numba-cuda 0.0.1__py3-none-any.whl → 0.0.12__py3-none-any.whl

numba_cuda/numba/cuda/errors.py

@@ -0,0 +1,59 @@
+import numbers
+from numba.core.errors import LoweringError
+
+
+class KernelRuntimeError(RuntimeError):
+    def __init__(self, msg, tid=None, ctaid=None):
+        self.tid = tid
+        self.ctaid = ctaid
+        self.msg = msg
+        t = ("An exception was raised in thread=%s block=%s\n"
+             "\t%s")
+        msg = t % (self.tid, self.ctaid, self.msg)
+        super(KernelRuntimeError, self).__init__(msg)
+
+
+class CudaLoweringError(LoweringError):
+    pass
+
+
+_launch_help_url = ("https://numba.readthedocs.io/en/stable/cuda/"
+                    "kernels.html#kernel-invocation")
+missing_launch_config_msg = """
+Kernel launch configuration was not specified. Use the syntax:
+
+kernel_function[blockspergrid, threadsperblock](arg0, arg1, ..., argn)
+
+See {} for help.
+
+""".format(_launch_help_url)
+
+
+def normalize_kernel_dimensions(griddim, blockdim):
+    """
+    Normalize and validate the user-supplied kernel dimensions.
+    """
+
+    def check_dim(dim, name):
+        if not isinstance(dim, (tuple, list)):
+            dim = [dim]
+        else:
+            dim = list(dim)
+        if len(dim) > 3:
+            raise ValueError('%s must be a sequence of 1, 2 or 3 integers, '
+                             'got %r' % (name, dim))
+        for v in dim:
+            if not isinstance(v, numbers.Integral):
+                raise TypeError('%s must be a sequence of integers, got %r'
+                                % (name, dim))
+        while len(dim) < 3:
+            dim.append(1)
+        return tuple(dim)
+
+    if None in (griddim, blockdim):
+        raise ValueError(missing_launch_config_msg)
+
+    griddim = check_dim(griddim, 'griddim')
+    blockdim = check_dim(blockdim, 'blockdim')
+
+    return griddim, blockdim

numba_cuda/numba/cuda/extending.py

@@ -0,0 +1,7 @@
+"""
+Added for symmetry with the core API
+"""
+
+from numba.core.extending import intrinsic as _intrinsic
+
+intrinsic = _intrinsic(target='cuda')

numba_cuda/numba/cuda/initialize.py

@@ -0,0 +1,13 @@
+def initialize_all():
+    # Import models to register them with the data model manager
+    import numba.cuda.models  # noqa: F401
+
+    from numba.cuda.decorators import jit
+    from numba.cuda.dispatcher import CUDADispatcher
+    from numba.core.target_extension import (target_registry,
+                                             dispatcher_registry,
+                                             jit_registry)
+
+    cuda_target = target_registry["cuda"]
+    jit_registry[cuda_target] = jit
+    dispatcher_registry[cuda_target] = CUDADispatcher

