numba-cuda 0.1.0__tar.gz → 0.3.0__tar.gz

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: numba-cuda
-Version: 0.1.0
+Version: 0.3.0
 Summary: CUDA target for Numba
 Author: Anaconda Inc., NVIDIA Corporation
 License: BSD 2-clause

numba_cuda-0.3.0/numba_cuda/VERSION

@@ -0,0 +1 @@
+0.3.0

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda/numba/cuda/cudadecl.py

@@ -508,6 +508,11 @@ class Cuda_atomic_cas(AbstractTemplate):
             return signature(dty, ary, idx, dty, dty)
 
 
+@register_global(breakpoint)
+class Cuda_breakpoint(ConcreteTemplate):
+    cases = [signature(types.none)]
+
+
 @register
 class Cuda_nanosleep(ConcreteTemplate):
     key = cuda.nanosleep

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda/numba/cuda/cudadrv/driver.py

@@ -42,6 +42,11 @@ from .mappings import FILE_EXTENSION_MAP
 from .linkable_code import LinkableCode, LTOIR, Fatbin, Object
 from numba.cuda.cudadrv import enums, drvapi, nvrtc
 
+try:
+    from pynvjitlink.api import NvJitLinker, NvJitLinkError
+except ImportError:
+    NvJitLinker, NvJitLinkError = None, None
+
 USE_NV_BINDING = config.CUDA_USE_NVIDIA_BINDING
 
 if USE_NV_BINDING:
@@ -92,20 +97,6 @@ ENABLE_PYNVJITLINK = (
 if not hasattr(config, "CUDA_ENABLE_PYNVJITLINK"):
     config.CUDA_ENABLE_PYNVJITLINK = ENABLE_PYNVJITLINK
 
-if ENABLE_PYNVJITLINK:
-    try:
-        from pynvjitlink.api import NvJitLinker, NvJitLinkError
-    except ImportError:
-        raise ImportError(
-            "Using pynvjitlink requires the pynvjitlink package to be available"
-        )
-
-    if config.CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY:
-        raise ValueError(
-            "Can't set CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY and "
-            "CUDA_ENABLE_PYNVJITLINK at the same time"
-        )
-
 
 def make_logger():
     logger = logging.getLogger(__name__)
@@ -3061,6 +3052,17 @@ class PyNvJitLinker(Linker):
         lto=False,
         additional_flags=None,
     ):
+        if NvJitLinker is None:
+            raise ImportError(
+                "Using pynvjitlink requires the pynvjitlink package to be "
+                "available"
+            )
+
+        if config.CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY:
+            raise ValueError(
+                "Can't set CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY and "
+                "CUDA_ENABLE_PYNVJITLINK at the same time"
+            )
 
         if cc is None:
             raise RuntimeError("PyNvJitLinker requires CC to be specified")

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda/numba/cuda/cudaimpl.py

@@ -934,6 +934,14 @@ def ptx_atomic_cas(context, builder, sig, args):
 
 # -----------------------------------------------------------------------------
 
+
+@lower(breakpoint)
+def ptx_brkpt(context, builder, sig, args):
+    brkpt = ir.InlineAsm(ir.FunctionType(ir.VoidType(), []),
+                         "brkpt;", '', side_effect=True)
+    builder.call(brkpt, ())
+
+
 @lower(stubs.nanosleep, types.uint32)
 def ptx_nanosleep(context, builder, sig, args):
     nanosleep = ir.InlineAsm(ir.FunctionType(ir.VoidType(), [ir.IntType(32)]),

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda/numba/cuda/dispatcher.py

@@ -37,6 +37,8 @@ cuda_fp16_math_funcs = ['hsin', 'hcos',
                         'hrcp', 'hrint',
                         'htrunc', 'hdiv']
 
+reshape_funcs = ['nocopy_empty_reshape', 'numba_attempt_nocopy_reshape']
+
 
 class _Kernel(serialize.ReduceMixin):
     '''
@@ -117,25 +119,43 @@ class _Kernel(serialize.ReduceMixin):
         if not link:
             link = []
 
