feat[cartesian]: Setting extra compile args and DaCe block size via e…

…nv variables (#1462)

* Set default block size using env variable.

* Fix nvcc flags.

* Set extra compile args using env variable.

* Rename env variable.

* Rename env variables.

* Run pre-commit.

* Use typing.List for compatibility with py3.8.

* Fix dace compilation.

* Address review comments.

* Shift statement outside if block.

src/gt4py/cartesian/backend/dace_backend.py

@@ -544,9 +544,14 @@ def keep_line(line):
     def apply(cls, stencil_ir: gtir.Stencil, builder: "StencilBuilder", sdfg: dace.SDFG):
         self = cls()
         with dace.config.temporary_config():
+            # To prevent conflict with 3rd party usage of DaCe config always make sure that any
+            #  changes be under the temporary_config manager
             if gt_config.GT4PY_USE_HIP:
                 dace.config.Config.set("compiler", "cuda", "backend", value="hip")
             dace.config.Config.set("compiler", "cuda", "max_concurrent_streams", value=-1)
+            dace.config.Config.set(
+                "compiler", "cuda", "default_block_size", value=gt_config.DACE_DEFAULT_BLOCK_SIZE
+            )
             dace.config.Config.set("compiler", "cpu", "openmp_sections", value=False)
             code_objects = sdfg.generate_code()
         is_gpu = "CUDA" in {co.title for co in code_objects}

src/gt4py/cartesian/caching.py

@@ -312,6 +312,8 @@ def stencil_id(self) -> StencilID:
             "api_annotations": f"[{', '.join(self._extract_api_annotations())}]",
             **self._externals,
         }
+        if self.builder.backend.name == "dace:gpu":
+            fingerprint["default_block_size"] = gt_config.DACE_DEFAULT_BLOCK_SIZE
 
         # typeignore because attrclass StencilID has generated constructor
         return StencilID(  # type: ignore

src/gt4py/cartesian/config.py

@@ -14,7 +14,7 @@
 
 import multiprocessing
 import os
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional
 
 import gridtools_cpp
 
@@ -49,6 +49,10 @@
 GT_CPP_TEMPLATE_DEPTH: int = 1024
 
 # Settings dict
+GT4PY_EXTRA_COMPILE_ARGS: str = os.environ.get("GT4PY_EXTRA_COMPILE_ARGS", "")
+extra_compile_args: List[str] = (
+    list(GT4PY_EXTRA_COMPILE_ARGS.split(" ")) if GT4PY_EXTRA_COMPILE_ARGS else []
+)
 build_settings: Dict[str, Any] = {
     "boost_include_path": os.path.join(BOOST_ROOT, "include"),
     "cuda_bin_path": os.path.join(CUDA_ROOT, "bin"),
@@ -57,10 +61,7 @@
     "gt_include_path": os.environ.get("GT_INCLUDE_PATH", GT_INCLUDE_PATH),
     "openmp_cppflags": os.environ.get("OPENMP_CPPFLAGS", "-fopenmp").split(),
     "openmp_ldflags": os.environ.get("OPENMP_LDFLAGS", "-fopenmp").split(),
-    "extra_compile_args": {
-        "cxx": [],
-        "cuda": [],
-    },
+    "extra_compile_args": {"cxx": extra_compile_args, "cuda": extra_compile_args},
     "extra_link_args": [],
     "parallel_jobs": multiprocessing.cpu_count(),
     "cpp_template_depth": os.environ.get("GT_CPP_TEMPLATE_DEPTH", GT_CPP_TEMPLATE_DEPTH),
@@ -83,3 +84,5 @@
 code_settings: Dict[str, Any] = {"root_package_name": "_GT_"}
 
 os.environ.setdefault("DACE_CONFIG", os.path.join(os.path.abspath("."), ".dace.conf"))
+
+DACE_DEFAULT_BLOCK_SIZE: str = os.environ.get("DACE_DEFAULT_BLOCK_SIZE", "64,8,1")