numba_cuda/numba/cuda/intrinsic_wrapper.py

@@ -0,0 +1,77 @@
+from .decorators import jit
+import numba
+
+
+@jit(device=True)
+def all_sync(mask, predicate):
+    """
+    If for all threads in the masked warp the predicate is true, then
+    a non-zero value is returned, otherwise 0 is returned.
+    """
+    return numba.cuda.vote_sync_intrinsic(mask, 0, predicate)[1]
+
+
+@jit(device=True)
+def any_sync(mask, predicate):
+    """
+    If for any thread in the masked warp the predicate is true, then
+    a non-zero value is returned, otherwise 0 is returned.
+    """
+    return numba.cuda.vote_sync_intrinsic(mask, 1, predicate)[1]
+
+
+@jit(device=True)
+def eq_sync(mask, predicate):
+    """
+    If for all threads in the masked warp the boolean predicate is the same,
+    then a non-zero value is returned, otherwise 0 is returned.
+    """
+    return numba.cuda.vote_sync_intrinsic(mask, 2, predicate)[1]
+
+
+@jit(device=True)
+def ballot_sync(mask, predicate):
+    """
+    Returns a mask of all threads in the warp whose predicate is true,
+    and are within the given mask.
+    """
+    return numba.cuda.vote_sync_intrinsic(mask, 3, predicate)[0]
+
+
+@jit(device=True)
+def shfl_sync(mask, value, src_lane):
+    """
+    Shuffles value across the masked warp and returns the value
+    from src_lane. If this is outside the warp, then the
+    given value is returned.
+    """
+    return numba.cuda.shfl_sync_intrinsic(mask, 0, value, src_lane, 0x1f)[0]
+
+
+@jit(device=True)
+def shfl_up_sync(mask, value, delta):
+    """
+    Shuffles value across the masked warp and returns the value
+    from (laneid - delta). If this is outside the warp, then the
+    given value is returned.
+    """
+    return numba.cuda.shfl_sync_intrinsic(mask, 1, value, delta, 0)[0]
+
+
+@jit(device=True)
+def shfl_down_sync(mask, value, delta):
+    """
+    Shuffles value across the masked warp and returns the value
+    from (laneid + delta). If this is outside the warp, then the
+    given value is returned.
+    """
+    return numba.cuda.shfl_sync_intrinsic(mask, 2, value, delta, 0x1f)[0]
+
+
+@jit(device=True)
+def shfl_xor_sync(mask, value, lane_mask):
+    """
+    Shuffles value across the masked warp and returns the value
+    from (laneid ^ lane_mask).
+    """
+    return numba.cuda.shfl_sync_intrinsic(mask, 3, value, lane_mask, 0x1f)[0]