+        asm = lib.get_asm_str()
+
         # A kernel needs cooperative launch if grid_sync is being used.
-        self.cooperative = 'cudaCGGetIntrinsicHandle' in lib.get_asm_str()
+        self.cooperative = 'cudaCGGetIntrinsicHandle' in asm
         # We need to link against cudadevrt if grid sync is being used.
         if self.cooperative:
             lib.needs_cudadevrt = True
 
-        basedir = os.path.dirname(os.path.abspath(__file__))
-        asm = lib.get_asm_str()
+        def link_to_library_functions(library_functions, library_path,
+                                      prefix=None):
+            """
+            Dynamically links to library functions by searching for their names
+            in the specified library and linking to the corresponding source
+            file.
+            """
+            if prefix is not None:
+                library_functions = [f"{prefix}{fn}" for fn in
+                                     library_functions]
 
-        res = [fn for fn in cuda_fp16_math_funcs
-               if (f'__numba_wrapper_{fn}' in asm)]
+            found_functions = [fn for fn in library_functions
+                               if f'{fn}' in asm]
 
-        if res:
-            # Path to the source containing the foreign function
-            functions_cu_path = os.path.join(basedir,
-                                             'cpp_function_wrappers.cu')
-            link.append(functions_cu_path)
+            if found_functions:
+                basedir = os.path.dirname(os.path.abspath(__file__))
+                source_file_path = os.path.join(basedir, library_path)
+                link.append(source_file_path)
 
-        link = self.maybe_link_nrt(link, tgt_ctx, asm)
+            return found_functions
+
+        # Link to the helper library functions if needed
+        link_to_library_functions(reshape_funcs, 'reshape_funcs.cu')
+        # Link to the CUDA FP16 math library functions if needed
+        link_to_library_functions(cuda_fp16_math_funcs,
+                                  'cpp_function_wrappers.cu',
+                                  '__numba_wrapper_')
+
+        self.maybe_link_nrt(link, tgt_ctx, asm)
 
         for filepath in link:
             lib.add_linking_file(filepath)
@@ -160,7 +180,7 @@ class _Kernel(serialize.ReduceMixin):
 
     def maybe_link_nrt(self, link, tgt_ctx, asm):
         if not tgt_ctx.enable_nrt:
-            return link
+            return
 
         all_nrt = "|".join(self.NRT_functions)
         pattern = (
@@ -175,8 +195,6 @@ class _Kernel(serialize.ReduceMixin):
             nrt_path = os.path.join(basedir, 'runtime', 'nrt.cu')
             link.append(nrt_path)
 
-        return link
-
     @property
     def library(self):
         return self._codelibrary

