Support Functions Typed with __nv_bfloat16_raw (#262)

Currently Numbast's type system does not recognize `__nv_bfloat16_raw`
data type. Certain CUDA C++ functions takes the raw type as input, so we
also provide those support here.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

## Release Notes

* **New Features**
* Added support for bfloat16 raw type conversions, enabling
bidirectional conversion between bfloat16 values and their raw
representations in CUDA kernels.
* Extended type system to recognize and properly handle the new bfloat16
raw variant.

* **Tests**
* Added comprehensive test coverage for bfloat16 raw conversion
functionality.

<sub>✏️ Tip: You can customize this high-level summary in your review
settings.</sub>

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Michael Wang <[email protected]>

Author: isVoid

numbast/src/numbast/static/renderer.py

@@ -110,6 +110,12 @@ def _try_import_numba_type(cls, typ: str):
             cls.Imports.add("from numba.cuda.types import bfloat16")
             cls._imported_numba_types.add(typ)
 
+        elif typ == "__nv_bfloat16_raw":
+            cls.Imports.add(
+                "from numba.cuda._internal.cuda_bf16 import _type_unnamed1405307 as bfloat16_raw_type"
+            )
+            cls._imported_numba_types.add(typ)
+
         elif typ in vector_types:
             # CUDA target specific types
             cls.Imports.add("from numba.cuda.vector_types import vector_types")

numbast/src/numbast/static/tests/data/bf16.cuh

@@ -10,4 +10,12 @@ nv_bfloat16 inline __device__ add(nv_bfloat16 a, nv_bfloat16 b) {
   return a + b;
 }
 
+__nv_bfloat16_raw inline __device__ bf16_to_raw(nv_bfloat16 a) {
+  return __nv_bfloat16_raw(a);
+}
+
+nv_bfloat16 inline __device__ bf16_from_raw(__nv_bfloat16_raw a) {
+  return __nv_bfloat16(a);
+}
+
 #endif

numbast/src/numbast/static/tests/test_bf16_support.py

@@ -23,3 +23,27 @@ def kernel(arr):
 
     # Check that bfloat16 is imported
     assert "from numba.cuda.types import bfloat16" in res1["src"]
+
+
+def test_bindings_from_bf16raw(make_binding):
+    res = make_binding("bf16.cuh", {}, {})
+
+    binding = res["bindings"]
+
+    bf16_from_raw = binding["bf16_from_raw"]
+    bf16_to_raw = binding["bf16_to_raw"]
+
+    @cuda.jit
+    def kernel(arr):
+        x = bfloat16(3.14)
+
+        x_raw = bf16_to_raw(x)
+        x2 = bf16_from_raw(x_raw)
+
+        arr[0] = float32(x2)
+
+    arr = cuda.device_array((1,), dtype="float32")
+
+    kernel[1, 1](arr)
+
+    assert pytest.approx(arr[0], 1e-2) == 3.14

numbast/src/numbast/static/types.py

@@ -59,6 +59,10 @@ def to_numba_type_str(ty: str):
         BaseRenderer._try_import_numba_type("__nv_bfloat16")
         return "bfloat16"
 
+    if ty == "__nv_bfloat16_raw":
+        BaseRenderer._try_import_numba_type("__nv_bfloat16_raw")
+        return "bfloat16_raw_type"
+
     if ty.endswith("*"):
         base_ty = ty.rstrip("*").rstrip(" ")
         ptr_ty_str = f"CPointer({to_numba_type_str(base_ty)})"

numbast/src/numbast/types.py

@@ -9,6 +9,8 @@
 from numba.cuda.vector_types import vector_types
 from numba.misc.special import typeof
 
+from numba.cuda._internal.cuda_bf16 import _type_unnamed1405307
+
 
 class FunctorType(nbtypes.Type):
     def __init__(self, name):
@@ -83,8 +85,8 @@ def register_enum_type(cxx_name: str, e: IntEnum):
 
 
 def to_numba_type(ty: str):
-    if ty == "__nv_bfloat16":
-        return bfloat16
+    if ty == "__nv_bfloat16_raw":
+        return _type_unnamed1405307
 
     if "FunctorType" in ty:
         return FunctorType(ty[:-11])