[SPIRV] Switch to new pass generation tablegen definitions. (#18214)

This is mostly NFC. The revision applies additional cleanups:

- Remove two outdated passes: SPIRVEmulateBf16 and
SPIRVFoldProcessorIDUses. They do not have any implementation. Removing
the declaration is free.
- Combine private members and options in SPIRVTileAndPromote.
- Switch a couple of passes to follow `create.*Pass` naming convention.

---------

Signed-off-by: hanhanW <[email protected]>

compiler/src/iree/compiler/Codegen/SPIRV/BUILD.bazel

@@ -28,7 +28,6 @@ iree_gentbl_cc_library(
 iree_compiler_cc_library(
     name = "PassHeaders",
     hdrs = [
-        "PassDetail.h",
         "Passes.h",
         "Passes.h.inc",
     ],

compiler/src/iree/compiler/Codegen/SPIRV/CMakeLists.txt

@@ -23,7 +23,6 @@ iree_cc_library(
   NAME
     PassHeaders
   HDRS
-    "PassDetail.h"
     "Passes.h"
     "Passes.h.inc"
   DEPS

compiler/src/iree/compiler/Codegen/SPIRV/ConvertToSPIRVPass.cpp

@@ -17,7 +17,6 @@
 #include <cstdint>
 #include <tuple>
 
-#include "iree/compiler/Codegen/SPIRV/PassDetail.h"
 #include "iree/compiler/Codegen/SPIRV/Passes.h"
 #include "iree/compiler/Codegen/SPIRV/Utils.h"
 #include "iree/compiler/Codegen/Utils/Utils.h"
@@ -64,6 +63,9 @@
 
 namespace mlir::iree_compiler {
 
+#define GEN_PASS_DEF_CONVERTTOSPIRVPASS
+#include "iree/compiler/Codegen/SPIRV/Passes.h.inc"
+
 namespace {
 
 //===----------------------------------------------------------------------===//
@@ -470,14 +472,17 @@ struct RemoveIdentityConversionCast final
 /// Converts remaining interface ops into SPIR-V global variables, GPU processor
 /// ID ops into SPIR-V global variables, loop/standard ops into corresponding
 /// SPIR-V ops.
-class ConvertToSPIRVPass final : public ConvertToSPIRVBase<ConvertToSPIRVPass> {
+class ConvertToSPIRVPass final
+    : public impl::ConvertToSPIRVPassBase<ConvertToSPIRVPass> {
 public:
+  using impl::ConvertToSPIRVPassBase<
+      ConvertToSPIRVPass>::ConvertToSPIRVPassBase;
+  explicit ConvertToSPIRVPass(unsigned indexBits) : indexBits(indexBits) {}
+
   void getDependentDialects(DialectRegistry &registry) const override {
     registry.insert<spirv::SPIRVDialect>();
   }
 
-  explicit ConvertToSPIRVPass(unsigned indexBits) : indexBits(indexBits) {}
-
   LogicalResult initializeOptions(
       StringRef options,
       function_ref<LogicalResult(const Twine &)> errorHandler) override {

compiler/src/iree/compiler/Codegen/SPIRV/Passes.cpp

@@ -201,7 +201,7 @@ static void addMemRefLoweringPasses(OpPassManager &modulePassManager) {
       // Turn scalar load/store from memrefs into vectorized ones if possible.
       // This gives better memory access patterns, which is very important for
       // perf.
-      .addPass(createSPIRVVectorizeLoadStore)
+      .addPass(createSPIRVVectorizeLoadStorePass)
       // Perform optimizations that need to across the scf.for region boundary.
       .addPass(createForOpCanonicalizationPass)
       // Perform various vector-level cross-op optimizations like load-store
@@ -385,8 +385,9 @@ void addSPIRVCooperativeMatrixVectorizePassPipeline(
   addBufferizePasses(funcPassManager, gpuAllocateWorkgroupMemoryFn);
 
   // Tile to GPU workgroups and promote.
-  funcPassManager.addPass(createSPIRVTileAndPromotePass(
-      /*promoteCMatrix=*/true, /*skipThreadLevel=*/true));
+  funcPassManager.addPass(
+      createSPIRVTileAndPromotePass(SPIRVTileAndPromotePassOptions{
+          /*promoteCMatrix=*/true, /*skipThreadLevel=*/true}));
   funcPassManager.addPass(createRemoveSingleIterationLoopPass());
   // Run canonicalization patterns to propagate constant shape sizes after
   // removing trip-one loops.
@@ -443,7 +444,7 @@ void addSPIRVCooperativeMatrixVectorizePassPipeline(
   funcPassManager.addPass(createForOpCanonicalizationPass());
   funcPassManager.addPass(createCanonicalizerPass());
   funcPassManager.addPass(createCSEPass());
-  funcPassManager.addPass(createSPIRVVectorToGPUSubgroupMMAOpsPass());
+  funcPassManager.addPass(createSPIRVVectorToGPUSubgroupMMAPass());
   funcPassManager.addPass(createCanonicalizerPass());
   funcPassManager.addPass(createCSEPass());
   addSPIRVVectorLoweringPasses(funcPassManager);

compiler/src/iree/compiler/Codegen/SPIRV/Passes.h

@@ -65,116 +65,6 @@ void buildSPIRVCodegenPassPipeline(OpPassManager &variantPassManager);
 /// Populates passes needed to link HAL executables across SPIRV targets.
 void buildSPIRVLinkingPassPipeline(OpPassManager &modulePassManager);
 