numba_cuda-0.3.0/numba_cuda/numba/cuda/reshape_funcs.cu

@@ -0,0 +1,151 @@
+/*
+ * Handle reshaping of zero-sized array.
+ * See numba_attempt_nocopy_reshape() below.
+ */
+#define NPY_MAXDIMS 32
+
+typedef long long int npy_intp;
+
+extern "C" __device__ int
+nocopy_empty_reshape(npy_intp nd, const npy_intp *dims, const npy_intp *strides,
+                     npy_intp newnd, const npy_intp *newdims,
+                     npy_intp *newstrides, npy_intp itemsize,
+                     int is_f_order)
+{
+    int i;
+    /* Just make the strides vaguely reasonable
+     * (they can have any value in theory).
+     */
+    for (i = 0; i < newnd; i++)
+        newstrides[i] = itemsize;
+    return 1;  /* reshape successful */
+}
+
+/*
+ * Straight from Numpy's _attempt_nocopy_reshape()
+ * (np/core/src/multiarray/shape.c).
+ * Attempt to reshape an array without copying data
+ *
+ * This function should correctly handle all reshapes, including
+ * axes of length 1. Zero strides should work but are untested.
+ *
+ * If a copy is needed, returns 0
+ * If no copy is needed, returns 1 and fills `npy_intp *newstrides`
+ *     with appropriate strides
+ */
+extern "C" __device__ int
+numba_attempt_nocopy_reshape(npy_intp nd, const npy_intp *dims, const npy_intp *strides,
+                             npy_intp newnd, const npy_intp *newdims,
+                             npy_intp *newstrides, npy_intp itemsize,
+                             int is_f_order)
+{
+    int oldnd;
+    npy_intp olddims[NPY_MAXDIMS];
+    npy_intp oldstrides[NPY_MAXDIMS];
+    npy_intp np, op, last_stride;
+    int oi, oj, ok, ni, nj, nk;
+
+    oldnd = 0;
+    /*
+     * Remove axes with dimension 1 from the old array. They have no effect
+     * but would need special cases since their strides do not matter.
+     */
+    for (oi = 0; oi < nd; oi++) {
+        if (dims[oi]!= 1) {
+            olddims[oldnd] = dims[oi];
+            oldstrides[oldnd] = strides[oi];
+            oldnd++;
+        }
+    }
+
+    np = 1;
+    for (ni = 0; ni < newnd; ni++) {
+        np *= newdims[ni];
+    }
+    op = 1;
+    for (oi = 0; oi < oldnd; oi++) {
+        op *= olddims[oi];
+    }
+    if (np != op) {
+        /* different total sizes; no hope */
+        return 0;
+    }
+
+    if (np == 0) {
+        /* the Numpy code does not handle 0-sized arrays */
+        return nocopy_empty_reshape(nd, dims, strides,
+                                    newnd, newdims, newstrides,
+                                    itemsize, is_f_order);
+    }
+
+    /* oi to oj and ni to nj give the axis ranges currently worked with */
+    oi = 0;
+    oj = 1;
+    ni = 0;
+    nj = 1;
+    while (ni < newnd && oi < oldnd) {
+        np = newdims[ni];
+        op = olddims[oi];
+
+        while (np != op) {
+            if (np < op) {
+                /* Misses trailing 1s, these are handled later */
+                np *= newdims[nj++];
+            } else {
+                op *= olddims[oj++];
+            }
+        }
+
+        /* Check whether the original axes can be combined */
+        for (ok = oi; ok < oj - 1; ok++) {
+            if (is_f_order) {
+                if (oldstrides[ok+1] != olddims[ok]*oldstrides[ok]) {
+                     /* not contiguous enough */
+                    return 0;
+                }
+            }
+            else {
+                /* C order */
+                if (oldstrides[ok] != olddims[ok+1]*oldstrides[ok+1]) {
+                    /* not contiguous enough */
+                    return 0;
+                }
+            }
+        }
+
+        /* Calculate new strides for all axes currently worked with */
+        if (is_f_order) {
+            newstrides[ni] = oldstrides[oi];
+            for (nk = ni + 1; nk < nj; nk++) {
+                newstrides[nk] = newstrides[nk - 1]*newdims[nk - 1];
+            }
+        }
+        else {
+            /* C order */
+            newstrides[nj - 1] = oldstrides[oj - 1];
+            for (nk = nj - 1; nk > ni; nk--) {
+                newstrides[nk - 1] = newstrides[nk]*newdims[nk];
+            }
+        }
+        ni = nj++;
+        oi = oj++;
+    }
+
+    /*
+     * Set strides corresponding to trailing 1s of the new shape.
+     */
+    if (ni >= 1) {
+        last_stride = newstrides[ni - 1];
+    }
+    else {
+        last_stride = itemsize;
+    }
+    if (is_f_order) {
+        last_stride *= newdims[ni - 1];
+    }
+    for (nk = ni; nk < newnd; nk++) {
+        newstrides[nk] = last_stride;
+    }
+
+    return 1;
+}

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda/numba/cuda/tests/cudadrv/test_nvjitlink.py

