numba-cuda 0.0.0__py3-none-any.whl → 0.0.12__py3-none-any.whl

numba_cuda/numba/cuda/tests/nocuda/test_nvvm.py

@@ -0,0 +1,54 @@
+from numba.cuda.cudadrv import nvvm
+from numba.cuda.testing import skip_on_cudasim
+from numba.core import utils
+
+from llvmlite import ir
+from llvmlite import binding as llvm
+
+import unittest
+
+
+original = "call void @llvm.memset.p0i8.i64(" \
+           "i8* align 4 %arg.x.41, i8 0, i64 %0, i1 false)"
+
+missing_align = "call void @llvm.memset.p0i8.i64(" \
+                "i8* %arg.x.41, i8 0, i64 %0, i1 false)"
+
+
+@skip_on_cudasim('libNVVM not supported in simulator')
+@unittest.skipIf(utils.MACHINE_BITS == 32, "CUDA not support for 32-bit")
+@unittest.skipIf(not nvvm.is_available(), "No libNVVM")
+class TestNvvmWithoutCuda(unittest.TestCase):
+    def test_nvvm_accepts_encoding(self):
+        # Test that NVVM will accept a constant containing all possible 8-bit
+        # characters. Taken from the test case added in llvmlite PR #53:
+        #
+        #     https://github.com/numba/llvmlite/pull/53
+        #
+        # This test case is included in Numba to ensure that the encoding used
+        # by llvmlite (e.g. utf-8, latin1, etc.) does not result in an input to
+        # NVVM that it cannot parse correctly
+
+        # Create a module with a constant containing all 8-bit characters
+        c = ir.Constant(ir.ArrayType(ir.IntType(8), 256),
+                        bytearray(range(256)))
+        m = ir.Module()
+        m.triple = 'nvptx64-nvidia-cuda'
+        nvvm.add_ir_version(m)
+        gv = ir.GlobalVariable(m, c.type, "myconstant")
+        gv.global_constant = True
+        gv.initializer = c
+        m.data_layout = nvvm.NVVM().data_layout
+
+        # Parse with LLVM then dump the parsed module into NVVM
+        parsed = llvm.parse_assembly(str(m))
+        ptx = nvvm.compile_ir(str(parsed))
+
+        # Ensure all characters appear in the generated constant array.
+        elements = ", ".join([str(i) for i in range(256)])
+        myconstant = f"myconstant[256] = {{{elements}}}".encode('utf-8')
+        self.assertIn(myconstant, ptx)
+
+
+if __name__ == '__main__':
+    unittest.main()

numba_cuda/numba/cuda/types.py

@@ -0,0 +1,37 @@
+from numba.core import types
+
+
+class Dim3(types.Type):
+    """
+    A 3-tuple (x, y, z) representing the position of a block or thread.
+    """
+    def __init__(self):
+        super().__init__(name='Dim3')
+
+
+class GridGroup(types.Type):
+    """
+    The grid of all threads in a cooperative kernel launch.
+    """
+    def __init__(self):
+        super().__init__(name='GridGroup')
+
+
+dim3 = Dim3()
+grid_group = GridGroup()
+
+
+class CUDADispatcher(types.Dispatcher):
+    """The type of CUDA dispatchers"""
+    # This type exists (instead of using types.Dispatcher as the type of CUDA
+    # dispatchers) so that we can have an alternative lowering for them to the
+    # lowering of CPU dispatchers - the CPU target lowers all dispatchers as a
+    # constant address, but we need to lower to a dummy value because it's not
+    # generally valid to use the address of CUDA kernels and functions.
+    #
+    # Notes: it may be a bug in the CPU target that it lowers all dispatchers to
+    # a constant address - it should perhaps only lower dispatchers acting as
+    # first-class functions to a constant address. Even if that bug is fixed, it
+    # is still probably a good idea to have a separate type for CUDA
+    # dispatchers, and this type might get other differentiation from the CPU
+    # dispatcher type in future.