Load ukernel bitcode as executable_object at the time of lowering to ukernels.

build_tools/bazel_to_cmake/bazel_to_cmake_converter.py

@@ -616,7 +616,7 @@ def iree_amdgpu_bitcode_library(self, name, gpu_arch, srcs, copts=None, out=None
             "GPU_ARCH", gpu_arch, quote=False
         )
         srcs_block = self._convert_srcs_block(srcs)
-        out_block = self._convert_string_arg_block("OUT", out, quote=False)
+        out_block = self._convert_string_arg_block("OUT", out, quote=True)
         copts_block = self._convert_string_list_block("COPTS", copts, sort=False)
 
         self._converter.body += (

compiler/plugins/target/ROCM/BUILD.bazel

@@ -27,10 +27,7 @@ iree_compiler_cc_library(
         "ROCMTargetUtils.h",
     ],
     deps = [
-        "//compiler/plugins/target/ROCM/builtins/ukernel:iree_uk_amdgpu_gfx1030",
-        "//compiler/plugins/target/ROCM/builtins/ukernel:iree_uk_amdgpu_gfx1100",
-        "//compiler/plugins/target/ROCM/builtins/ukernel:iree_uk_amdgpu_gfx90a",
-        "//compiler/plugins/target/ROCM/builtins/ukernel:iree_uk_amdgpu_gfx942",
+        "//compiler/plugins/target/ROCM/builtins/ukernel:iree_uk_amdgpu_bitcode",
         "//compiler/src/iree/compiler/Codegen/Dialect/Codegen/IR:IREECodegenDialect",
         "//compiler/src/iree/compiler/Codegen/Dialect/GPU/IR:IREEGPUDialect",
         "//compiler/src/iree/compiler/Codegen/Dialect/GPU/TargetUtils:KnownTargets",

compiler/plugins/target/ROCM/CMakeLists.txt

@@ -64,10 +64,7 @@ iree_cc_library(
     iree::compiler::Dialect::HAL::Utils::LLVMLinkerUtils
     iree::compiler::PluginAPI
     iree::compiler::Utils
-    iree::compiler::plugins::target::ROCM::builtins::ukernel::iree_uk_amdgpu_gfx1030
-    iree::compiler::plugins::target::ROCM::builtins::ukernel::iree_uk_amdgpu_gfx1100
-    iree::compiler::plugins::target::ROCM::builtins::ukernel::iree_uk_amdgpu_gfx90a
-    iree::compiler::plugins::target::ROCM::builtins::ukernel::iree_uk_amdgpu_gfx942
+    iree::compiler::plugins::target::ROCM::builtins::ukernel::iree_uk_amdgpu_bitcode
     iree::schemas::amdgpu_executable_def_c_fbs
     iree::schemas::executable_debug_info_c_fbs
     iree::schemas::hip_executable_def_c_fbs

compiler/plugins/target/ROCM/ROCMTarget.cpp

@@ -8,6 +8,7 @@
 
 #include <cstdint>
 
+#include "compiler/plugins/target/ROCM/builtins/ukernel/iree_uk_amdgpu_bitcode.h"
 #include "iree/compiler/Codegen/Dialect/Codegen/IR/IREECodegenDialect.h"
 #include "iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.h"
 #include "iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUDialect.h"
@@ -21,6 +22,7 @@
 #include "iree/compiler/Dialect/HAL/Utils/ExecutableDebugInfoUtils.h"
 #include "iree/compiler/Dialect/HAL/Utils/LLVMLinkerUtils.h"
 #include "iree/compiler/PluginAPI/Client.h"
+#include "iree/compiler/Utils/EmbeddedDataDirectory.h"
 #include "iree/compiler/Utils/FlatbufferUtils.h"
 #include "iree/compiler/Utils/ToolUtils.h"
 #include "iree/schemas/amdgpu_executable_def_builder.h"
@@ -206,6 +208,7 @@ static std::string translateModuleToISA(llvm::Module &module,
   }
   return targetISA;
 }
+
 } // namespace
 
 class ROCMTargetBackend final : public TargetBackend {
@@ -513,20 +516,6 @@ class ROCMTargetBackend final : public TargetBackend {
         return failure();
       }
 