numba_cuda/numba/cuda/intrinsics.py

@@ -0,0 +1,198 @@
+from llvmlite import ir
+
+from numba import cuda, types
+from numba.core import cgutils
+from numba.core.errors import RequireLiteralValue
+from numba.core.typing import signature
+from numba.core.extending import overload_attribute
+from numba.cuda import nvvmutils
+from numba.cuda.extending import intrinsic
+
+
+#-------------------------------------------------------------------------------
+# Grid functions
+
+def _type_grid_function(ndim):
+    val = ndim.literal_value
+    if val == 1:
+        restype = types.int64
+    elif val in (2, 3):
+        restype = types.UniTuple(types.int64, val)
+    else:
+        raise ValueError('argument can only be 1, 2, 3')
+
+    return signature(restype, types.int32)
+
+
+@intrinsic
+def grid(typingctx, ndim):
+    '''grid(ndim)
+
+    Return the absolute position of the current thread in the entire grid of
+    blocks.  *ndim* should correspond to the number of dimensions declared when
+    instantiating the kernel. If *ndim* is 1, a single integer is returned.
+    If *ndim* is 2 or 3, a tuple of the given number of integers is returned.
+
+    Computation of the first integer is as follows::
+
+        cuda.threadIdx.x + cuda.blockIdx.x * cuda.blockDim.x
+
+    and is similar for the other two indices, but using the ``y`` and ``z``
+    attributes.
+    '''
+
+    if not isinstance(ndim, types.IntegerLiteral):
+        raise RequireLiteralValue(ndim)
+
+    sig = _type_grid_function(ndim)
+
+    def codegen(context, builder, sig, args):
+        restype = sig.return_type
+        if restype == types.int64:
+            return nvvmutils.get_global_id(builder, dim=1)
+        elif isinstance(restype, types.UniTuple):
+            ids = nvvmutils.get_global_id(builder, dim=restype.count)
+            return cgutils.pack_array(builder, ids)
+
+    return sig, codegen
+
+
+@intrinsic
+def gridsize(typingctx, ndim):
+    '''gridsize(ndim)
+
+    Return the absolute size (or shape) in threads of the entire grid of
+    blocks. *ndim* should correspond to the number of dimensions declared when
+    instantiating the kernel. If *ndim* is 1, a single integer is returned.
+    If *ndim* is 2 or 3, a tuple of the given number of integers is returned.
+
+    Computation of the first integer is as follows::
+
+        cuda.blockDim.x * cuda.gridDim.x
+
+    and is similar for the other two indices, but using the ``y`` and ``z``
+    attributes.
+    '''
+
+    if not isinstance(ndim, types.IntegerLiteral):
+        raise RequireLiteralValue(ndim)
+
+    sig = _type_grid_function(ndim)
+
+    def _nthreads_for_dim(builder, dim):
+        i64 = ir.IntType(64)
+        ntid = nvvmutils.call_sreg(builder, f"ntid.{dim}")
+        nctaid = nvvmutils.call_sreg(builder, f"nctaid.{dim}")
+        return builder.mul(builder.sext(ntid, i64), builder.sext(nctaid, i64))
+
+    def codegen(context, builder, sig, args):
+        restype = sig.return_type
+        nx = _nthreads_for_dim(builder, 'x')
+
+        if restype == types.int64:
+            return nx
+        elif isinstance(restype, types.UniTuple):
+            ny = _nthreads_for_dim(builder, 'y')
+
+            if restype.count == 2:
+                return cgutils.pack_array(builder, (nx, ny))
+            elif restype.count == 3:
+                nz = _nthreads_for_dim(builder, 'z')
+                return cgutils.pack_array(builder, (nx, ny, nz))
+
+    return sig, codegen
+
+
+@intrinsic
+def _warpsize(typingctx):
+    sig = signature(types.int32)
+
+    def codegen(context, builder, sig, args):
+        return nvvmutils.call_sreg(builder, 'warpsize')
+
+    return sig, codegen
+
+
+@overload_attribute(types.Module(cuda), 'warpsize', target='cuda')
+def cuda_warpsize(mod):
+    '''
+    The size of a warp. All architectures implemented to date have a warp size
+    of 32.
+    '''
+    def get(mod):
+        return _warpsize()
+    return get
+
+
+#-------------------------------------------------------------------------------
+# syncthreads
+
+@intrinsic
+def syncthreads(typingctx):
+    '''
+    Synchronize all threads in the same thread block.  This function implements
+    the same pattern as barriers in traditional multi-threaded programming: this
+    function waits until all threads in the block call it, at which point it
+    returns control to all its callers.
+    '''
+    sig = signature(types.none)
+
+    def codegen(context, builder, sig, args):
+        fname = 'llvm.nvvm.barrier0'
+        lmod = builder.module
+        fnty = ir.FunctionType(ir.VoidType(), ())
+        sync = cgutils.get_or_insert_function(lmod, fnty, fname)
+        builder.call(sync, ())
+        return context.get_dummy_value()
+
+    return sig, codegen
+
+
+def _syncthreads_predicate(typingctx, predicate, fname):
+    if not isinstance(predicate, types.Integer):
+        return None
+
+    sig = signature(types.i4, types.i4)
+
+    def codegen(context, builder, sig, args):
+        fnty = ir.FunctionType(ir.IntType(32), (ir.IntType(32),))
+        sync = cgutils.get_or_insert_function(builder.module, fnty, fname)
+        return builder.call(sync, args)
+
+    return sig, codegen
+
+
+@intrinsic
+def syncthreads_count(typingctx, predicate):
+    '''
+    syncthreads_count(predicate)
+
+    An extension to numba.cuda.syncthreads where the return value is a count
+    of the threads where predicate is true.
+    '''
+    fname = 'llvm.nvvm.barrier0.popc'
+    return _syncthreads_predicate(typingctx, predicate, fname)
+
+
+@intrinsic
+def syncthreads_and(typingctx, predicate):
+    '''
+    syncthreads_and(predicate)
+
+    An extension to numba.cuda.syncthreads where 1 is returned if predicate is
+    true for all threads or 0 otherwise.
+    '''
+    fname = 'llvm.nvvm.barrier0.and'
+    return _syncthreads_predicate(typingctx, predicate, fname)
+
+
+@intrinsic
+def syncthreads_or(typingctx, predicate):
+    '''
+    syncthreads_or(predicate)
+
+    An extension to numba.cuda.syncthreads where 1 is returned if predicate is
+    true for any thread or 0 otherwise.
+    '''
+    fname = 'llvm.nvvm.barrier0.or'
+    return _syncthreads_predicate(typingctx, predicate, fname)