@@ -2,6 +2,18 @@ from numba.cuda.testing import unittest
 from numba.cuda.testing import skip_on_cudasim
 from numba.cuda.testing import CUDATestCase
 from numba.cuda.cudadrv.driver import PyNvJitLinker
+from numba.cuda import get_current_device
+
+from numba import cuda
+from numba import config
+from numba.tests.support import run_in_subprocess, override_config
+
+try:
+    import pynvjitlink  # noqa: F401
+    PYNVJITLINK_INSTALLED = True
+except ImportError:
+    PYNVJITLINK_INSTALLED = False
+
 
 import itertools
 import os
@@ -9,9 +21,6 @@ import io
 import contextlib
 import warnings
 
-from numba.cuda import get_current_device
-from numba import cuda
-from numba import config
 
 TEST_BIN_DIR = os.getenv("NUMBA_CUDA_TEST_BIN_DIR")
 if TEST_BIN_DIR:
@@ -251,5 +260,69 @@ class TestLinker(CUDATestCase):
                 pass
 
 
+@unittest.skipIf(
+    not PYNVJITLINK_INSTALLED, reason="Pynvjitlink is not installed"
+)
+class TestLinkerUsage(CUDATestCase):
+    """Test that whether pynvjitlink can be enabled by both environment variable
+    and modification of config at runtime.
+    """
+
+    src = """if 1:
+        import os
+        from numba import cuda, config
+
+        {config}
+
+        TEST_BIN_DIR = os.getenv("NUMBA_CUDA_TEST_BIN_DIR")
+        if TEST_BIN_DIR:
+            test_device_functions_cubin = os.path.join(
+                TEST_BIN_DIR, "test_device_functions.cubin"
+            )
+
+        sig = "uint32(uint32, uint32)"
+        add_from_numba = cuda.declare_device("add_from_numba", sig)
+
+        @cuda.jit(link=[test_device_functions_cubin], lto=True)
+        def kernel(result):
+            result[0] = add_from_numba(1, 2)
+
+        result = cuda.device_array(1)
+        kernel[1, 1](result)
+        assert result[0] == 3
+        """
+
+    def test_linker_enabled_envvar(self):
+        env = os.environ.copy()
+        env['NUMBA_CUDA_ENABLE_PYNVJITLINK'] = "1"
+        run_in_subprocess(self.src.format(config=""), env=env)
+
+    def test_linker_disabled_envvar(self):
+        env = os.environ.copy()
+        env.pop('NUMBA_CUDA_ENABLE_PYNVJITLINK', None)
+        with self.assertRaisesRegex(
+            AssertionError, "LTO and additional flags require PyNvJitLinker"
+        ):
+            # Actual error raised is `ValueError`, but `run_in_subprocess`
+            # reraises as AssertionError.
+            run_in_subprocess(self.src.format(config=""), env=env)
+
+    def test_linker_enabled_config(self):
+        env = os.environ.copy()
+        env.pop('NUMBA_CUDA_ENABLE_PYNVJITLINK', None)
+        run_in_subprocess(self.src.format(
+            config="config.CUDA_ENABLE_PYNVJITLINK = True"), env=env)
+
+    def test_linker_disabled_config(self):
+        env = os.environ.copy()
+        env.pop('NUMBA_CUDA_ENABLE_PYNVJITLINK', None)
+        with override_config("CUDA_ENABLE_PYNVJITLINK", False):
+            with self.assertRaisesRegex(
+                AssertionError, "LTO and additional flags require PyNvJitLinker"
+            ):
+                run_in_subprocess(self.src.format(
+                    config="config.CUDA_ENABLE_PYNVJITLINK = False"), env=env)
+
+
 if __name__ == "__main__":
     unittest.main()

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda/numba/cuda/tests/cudapy/test_array.py

@@ -12,6 +12,31 @@ else:
                             cuda.pinned_array_like)
 
