[Codegen][Tuner] retire the C/Python binding for querying mma intrinsic. NFC. (#21816)

Since we've already exposed the IREE GPU python binding for target info
in #21782 and #21812, therefore we can now retire the legacy C/Python
binding for querying mma intrinsic. This also allows us to fully remove
the `TestLLVMGPUQueryMMAPass ` including the utility function and
associated test file.

---------

Signed-off-by: Bangtian Liu <[email protected]>

Author: bangtianliu

compiler/bindings/c/iree/compiler/dialects/iree_codegen.h

@@ -75,10 +75,6 @@ MLIR_CAPI_EXPORTED void ireeCodegenGetTunerRootOps(MlirModule module,
                                                    size_t *numOps,
                                                    MlirOperation *rootOps);
 
-MLIR_CAPI_EXPORTED void ireeCodegenQueryMMAIntrinsics(MlirOperation op,
-                                                      size_t *numIntrinsics,
-                                                      uint32_t *mmaIntrinsics);
-
 struct ireeCodegenAttentionOpDetail {
   MlirAttribute batch;
   MlirAttribute m;

compiler/bindings/python/IREECompilerDialectsModule.cpp

@@ -38,23 +38,6 @@ ireeCodegenGetExecutableVariantOpsBinding(MlirModule module) {
   return ops;
 }
 
-static std::vector<py::object>
-ireeCodegenQueryMMAIntrinsicsBinding(MlirOperation op) {
-  size_t numMMAs = 0;
-  ireeCodegenQueryMMAIntrinsics(op, &numMMAs, nullptr);
-  std::vector<uint32_t> mmaIntrinsics(numMMAs);
-  ireeCodegenQueryMMAIntrinsics(op, &numMMAs, mmaIntrinsics.data());
-
-  py::object mmaIntrinsicEnum =
-      py::module_::import_(kGpuModuleImportPath).attr(kMMAIntrinsicEnumName);
-  std::vector<py::object> mmaList(numMMAs);
-  for (size_t i = 0; i < numMMAs; ++i) {
-    mmaList[i] = mmaIntrinsicEnum(mmaIntrinsics[i]);
-  }
-
-  return mmaList;
-}
-
 static std::vector<MlirOperation>
 ireeCodegenGetTunerRootOpsBinding(MlirModule module) {
   size_t numOps = 0;
@@ -648,15 +631,6 @@ NB_MODULE(_ireeCompilerDialects, m) {
       "Gets the executable variant operations from a module.",
       py::arg("module"));
 
-  //===-------------------------------------------------------------------===//
-  // Binding to utility function queryMMAIntrinsics
-  //===-------------------------------------------------------------------===//
-
-  iree_codegen_module.def(
-      "query_mma_intrinsics", &ireeCodegenQueryMMAIntrinsicsBinding,
-      "Queries the MMA intrinsics from an executable variant op.",
-      py::arg("op"));
-
   //===-------------------------------------------------------------------===//
   // Binding to utility function ireeCodegenGetTunerRootOps
   //===-------------------------------------------------------------------===//

compiler/src/iree/compiler/API/Internal/IREECodegenDialectCAPI.cpp

@@ -30,7 +30,6 @@ using mlir::iree_compiler::IREE::Codegen::DispatchLoweringPassPipeline;
 using mlir::iree_compiler::IREE::Codegen::DispatchLoweringPassPipelineAttr;
 using mlir::iree_compiler::IREE::Codegen::LoweringConfigAttrInterface;
 using mlir::iree_compiler::IREE::Codegen::TranslationInfoAttr;
-using mlir::iree_compiler::IREE::GPU::MMAIntrinsic;
 using mlir::iree_compiler::IREE::HAL::ExecutableVariantOp;
 
 bool ireeAttributeIsACodegenDispatchLoweringPassPipelineAttr(
@@ -181,29 +180,6 @@ void ireeCodegenGetExecutableVariantOps(MlirModule module, size_t *numOps,
   }
 }
 
-void ireeCodegenQueryMMAIntrinsics(MlirOperation op, size_t *numIntrinsics,
-                                   uint32_t *mmaIntrinsics) {
-  assert(numIntrinsics && "numIntrinsics cannot be nullptr");
-
-  mlir::Operation *mlirOp = unwrap(op);
-  auto variantOp = llvm::dyn_cast_if_present<ExecutableVariantOp>(mlirOp);
-  assert(variantOp && "operation is not a ExecutableVariantOp");
-
-  llvm::SmallVector<MMAIntrinsic> intrinsics =
-      mlir::iree_compiler::queryMMAIntrinsics(variantOp);
-  if (!mmaIntrinsics) {
-    *numIntrinsics = intrinsics.size();
-    return;
-  }
-
-  assert(*numIntrinsics == intrinsics.size() &&
-         "*numIntrinsics must match the number of elements in the intrinsics");
-
-  for (size_t i = 0, e = intrinsics.size(); i < e; ++i) {
-    mmaIntrinsics[i] = static_cast<uint32_t>(intrinsics[i]);
-  }
-}
-
 void ireeCodegenGetTunerRootOps(MlirModule module, size_t *numOps,
                                 MlirOperation *rootOps) {
   assert(!mlirModuleIsNull(module) && "module cannot be nullptr");