-      // Link module to any enabled ukernels.
-      StringRef bitcodeDirectory = options.bitcodeDirectory;
-      StringRef enabledUkernels;
-      if (auto attr = getConfigStringAttr(targetAttr, "ukernels"))
-        enabledUkernels = attr->getValue();
-      if (!enabledUkernels.empty() && enabledUkernels != "none") {
-        if (failed(linkUkernelBitcodeFiles(
-                variantOp.getLoc(), llvmModule.get(), enabledUkernels,
-                targetArch, bitcodeDirectory, llvm::Linker::OverrideFromSrc,
-                *targetMachine))) {
-          return failure();
-        }
-      }
-
       // Link bitcode (*.bc) object attrs specified by the input program.
       // Note that this happens after the command-line files so that the command
       // line ones override the symbols coming from the embedded files.
@@ -548,14 +537,15 @@ class ROCMTargetBackend final : public TargetBackend {
       }
 
       // Link module to HIP device library.
-      if (bitcodeDirectory.empty()) {
+      if (options.bitcodeDirectory.empty()) {
         return variantOp.emitError()
                << "cannot find ROCM bitcode files. Check your installation "
                   "consistency and in the worst case, set "
                   "--iree-hip-bc-dir= to a path on your system.";
       }
       if (failed(linkHIPBitcodeIfNeeded(variantOp.getLoc(), llvmModule.get(),
-                                        targetArch, bitcodeDirectory))) {
+                                        targetArch,
+                                        options.bitcodeDirectory))) {
         return failure();
       }
 
@@ -881,6 +871,7 @@ class HIPTargetDevice final : public TargetDevice {
 };
 
 namespace {
+
 struct ROCMSession final
     : PluginSession<ROCMSession, ROCMOptions,
                     PluginActivationPolicy::DefaultActivated> {
@@ -910,10 +901,23 @@ struct ROCMSession final
 
 } // namespace mlir::iree_compiler::IREE::HAL
 
+// Iterate over ukernel bitcode embedded-data files, and insert them into the
+// EmbeddedDataDirectory singleton.
+static void addAMDGPUUkernelBitcodeToGlobalEmbeddedDataDirectory() {
+  using mlir::iree_compiler::EmbeddedDataDirectory;
+  EmbeddedDataDirectory::withGlobal([](EmbeddedDataDirectory &dir) {
+    const iree_file_toc_t *toc = iree_uk_amdgpu_bitcode_create();
+    for (size_t i = 0; i < iree_uk_amdgpu_bitcode_size(); ++i) {
+      dir.addFile(toc[i].name, llvm::StringRef{toc[i].data, toc[i].size});
+    }
+  });
+}
+
 extern "C" bool iree_register_compiler_plugin_hal_target_rocm(
     mlir::iree_compiler::PluginRegistrar *registrar) {
   registrar->registerPlugin<mlir::iree_compiler::IREE::HAL::ROCMSession>(
       "hal_target_rocm");
+  addAMDGPUUkernelBitcodeToGlobalEmbeddedDataDirectory();
   return true;
 }
 

compiler/plugins/target/ROCM/ROCMTargetUtils.cpp

@@ -6,10 +6,6 @@
 
 #include "compiler/plugins/target/ROCM/ROCMTargetUtils.h"
 
-#include "compiler/plugins/target/ROCM/builtins/ukernel/iree_uk_amdgpu_gfx1030.h"
-#include "compiler/plugins/target/ROCM/builtins/ukernel/iree_uk_amdgpu_gfx1100.h"
-#include "compiler/plugins/target/ROCM/builtins/ukernel/iree_uk_amdgpu_gfx90a.h"
-#include "compiler/plugins/target/ROCM/builtins/ukernel/iree_uk_amdgpu_gfx942.h"
 #include "iree/compiler/Codegen/Utils/GPUUtils.h"
 #include "iree/compiler/Dialect/HAL/Utils/LLVMLinkerUtils.h"
 #include "iree/compiler/Utils/ToolUtils.h"
@@ -185,43 +181,6 @@ LogicalResult linkHIPBitcodeIfNeeded(Location loc, llvm::Module *module,
   return linkWithBitcodeFiles(loc, module, bitcodePaths);
 }
 
