[Intel] Remove DecomposeUnsupportedConversions (#3734)

This is to prepare for upstream commit
triton-lang/triton@dd6a540.

Signed-off-by: Whitney Tsang <[email protected]>

Author: whitneywhtsang

test/TritonIntelGPU/tritonintlgpu-decompose-unsupported-cvt-layout.mlir

@@ -332,10 +332,6 @@ def make_llir(src, metadata, options):
             srcMgr = llvm.source_mgr()
             ir.source_mgr_diag(srcMgr, mod.context)
             mod.context.printOpOnDiagnostic(True)
-        # FIXME: Advanced path drops tensor layouts, so this will crash on some
-        # operations being used, e.g., convert_layout.
-        if os.getenv("TRITON_INTEL_REDUCE_TRANSPOSE", "0") != "1":
-            intel.passes.ttgpuir.add_decompose_unsupported_conversions(pm)
         passes.convert.add_scf_to_cf(pm)
         passes.convert.add_index_to_llvmir(pm)
         # FIXME: Advanced path uses custom type conversion and needs hacky

third_party/intel/backend/compiler.py

@@ -3,11 +3,6 @@
 
 include "mlir/Pass/PassBase.td"
 
-def IntelDecomposeUnsupportedConversions
-    : Pass<"intel-decompose-unsupported-conversions", "mlir::ModuleOp"> {
-  let summary = "Decompose conversions that are not supported by TritonGPU -> LLVM";
-}
-
 def ConvertTritonIntelGPUToLLVM
     : Pass<"convert-triton-intel-gpu-to-llvm", "mlir::ModuleOp"> {
   let summary = "Convert TritonIntelGPU to LLVM";

third_party/intel/include/TritonIntelGPUToLLVM/Passes.td

@@ -3,7 +3,6 @@ add_triton_library(TritonIntelGPUToLLVM
     BF16Casts.cpp
     ControlFlowOpToLLVM.cpp
     ConvertLayoutOpToLLVM.cpp
-    DecomposeUnsupportedConversions.cpp
     DotOpToLLVM/DPAS.cpp
     DotOpToLLVM/FMA.cpp
     DotOpToLLVM.cpp

third_party/intel/lib/TritonIntelGPUToLLVM/CMakeLists.txt

@@ -81,8 +81,6 @@ void init_triton_intel_passes_ttgpuir(py::module &&m) {
                      gpu::intel::createTritonIntelGPUAccelerateMatmul);
   ADD_PASS_WRAPPER_0("add_rewrite_stack_ptr",
                      gpu::intel::createTritonIntelGPURewriteStackPtr);
-  ADD_PASS_WRAPPER_0("add_decompose_unsupported_conversions",
-                     gpu::intel::createIntelDecomposeUnsupportedConversions);
   ADD_PASS_WRAPPER_OPT_2("add_pipeline",
                          gpu::intel::createTritonIntelGPUPipeline, int, bool);
   ADD_PASS_WRAPPER_0("add_remove_layout_conversions",