numba_cuda/numba/cuda/kernels/reduction.py

@@ -0,0 +1,262 @@
+"""
+A library written in CUDA Python for generating reduction kernels
+"""
+
+from numba.np.numpy_support import from_dtype
+
+
+_WARPSIZE = 32
+_NUMWARPS = 4
+
+
+def _gpu_reduce_factory(fn, nbtype):
+    from numba import cuda
+
+    reduce_op = cuda.jit(device=True)(fn)
+    inner_sm_size = _WARPSIZE + 1   # plus one to avoid SM collision
+    max_blocksize = _NUMWARPS * _WARPSIZE
+
+    @cuda.jit(device=True)
+    def inner_warp_reduction(sm_partials, init):
+        """
+        Compute reduction within a single warp
+        """
+        tid = cuda.threadIdx.x
+        warpid = tid // _WARPSIZE
+        laneid = tid % _WARPSIZE
+
+        sm_this = sm_partials[warpid, :]
+        sm_this[laneid] = init
+        cuda.syncwarp()
+
+        width = _WARPSIZE // 2
+        while width:
+            if laneid < width:
+                old = sm_this[laneid]
+                sm_this[laneid] = reduce_op(old, sm_this[laneid + width])
+            cuda.syncwarp()
+            width //= 2
+
+    @cuda.jit(device=True)
+    def device_reduce_full_block(arr, partials, sm_partials):
+        """
+        Partially reduce `arr` into `partials` using `sm_partials` as working
+        space.  The algorithm goes like:
+
+            array chunks of 128:  |   0 | 128 | 256 | 384 | 512 |
+                        block-0:  |   x |     |     |   x |     |
+                        block-1:  |     |   x |     |     |   x |
+                        block-2:  |     |     |   x |     |     |
+
+        The array is divided into chunks of 128 (size of a threadblock).
+        The threadblocks consumes the chunks in roundrobin scheduling.
+        First, a threadblock loads a chunk into temp memory.  Then, all
+        subsequent chunks are combined into the temp memory.
+
+        Once all chunks are processed.  Inner-block reduction is performed
+        on the temp memory.  So that, there will just be one scalar result
+        per block.  The result from each block is stored to `partials` at
+        the dedicated slot.
+        """
+        tid = cuda.threadIdx.x
+        blkid = cuda.blockIdx.x
+        blksz = cuda.blockDim.x
+        gridsz = cuda.gridDim.x
+
+        # block strided loop to compute the reduction
+        start = tid + blksz * blkid
+        stop = arr.size
+        step = blksz * gridsz
+
+        # load first value
+        tmp = arr[start]
+        # loop over all values in block-stride
+        for i in range(start + step, stop, step):
+            tmp = reduce_op(tmp, arr[i])
+
+        cuda.syncthreads()
+        # inner-warp reduction
+        inner_warp_reduction(sm_partials, tmp)
+
+        cuda.syncthreads()
+        # at this point, only the first slot for each warp in tsm_partials
+        # is valid.
+
+        # finish up block reduction
+        # warning: this is assuming 4 warps.
+        # assert numwarps == 4
+        if tid < 2:
+            sm_partials[tid, 0] = reduce_op(sm_partials[tid, 0],
+                                            sm_partials[tid + 2, 0])
+            cuda.syncwarp()
+        if tid == 0:
+            partials[blkid] = reduce_op(sm_partials[0, 0], sm_partials[1, 0])
+
+    @cuda.jit(device=True)
+    def device_reduce_partial_block(arr, partials, sm_partials):
+        """
+        This computes reduction on `arr`.
+        This device function must be used by 1 threadblock only.
+        The blocksize must match `arr.size` and must not be greater than 128.
+        """
+        tid = cuda.threadIdx.x
+        blkid = cuda.blockIdx.x
+        blksz = cuda.blockDim.x
+        warpid = tid // _WARPSIZE
+        laneid = tid % _WARPSIZE
+
+        size = arr.size
+        # load first value
+        tid = cuda.threadIdx.x
+        value = arr[tid]
+        sm_partials[warpid, laneid] = value
+
+        cuda.syncthreads()
+
+        if (warpid + 1) * _WARPSIZE < size:
+            # fully populated warps
+            inner_warp_reduction(sm_partials, value)
+        else:
+            # partially populated warps
+            # NOTE: this uses a very inefficient sequential algorithm
+            if laneid == 0:
+                sm_this = sm_partials[warpid, :]
+                base = warpid * _WARPSIZE
+                for i in range(1, size - base):
+                    sm_this[0] = reduce_op(sm_this[0], sm_this[i])
+
+        cuda.syncthreads()
+        # finish up
+        if tid == 0:
+            num_active_warps = (blksz + _WARPSIZE - 1) // _WARPSIZE
+
+            result = sm_partials[0, 0]
+            for i in range(1, num_active_warps):
+                result = reduce_op(result, sm_partials[i, 0])
+
+            partials[blkid] = result
+
+    def gpu_reduce_block_strided(arr, partials, init, use_init):
+        """
