[HostIr] Refactor hir optimization pass

CMakeLists.txt

@@ -150,6 +150,7 @@ list(APPEND NVFUSER_SRCS
   ${NVFUSER_SRCS_DIR}/host_ir/executor.cpp
   ${NVFUSER_SRCS_DIR}/host_ir/host_ir.cpp
   ${NVFUSER_SRCS_DIR}/host_ir/lower.cpp
+  ${NVFUSER_SRCS_DIR}/host_ir/lower_to_communication.cpp
   ${NVFUSER_SRCS_DIR}/id_model/circular_buffer_indexing.cpp
   ${NVFUSER_SRCS_DIR}/id_model/contiguity.cpp
   ${NVFUSER_SRCS_DIR}/id_model/id_model.cpp
@@ -216,6 +217,7 @@ list(APPEND NVFUSER_SRCS
   ${NVFUSER_SRCS_DIR}/preseg_passes/remove_empty.cpp
   ${NVFUSER_SRCS_DIR}/preseg_passes/reorder_sharded_axis.cpp
   ${NVFUSER_SRCS_DIR}/preseg_passes/segment_inplace_update.cpp
+  ${NVFUSER_SRCS_DIR}/host_ir/pass/convert_op_to_communication.cpp
   ${NVFUSER_SRCS_DIR}/host_ir/pass/stream_parallel_type.cpp
   ${NVFUSER_SRCS_DIR}/host_ir/pass/insert_deallocations.cpp
   ${NVFUSER_SRCS_DIR}/preseg_passes/translate_no_reduction_matmul_to_mul_squeeze.cpp

csrc/host_ir/executor.cpp

@@ -12,6 +12,8 @@
 #include <fusion_profiler.h>
 #include <host_ir/executor.h>
 #include <host_ir/lower.h>
+#include <host_ir/lower_to_communication.h>
+#include <host_ir/pass/convert_op_to_communication.h>
 #include <instrumentation.h>
 #include <ir/iostream.h>
 #include <ir/utils.h>
@@ -73,9 +75,8 @@ void HostIrExecutor::compile(Fusion* fusion) {
     std::vector<Expr*> exprs = fusion->exprs();
     DeviceIdxType my_device_idx = communicator_ ? communicator_->deviceId() : 0;
     for (Expr* e : exprs) {
-      HostIrLower lower;
-      std::vector<Expr*> communications =
-          lower.lower(cloner.clone(e), my_device_idx);
+      std::vector<Expr*> communications = convertSingleOpToCommunication(
+          cloner.clone(e), my_device_idx, HostIrLowerParams());
       for (auto* communication : communications) {
         host_ir_container_->pushBackTopLevelExprs(communication);
       }