 
+def array_reshape1d(arr, newshape, got):
+    y = arr.reshape(newshape)
+    for i in range(y.shape[0]):
+        got[i] = y[i]
+
+
+def array_reshape2d(arr, newshape, got):
+    y = arr.reshape(newshape)
+    for i in range(y.shape[0]):
+        for j in range(y.shape[1]):
+            got[i, j] = y[i, j]
+
+
+def array_reshape3d(arr, newshape, got):
+    y = arr.reshape(newshape)
+    for i in range(y.shape[0]):
+        for j in range(y.shape[1]):
+            for k in range(y.shape[2]):
+                got[i, j, k] = y[i, j, k]
+
+
+def array_reshape(arr, newshape):
+    return arr.reshape(newshape)
+
+
 class TestCudaArray(CUDATestCase):
     def test_gpu_array_zero_length(self):
         x = np.arange(0)
@@ -255,6 +280,54 @@ class TestCudaArray(CUDATestCase):
 
         self.assertEqual(1, len(func.overloads))
 
+    def test_array_reshape(self):
+        def check(pyfunc, kernelfunc, arr, shape):
+            kernel = cuda.jit(kernelfunc)
+            expected = pyfunc(arr, shape)
+            got = np.zeros(expected.shape, dtype=arr.dtype)
+            kernel[1, 1](arr, shape, got)
+            self.assertPreciseEqual(got, expected)
+
+        def check_only_shape(kernelfunc, arr, shape, expected_shape):
+            kernel = cuda.jit(kernelfunc)
+            got = np.zeros(expected_shape, dtype=arr.dtype)
+            kernel[1, 1](arr, shape, got)
+            self.assertEqual(got.shape, expected_shape)
+            self.assertEqual(got.size, arr.size)
+
+        # 0-sized arrays
+        def check_empty(arr):
+            check(array_reshape, array_reshape1d, arr, 0)
+            check(array_reshape, array_reshape1d, arr, (0,))
+            check(array_reshape, array_reshape3d, arr, (1, 0, 2))
+            check_only_shape(array_reshape2d, arr, (0, -1), (0, 0))
+            check_only_shape(array_reshape2d, arr, (4, -1), (4, 0))
+            check_only_shape(array_reshape3d, arr, (-1, 0, 4), (0, 0, 4))
+
+        # C-contiguous
+        arr = np.arange(24)
+        check(array_reshape, array_reshape1d, arr, (24,))
+        check(array_reshape, array_reshape2d, arr, (4, 6))
+        check(array_reshape, array_reshape2d, arr, (8, 3))
+        check(array_reshape, array_reshape3d, arr, (8, 1, 3))
+
+        arr = np.arange(24).reshape((1, 8, 1, 1, 3, 1))
+        check(array_reshape, array_reshape1d, arr, (24,))
+        check(array_reshape, array_reshape2d, arr, (4, 6))
+        check(array_reshape, array_reshape2d, arr, (8, 3))
+        check(array_reshape, array_reshape3d, arr, (8, 1, 3))
+
+        # Test negative shape value
+        arr = np.arange(25).reshape(5,5)
+        check(array_reshape, array_reshape1d, arr, -1)
+        check(array_reshape, array_reshape1d, arr, (-1,))
+        check(array_reshape, array_reshape2d, arr, (-1, 5))
+        check(array_reshape, array_reshape3d, arr, (5, -1, 5))
+        check(array_reshape, array_reshape3d, arr, (5, 5, -1))
+
+        arr = np.array([])
+        check_empty(arr)
+
 
 if __name__ == '__main__':
     unittest.main()

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: numba-cuda
-Version: 0.1.0
+Version: 0.3.0
 Summary: CUDA target for Numba
 Author: Anaconda Inc., NVIDIA Corporation
 License: BSD 2-clause

{numba_cuda-0.1.0 → numba_cuda-0.3.0}/numba_cuda.egg-info/SOURCES.txt

@@ -43,6 +43,7 @@ numba_cuda/numba/cuda/models.py
 numba_cuda/numba/cuda/nvvmutils.py
 numba_cuda/numba/cuda/printimpl.py
 numba_cuda/numba/cuda/random.py
+numba_cuda/numba/cuda/reshape_funcs.cu
 numba_cuda/numba/cuda/simulator_init.py
 numba_cuda/numba/cuda/stubs.py
 numba_cuda/numba/cuda/target.py

numba_cuda-0.1.0/numba_cuda/VERSION

@@ -1 +0,0 @@
-0.1.0