+        Perform reductions on *arr* and writing out partial reduction result
+        into *partials*.  The length of *partials* is determined by the
+        number of threadblocks. The initial value is set with *init*.
+
+        Launch config:
+
+        Blocksize must be multiple of warpsize and it is limited to 4 warps.
+        """
+        tid = cuda.threadIdx.x
+
+        sm_partials = cuda.shared.array((_NUMWARPS, inner_sm_size),
+                                        dtype=nbtype)
+        if cuda.blockDim.x == max_blocksize:
+            device_reduce_full_block(arr, partials, sm_partials)
+        else:
+            device_reduce_partial_block(arr, partials, sm_partials)
+        # deal with the initializer
+        if use_init and tid == 0 and cuda.blockIdx.x == 0:
+            partials[0] = reduce_op(partials[0], init)
+
+    return cuda.jit(gpu_reduce_block_strided)
+
+
+class Reduce(object):
+    """Create a reduction object that reduces values using a given binary
+    function. The binary function is compiled once and cached inside this
+    object. Keeping this object alive will prevent re-compilation.
+    """
+
+    _cache = {}
+
+    def __init__(self, functor):
+        """
+        :param functor: A function implementing a binary operation for
+                        reduction. It will be compiled as a CUDA device
+                        function using ``cuda.jit(device=True)``.
+        """
+        self._functor = functor
+
+    def _compile(self, dtype):
+        key = self._functor, dtype
+        if key in self._cache:
+            kernel = self._cache[key]
+        else:
+            kernel = _gpu_reduce_factory(self._functor, from_dtype(dtype))
+            self._cache[key] = kernel
+        return kernel
+
+    def __call__(self, arr, size=None, res=None, init=0, stream=0):
+        """Performs a full reduction.
+
+        :param arr: A host or device array.
+        :param size: Optional integer specifying the number of elements in
+                    ``arr`` to reduce. If this parameter is not specified, the
+                    entire array is reduced.
+        :param res: Optional device array into which to write the reduction
+                    result to. The result is written into the first element of
+                    this array. If this parameter is specified, then no
+                    communication of the reduction output takes place from the
+                    device to the host.
+        :param init: Optional initial value for the reduction, the type of which
+                    must match ``arr.dtype``.
+        :param stream: Optional CUDA stream in which to perform the reduction.
+                    If no stream is specified, the default stream of 0 is
+                    used.
+        :return: If ``res`` is specified, ``None`` is returned. Otherwise, the
+                result of the reduction is returned.
+        """
+        from numba import cuda
+
+        # ensure 1d array
+        if arr.ndim != 1:
+            raise TypeError("only support 1D array")
+
+        # adjust array size
+        if size is not None:
+            arr = arr[:size]
+
+        init = arr.dtype.type(init)  # ensure the right type
+
+        # return `init` if `arr` is empty
+        if arr.size < 1:
+            return init
+
+        kernel = self._compile(arr.dtype)
+
+        # Perform the reduction on the GPU
+        blocksize = _NUMWARPS * _WARPSIZE
+        size_full = (arr.size // blocksize) * blocksize
+        size_partial = arr.size - size_full
+        full_blockct = min(size_full // blocksize, _WARPSIZE * 2)
+
+        # allocate size of partials array
+        partials_size = full_blockct
+        if size_partial:
+            partials_size += 1
+        partials = cuda.device_array(shape=partials_size, dtype=arr.dtype)
+
+        if size_full:
+            # kernel for the fully populated threadblocks
+            kernel[full_blockct, blocksize, stream](arr[:size_full],
+                                                    partials[:full_blockct],
+                                                    init,
+                                                    True)
+
+        if size_partial:
+            # kernel for partially populated threadblocks
+            kernel[1, size_partial, stream](arr[size_full:],
+                                            partials[full_blockct:],
+                                            init,
+                                            not full_blockct)
+
+        if partials.size > 1:
+            # finish up
+            kernel[1, partials_size, stream](partials, partials, init, False)
+
+        # handle return value
+        if res is not None:
+            res[:1].copy_to_device(partials[:1], stream=stream)
+            return
+        else:
+            return partials[0]

