Remove legacy sync path (#19714)

This removes all the needless blocking operations that were occurring on hip. Removal should boost speed and allow multi device parallelism.
iree-org · Jan 16, 2025 · 3e8c81c · 3e8c81c
1 parent f6f6388
commit 3e8c81c
Show file tree

Hide file tree

Showing 10 changed files with 1 addition and 469 deletions.
diff --git a/compiler/plugins/target/ROCM/ROCMTarget.cpp b/compiler/plugins/target/ROCM/ROCMTarget.cpp
@@ -60,14 +60,12 @@ namespace mlir::iree_compiler::IREE::HAL {
 namespace {
 
 // TODO(#18792): rename flags back to iree-rocm- as they are not HIP-specific.
-// Only iree-hip-legacy-sync applies uniquely to HIP.
 struct ROCMOptions {
   std::string target = "";
   std::string targetFeatures = "";
   std::string bitcodeDirectory = getDefaultBitcodeDirectory();
   int wavesPerEu = 0;
   std::string enableROCMUkernels = "none";
-  bool legacySync = true;
   bool slpVectorization = true;
   bool globalISel = false;
 
@@ -107,9 +105,7 @@ struct ROCMOptions {
         cl::desc("Enables microkernels in the HIP compiler backend. May be "
                  "`default`, `none`, `all`, or a comma-separated list of "
                  "specific unprefixed microkernels to enable, e.g. `mmt4d`."));
-    binder.opt<bool>("iree-hip-legacy-sync", legacySync, cl::cat(category),
-                     cl::desc("Enables 'legacy-sync' mode, which is required "
-                              "for inline execution."));
+
     binder.list<std::string>(
         "iree-hip-pass-plugin-path", passPlugins,
         cl::desc("LLVM pass plugins are out of tree libraries that implement "
@@ -876,12 +872,6 @@ class HIPTargetDevice final : public TargetDevice {
     Builder b(context);
 
     SmallVector<NamedAttribute> deviceConfigAttrs;
-    if (options.legacySync) {
-      // Indicates that the runtime HAL driver operates only in the legacy
-      // synchronous mode.
-      deviceConfigAttrs.emplace_back(b.getStringAttr("legacy_sync"),
-                                     b.getUnitAttr());
-    }
     auto deviceConfigAttr = b.getDictionaryAttr(deviceConfigAttrs);
 
     SmallVector<NamedAttribute> executableConfigAttrs;

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/BUILD.bazel b/compiler/src/iree/compiler/Dialect/HAL/Transforms/BUILD.bazel
@@ -23,7 +23,6 @@ iree_compiler_cc_library(
         "DumpExecutableBenchmarks.cpp",
         "DumpExecutableSources.cpp",
         "ElideRedundantCommands.cpp",
-        "FixupLegacySync.cpp",
         "HoistExecutableObjects.cpp",
         "InitializeDevices.cpp",
         "InlineMemoizeRegions.cpp",

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/CMakeLists.txt b/compiler/src/iree/compiler/Dialect/HAL/Transforms/CMakeLists.txt
@@ -24,7 +24,6 @@ iree_cc_library(
     "DumpExecutableBenchmarks.cpp"
     "DumpExecutableSources.cpp"
     "ElideRedundantCommands.cpp"
-    "FixupLegacySync.cpp"
     "HoistExecutableObjects.cpp"
     "InitializeDevices.cpp"
     "InlineMemoizeRegions.cpp"

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/FixupLegacySync.cpp b/compiler/src/iree/compiler/Dialect/HAL/Transforms/FixupLegacySync.cpp
diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/Passes.cpp b/compiler/src/iree/compiler/Dialect/HAL/Transforms/Passes.cpp
@@ -466,10 +466,6 @@ void buildHALTransformPassPipeline(OpPassManager &passManager,
     passManager.addPass(IREE::HAL::createOutlineMemoizeRegionsPass());
   }
 
-  // If any devices require the legacy synchronous execution behavior then
-  // make all async operations blocking.
-  passManager.addPass(IREE::HAL::createFixupLegacySyncPass());
-
   // Prune unused executables and their contents.
   passManager.addPass(IREE::HAL::createPruneExecutablesPass());
 

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/Passes.td b/compiler/src/iree/compiler/Dialect/HAL/Transforms/Passes.td
@@ -187,24 +187,6 @@ def VerifyDevicesPass :
   ];
 }
 
-def FixupLegacySyncPass :
-    Pass<"iree-hal-fixup-legacy-sync", "mlir::ModuleOp"> {
-  let summary = "Applies fixups to the program for when using legacy HAL devices.";
-  let description = [{
-    Applies fixups to the program for when using legacy HAL devices that only
-    support synchronous execution. Once all devices support async this will be
-    removed.
-
-    NOTE: this pass only exists for backwards compatibility with legacy HAL
-    drivers. It will be removed once all have migrated to the modern async APIs.
-  }];
-  let dependentDialects = [
-    "mlir::arith::ArithDialect",
-    "IREE::HAL::HALDialect",
-    "IREE::Util::UtilDialect",
-  ];
-}
-
 def InlineMemoizeRegionsPass :
     Pass<"iree-hal-inline-memoize-regions", ""> {
   let summary = "Inlines `hal.device.memoize` regions into their parent region.";

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/BUILD.bazel b/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/BUILD.bazel
@@ -23,7 +23,6 @@ iree_lit_test_suite(
             "dump_executable_benchmarks.mlir",
             "dump_executable_sources.mlir",
             "elide_redundant_commands.mlir",
-            "fixup_legacy_sync.mlir",
             "hoist_executable_objects.mlir",
             "initialize_devices.mlir",
             "inline_memoize_regions.mlir",

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/CMakeLists.txt b/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/CMakeLists.txt
@@ -21,7 +21,6 @@ iree_lit_test_suite(
     "dump_executable_benchmarks.mlir"
     "dump_executable_sources.mlir"
     "elide_redundant_commands.mlir"
-    "fixup_legacy_sync.mlir"
     "hoist_executable_objects.mlir"
     "initialize_devices.mlir"
     "inline_memoize_regions.mlir"