-//===---------------------------------------------------------------------===//
-// SPIR-V passes
-//===---------------------------------------------------------------------===//
-
-/// Pass to perform the final conversion to SPIR-V dialect.
-///
-/// This pass converts remaining interface ops into SPIR-V global variables,
-/// GPU processor ID ops into SPIR-V global variables, loop/standard ops into
-/// corresponding SPIR-V ops.
-std::unique_ptr<OperationPass<ModuleOp>>
-createConvertToSPIRVPass(unsigned indexWidth = 32);
-
-/// Annotates the innermost Winograd loops with the spirv distribute
-/// attribute.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVAnnotateWinogradLoopsPass();
-
-/// Breaks down large vectors not natively supported by SPIR-V.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVBreakDownLargeVectorPass();
-
-// Converts #iree_gpu.target into #spirv.target_env.
-std::unique_ptr<OperationPass<ModuleOp>> createSPIRVConvertGPUTargetPass();
-
-/// Emulates bfloat 16 ops with 32-bit float ops.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVEmulateBf16Pass();
-
-/// Emulates 64-bit integer ops with 32-bit integer ops.
-std::unique_ptr<InterfacePass<FunctionOpInterface>> createSPIRVEmulateI64Pass();
-
-/// Turns static shaped storage buffer subspan ops into dynamic shaped ones.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVEraseStorageBufferStaticShapePass();
-
-/// Pass to perform final vector ops lowering to meet SPIR-V requirements.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVFinalVectorLoweringPass();
-
-/// Creates a pass to fold processor ID uses where possible.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVFoldProcessorIDUsesPass();
-
-/// Pass to perform initial vector ops lowering to meet SPIR-V requirements.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVInitialVectorLoweringPass();
-
-/// Links SPIR-V HAL executables within the top-level program module.
-std::unique_ptr<OperationPass<ModuleOp>> createSPIRVLinkExecutablesPass();
-
-/// Pass to set the lowering strategy for the target variant.
-std::unique_ptr<OperationPass<ModuleOp>>
-createSPIRVSelectLoweringStrategyPass();
-
-/// Main pass to lower executables to scalar + vector code on SPIR-V path.
-/// Invokes one of the pass pipelines that translate the executable to
-/// scalar + vector code.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVLowerExecutableTargetPass();
-
-/// Pass to lower executables using Transform dialect on the SPIR-V backend.
-/// This shouldnt be a separate pass, but it is since there are some
-/// extra spir-v passes that need to be run as well.
-std::unique_ptr<OperationPass<ModuleOp>>
-createSPIRVLowerExecutableUsingTransformDialectPass();
-
-/// Pass to map MemRef memory spaces to SPIR-V storage classes.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVMapMemRefStorageClassPass();
-
-/// Pass to materialize SPIR-V target requirements of hal.exectuable.variant
-/// ops into hal.executable.condition regions.
-std::unique_ptr<OperationPass<IREE::HAL::ExecutableVariantOp>>
-createSPIRVMaterializeExecutableConditionsPass();
-
-/// Pass to tile and distribute Linalg ops with buffer semantics to
-/// invocations.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVTileAndDistributePass();
-
-/// Pass to promote Linalg ops with buffer semantics to use workgroup memory
-/// and then tile to invocations.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVTileAndPromotePass(bool promoteCMatrix = false,
-                              bool skipThreadLevel = false);
-
-/// Pass to tile Linalg ops with buffer semantics suitable for lowering to
-/// SPIR-V cooperative ops.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVTileToCooperativeOpsPass();
-
-// Trims the SPIR-V target environment of a HAL executable variant to the
-// minimal requirement per the compiled spirv.module op needs.
-std::unique_ptr<OperationPass<IREE::HAL::ExecutableVariantOp>>
-createSPIRVTrimExecutableTargetEnvPass();
-
-/// Converts vector ops to gpu subgroup MMA ops.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVVectorToGPUSubgroupMMAOpsPass();
-
-/// Converts memref of scalar to memref of vector of efficent size. This will
-/// allow to convert memory accesses to vector load/store in SPIR-V without
-/// having pointer bitcast.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVVectorizeLoadStore();
-
-/// Pass to do vectorization suitable for lowering to SPIR-V cooperative ops.
-std::unique_ptr<InterfacePass<FunctionOpInterface>>
-createSPIRVVectorizeToCooperativeOpsPass();
-
 /// Pass pipeline to lower IREE HAL executables by tiling and distributing to
 /// workgroups and invocations and vectorizing. Each invocation handles a
 /// vector.
@@ -200,10 +90,25 @@ LogicalResult verifySPIRVMatmulPromoteVectorizePassPipeline(
     IREE::Codegen::TranslationInfoAttr translationInfo,
     ArrayRef<int64_t> workgroupSize);
 
+//===---------------------------------------------------------------------===//
+// Wrappers that not use tablegen options.
+//===---------------------------------------------------------------------===//
+
+/// Pass to perform the final conversion to SPIR-V dialect.
+///
+/// This pass converts remaining interface ops into SPIR-V global variables,
+/// GPU processor ID ops into SPIR-V global variables, loop/standard ops into
+/// corresponding SPIR-V ops.
+std::unique_ptr<OperationPass<ModuleOp>>
+createConvertToSPIRVPass(unsigned indexWidth);
+
 //----------------------------------------------------------------------------//
 // Registration
 //----------------------------------------------------------------------------//
 
+#define GEN_PASS_DECL
+#include "iree/compiler/Codegen/SPIRV/Passes.h.inc" // IWYU pragma: keep
+
 void registerCodegenSPIRVPasses();
 
 } // namespace mlir::iree_compiler