numba-cuda 0.9.0__py3-none-any.whl → 0.10.1__py3-none-any.whl

numba_cuda/VERSION

@@ -1 +1 @@
-0.9.0
+0.10.1

numba_cuda/numba/cuda/compiler.py

@@ -40,6 +40,7 @@ from numba.cuda.api import get_current_device
 from numba.cuda.cudadrv import nvvm
 from numba.cuda.descriptor import cuda_target
 from numba.cuda.target import CUDACABICallConv
+from numba.cuda import lowering
 
 
 def _nvvm_options_type(x):
@@ -163,6 +164,18 @@ class CreateLibrary(LoweringPass):
         return True
 
 
+@register_pass(mutates_CFG=True, analysis_only=False)
+class CUDANativeLowering(NativeLowering):
+    """Lowering pass for a CUDA native function IR described solely in terms of
+    Numba's standard `numba.core.ir` nodes."""
+
+    _name = "cuda_native_lowering"
+
+    @property
+    def lowering_class(self):
+        return lowering.CUDALower
+
+
 class CUDABytecodeInterpreter(Interpreter):
     # Based on the superclass implementation, but names the resulting variable
     # "$bool<N>" instead of "bool<N>" - see Numba PR #9888:
@@ -251,7 +264,7 @@ class CUDACompiler(CompilerBase):
 
         # lower
         pm.add_pass(CreateLibrary, "create library")
-        pm.add_pass(NativeLowering, "native lowering")
+        pm.add_pass(CUDANativeLowering, "cuda native lowering")
         pm.add_pass(CUDABackend, "cuda backend")
 
         pm.finalize()
@@ -265,7 +278,7 @@ def compile_cuda(
     args,
     debug=False,
     lineinfo=False,
-    inline=False,
+    forceinline=False,
     fastmath=False,
     nvvm_options=None,
     cc=None,
@@ -303,7 +316,7 @@ def compile_cuda(
     else:
         flags.error_model = "numpy"
 
-    if inline:
+    if forceinline:
         flags.forceinline = True
     if fastmath:
         flags.fastmath = True
@@ -561,6 +574,7 @@ def compile(
     abi="c",
     abi_info=None,
     output="ptx",
+    forceinline=False,
 ):
     """Compile a Python function to PTX or LTO-IR for a given set of argument
     types.
@@ -601,6 +615,11 @@ def compile(
     :type abi_info: dict
     :param output: Type of output to generate, either ``"ptx"`` or ``"ltoir"``.
     :type output: str
+    :param forceinline: Enables inlining at the NVVM IR level when set to
+                        ``True``. This is accomplished by adding the
+                        ``alwaysinline`` function attribute to the function
+                        definition. This is only valid when the output is
+                        ``"ltoir"``.
     :return: (code, resty): The compiled code and inferred return type
     :rtype: tuple
     """
@@ -613,6 +632,12 @@ def compile(
     if output not in ("ptx", "ltoir"):
         raise NotImplementedError(f"Unsupported output type: {output}")
 
+    if forceinline and not device:
+        raise ValueError("Cannot force-inline kernels")
+
+    if forceinline and output != "ltoir":
+        raise ValueError("Can only designate forced inlining in LTO-IR")
+
     debug = config.CUDA_DEBUGINFO_DEFAULT if debug is None else debug
     opt = (config.OPT != 0) if opt is None else opt
 
@@ -647,6 +672,7 @@ def compile(
         fastmath=fastmath,
         nvvm_options=nvvm_options,
         cc=cc,
+        forceinline=forceinline,
     )
     resty = cres.signature.return_type
 
@@ -686,6 +712,7 @@ def compile_for_current_device(
     abi="c",
     abi_info=None,
     output="ptx",
+    forceinline=False,
 ):
     """Compile a Python function to PTX or LTO-IR for a given signature for the
     current device's compute capabilility. This calls :func:`compile` with an
@@ -703,6 +730,7 @@ def compile_for_current_device(
         abi=abi,
         abi_info=abi_info,
         output=output,
+        forceinline=forceinline,
     )
 
 
@@ -717,6 +745,7 @@ def compile_ptx(
     opt=None,
     abi="numba",
     abi_info=None,
+    forceinline=False,
 ):
     """Compile a Python function to PTX for a given signature. See
     :func:`compile`. The defaults for this function are to compile a kernel
@@ -734,6 +763,7 @@ def compile_ptx(
         abi=abi,
         abi_info=abi_info,
         output="ptx",
+        forceinline=forceinline,
     )
 
 
@@ -747,6 +777,7 @@ def compile_ptx_for_current_device(
     opt=None,
     abi="numba",
     abi_info=None,
+    forceinline=False,
 ):
     """Compile a Python function to PTX for a given signature for the current
     device's compute capabilility. See :func:`compile_ptx`."""
@@ -762,6 +793,7 @@ def compile_ptx_for_current_device(
         opt=opt,
         abi=abi,
         abi_info=abi_info,
+        forceinline=forceinline,
     )