numba_cuda/numba/cuda/kernels/transpose.py

@@ -0,0 +1,65 @@
+from numba import cuda
+from numba.cuda.cudadrv.driver import driver
+import math
+from numba.np import numpy_support as nps
+
+
+def transpose(a, b=None):
+    """Compute the transpose of 'a' and store it into 'b', if given,
+    and return it. If 'b' is not given, allocate a new array
+    and return that.
+
+    This implements the algorithm documented in
+    http://devblogs.nvidia.com/parallelforall/efficient-matrix-transpose-cuda-cc/
+
+    :param a: an `np.ndarray` or a `DeviceNDArrayBase` subclass. If already on
+        the device its stream will be used to perform the transpose (and to copy
+        `b` to the device if necessary).
+    """
+
+    # prefer `a`'s stream if
+    stream = getattr(a, 'stream', 0)
+
+    if not b:
+        cols, rows = a.shape
+        strides = a.dtype.itemsize * cols, a.dtype.itemsize
+        b = cuda.cudadrv.devicearray.DeviceNDArray(
+            (rows, cols),
+            strides,
+            dtype=a.dtype,
+            stream=stream)
+
+    dt = nps.from_dtype(a.dtype)
+
+    tpb = driver.get_device().MAX_THREADS_PER_BLOCK
+    # we need to factor available threads into x and y axis
+    tile_width = int(math.pow(2, math.log(tpb, 2) / 2))
+    tile_height = int(tpb / tile_width)
+
+    tile_shape = (tile_height, tile_width + 1)
+
+    @cuda.jit
+    def kernel(input, output):
+
+        tile = cuda.shared.array(shape=tile_shape, dtype=dt)
+
+        tx = cuda.threadIdx.x
+        ty = cuda.threadIdx.y
+        bx = cuda.blockIdx.x * cuda.blockDim.x
+        by = cuda.blockIdx.y * cuda.blockDim.y
+        x = by + tx
+        y = bx + ty
+
+        if by + ty < input.shape[0] and bx + tx < input.shape[1]:
+            tile[ty, tx] = input[by + ty, bx + tx]
+        cuda.syncthreads()
+        if y < output.shape[0] and x < output.shape[1]:
+            output[y, x] = tile[tx, ty]
+
+    # one block per tile, plus one for remainders
+    blocks = int(b.shape[0] / tile_height + 1), int(b.shape[1] / tile_width + 1)
+    # one thread per tile element
+    threads = tile_height, tile_width
+    kernel[blocks, threads, stream](a, b)
+
+    return b