-static std::tuple<const iree_file_toc_t *, int>
-getUkernelBitcodeTOC(StringRef gpuArch) {
-  return llvm::StringSwitch<std::tuple<const iree_file_toc_t *, int>>(gpuArch)
-      .Case("gfx90a",
-            {iree_uk_amdgpu_gfx90a_create(), iree_uk_amdgpu_gfx90a_size()})
-      .Case("gfx942",
-            {iree_uk_amdgpu_gfx942_create(), iree_uk_amdgpu_gfx942_size()})
-      .Case("gfx1030",
-            {iree_uk_amdgpu_gfx1030_create(), iree_uk_amdgpu_gfx1030_size()})
-      .Case("gfx1100",
-            {iree_uk_amdgpu_gfx1100_create(), iree_uk_amdgpu_gfx1100_size()})
-      .Default({nullptr, 0});
-}
-
-// Links optimized Ukernel bitcode into the given module if the module needs it.
-LogicalResult linkUkernelBitcodeFiles(Location loc, llvm::Module *module,
-                                      StringRef enabledUkernelsStr,
-                                      StringRef targetChip,
-                                      StringRef bitcodePath,
-                                      unsigned linkerFlags,
-                                      llvm::TargetMachine &targetMachine) {
-  auto [toc, toc_size] = getUkernelBitcodeTOC(targetChip);
-  if (!toc) {
-    return failure();
-  }
-
-  llvm::Linker linker(*module);
-  for (int i = 0; i < toc_size; ++i) {
-    if (failed(linkBitcodeFile(loc, linker, linkerFlags, toc[i].name,
-                               llvm::StringRef(toc[i].data, toc[i].size),
-                               targetMachine, module->getContext())))
-      return failure();
-  }
-
-  return success();
-}
-
 // Link object file using lld lnker to generate code object
 // Inspiration from this section comes from LLVM-PROJECT-MLIR by
 // ROCmSoftwarePlatform

compiler/plugins/target/ROCM/builtins/ukernel/BUILD.bazel

@@ -49,19 +49,20 @@ argmax_types = [
         "iree_uk_amdgpu_argmax_%s.c" % type,
         "common.h",
     ],
+    out = "iree_uk_amdgpu_argmax_%s.%s.bc" % (type, gpu_arch),
     gpu_arch = gpu_arch,
 ) for type in argmax_types for gpu_arch in gpu_archs]
 
-argmax_bc_files = {gpu_arch: [
-    ":iree_uk_amdgpu_argmax_%s.c.%s.bc" % (type, gpu_arch)
+argmax_bc_files = [
+    ":iree_uk_amdgpu_argmax_%s.%s.bc" % (type, gpu_arch)
     for type in argmax_types
-] for gpu_arch in gpu_archs}
+    for gpu_arch in gpu_archs
+]
 
-[iree_c_embed_data(
-    name = "iree_uk_amdgpu_%s" % gpu_arch,
-    srcs = argmax_bc_files[gpu_arch],
-    c_file_output = "iree_uk_amdgpu_%s.c" % gpu_arch,
+iree_c_embed_data(
+    name = "iree_uk_amdgpu_bitcode",
+    srcs = argmax_bc_files,
+    c_file_output = "iree_uk_amdgpu_bitcode.c",
     flatten = True,
-    h_file_output = "iree_uk_amdgpu_%s.h" % gpu_arch,
-    identifier = "iree_uk_amdgpu_%s" % gpu_arch,
-) for gpu_arch in gpu_archs]
+    h_file_output = "iree_uk_amdgpu_bitcode.h",
+)