8*">
+// CHECK-NEXT: %10 = llvm.constant(66 : i8) : !llvm.i8
+// CHECK-NEXT: llvm.store %10, %9 : !llvm<"i8*">
+// CHECK-NEXT: %11 = llvm.constant(3 : i32) : !llvm.i32
+// CHECK-NEXT: %12 = llvm.getelementptr %1[%11] : (!llvm<"i8*">, !llvm.i32) -> !llvm<"i8*">
+// CHECK-NEXT: %13 = llvm.constant(73 : i8) : !llvm.i8
+// CHECK-NEXT: llvm.store %13, %12 : !llvm<"i8*">
+// CHECK-NEXT: %14 = llvm.constant(4 : i32) : !llvm.i32
+// CHECK-NEXT: %15 = llvm.getelementptr %1[%14] : (!llvm<"i8*">, !llvm.i32) -> !llvm<"i8*">
+// CHECK-NEXT: %16 = llvm.constant(78 : i8) : !llvm.i8
+// CHECK-NEXT: llvm.store %16, %15 : !llvm<"i8*">
+// CHECK-NEXT: llvm.return %1 : !llvm<"i8*">