compiler/src/iree/compiler/API/api_exports.c

@@ -30,7 +30,6 @@ extern void ireeCodegenGetExecutableVariantOps();
 extern void ireeGPUGetSingleSubgroupLayout();
 extern void ireeCodegenGetTunerRootOps();
 extern void ireeCodegenGetAttentionOpDetail();
-extern void ireeCodegenQueryMMAIntrinsics();
 extern void ireeCodegenMlirOperationIsACodegenAttentionOp();
 extern void ireeCodegenTranslationInfoAttrGet();
 extern void ireeCodegenTranslationInfoAttrGetParameters();
@@ -947,7 +946,6 @@ uintptr_t __iree_compiler_hidden_force_extern() {
   x += (uintptr_t)&ireeGPUGetSingleSubgroupLayout;
   x += (uintptr_t)&ireeCodegenGetTunerRootOps;
   x += (uintptr_t)&ireeCodegenGetAttentionOpDetail;
-  x += (uintptr_t)&ireeCodegenQueryMMAIntrinsics;
   x += (uintptr_t)&ireeCodegenMlirOperationIsACodegenAttentionOp;
   x += (uintptr_t)&ireeCodegenTranslationInfoAttrGet;
   x += (uintptr_t)&ireeCodegenTranslationInfoAttrGetParameters;

compiler/src/iree/compiler/API/api_exports.def

@@ -20,7 +20,6 @@ EXPORTS
   ireeGPUGetSingleSubgroupLayout
   ireeCodegenGetTunerRootOps
   ireeCodegenGetAttentionOpDetail
-  ireeCodegenQueryMMAIntrinsics
   ireeCodegenMlirOperationIsACodegenAttentionOp
   ireeCodegenTranslationInfoAttrGet
   ireeCodegenTranslationInfoAttrGetParameters

compiler/src/iree/compiler/API/api_exports.ld

@@ -21,7 +21,6 @@ VER_0 {
     ireeGPUGetSingleSubgroupLayout;
     ireeCodegenGetTunerRootOps;
     ireeCodegenGetAttentionOpDetail;
-    ireeCodegenQueryMMAIntrinsics;
     ireeCodegenMlirOperationIsACodegenAttentionOp;
     ireeCodegenTranslationInfoAttrGet;
     ireeCodegenTranslationInfoAttrGetParameters;

compiler/src/iree/compiler/API/api_exports.macos.lst

@@ -19,7 +19,6 @@ _ireeCodegenGetExecutableVariantOps
 _ireeGPUGetSingleSubgroupLayout
 _ireeCodegenGetTunerRootOps
 _ireeCodegenGetAttentionOpDetail
-_ireeCodegenQueryMMAIntrinsics
 _ireeCodegenMlirOperationIsACodegenAttentionOp
 _ireeCodegenTranslationInfoAttrGet
 _ireeCodegenTranslationInfoAttrGetParameters

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

@@ -115,7 +115,6 @@ iree_compiler_cc_library(
         "ROCDLConfigureBufferInstructions.cpp",
         "ROCDLLowerExecutableTarget.cpp",
         "ROCDLPrefetching.cpp",
-        "TestLLVMGPUQueryMMAPass.cpp",
         "Verifiers.cpp",
     ],
     hdrs = [

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

@@ -94,7 +94,6 @@ iree_cc_library(
     "ROCDLConfigureBufferInstructions.cpp"
     "ROCDLLowerExecutableTarget.cpp"
     "ROCDLPrefetching.cpp"
-    "TestLLVMGPUQueryMMAPass.cpp"
     "Verifiers.cpp"
   DEPS
     ::PassHeaders

compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.td

@@ -137,9 +137,4 @@ def TestLLVMGPUScalarizeMathOpPass :
   let summary = "Test pass for several legalization patterns.";
 }
 
-def TestLLVMGPUQueryMMAPass :
-    Pass<"iree-test-llvmgpu-query-mma", "ModuleOp"> {
-  let summary = "Test pass for querying the supported mma intrinsic instructions.";
-}
-
 #endif // IREE_CODEGEN_LLVMGPU_PASSES